/home/liu/actions-runner/_work/ccv/ccv/test/int/nnc/cudnn.tests.c
Line | Count | Source |
1 | | #include "case.h" |
2 | | #include "ccv_case.h" |
3 | | #include "ccv_nnc_case.h" |
4 | | #include <ccv.h> |
5 | | #include <nnc/ccv_nnc.h> |
6 | | #include <nnc/ccv_nnc_easy.h> |
7 | | #include <3rdparty/dsfmt/dSFMT.h> |
8 | | |
9 | | TEST_SETUP() |
10 | | { |
11 | | ccv_nnc_init(); |
12 | | } |
13 | | |
14 | 46.2M | #define INPUT_DIM (3) |
15 | 231M | #define OUTPUT_DIM (96) |
16 | | |
17 | 91.5M | #define INPUT_SIZE (224) |
18 | 308M | #define OUTPUT_SIZE (112) |
19 | | |
20 | 1.07M | #define KERNEL_SIZE (7) |
21 | | |
22 | | #define BATCH_SIZE (16) |
23 | | |
24 | | TEST_CASE("cudnn forward convolution") |
25 | 1 | { |
26 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
27 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
28 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
29 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
30 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
31 | 1 | assert(cmd.backend >= 0); |
32 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, b->info); |
33 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, b->info) == 0); |
34 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
35 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
36 | | // configure the inlets. |
37 | 1 | dsfmt_t dsfmt; |
38 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
39 | 1 | int i; |
40 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
41 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
42 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
43 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
44 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
45 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
46 | | // Copy generated matrix values over to GPU. |
47 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
48 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
49 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
50 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
51 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
52 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
53 | 1 | assert(move.backend >= 0); |
54 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
55 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
56 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
57 | | |
58 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
59 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
60 | 1 | assert(transform.backend >= 0); |
61 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
62 | 1 | ccv_nnc_cmd_exec(transform, ccv_nnc_no_hint, 0, TENSOR_LIST(gw), TENSOR_LIST(gwo), stream_context); |
63 | 1 | ccv_nnc_stream_context_wait(stream_context); |
64 | 1 | ccv_nnc_tensor_free(gw); |
65 | | |
66 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
67 | 1 | assert(cmd.backend >= 0); |
68 | 1 | cmd.algorithm = -1; |
69 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context); |
70 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context)); |
71 | 1 | ccv_nnc_stream_context_wait(stream_context); |
72 | 1 | ccv_nnc_stream_context_free(stream_context); |
73 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
74 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
75 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
76 | 1 | ccv_nnc_tensor_free(c); |
77 | 1 | ccv_nnc_tensor_free(gc); |
78 | 1 | ccv_nnc_tensor_free(bias); |
79 | 1 | ccv_nnc_tensor_free(w); |
80 | 1 | ccv_nnc_tensor_free(b); |
81 | 1 | ccv_nnc_tensor_free(a); |
82 | 1 | ccv_nnc_tensor_free(gbias); |
83 | 1 | ccv_nnc_tensor_free(gwo); |
84 | 1 | ccv_nnc_tensor_free(ga); |
85 | 1 | } |
86 | | |
87 | | TEST_CASE("cudnn forward convolution in nchw format") |
88 | 1 | { |
89 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
90 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
91 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, OUTPUT_DIM, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
92 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
93 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
94 | 1 | assert(cmd.backend >= 0); |
95 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, b->info); |
96 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, b->info) == 0); |
97 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
98 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, OUTPUT_DIM), 0); |
99 | | // configure the inlets. |
100 | 1 | dsfmt_t dsfmt; |
101 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
102 | 1 | int i; |
103 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
104 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
105 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
106 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
107 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
108 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
109 | | // Copy generated matrix values over to GPU. |
110 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
111 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
112 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM), 0); |
113 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
114 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
115 | 1 | assert(move.backend >= 0); |
116 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
117 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
118 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, OUTPUT_DIM, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
119 | | |
120 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
121 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
122 | 1 | assert(transform.backend >= 0); |
123 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
124 | 1 | assert(cmd.backend >= 0); |
125 | 1 | cmd.algorithm = -1; |
126 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), 0); |
127 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), 0)); |
128 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, OUTPUT_DIM, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
129 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
130 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 1e-5, "output from cudnn should match from CPU"); |
131 | 1 | ccv_nnc_tensor_free(c); |
132 | 1 | ccv_nnc_tensor_free(gc); |
133 | 1 | ccv_nnc_tensor_free(bias); |
134 | 1 | ccv_nnc_tensor_free(w); |
135 | 1 | ccv_nnc_tensor_free(b); |
136 | 1 | ccv_nnc_tensor_free(a); |
137 | 1 | ccv_nnc_tensor_free(gbias); |
138 | 1 | ccv_nnc_tensor_free(gw); |
139 | 1 | ccv_nnc_tensor_free(ga); |
140 | 1 | } |
141 | | |
142 | | TEST_CASE("cudnn forward convolution in half precision") |
143 | 1 | { |
144 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
145 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
146 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
147 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
148 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
149 | 1 | assert(cmd.backend >= 0); |
150 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, b->info); |
151 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, b->info) == 0); |
152 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
153 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
154 | | // configure the inlets. |
155 | 1 | dsfmt_t dsfmt; |
156 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
157 | 1 | int i; |
158 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
159 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
160 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
161 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
162 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
163 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
164 | 1 | ccv_nnc_tensor_t* a1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
165 | 1 | ccv_nnc_tensor_t* w1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
166 | 1 | ccv_nnc_tensor_t* bias1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM), 0); |
167 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(a1, w1, bias1), 0); |
168 | | // Copy generated matrix values over to GPU. |
169 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
170 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
171 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 16F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
172 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM), 0); |
173 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a1, w1, bias1), TENSOR_LIST(ga, gw, gbias), 0); |
174 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
175 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
176 | | |
177 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
178 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
179 | 1 | assert(transform.backend >= 0); |
180 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
181 | 1 | ccv_nnc_cmd_exec(transform, ccv_nnc_no_hint, 0, TENSOR_LIST(gw), TENSOR_LIST(gwo), stream_context); |
182 | 1 | ccv_nnc_stream_context_wait(stream_context); |
183 | 1 | ccv_nnc_tensor_free(gw); |
184 | | |
185 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
186 | 1 | assert(cmd.backend >= 0); |
187 | 1 | cmd.algorithm = -1; |
188 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 512 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context); |
189 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context)); |
190 | 1 | ccv_nnc_stream_context_wait(stream_context); |
191 | 1 | ccv_nnc_stream_context_free(stream_context); |
192 | 1 | ccv_nnc_tensor_t* c1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
193 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c1), 0); |
194 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
195 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c1), TENSOR_LIST(c), 0); |
196 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 5e-3, "output from cudnn should match from CPU"); |
197 | 1 | ccv_nnc_tensor_free(c); |
198 | 1 | ccv_nnc_tensor_free(gc); |
199 | 1 | ccv_nnc_tensor_free(bias); |
200 | 1 | ccv_nnc_tensor_free(w); |
201 | 1 | ccv_nnc_tensor_free(b); |
202 | 1 | ccv_nnc_tensor_free(a); |
203 | 1 | ccv_nnc_tensor_free(c1); |
204 | 1 | ccv_nnc_tensor_free(bias1); |
205 | 1 | ccv_nnc_tensor_free(w1); |
206 | 1 | ccv_nnc_tensor_free(a1); |
207 | 1 | ccv_nnc_tensor_free(gbias); |
208 | 1 | ccv_nnc_tensor_free(gwo); |
209 | 1 | ccv_nnc_tensor_free(ga); |
210 | 1 | } |
211 | | |
212 | | TEST_CASE("cudnn forward convolution in half precision with palettize weights") |
213 | 1 | { |
214 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
215 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
216 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
217 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
218 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
219 | 1 | assert(cmd.backend >= 0); |
220 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, b->info); |
221 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, b->info) == 0); |
222 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
223 | 1 | ccv_nnc_tensor_t* wo = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
224 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
225 | | // configure the inlets. |
226 | 1 | dsfmt_t dsfmt; |
227 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
228 | 1 | int i; |
229 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
230 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
231 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
232 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
233 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
234 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
235 | 1 | ccv_nnc_cmd_exec(CMD_FORMAT_TRANSFORM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(w), TENSOR_LIST(wo), 0); |
236 | 1 | ccv_nnc_tensor_t* a1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
237 | 1 | ccv_nnc_tensor_t* w1o = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
238 | 1 | ccv_nnc_tensor_t* bias1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM), 0); |
239 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, wo, bias), TENSOR_LIST(a1, w1o, bias1), 0); |
240 | 1 | ccv_nnc_tensor_t* pw1o = ccv_nnc_tensor_new(0, ccv_nnc_tensor_palettize(CPU_TENSOR_NCHW(16F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 6, 1280), 0); |
241 | 1 | (void)ccv_nnc_palettize(w1o->data.u8, CCV_16F, CCV_TENSOR_CPU_MEMORY, ccv_nnc_tensor_count(w1o->info), 6, 1280, pw1o->data.u8, ccv_nnc_tensor_data_size_without_padding(pw1o->info)); |
242 | | // Copy generated matrix values over to GPU. |
243 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
244 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, ccv_nnc_tensor_palettize(GPU_TENSOR_NCHW(000, 16F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 6, 1280), 0); |
245 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM), 0); |
246 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a1, pw1o, bias1), TENSOR_LIST(ga, gwo, gbias), 0); |
247 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
248 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
249 | | |
250 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
251 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
252 | 1 | assert(cmd.backend >= 0); |
253 | 1 | cmd.algorithm = -1; |
254 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 512 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context); |
255 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context)); |
256 | 1 | ccv_nnc_stream_context_wait(stream_context); |
257 | 1 | ccv_nnc_stream_context_free(stream_context); |
258 | 1 | ccv_nnc_tensor_t* c1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
259 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c1), 0); |
260 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
261 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c1), TENSOR_LIST(c), 0); |
262 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 5e-3, "output from cudnn should match from CPU"); |
263 | 1 | ccv_nnc_tensor_free(c); |
264 | 1 | ccv_nnc_tensor_free(gc); |
265 | 1 | ccv_nnc_tensor_free(bias); |
266 | 1 | ccv_nnc_tensor_free(w); |
267 | 1 | ccv_nnc_tensor_free(wo); |
268 | 1 | ccv_nnc_tensor_free(b); |
269 | 1 | ccv_nnc_tensor_free(a); |
270 | 1 | ccv_nnc_tensor_free(c1); |
271 | 1 | ccv_nnc_tensor_free(bias1); |
272 | 1 | ccv_nnc_tensor_free(w1o); |
273 | 1 | ccv_nnc_tensor_free(pw1o); |
274 | 1 | ccv_nnc_tensor_free(a1); |
275 | 1 | ccv_nnc_tensor_free(gbias); |
276 | 1 | ccv_nnc_tensor_free(gwo); |
277 | 1 | ccv_nnc_tensor_free(ga); |
278 | 1 | } |
279 | | |
280 | | TEST_CASE("cudnn forward convolution with dilation 2, 3") |
281 | 1 | { |
282 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
283 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
284 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
285 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
286 | 1 | cmd.info.convolution.dilation[0] = 2; |
287 | 1 | cmd.info.convolution.dilation[1] = 3; |
288 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
289 | 1 | assert(cmd.backend >= 0); |
290 | 1 | ccv_nnc_cmd_param_t modified_cmd = cmd.info; |
291 | 1 | modified_cmd.size.dim[0] = (cmd.info.size.dim[0] - 1) * ccv_max(cmd.info.convolution.dilation[0], 1) + 1; |
292 | 1 | modified_cmd.size.dim[1] = (cmd.info.size.dim[1] - 1) * ccv_max(cmd.info.convolution.dilation[1], 1) + 1; |
293 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(modified_cmd, a->info, b->info); |
294 | 1 | assert(ccv_nnc_hint_verify(hint, modified_cmd, a->info, b->info) == 0); |
295 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
296 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
297 | | // configure the inlets. |
298 | 1 | dsfmt_t dsfmt; |
299 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
300 | 1 | int i; |
301 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
302 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
303 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
304 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
305 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
306 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
307 | | // Copy generated matrix values over to GPU. |
308 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
309 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
310 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
311 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
312 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
313 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
314 | 1 | assert(move.backend >= 0); |
315 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
316 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
317 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
318 | | |
319 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
320 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
321 | 1 | assert(transform.backend >= 0); |
322 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
323 | 1 | ccv_nnc_cmd_exec(transform, ccv_nnc_no_hint, 0, TENSOR_LIST(gw), TENSOR_LIST(gwo), stream_context); |
324 | 1 | ccv_nnc_stream_context_wait(stream_context); |
325 | 1 | ccv_nnc_tensor_free(gw); |
326 | | |
327 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
328 | 1 | assert(cmd.backend >= 0); |
329 | 1 | cmd.algorithm = -1; |
330 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context); |
331 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context)); |
332 | 1 | ccv_nnc_stream_context_wait(stream_context); |
333 | 1 | ccv_nnc_stream_context_free(stream_context); |
334 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
335 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
336 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
337 | 1 | ccv_nnc_tensor_free(c); |
338 | 1 | ccv_nnc_tensor_free(gc); |
339 | 1 | ccv_nnc_tensor_free(bias); |
340 | 1 | ccv_nnc_tensor_free(w); |
341 | 1 | ccv_nnc_tensor_free(b); |
342 | 1 | ccv_nnc_tensor_free(a); |
343 | 1 | ccv_nnc_tensor_free(gbias); |
344 | 1 | ccv_nnc_tensor_free(gwo); |
345 | 1 | ccv_nnc_tensor_free(ga); |
346 | 1 | } |
347 | | |
348 | | TEST_CASE("cudnn forward convolution 3d") |
349 | 1 | { |
350 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
351 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, 5, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
352 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, 3, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
353 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
354 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, b->info); |
355 | 1 | hint.stride.dim[0] = 2; |
356 | 1 | hint.border.begin[0] = 1; |
357 | 1 | hint.border.end[0] = 1; |
358 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, b->info) == 0); |
359 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
360 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
361 | | // configure the inlets. |
362 | 1 | dsfmt_t dsfmt; |
363 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
364 | 1 | int i; |
365 | 42.3k | for (i = 0; i < INPUT_DIM * 3 * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++42.3k ) |
366 | 42.3k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
367 | 12.0M | for (i = 0; i < 5 * INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++12.0M ) |
368 | 12.0M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
369 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
370 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
371 | | // Copy generated matrix values over to GPU. |
372 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, 5, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
373 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
374 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE), 0); |
375 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
376 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
377 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
378 | 1 | assert(move.backend >= 0); |
379 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
380 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, 3, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
381 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
382 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
383 | 1 | assert(transform.backend >= 0); |
384 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
385 | 1 | ccv_nnc_cmd_exec(transform, ccv_nnc_no_hint, 0, TENSOR_LIST(gw), TENSOR_LIST(gwo), stream_context); |
386 | 1 | ccv_nnc_stream_context_wait(stream_context); |
387 | 1 | ccv_nnc_tensor_free(gw); |
388 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
389 | 1 | assert(cmd.backend >= 0); |
390 | 1 | cmd.algorithm = -1; |
391 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context); |
392 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context)); |
393 | 1 | ccv_nnc_stream_context_wait(stream_context); |
394 | 1 | ccv_nnc_stream_context_free(stream_context); |
395 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, 3, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
396 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
397 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
398 | 1 | assert(cmd.backend >= 0); |
399 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
400 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * 3 * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
401 | 1 | ccv_nnc_tensor_free(c); |
402 | 1 | ccv_nnc_tensor_free(gc); |
403 | 1 | ccv_nnc_tensor_free(bias); |
404 | 1 | ccv_nnc_tensor_free(w); |
405 | 1 | ccv_nnc_tensor_free(b); |
406 | 1 | ccv_nnc_tensor_free(a); |
407 | 1 | ccv_nnc_tensor_free(gbias); |
408 | 1 | ccv_nnc_tensor_free(gwo); |
409 | 1 | ccv_nnc_tensor_free(ga); |
410 | 1 | } |
411 | | |
412 | | TEST_CASE("cudnn forward convolution 3d in nchw format") |
413 | 1 | { |
414 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
415 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, INPUT_DIM, 5, INPUT_SIZE, INPUT_SIZE), 0); |
416 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, OUTPUT_DIM, 3, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
417 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(1, OUTPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
418 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, b->info); |
419 | 1 | hint.stride.dim[0] = 2; |
420 | 1 | hint.border.begin[0] = 1; |
421 | 1 | hint.border.end[0] = 1; |
422 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, b->info) == 0); |
423 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, OUTPUT_DIM, INPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE), 0); |
424 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, OUTPUT_DIM), 0); |
425 | | // configure the inlets. |
426 | 1 | dsfmt_t dsfmt; |
427 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
428 | 1 | int i; |
429 | 42.3k | for (i = 0; i < 3 * INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++42.3k ) |
430 | 42.3k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
431 | 12.0M | for (i = 0; i < 5 * INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++12.0M ) |
432 | 12.0M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
433 | 97 | for (i = 0; i < OUTPUT_DIM; i++96 ) |
434 | 96 | bias->data.f32[i] = (float)i / OUTPUT_DIM; |
435 | | // Copy generated matrix values over to GPU. |
436 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, INPUT_DIM, 5, INPUT_SIZE, INPUT_SIZE), 0); |
437 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, 3, KERNEL_SIZE, KERNEL_SIZE), 0); |
438 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM), 0); |
439 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
440 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
441 | 1 | assert(move.backend >= 0); |
442 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
443 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, OUTPUT_DIM, 3, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
444 | | |
445 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
446 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
447 | 1 | assert(transform.backend >= 0); |
448 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
449 | 1 | assert(cmd.backend >= 0); |
450 | 1 | cmd.algorithm = -1; |
451 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), 0); |
452 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), 0)); |
453 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, OUTPUT_DIM, 3, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
454 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
455 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
456 | 1 | assert(cmd.backend >= 0); |
457 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
458 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * 3 * OUTPUT_DIM * OUTPUT_SIZE * OUTPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
459 | 1 | ccv_nnc_tensor_free(c); |
460 | 1 | ccv_nnc_tensor_free(gc); |
461 | 1 | ccv_nnc_tensor_free(bias); |
462 | 1 | ccv_nnc_tensor_free(w); |
463 | 1 | ccv_nnc_tensor_free(b); |
464 | 1 | ccv_nnc_tensor_free(a); |
465 | 1 | ccv_nnc_tensor_free(gbias); |
466 | 1 | ccv_nnc_tensor_free(gw); |
467 | 1 | ccv_nnc_tensor_free(ga); |
468 | 1 | } |
469 | | |
470 | | TEST_CASE("cudnn backward convolution") |
471 | 1 | { |
472 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
473 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
474 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
475 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
476 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_BACKWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
477 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
478 | 1 | assert(cmd.backend >= 0); |
479 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, g->info); |
480 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, g->info) == 0); |
481 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
482 | 1 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
483 | 1 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
484 | | // configure the inlets. |
485 | 1 | dsfmt_t dsfmt; |
486 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
487 | 1 | int i; |
488 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
489 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
490 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
491 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
492 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
493 | 19.2M | g->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / OUTPUT_DIM; // (OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM); |
494 | | // Copy generated matrix values over to GPU. |
495 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
496 | 1 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
497 | 1 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
498 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
499 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
500 | 1 | ccv_nnc_tensor_t* gdw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
501 | 1 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
502 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
503 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
504 | 1 | assert(move.backend >= 0); |
505 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, g), TENSOR_LIST(ga, gw, gg), 0); |
506 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(g, a, w), TENSOR_LIST(h, dw, dbias), 0); |
507 | | |
508 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
509 | 1 | assert(cmd.backend >= 0); |
510 | 1 | cmd.algorithm = -1; |
511 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
512 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(gg, ga, gw), TENSOR_LIST(gh, gdw, gdbias), stream_context); |
513 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(gg, ga, gw), TENSOR_LIST(gh, gdw, gdbias), stream_context)); |
514 | 1 | ccv_nnc_stream_context_wait(stream_context); |
515 | 1 | ccv_nnc_stream_context_free(stream_context); |
516 | 1 | ccv_nnc_tensor_t* ch = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
517 | 1 | ccv_nnc_tensor_t* cdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
518 | 1 | ccv_nnc_tensor_t* cdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
519 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdw, gdbias), TENSOR_LIST(ch, cdw, cdbias), 0); |
520 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dw->data.f32, cdw->data.f32, INPUT_DIM * OUTPUT_DIM * KERNEL_SIZE * KERNEL_SIZE, 5e-1, "output from cudnn should match from CPU"); |
521 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dbias->data.f32, cdbias->data.f32, OUTPUT_DIM, 5e-1, "output from cudnn should match from CPU"); |
522 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, h->data.f32, ch->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
523 | 1 | ccv_nnc_tensor_free(h); |
524 | 1 | ccv_nnc_tensor_free(gh); |
525 | 1 | ccv_nnc_tensor_free(w); |
526 | 1 | ccv_nnc_tensor_free(g); |
527 | 1 | ccv_nnc_tensor_free(a); |
528 | 1 | ccv_nnc_tensor_free(gbias); |
529 | 1 | ccv_nnc_tensor_free(gdbias); |
530 | 1 | ccv_nnc_tensor_free(gdw); |
531 | 1 | ccv_nnc_tensor_free(gw); |
532 | 1 | ccv_nnc_tensor_free(gg); |
533 | 1 | ccv_nnc_tensor_free(ga); |
534 | 1 | ccv_nnc_tensor_free(ch); |
535 | 1 | ccv_nnc_tensor_free(cdw); |
536 | 1 | ccv_nnc_tensor_free(cdbias); |
537 | 1 | } |
538 | | |
539 | | TEST_CASE("cudnn backward convolution in nchw format") |
540 | 1 | { |
541 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
542 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
543 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
544 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
545 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_BACKWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
546 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
547 | 1 | assert(cmd.backend >= 0); |
548 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, g->info); |
549 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, g->info) == 0); |
550 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
551 | 1 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
552 | 1 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
553 | | // configure the inlets. |
554 | 1 | dsfmt_t dsfmt; |
555 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
556 | 1 | int i; |
557 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
558 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
559 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
560 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
561 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
562 | 19.2M | g->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / OUTPUT_DIM; // (OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM); |
563 | | // Copy generated matrix values over to GPU. |
564 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
565 | 1 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
566 | 1 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
567 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
568 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
569 | 1 | ccv_nnc_tensor_t* gdw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
570 | 1 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
571 | 1 | ccv_nnc_tensor_t* gao = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
572 | 1 | ccv_nnc_tensor_t* ggo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, OUTPUT_DIM, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
573 | 1 | ccv_nnc_tensor_t* gho = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
574 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
575 | 1 | ccv_nnc_tensor_t* gbiaso = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM), 0); |
576 | 1 | ccv_nnc_tensor_t* gdwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
577 | 1 | ccv_nnc_tensor_t* gdbiaso = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM), 0); |
578 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, g), TENSOR_LIST(ga, gw, gg), 0); |
579 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(g, a, w), TENSOR_LIST(h, dw, dbias), 0); |
580 | 1 | ccv_nnc_cmd_exec(CMD_FORMAT_TRANSFORM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gw, gg), TENSOR_LIST(gao, gwo, ggo), 0); |
581 | | |
582 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
583 | 1 | assert(cmd.backend >= 0); |
584 | 1 | cmd.algorithm = -1; |
585 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
586 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ggo, gao, gwo), TENSOR_LIST(gho, gdwo, gdbiaso), stream_context); |
587 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ggo, gao, gwo), TENSOR_LIST(gho, gdwo, gdbiaso), stream_context)); |
588 | 1 | ccv_nnc_stream_context_wait(stream_context); |
589 | 1 | ccv_nnc_stream_context_free(stream_context); |
590 | 1 | ccv_nnc_tensor_t* ch = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
591 | 1 | ccv_nnc_tensor_t* cdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
592 | 1 | ccv_nnc_tensor_t* cdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
593 | 1 | ccv_nnc_cmd_exec(CMD_FORMAT_TRANSFORM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gho, gdwo, gdbiaso), TENSOR_LIST(gh, gdw, gdbias), 0); |
594 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdw, gdbias), TENSOR_LIST(ch, cdw, cdbias), 0); |
595 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dw->data.f32, cdw->data.f32, INPUT_DIM * OUTPUT_DIM * KERNEL_SIZE * KERNEL_SIZE, 5e-1, "output from cudnn should match from CPU"); |
596 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dbias->data.f32, cdbias->data.f32, OUTPUT_DIM, 5e-1, "output from cudnn should match from CPU"); |
597 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, h->data.f32, ch->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
598 | 1 | ccv_nnc_tensor_free(gao); |
599 | 1 | ccv_nnc_tensor_free(ggo); |
600 | 1 | ccv_nnc_tensor_free(gho); |
601 | 1 | ccv_nnc_tensor_free(gwo); |
602 | 1 | ccv_nnc_tensor_free(gbiaso); |
603 | 1 | ccv_nnc_tensor_free(gdwo); |
604 | 1 | ccv_nnc_tensor_free(gdbiaso); |
605 | 1 | ccv_nnc_tensor_free(h); |
606 | 1 | ccv_nnc_tensor_free(gh); |
607 | 1 | ccv_nnc_tensor_free(w); |
608 | 1 | ccv_nnc_tensor_free(g); |
609 | 1 | ccv_nnc_tensor_free(a); |
610 | 1 | ccv_nnc_tensor_free(gbias); |
611 | 1 | ccv_nnc_tensor_free(gdbias); |
612 | 1 | ccv_nnc_tensor_free(gdw); |
613 | 1 | ccv_nnc_tensor_free(gw); |
614 | 1 | ccv_nnc_tensor_free(gg); |
615 | 1 | ccv_nnc_tensor_free(ga); |
616 | 1 | ccv_nnc_tensor_free(ch); |
617 | 1 | ccv_nnc_tensor_free(cdw); |
618 | 1 | ccv_nnc_tensor_free(cdbias); |
619 | 1 | } |
620 | | |
621 | | TEST_CASE("cudnn backward convolution in half precision") |
622 | 1 | { |
623 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
624 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
625 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
626 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
627 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_BACKWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
628 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
629 | 1 | assert(cmd.backend >= 0); |
630 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, a->info, g->info); |
631 | 1 | assert(ccv_nnc_hint_verify(hint, cmd.info, a->info, g->info) == 0); |
632 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
633 | 1 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
634 | 1 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
635 | | // configure the inlets. |
636 | 1 | dsfmt_t dsfmt; |
637 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
638 | 1 | int i; |
639 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
640 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
641 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
642 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
643 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
644 | 19.2M | g->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / OUTPUT_DIM; // (OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM); |
645 | | // Copy generated matrix values over to GPU. |
646 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
647 | 1 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
648 | 1 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
649 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
650 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM), 0); |
651 | 1 | ccv_nnc_tensor_t* gdw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
652 | 1 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM), 0); |
653 | 1 | ccv_nnc_tensor_t* a16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
654 | 1 | ccv_nnc_tensor_t* g16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
655 | 1 | ccv_nnc_tensor_t* w16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
656 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, g), TENSOR_LIST(a16, w16, g16), 0); |
657 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a16, w16, g16), TENSOR_LIST(ga, gw, gg), 0); |
658 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(g, a, w), TENSOR_LIST(h, dw, dbias), 0); |
659 | | |
660 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
661 | 1 | assert(cmd.backend >= 0); |
662 | 1 | cmd.algorithm = -1; |
663 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
664 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 512 * 1024 * 1024, hint, 0, TENSOR_LIST(gg, ga, gw), TENSOR_LIST(gh, gdw, gdbias), stream_context); |
665 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(gg, ga, gw), TENSOR_LIST(gh, gdw, gdbias), stream_context)); |
666 | 1 | ccv_nnc_stream_context_wait(stream_context); |
667 | 1 | ccv_nnc_stream_context_free(stream_context); |
668 | 1 | ccv_nnc_tensor_t* ch16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
669 | 1 | ccv_nnc_tensor_t* cdw16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
670 | 1 | ccv_nnc_tensor_t* cdbias16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM), 0); |
671 | 1 | ccv_nnc_tensor_t* ch = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
672 | 1 | ccv_nnc_tensor_t* cdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
673 | 1 | ccv_nnc_tensor_t* cdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
674 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdw, gdbias), TENSOR_LIST(ch16, cdw16, cdbias16), 0); |
675 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ch16, cdw16, cdbias16), TENSOR_LIST(ch, cdw, cdbias), 0); |
676 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dw->data.f32, cdw->data.f32, INPUT_DIM * OUTPUT_DIM * KERNEL_SIZE * KERNEL_SIZE, 12, "output from cudnn should match from CPU"); |
677 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dbias->data.f32, cdbias->data.f32, OUTPUT_DIM, 5, "output from cudnn should match from CPU"); |
678 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, h->data.f32, ch->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
679 | 1 | ccv_nnc_tensor_free(h); |
680 | 1 | ccv_nnc_tensor_free(gh); |
681 | 1 | ccv_nnc_tensor_free(w); |
682 | 1 | ccv_nnc_tensor_free(g); |
683 | 1 | ccv_nnc_tensor_free(a); |
684 | 1 | ccv_nnc_tensor_free(w16); |
685 | 1 | ccv_nnc_tensor_free(g16); |
686 | 1 | ccv_nnc_tensor_free(a16); |
687 | 1 | ccv_nnc_tensor_free(gbias); |
688 | 1 | ccv_nnc_tensor_free(gdbias); |
689 | 1 | ccv_nnc_tensor_free(gdw); |
690 | 1 | ccv_nnc_tensor_free(gw); |
691 | 1 | ccv_nnc_tensor_free(gg); |
692 | 1 | ccv_nnc_tensor_free(ga); |
693 | 1 | ccv_nnc_tensor_free(ch); |
694 | 1 | ccv_nnc_tensor_free(cdw); |
695 | 1 | ccv_nnc_tensor_free(cdbias); |
696 | 1 | ccv_nnc_tensor_free(ch16); |
697 | 1 | ccv_nnc_tensor_free(cdw16); |
698 | 1 | ccv_nnc_tensor_free(cdbias16); |
699 | 1 | } |
700 | | |
701 | | TEST_CASE("cudnn backward convolution with dilation 2, 3") |
702 | 1 | { |
703 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
704 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
705 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
706 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
707 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_BACKWARD(1, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM); |
708 | 1 | cmd.info.convolution.dilation[0] = 2; |
709 | 1 | cmd.info.convolution.dilation[1] = 3; |
710 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
711 | 1 | assert(cmd.backend >= 0); |
712 | 1 | ccv_nnc_cmd_param_t modified_cmd = cmd.info; |
713 | 1 | modified_cmd.size.dim[0] = (modified_cmd.size.dim[0] - 1) * cmd.info.convolution.dilation[0] + 1; |
714 | 1 | modified_cmd.size.dim[1] = (modified_cmd.size.dim[1] - 1) * cmd.info.convolution.dilation[1] + 1; |
715 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(modified_cmd, a->info, g->info); |
716 | 1 | assert(ccv_nnc_hint_verify(hint, modified_cmd, a->info, g->info) == 0); |
717 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
718 | 1 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
719 | 1 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
720 | | // configure the inlets. |
721 | 1 | dsfmt_t dsfmt; |
722 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
723 | 1 | int i; |
724 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
725 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
726 | 2.40M | for (i = 0; i < INPUT_SIZE * INPUT_SIZE * INPUT_DIM * ccv_max(1, BATCH_SIZE); i++2.40M ) |
727 | 2.40M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
728 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
729 | 19.2M | g->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / OUTPUT_DIM; // (OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM); |
730 | | // Copy generated matrix values over to GPU. |
731 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
732 | 1 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
733 | 1 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
734 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
735 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
736 | 1 | ccv_nnc_tensor_t* gdw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
737 | 1 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM), 0); |
738 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
739 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
740 | 1 | assert(move.backend >= 0); |
741 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, g), TENSOR_LIST(ga, gw, gg), 0); |
742 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(g, a, w), TENSOR_LIST(h, dw, dbias), 0); |
743 | | |
744 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
745 | 1 | assert(cmd.backend >= 0); |
746 | 1 | cmd.algorithm = -1; |
747 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
748 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(gg, ga, gw), TENSOR_LIST(gh, gdw, gdbias), stream_context); |
749 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(gg, ga, gw), TENSOR_LIST(gh, gdw, gdbias), stream_context)); |
750 | 1 | ccv_nnc_stream_context_wait(stream_context); |
751 | 1 | ccv_nnc_stream_context_free(stream_context); |
752 | 1 | ccv_nnc_tensor_t* ch = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
753 | 1 | ccv_nnc_tensor_t* cdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
754 | 1 | ccv_nnc_tensor_t* cdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM), 0); |
755 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdw, gdbias), TENSOR_LIST(ch, cdw, cdbias), 0); |
756 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dw->data.f32, cdw->data.f32, INPUT_DIM * OUTPUT_DIM * KERNEL_SIZE * KERNEL_SIZE, 5e-1, "output from cudnn should match from CPU"); |
757 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dbias->data.f32, cdbias->data.f32, OUTPUT_DIM, 5e-1, "output from cudnn should match from CPU"); |
758 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, h->data.f32, ch->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 1e-4, "output from cudnn should match from CPU"); |
759 | 1 | ccv_nnc_tensor_free(h); |
760 | 1 | ccv_nnc_tensor_free(gh); |
761 | 1 | ccv_nnc_tensor_free(w); |
762 | 1 | ccv_nnc_tensor_free(g); |
763 | 1 | ccv_nnc_tensor_free(a); |
764 | 1 | ccv_nnc_tensor_free(gbias); |
765 | 1 | ccv_nnc_tensor_free(gdbias); |
766 | 1 | ccv_nnc_tensor_free(gdw); |
767 | 1 | ccv_nnc_tensor_free(gw); |
768 | 1 | ccv_nnc_tensor_free(gg); |
769 | 1 | ccv_nnc_tensor_free(ga); |
770 | 1 | ccv_nnc_tensor_free(ch); |
771 | 1 | ccv_nnc_tensor_free(cdw); |
772 | 1 | ccv_nnc_tensor_free(cdbias); |
773 | 1 | } |
774 | | |
775 | | TEST_CASE("compare batch norm with cudnn") |
776 | 1 | { |
777 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
778 | 1 | ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
779 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
780 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
781 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host x"); |
782 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
783 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
784 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host y"); |
785 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "scale"); |
786 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "bias"); |
787 | 1 | ccv_nnc_tensor_symbol_t bmean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
788 | 1 | ccv_nnc_tensor_symbol_t bvar = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
789 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
790 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
791 | 1 | ccv_nnc_tensor_symbol_t bmean_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
792 | 1 | ccv_nnc_tensor_symbol_t bvar_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
793 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_mean"); |
794 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_inv_std"); |
795 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
796 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(scale), "set_scale"); |
797 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(bias), "set_bias"); |
798 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(bx, scale, bias, bmean, bvar), TENSOR_SYMBOL_LIST(by, bmean_out, bvar_out, saved_mean, saved_inv_std), "batch_norm"); |
799 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
800 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
801 | 1 | ccv_nnc_graph_t* graph = 0; |
802 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
803 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
804 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
805 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
806 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
807 | 1 | dsfmt_t dsfmt; |
808 | 1 | float xdata[2 * 2 * 2 * 10]; |
809 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
810 | 1 | int i; |
811 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
812 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
813 | 80 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
814 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
815 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
816 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
817 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
818 | 1 | ccv_nnc_graph_free(graph); |
819 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
820 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
821 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
822 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "scale"); |
823 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "bias"); |
824 | 1 | ccv_nnc_tensor_symbol_t cmean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
825 | 1 | ccv_nnc_tensor_symbol_t cvar = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
826 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
827 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
828 | 1 | ccv_nnc_tensor_symbol_t cmean_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
829 | 1 | ccv_nnc_tensor_symbol_t cvar_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
830 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_mean"); |
831 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_inv_std"); |
832 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(cscale), "set_scale"); |
833 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(cbias), "set_bias"); |
834 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(cx, cscale, cbias, cmean, cvar), TENSOR_SYMBOL_LIST(cy, cmean_out, cvar_out, csaved_mean, csaved_inv_std), "batch_norm"); |
835 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
836 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
837 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
838 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
839 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
840 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
841 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 2 * 2 * 10); |
842 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
843 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
844 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 2 * 2 * 10, 1e-5, "batch norm result from cudnn should match the one from reference implementation"); |
845 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
846 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
847 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
848 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
849 | 1 | ccv_nnc_graph_free(cpu_graph); |
850 | 1 | } |
851 | | |
852 | | TEST_CASE("compare batch norm with cudnn in half precision") |
853 | 1 | { |
854 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
855 | 1 | ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
856 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
857 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
858 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host x"); |
859 | 1 | ccv_nnc_tensor_symbol_t x16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 2, 2, 2, 10), "x in half precision"); |
860 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 2, 2, 2, 10), "x"); |
861 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 2, 2, 2, 10), "y"); |
862 | 1 | ccv_nnc_tensor_symbol_t y16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 2, 2, 2, 10), "y in half precision"); |
863 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host y"); |
864 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "scale"); |
865 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "bias"); |
866 | 1 | ccv_nnc_tensor_symbol_t bmean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
867 | 1 | ccv_nnc_tensor_symbol_t bvar = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
868 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
869 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
870 | 1 | ccv_nnc_tensor_symbol_t bmean_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
871 | 1 | ccv_nnc_tensor_symbol_t bvar_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
872 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_mean"); |
873 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_inv_std"); |
874 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(x16), "convert x"); |
875 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x16), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
876 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(scale), "set_scale"); |
877 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(bias), "set_bias"); |
878 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(bx, scale, bias, bmean, bvar), TENSOR_SYMBOL_LIST(by, bmean_out, bvar_out, saved_mean, saved_inv_std), "batch_norm"); |
879 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y16), "transfer y"); |
880 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(y16), TENSOR_SYMBOL_LIST(y), "convert y"); |
881 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
882 | 1 | ccv_nnc_graph_t* graph = 0; |
883 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
884 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
885 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
886 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
887 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
888 | 1 | dsfmt_t dsfmt; |
889 | 1 | float xdata[2 * 2 * 2 * 10]; |
890 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
891 | 1 | int i; |
892 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
893 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
894 | 80 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
895 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
896 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
897 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
898 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
899 | 1 | ccv_nnc_graph_free(graph); |
900 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
901 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
902 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
903 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "scale"); |
904 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "bias"); |
905 | 1 | ccv_nnc_tensor_symbol_t cmean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
906 | 1 | ccv_nnc_tensor_symbol_t cvar = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
907 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
908 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
909 | 1 | ccv_nnc_tensor_symbol_t cmean_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
910 | 1 | ccv_nnc_tensor_symbol_t cvar_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
911 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_mean"); |
912 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_inv_std"); |
913 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(cscale), "set_scale"); |
914 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(cbias), "set_bias"); |
915 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(cx, cscale, cbias, cmean, cvar), TENSOR_SYMBOL_LIST(cy, cmean_out, cvar_out, csaved_mean, csaved_inv_std), "batch_norm"); |
916 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
917 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
918 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
919 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
920 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
921 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
922 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 2 * 2 * 10); |
923 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
924 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
925 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 2 * 2 * 10, 1e-3, "batch norm result from cudnn should match the one from reference implementation"); |
926 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
927 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
928 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
929 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
930 | 1 | ccv_nnc_graph_free(cpu_graph); |
931 | 1 | } |
932 | | |
933 | | TEST_CASE("compare batch norm gradient with cudnn") |
934 | 1 | { |
935 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
936 | 1 | ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
937 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
938 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
939 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
940 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
941 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "scale"); |
942 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "bias"); |
943 | 1 | ccv_nnc_tensor_symbol_t bmean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
944 | 1 | ccv_nnc_tensor_symbol_t bvar = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
945 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
946 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
947 | 1 | ccv_nnc_tensor_symbol_t bmean_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
948 | 1 | ccv_nnc_tensor_symbol_t bvar_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
949 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_mean"); |
950 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_inv_std"); |
951 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(scale), "set_scale"); |
952 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(bias), "set_bias"); |
953 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(bx, scale, bias, bmean, bvar), TENSOR_SYMBOL_LIST(by, bmean_out, bvar_out, saved_mean, saved_inv_std), "batch_norm"); |
954 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
955 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx, scale, bias), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
956 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
957 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
958 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
959 | 1 | ccv_nnc_graph_t* graph = 0; |
960 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
961 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
962 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
963 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
964 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
965 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
966 | 1 | dsfmt_t dsfmt; |
967 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
968 | 1 | int i; |
969 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
970 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
971 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
972 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
973 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
974 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
975 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
976 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
977 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
978 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
979 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
980 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
981 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
982 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
983 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
984 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
985 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
986 | 1 | ccv_nnc_graph_free(graph); |
987 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
988 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
989 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
990 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "scale"); |
991 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "bias"); |
992 | 1 | ccv_nnc_tensor_symbol_t cmean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
993 | 1 | ccv_nnc_tensor_symbol_t cvar = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
994 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
995 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
996 | 1 | ccv_nnc_tensor_symbol_t cmean_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
997 | 1 | ccv_nnc_tensor_symbol_t cvar_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
998 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_mean"); |
999 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_inv_std"); |
1000 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(cscale), "set_scale"); |
1001 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(cbias), "set_bias"); |
1002 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(cx, cscale, cbias, cmean, cvar), TENSOR_SYMBOL_LIST(cy, cmean_out, cvar_out, csaved_mean, csaved_inv_std), "batch_norm"); |
1003 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1004 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx, cscale, cbias), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1005 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1006 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1007 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1008 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1009 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1010 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1011 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1012 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1013 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1014 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1015 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1016 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1017 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1018 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "batch norm gradient result from cudnn should match the one from reference implementation"); |
1019 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1020 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1021 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1022 | 1 | ccv_nnc_graph_free(cpu_graph); |
1023 | 1 | ccv_nnc_tensor_free(x_tensor); |
1024 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1025 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1026 | 1 | } |
1027 | | |
1028 | | TEST_CASE("compare batch norm gradient with cudnn in half precision") |
1029 | 1 | { |
1030 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1031 | 1 | ccv_nnc_cmd_ok(CCV_NNC_BATCH_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1032 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1033 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1034 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 2, 2, 2, 10), "x"); |
1035 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 2, 2, 2, 10), "y"); |
1036 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "scale"); |
1037 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "bias"); |
1038 | 1 | ccv_nnc_tensor_symbol_t bmean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
1039 | 1 | ccv_nnc_tensor_symbol_t bvar = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
1040 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
1041 | 1 | ccv_nnc_tensor_symbol_set_flags(symbolic_graph, bvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
1042 | 1 | ccv_nnc_tensor_symbol_t bmean_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "mean"); |
1043 | 1 | ccv_nnc_tensor_symbol_t bvar_out = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "var"); |
1044 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_mean"); |
1045 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10), "saved_inv_std"); |
1046 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(scale), "set_scale"); |
1047 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(bias), "set_bias"); |
1048 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(bx, scale, bias, bmean, bvar), TENSOR_SYMBOL_LIST(by, bmean_out, bvar_out, saved_mean, saved_inv_std), "batch_norm"); |
1049 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1050 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx, scale, bias), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1051 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1052 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1053 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1054 | 1 | ccv_nnc_graph_t* graph = 0; |
1055 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1056 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1057 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1058 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1059 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1060 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1061 | 1 | dsfmt_t dsfmt; |
1062 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1063 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 2, 2, 2, 10), 0); |
1064 | 1 | int i; |
1065 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1066 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1067 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1068 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
1069 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(bx_tensor), 0); |
1070 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1071 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1072 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 2, 2, 2, 10), 0); |
1073 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1074 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1075 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1076 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
1077 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dby_tensor), 0); |
1078 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1079 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1080 | 1 | ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 2, 2, 2, 10), 0); |
1081 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1082 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx16_tensor), 0); |
1083 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0); |
1084 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1085 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1086 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1087 | 1 | ccv_nnc_graph_free(graph); |
1088 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1089 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1090 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1091 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "scale"); |
1092 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "bias"); |
1093 | 1 | ccv_nnc_tensor_symbol_t cmean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
1094 | 1 | ccv_nnc_tensor_symbol_t cvar = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
1095 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cmean, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
1096 | 1 | ccv_nnc_tensor_symbol_set_flags(cpu_symbolic_graph, cvar, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); |
1097 | 1 | ccv_nnc_tensor_symbol_t cmean_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "mean"); |
1098 | 1 | ccv_nnc_tensor_symbol_t cvar_out = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "var"); |
1099 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_mean"); |
1100 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 10), "saved_inv_std"); |
1101 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(1), 0, 0, TENSOR_SYMBOL_LIST(cscale), "set_scale"); |
1102 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_SET_FORWARD(0), 0, 0, TENSOR_SYMBOL_LIST(cbias), "set_bias"); |
1103 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_BATCH_NORM_FORWARD(1e-4, 0, 0.9, 0, 1, 2), TENSOR_SYMBOL_LIST(cx, cscale, cbias, cmean, cvar), TENSOR_SYMBOL_LIST(cy, cmean_out, cvar_out, csaved_mean, csaved_inv_std), "batch_norm"); |
1104 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1105 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx, cscale, cbias), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1106 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1107 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1108 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1109 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1110 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1111 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1112 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1113 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1114 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1115 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1116 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1117 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1118 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1119 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dx_tensor->data.f32, dcx_tensor->data.f32, 2 * 2 * 2 * 10, 2e-3, "batch norm result from cudnn should match the one from reference implementation"); |
1120 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1121 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1122 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1123 | 1 | ccv_nnc_graph_free(cpu_graph); |
1124 | 1 | ccv_nnc_tensor_free(x_tensor); |
1125 | 1 | ccv_nnc_tensor_free(x16_tensor); |
1126 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1127 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
1128 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1129 | 1 | ccv_nnc_tensor_free(dx16_tensor); |
1130 | 1 | } |
1131 | | |
1132 | | TEST_CASE("compare layer norm with cudnn") |
1133 | 1 | { |
1134 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1135 | 1 | ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1136 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1137 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1138 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host x"); |
1139 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
1140 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
1141 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host y"); |
1142 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "scale"); |
1143 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "bias"); |
1144 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_mean"); |
1145 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
1146 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
1147 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "layer_norm"); |
1148 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
1149 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1150 | 1 | ccv_nnc_graph_t* graph = 0; |
1151 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1152 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1153 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1154 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1155 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1156 | 1 | dsfmt_t dsfmt; |
1157 | 1 | float xdata[2 * 2 * 2 * 10]; |
1158 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
1159 | 1 | int i; |
1160 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1161 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1162 | 80 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1163 | 1 | float scaledata[1 * 2 * 2 * 10]; |
1164 | 1 | float biasdata[1 * 2 * 2 * 10]; |
1165 | 41 | for (i = 0; i < 1 * 2 * 2 * 10; i++40 ) |
1166 | 40 | { |
1167 | 40 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1168 | 40 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1169 | 40 | } |
1170 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1171 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1172 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1173 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1174 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
1175 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1176 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1177 | 1 | ccv_nnc_graph_free(graph); |
1178 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1179 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1180 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1181 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "scale"); |
1182 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "bias"); |
1183 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_mean"); |
1184 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
1185 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "layer_norm"); |
1186 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1187 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1188 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1189 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1190 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1191 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1192 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 2 * 2 * 10); |
1193 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1194 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 2 * 2 * 10); |
1195 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1196 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 2 * 2 * 10); |
1197 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1198 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
1199 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 2 * 2 * 10, 1e-5, "layer norm result from cudnn should match the one from reference implementation"); |
1200 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1201 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1202 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1203 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1204 | 1 | ccv_nnc_graph_free(cpu_graph); |
1205 | 1 | } |
1206 | | |
1207 | | TEST_CASE("compare layer norm gradient with cudnn") |
1208 | 1 | { |
1209 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1210 | 1 | ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1211 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1212 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1213 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
1214 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
1215 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "scale"); |
1216 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "bias"); |
1217 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_mean"); |
1218 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
1219 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "layer_norm"); |
1220 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1221 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx, scale, bias), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1222 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1223 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1224 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1225 | 1 | ccv_nnc_graph_t* graph = 0; |
1226 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1227 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1228 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1229 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1230 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1231 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1232 | 1 | dsfmt_t dsfmt; |
1233 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1234 | 1 | int i; |
1235 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1236 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1237 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1238 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1239 | 1 | float scaledata[1 * 2 * 2 * 10]; |
1240 | 1 | float biasdata[1 * 2 * 2 * 10]; |
1241 | 41 | for (i = 0; i < 1 * 2 * 2 * 10; i++40 ) |
1242 | 40 | { |
1243 | 40 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1244 | 40 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1245 | 40 | } |
1246 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1247 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1248 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1249 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1250 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1251 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1252 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1253 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1254 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
1255 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1256 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1257 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1258 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
1259 | 1 | ccv_nnc_tensor_t* const dbscale_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, scale)); |
1260 | 1 | ccv_nnc_tensor_t* const dbbias_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bias)); |
1261 | 1 | ccv_nnc_tensor_t* const dscale_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1262 | 1 | ccv_nnc_tensor_t* const dbias_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1263 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbscale_tensor, dbbias_tensor), TENSOR_LIST(dscale_tensor, dbias_tensor), 0); |
1264 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1265 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1266 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1267 | 1 | ccv_nnc_graph_free(graph); |
1268 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1269 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1270 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1271 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "scale"); |
1272 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "bias"); |
1273 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_mean"); |
1274 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
1275 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "layer_norm"); |
1276 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1277 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx, cscale, cbias), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1278 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1279 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1280 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1281 | 1 | ccv_nnc_tensor_symbol_t dcscale = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cscale); |
1282 | 1 | ccv_nnc_tensor_symbol_t dcbias = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cbias); |
1283 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1284 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1285 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1286 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1287 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1288 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1289 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1290 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1291 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1292 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 2 * 2 * 10); |
1293 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1294 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 2 * 2 * 10); |
1295 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1296 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1297 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
1298 | 1 | ccv_nnc_tensor_t* const dcscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcscale); |
1299 | 1 | REQUIRE_TENSOR_EQ(dscale_tensor, dcscale_tensor, "layer norm scale gradient result from cudnn should match the one from reference implementation"); |
1300 | 1 | ccv_nnc_tensor_t* const dcbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcbias); |
1301 | 1 | REQUIRE_TENSOR_EQ(dbias_tensor, dcbias_tensor, "layer norm bias gradient result from cudnn should match the one from reference implementation"); |
1302 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1303 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1304 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1305 | 1 | ccv_nnc_graph_free(cpu_graph); |
1306 | 1 | ccv_nnc_tensor_free(x_tensor); |
1307 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1308 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1309 | 1 | ccv_nnc_tensor_free(dscale_tensor); |
1310 | 1 | ccv_nnc_tensor_free(dbias_tensor); |
1311 | 1 | } |
1312 | | |
1313 | | TEST_CASE("compare layer norm only gradient with cudnn") |
1314 | 1 | { |
1315 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1316 | 1 | ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1317 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1318 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1319 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
1320 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
1321 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "scale"); |
1322 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "bias"); |
1323 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_mean"); |
1324 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
1325 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "layer_norm"); |
1326 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1327 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1328 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1329 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1330 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1331 | 1 | ccv_nnc_graph_t* graph = 0; |
1332 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1333 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1334 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1335 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1336 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1337 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1338 | 1 | dsfmt_t dsfmt; |
1339 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1340 | 1 | int i; |
1341 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1342 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1343 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1344 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1345 | 1 | float scaledata[1 * 2 * 2 * 10]; |
1346 | 1 | float biasdata[1 * 2 * 2 * 10]; |
1347 | 41 | for (i = 0; i < 1 * 2 * 2 * 10; i++40 ) |
1348 | 40 | { |
1349 | 40 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1350 | 40 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1351 | 40 | } |
1352 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1353 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
1354 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1355 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1356 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1357 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1358 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1359 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1360 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
1361 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1362 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1363 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1364 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
1365 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1366 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1367 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1368 | 1 | ccv_nnc_graph_free(graph); |
1369 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1370 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1371 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1372 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "scale"); |
1373 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "bias"); |
1374 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_mean"); |
1375 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
1376 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "layer_norm"); |
1377 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1378 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1379 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1380 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1381 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1382 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1383 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1384 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1385 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1386 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1387 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1388 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1389 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1390 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1391 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 2 * 2 * 10); |
1392 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1393 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 2 * 2 * 10); |
1394 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1395 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1396 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
1397 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1398 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1399 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1400 | 1 | ccv_nnc_graph_free(cpu_graph); |
1401 | 1 | ccv_nnc_tensor_free(x_tensor); |
1402 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1403 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1404 | 1 | } |
1405 | | |
1406 | | TEST_CASE("compare layer norm with cudnn without scale / bias") |
1407 | 1 | { |
1408 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1409 | 1 | ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1410 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1411 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1412 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host x"); |
1413 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
1414 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
1415 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host y"); |
1416 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_mean"); |
1417 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
1418 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
1419 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "layer_norm"); |
1420 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
1421 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1422 | 1 | ccv_nnc_graph_t* graph = 0; |
1423 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1424 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1425 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1426 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1427 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1428 | 1 | dsfmt_t dsfmt; |
1429 | 1 | float xdata[2 * 2 * 2 * 10]; |
1430 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
1431 | 1 | int i; |
1432 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1433 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1434 | 80 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1435 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1436 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
1437 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1438 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1439 | 1 | ccv_nnc_graph_free(graph); |
1440 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1441 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1442 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1443 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_mean"); |
1444 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
1445 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "layer_norm"); |
1446 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1447 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1448 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1449 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1450 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1451 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1452 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 2 * 2 * 10); |
1453 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1454 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
1455 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 2 * 2 * 10, 1e-5, "layer norm result from cudnn should match the one from reference implementation"); |
1456 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1457 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1458 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1459 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1460 | 1 | ccv_nnc_graph_free(cpu_graph); |
1461 | 1 | } |
1462 | | |
1463 | | TEST_CASE("compare layer norm gradient with cudnn without scale / bias") |
1464 | 1 | { |
1465 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1466 | 1 | ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1467 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1468 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1469 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
1470 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
1471 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_mean"); |
1472 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
1473 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "layer_norm"); |
1474 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1475 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1476 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1477 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1478 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1479 | 1 | ccv_nnc_graph_t* graph = 0; |
1480 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1481 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1482 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1483 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1484 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1485 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1486 | 1 | dsfmt_t dsfmt; |
1487 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1488 | 1 | int i; |
1489 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1490 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1491 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1492 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1493 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1494 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1495 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1496 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1497 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1498 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
1499 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1500 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1501 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1502 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
1503 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1504 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1505 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1506 | 1 | ccv_nnc_graph_free(graph); |
1507 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1508 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1509 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1510 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_mean"); |
1511 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
1512 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 0, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "layer_norm"); |
1513 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1514 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1515 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1516 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1517 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1518 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1519 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1520 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1521 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1522 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1523 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1524 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1525 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1526 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1527 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1528 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
1529 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1530 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1531 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1532 | 1 | ccv_nnc_graph_free(cpu_graph); |
1533 | 1 | ccv_nnc_tensor_free(x_tensor); |
1534 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1535 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1536 | 1 | } |
1537 | | |
1538 | | TEST_CASE("compare layer norm only gradient with cudnn without scale / bias") |
1539 | 1 | { |
1540 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1541 | 1 | ccv_nnc_cmd_ok(CCV_NNC_LAYER_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1542 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1543 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1544 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
1545 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
1546 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_mean"); |
1547 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
1548 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "layer_norm"); |
1549 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1550 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1551 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1552 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1553 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1554 | 1 | ccv_nnc_graph_t* graph = 0; |
1555 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1556 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1557 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1558 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1559 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1560 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1561 | 1 | dsfmt_t dsfmt; |
1562 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1563 | 1 | int i; |
1564 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1565 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1566 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1567 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1568 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1569 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1570 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1571 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
1572 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1573 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
1574 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1575 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1576 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
1577 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
1578 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1579 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1580 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1581 | 1 | ccv_nnc_graph_free(graph); |
1582 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1583 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
1584 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
1585 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_mean"); |
1586 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
1587 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_LAYER_NORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "layer_norm"); |
1588 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1589 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1590 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1591 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1592 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1593 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1594 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1595 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1596 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1597 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1598 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1599 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1600 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
1601 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1602 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1603 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
1604 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1605 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1606 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1607 | 1 | ccv_nnc_graph_free(cpu_graph); |
1608 | 1 | ccv_nnc_tensor_free(x_tensor); |
1609 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1610 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1611 | 1 | } |
1612 | | |
1613 | | TEST_CASE("compare group norm with cudnn") |
1614 | 1 | { |
1615 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1616 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1617 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1618 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1619 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host x"); |
1620 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
1621 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
1622 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host y"); |
1623 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 2, 10), "scale"); |
1624 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 2, 10), "bias"); |
1625 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_mean"); |
1626 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_inv_std"); |
1627 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
1628 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
1629 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
1630 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1631 | 1 | ccv_nnc_graph_t* graph = 0; |
1632 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1633 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1634 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1635 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1636 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1637 | 1 | dsfmt_t dsfmt; |
1638 | 1 | float xdata[2 * 16 * 2 * 10]; |
1639 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
1640 | 1 | int i; |
1641 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1642 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1643 | 640 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1644 | 1 | float scaledata[1 * 16 * 2 * 10]; |
1645 | 1 | float biasdata[1 * 16 * 2 * 10]; |
1646 | 321 | for (i = 0; i < 1 * 16 * 2 * 10; i++320 ) |
1647 | 320 | { |
1648 | 320 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1649 | 320 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1650 | 320 | } |
1651 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1652 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1653 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1654 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1655 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
1656 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1657 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1658 | 1 | ccv_nnc_graph_free(graph); |
1659 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1660 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
1661 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
1662 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), "scale"); |
1663 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), "bias"); |
1664 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_mean"); |
1665 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_inv_std"); |
1666 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
1667 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1668 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1669 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1670 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1671 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1672 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1673 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 16 * 2 * 10); |
1674 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1675 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 16 * 2 * 10); |
1676 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1677 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 16 * 2 * 10); |
1678 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1679 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
1680 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 16 * 2 * 10, 1e-5, "layer norm result from cudnn should match the one from reference implementation"); |
1681 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1682 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1683 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1684 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1685 | 1 | ccv_nnc_graph_free(cpu_graph); |
1686 | 1 | } |
1687 | | |
1688 | | TEST_CASE("compare group norm gradient with cudnn") |
1689 | 1 | { |
1690 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1691 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1692 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1693 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1694 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
1695 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
1696 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 2, 10), "scale"); |
1697 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 2, 10), "bias"); |
1698 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_mean"); |
1699 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_inv_std"); |
1700 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
1701 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1702 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx, scale, bias), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1703 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1704 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1705 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1706 | 1 | ccv_nnc_graph_t* graph = 0; |
1707 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1708 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1709 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1710 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1711 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1712 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1713 | 1 | dsfmt_t dsfmt; |
1714 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1715 | 1 | int i; |
1716 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1717 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1718 | 640 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1719 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1720 | 1 | float scaledata[1 * 16 * 2 * 10]; |
1721 | 1 | float biasdata[1 * 16 * 2 * 10]; |
1722 | 321 | for (i = 0; i < 1 * 16 * 2 * 10; i++320 ) |
1723 | 320 | { |
1724 | 320 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1725 | 320 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1726 | 320 | } |
1727 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1728 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1729 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1730 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1731 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1732 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1733 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1734 | 640 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1735 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
1736 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1737 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1738 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1739 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
1740 | 1 | ccv_nnc_tensor_t* const dbscale_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, scale)); |
1741 | 1 | ccv_nnc_tensor_t* const dbbias_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bias)); |
1742 | 1 | ccv_nnc_tensor_t* const dscale_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1743 | 1 | ccv_nnc_tensor_t* const dbias_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1744 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbscale_tensor, dbbias_tensor), TENSOR_LIST(dscale_tensor, dbias_tensor), 0); |
1745 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1746 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1747 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1748 | 1 | ccv_nnc_graph_free(graph); |
1749 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1750 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
1751 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
1752 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), "scale"); |
1753 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), "bias"); |
1754 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_mean"); |
1755 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_inv_std"); |
1756 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
1757 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1758 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx, cscale, cbias), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1759 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1760 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1761 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1762 | 1 | ccv_nnc_tensor_symbol_t dcscale = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cscale); |
1763 | 1 | ccv_nnc_tensor_symbol_t dcbias = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cbias); |
1764 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1765 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1766 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1767 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1768 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1769 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
1770 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1771 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
1772 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1773 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 16 * 2 * 10); |
1774 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1775 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 16 * 2 * 10); |
1776 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1777 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1778 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
1779 | 1 | ccv_nnc_tensor_t* const dcscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcscale); |
1780 | 1 | REQUIRE_TENSOR_EQ(dscale_tensor, dcscale_tensor, "layer norm scale gradient result from cudnn should match the one from reference implementation"); |
1781 | 1 | ccv_nnc_tensor_t* const dcbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcbias); |
1782 | 1 | REQUIRE_TENSOR_EQ(dbias_tensor, dcbias_tensor, "layer norm bias gradient result from cudnn should match the one from reference implementation"); |
1783 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1784 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1785 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1786 | 1 | ccv_nnc_graph_free(cpu_graph); |
1787 | 1 | ccv_nnc_tensor_free(x_tensor); |
1788 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1789 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1790 | 1 | ccv_nnc_tensor_free(dscale_tensor); |
1791 | 1 | ccv_nnc_tensor_free(dbias_tensor); |
1792 | 1 | } |
1793 | | |
1794 | | TEST_CASE("compare group norm only gradient with cudnn") |
1795 | 1 | { |
1796 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1797 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1798 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1799 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1800 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
1801 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
1802 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 2, 10), "scale"); |
1803 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 2, 10), "bias"); |
1804 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_mean"); |
1805 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_inv_std"); |
1806 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
1807 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1808 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1809 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1810 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1811 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1812 | 1 | ccv_nnc_graph_t* graph = 0; |
1813 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1814 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1815 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1816 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1817 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1818 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1819 | 1 | dsfmt_t dsfmt; |
1820 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1821 | 1 | int i; |
1822 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1823 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1824 | 640 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1825 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1826 | 1 | float scaledata[1 * 16 * 2 * 10]; |
1827 | 1 | float biasdata[1 * 16 * 2 * 10]; |
1828 | 321 | for (i = 0; i < 1 * 16 * 2 * 10; i++320 ) |
1829 | 320 | { |
1830 | 320 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1831 | 320 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1832 | 320 | } |
1833 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1834 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), 0); |
1835 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1836 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1837 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1838 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
1839 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1840 | 640 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
1841 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
1842 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1843 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
1844 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1845 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
1846 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1847 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1848 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1849 | 1 | ccv_nnc_graph_free(graph); |
1850 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1851 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
1852 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
1853 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), "scale"); |
1854 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 2, 10), "bias"); |
1855 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_mean"); |
1856 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_inv_std"); |
1857 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
1858 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1859 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
1860 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1861 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
1862 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
1863 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1864 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1865 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1866 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1867 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1868 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
1869 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
1870 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
1871 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1872 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 16 * 2 * 10); |
1873 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1874 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 16 * 2 * 10); |
1875 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1876 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
1877 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
1878 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1879 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1880 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1881 | 1 | ccv_nnc_graph_free(cpu_graph); |
1882 | 1 | ccv_nnc_tensor_free(x_tensor); |
1883 | 1 | ccv_nnc_tensor_free(dy_tensor); |
1884 | 1 | ccv_nnc_tensor_free(dx_tensor); |
1885 | 1 | } |
1886 | | |
1887 | | TEST_CASE("compare group norm and reduce HW with cudnn") |
1888 | 1 | { |
1889 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1890 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1891 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1892 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1893 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host x"); |
1894 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
1895 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
1896 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host y"); |
1897 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 1, 1), "scale"); |
1898 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 1, 1), "bias"); |
1899 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_mean"); |
1900 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_inv_std"); |
1901 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
1902 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
1903 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
1904 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1905 | 1 | ccv_nnc_graph_t* graph = 0; |
1906 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1907 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1908 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1909 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1910 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1911 | 1 | dsfmt_t dsfmt; |
1912 | 1 | float xdata[2 * 16 * 2 * 10]; |
1913 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
1914 | 1 | int i; |
1915 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1916 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1917 | 640 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1918 | 1 | float scaledata[1 * 16]; |
1919 | 1 | float biasdata[1 * 16]; |
1920 | 17 | for (i = 0; i < 1 * 16; i++16 ) |
1921 | 16 | { |
1922 | 16 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1923 | 16 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
1924 | 16 | } |
1925 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), 0); |
1926 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), 0); |
1927 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
1928 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
1929 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
1930 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
1931 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
1932 | 1 | ccv_nnc_graph_free(graph); |
1933 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1934 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
1935 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
1936 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), "scale"); |
1937 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), "bias"); |
1938 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_mean"); |
1939 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_inv_std"); |
1940 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
1941 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1942 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
1943 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
1944 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
1945 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
1946 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
1947 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 16 * 2 * 10); |
1948 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
1949 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 16); |
1950 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
1951 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 16); |
1952 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
1953 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
1954 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 16 * 2 * 10, 1e-5, "layer norm result from cudnn should match the one from reference implementation"); |
1955 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
1956 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
1957 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
1958 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
1959 | 1 | ccv_nnc_graph_free(cpu_graph); |
1960 | 1 | } |
1961 | | |
1962 | | TEST_CASE("compare group norm gradient and reduce HW with cudnn") |
1963 | 1 | { |
1964 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1965 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
1966 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
1967 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1968 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
1969 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
1970 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 1, 1), "scale"); |
1971 | 1 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 16, 1, 1), "bias"); |
1972 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_mean"); |
1973 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_inv_std"); |
1974 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale, bias), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
1975 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1976 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx, scale, bias), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
1977 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1978 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
1979 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
1980 | 1 | ccv_nnc_graph_t* graph = 0; |
1981 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
1982 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
1983 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
1984 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
1985 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
1986 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
1987 | 1 | dsfmt_t dsfmt; |
1988 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
1989 | 1 | int i; |
1990 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
1991 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
1992 | 640 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
1993 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
1994 | 1 | float scaledata[1 * 16]; |
1995 | 1 | float biasdata[1 * 16]; |
1996 | 17 | for (i = 0; i < 1 * 16; i++16 ) |
1997 | 16 | { |
1998 | 16 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
1999 | 16 | biasdata[i] = dsfmt_genrand_open_close(&dsfmt); |
2000 | 16 | } |
2001 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), 0); |
2002 | 1 | ccv_nnc_tensor_t bias_tensor = ccv_nnc_tensor(biasdata, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), 0); |
2003 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor, &bias_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale), ccv_nnc_tensor_from_symbol(tensor_arena, bias)), 0); |
2004 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2005 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2006 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2007 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2008 | 640 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2009 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2010 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2011 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2012 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2013 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2014 | 1 | ccv_nnc_tensor_t* const dbscale_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, scale)); |
2015 | 1 | ccv_nnc_tensor_t* const dbbias_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bias)); |
2016 | 1 | ccv_nnc_tensor_t* const dscale_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), 0); |
2017 | 1 | ccv_nnc_tensor_t* const dbias_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), 0); |
2018 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbscale_tensor, dbbias_tensor), TENSOR_LIST(dscale_tensor, dbias_tensor), 0); |
2019 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2020 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2021 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2022 | 1 | ccv_nnc_graph_free(graph); |
2023 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2024 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
2025 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
2026 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), "scale"); |
2027 | 1 | ccv_nnc_tensor_symbol_t cbias = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 16, 1, 1), "bias"); |
2028 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_mean"); |
2029 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_inv_std"); |
2030 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale, cbias), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
2031 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2032 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx, cscale, cbias), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2033 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2034 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2035 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2036 | 1 | ccv_nnc_tensor_symbol_t dcscale = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cscale); |
2037 | 1 | ccv_nnc_tensor_symbol_t dcbias = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cbias); |
2038 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2039 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2040 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2041 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2042 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2043 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2044 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2045 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2046 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
2047 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 16); |
2048 | 1 | ccv_nnc_tensor_t* const cbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cbias); |
2049 | 1 | memcpy(cbias_tensor->data.f32, biasdata, sizeof(float) * 1 * 16); |
2050 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2051 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2052 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
2053 | 1 | ccv_nnc_tensor_t* const dcscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcscale); |
2054 | 1 | REQUIRE_TENSOR_EQ(dscale_tensor, dcscale_tensor, "layer norm scale gradient result from cudnn should match the one from reference implementation"); |
2055 | 1 | ccv_nnc_tensor_t* const dcbias_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcbias); |
2056 | 1 | REQUIRE_TENSOR_EQ(dbias_tensor, dcbias_tensor, "layer norm bias gradient result from cudnn should match the one from reference implementation"); |
2057 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2058 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2059 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2060 | 1 | ccv_nnc_graph_free(cpu_graph); |
2061 | 1 | ccv_nnc_tensor_free(x_tensor); |
2062 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2063 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2064 | 1 | ccv_nnc_tensor_free(dscale_tensor); |
2065 | 1 | ccv_nnc_tensor_free(dbias_tensor); |
2066 | 1 | } |
2067 | | |
2068 | | TEST_CASE("compare group norm with cudnn without scale / bias") |
2069 | 1 | { |
2070 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2071 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2072 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2073 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2074 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host x"); |
2075 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
2076 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
2077 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host y"); |
2078 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_mean"); |
2079 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_inv_std"); |
2080 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
2081 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
2082 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
2083 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2084 | 1 | ccv_nnc_graph_t* graph = 0; |
2085 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2086 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2087 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2088 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2089 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2090 | 1 | dsfmt_t dsfmt; |
2091 | 1 | float xdata[2 * 16 * 2 * 10]; |
2092 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
2093 | 1 | int i; |
2094 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2095 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2096 | 640 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
2097 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2098 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
2099 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2100 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2101 | 1 | ccv_nnc_graph_free(graph); |
2102 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2103 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
2104 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
2105 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_mean"); |
2106 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_inv_std"); |
2107 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
2108 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2109 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2110 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2111 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2112 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2113 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2114 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 16 * 2 * 10); |
2115 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2116 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
2117 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 16 * 2 * 10, 1e-5, "layer norm result from cudnn should match the one from reference implementation"); |
2118 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2119 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2120 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2121 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2122 | 1 | ccv_nnc_graph_free(cpu_graph); |
2123 | 1 | } |
2124 | | |
2125 | | TEST_CASE("compare group norm gradient with cudnn without scale / bias") |
2126 | 1 | { |
2127 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2128 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2129 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2130 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2131 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
2132 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
2133 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_mean"); |
2134 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_inv_std"); |
2135 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
2136 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2137 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2138 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2139 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2140 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2141 | 1 | ccv_nnc_graph_t* graph = 0; |
2142 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2143 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2144 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2145 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2146 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2147 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2148 | 1 | dsfmt_t dsfmt; |
2149 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2150 | 1 | int i; |
2151 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2152 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2153 | 640 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2154 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2155 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2156 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2157 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2158 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2159 | 640 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2160 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2161 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2162 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2163 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2164 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2165 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2166 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2167 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2168 | 1 | ccv_nnc_graph_free(graph); |
2169 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2170 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
2171 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
2172 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_mean"); |
2173 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_inv_std"); |
2174 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
2175 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2176 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2177 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2178 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2179 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2180 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2181 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2182 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2183 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2184 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2185 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2186 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2187 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2188 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2189 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2190 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
2191 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2192 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2193 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2194 | 1 | ccv_nnc_graph_free(cpu_graph); |
2195 | 1 | ccv_nnc_tensor_free(x_tensor); |
2196 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2197 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2198 | 1 | } |
2199 | | |
2200 | | TEST_CASE("compare group norm only gradient with cudnn without scale / bias") |
2201 | 1 | { |
2202 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2203 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2204 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2205 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2206 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
2207 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
2208 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_mean"); |
2209 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2, 10), "saved_inv_std"); |
2210 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
2211 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2212 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2213 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2214 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2215 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2216 | 1 | ccv_nnc_graph_t* graph = 0; |
2217 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2218 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2219 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2220 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2221 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2222 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2223 | 1 | dsfmt_t dsfmt; |
2224 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2225 | 1 | int i; |
2226 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2227 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2228 | 640 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2229 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2230 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2231 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2232 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2233 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2234 | 640 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2235 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2236 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2237 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2238 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2239 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2240 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2241 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2242 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2243 | 1 | ccv_nnc_graph_free(graph); |
2244 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2245 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
2246 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
2247 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_mean"); |
2248 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 2, 10), "saved_inv_std"); |
2249 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
2250 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2251 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2252 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2253 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2254 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2255 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2256 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2257 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2258 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2259 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2260 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2261 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2262 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2263 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2264 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2265 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
2266 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2267 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2268 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2269 | 1 | ccv_nnc_graph_free(cpu_graph); |
2270 | 1 | ccv_nnc_tensor_free(x_tensor); |
2271 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2272 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2273 | 1 | } |
2274 | | |
2275 | | TEST_CASE("compare group norm and reduce HW with cudnn without scale / bias") |
2276 | 1 | { |
2277 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2278 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2279 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2280 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2281 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host x"); |
2282 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
2283 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
2284 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "host y"); |
2285 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_mean"); |
2286 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_inv_std"); |
2287 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
2288 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
2289 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
2290 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2291 | 1 | ccv_nnc_graph_t* graph = 0; |
2292 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2293 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2294 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2295 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2296 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2297 | 1 | dsfmt_t dsfmt; |
2298 | 1 | float xdata[2 * 16 * 2 * 10]; |
2299 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
2300 | 1 | int i; |
2301 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2302 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2303 | 640 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
2304 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2305 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
2306 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2307 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2308 | 1 | ccv_nnc_graph_free(graph); |
2309 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2310 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
2311 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
2312 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_mean"); |
2313 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_inv_std"); |
2314 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
2315 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2316 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2317 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2318 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2319 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2320 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2321 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 16 * 2 * 10); |
2322 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2323 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
2324 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 16 * 2 * 10, 1e-5, "layer norm result from cudnn should match the one from reference implementation"); |
2325 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2326 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2327 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2328 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2329 | 1 | ccv_nnc_graph_free(cpu_graph); |
2330 | 1 | } |
2331 | | |
2332 | | TEST_CASE("compare group norm gradient and reduce HW with cudnn without scale / bias") |
2333 | 1 | { |
2334 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2335 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GROUP_NORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2336 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2337 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2338 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "x"); |
2339 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 16, 2, 10), "y"); |
2340 | 1 | ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_mean"); |
2341 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 4, 1, 1), "saved_inv_std"); |
2342 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_mean, saved_inv_std), "group_norm"); |
2343 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2344 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2345 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2346 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2347 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2348 | 1 | ccv_nnc_graph_t* graph = 0; |
2349 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2350 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2351 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2352 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2353 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2354 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2355 | 1 | dsfmt_t dsfmt; |
2356 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2357 | 1 | int i; |
2358 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2359 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2360 | 640 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2361 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2362 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2363 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2364 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2365 | 641 | for (i = 0; i < 2 * 16 * 2 * 10; i++640 ) |
2366 | 640 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2367 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2368 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2369 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2370 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), 0); |
2371 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2372 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2373 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2374 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2375 | 1 | ccv_nnc_graph_free(graph); |
2376 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2377 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "x"); |
2378 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 16, 2, 10), "y"); |
2379 | 1 | ccv_nnc_tensor_symbol_t csaved_mean = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_mean"); |
2380 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 4, 1, 1), "saved_inv_std"); |
2381 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_GROUP_NORM_FORWARD(1, 4, 1e-5, 0, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_mean, csaved_inv_std), "group_norm"); |
2382 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2383 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2384 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2385 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2386 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2387 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2388 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2389 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2390 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2391 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2392 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2393 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2394 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 16 * 2 * 10); |
2395 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2396 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2397 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "layer norm gradient result from cudnn should match the one from reference implementation"); |
2398 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2399 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2400 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2401 | 1 | ccv_nnc_graph_free(cpu_graph); |
2402 | 1 | ccv_nnc_tensor_free(x_tensor); |
2403 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2404 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2405 | 1 | } |
2406 | | |
2407 | | TEST_CASE("compare rmsnorm with cudnn") |
2408 | 1 | { |
2409 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2410 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2411 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2412 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2413 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host x"); |
2414 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
2415 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
2416 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host y"); |
2417 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "scale"); |
2418 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
2419 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
2420 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale), TENSOR_SYMBOL_LIST(by, saved_inv_std), "rmsnorm"); |
2421 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
2422 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2423 | 1 | ccv_nnc_graph_t* graph = 0; |
2424 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2425 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2426 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2427 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2428 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2429 | 1 | dsfmt_t dsfmt; |
2430 | 1 | float xdata[2 * 2 * 2 * 10]; |
2431 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
2432 | 1 | int i; |
2433 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2434 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2435 | 80 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
2436 | 1 | float scaledata[1 * 2 * 2 * 10]; |
2437 | 41 | for (i = 0; i < 1 * 2 * 2 * 10; i++40 ) |
2438 | 40 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
2439 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
2440 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale)), 0); |
2441 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2442 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
2443 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2444 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2445 | 1 | ccv_nnc_graph_free(graph); |
2446 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2447 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
2448 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
2449 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "scale"); |
2450 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
2451 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale), TENSOR_SYMBOL_LIST(cy, csaved_inv_std), "rmsnorm"); |
2452 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2453 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2454 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2455 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2456 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2457 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2458 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 2 * 2 * 10); |
2459 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
2460 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 2 * 2 * 10); |
2461 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2462 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
2463 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 2 * 2 * 10, 1e-5, "rmsnorm result from cudnn should match the one from reference implementation"); |
2464 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2465 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2466 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2467 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2468 | 1 | ccv_nnc_graph_free(cpu_graph); |
2469 | 1 | } |
2470 | | |
2471 | | TEST_CASE("compare rmsnorm gradient with cudnn") |
2472 | 1 | { |
2473 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2474 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2475 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2476 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2477 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
2478 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
2479 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "scale"); |
2480 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
2481 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale), TENSOR_SYMBOL_LIST(by, saved_inv_std), "rmsnorm"); |
2482 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2483 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx, scale), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2484 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2485 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2486 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2487 | 1 | ccv_nnc_graph_t* graph = 0; |
2488 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2489 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2490 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2491 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2492 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2493 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2494 | 1 | dsfmt_t dsfmt; |
2495 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2496 | 1 | int i; |
2497 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2498 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2499 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2500 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2501 | 1 | float scaledata[1 * 2 * 2 * 10]; |
2502 | 41 | for (i = 0; i < 1 * 2 * 2 * 10; i++40 ) |
2503 | 40 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
2504 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
2505 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale)), 0); |
2506 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2507 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2508 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2509 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2510 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2511 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2512 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2513 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2514 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2515 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2516 | 1 | ccv_nnc_tensor_t* const dbscale_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_for_backward(symbolic_graph, scale)); |
2517 | 1 | ccv_nnc_tensor_t* const dscale_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
2518 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbscale_tensor), TENSOR_LIST(dscale_tensor), 0); |
2519 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2520 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2521 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2522 | 1 | ccv_nnc_graph_free(graph); |
2523 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2524 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
2525 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
2526 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "scale"); |
2527 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
2528 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale), TENSOR_SYMBOL_LIST(cy, csaved_inv_std), "rmsnorm"); |
2529 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2530 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx, cscale), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2531 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2532 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2533 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2534 | 1 | ccv_nnc_tensor_symbol_t dcscale = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cscale); |
2535 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2536 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2537 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2538 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2539 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2540 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2541 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2542 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2543 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
2544 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 2 * 2 * 10); |
2545 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2546 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2547 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "rmsnorm gradient result from cudnn should match the one from reference implementation"); |
2548 | 1 | ccv_nnc_tensor_t* const dcscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcscale); |
2549 | 1 | REQUIRE_TENSOR_EQ(dscale_tensor, dcscale_tensor, "rmsnorm scale gradient result from cudnn should match the one from reference implementation"); |
2550 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2551 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2552 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2553 | 1 | ccv_nnc_graph_free(cpu_graph); |
2554 | 1 | ccv_nnc_tensor_free(x_tensor); |
2555 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2556 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2557 | 1 | ccv_nnc_tensor_free(dscale_tensor); |
2558 | 1 | } |
2559 | | |
2560 | | TEST_CASE("compare rmsnorm only gradient with cudnn") |
2561 | 1 | { |
2562 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2563 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2564 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2565 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2566 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
2567 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
2568 | 1 | ccv_nnc_tensor_symbol_t scale = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 1, 2, 2, 10), "scale"); |
2569 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
2570 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(bx, scale), TENSOR_SYMBOL_LIST(by, saved_inv_std), "rmsnorm"); |
2571 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2572 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2573 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2574 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2575 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2576 | 1 | ccv_nnc_graph_t* graph = 0; |
2577 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2578 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2579 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2580 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2581 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2582 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2583 | 1 | dsfmt_t dsfmt; |
2584 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2585 | 1 | int i; |
2586 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2587 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2588 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2589 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2590 | 1 | float scaledata[1 * 2 * 2 * 10]; |
2591 | 41 | for (i = 0; i < 1 * 2 * 2 * 10; i++40 ) |
2592 | 40 | scaledata[i] = dsfmt_genrand_open_close(&dsfmt); |
2593 | 1 | ccv_nnc_tensor_t scale_tensor = ccv_nnc_tensor(scaledata, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), 0); |
2594 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(&scale_tensor), TENSOR_LIST(ccv_nnc_tensor_from_symbol(tensor_arena, scale)), 0); |
2595 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2596 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2597 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2598 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2599 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2600 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2601 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2602 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2603 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2604 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2605 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2606 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2607 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2608 | 1 | ccv_nnc_graph_free(graph); |
2609 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2610 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
2611 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
2612 | 1 | ccv_nnc_tensor_symbol_t cscale = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 1, 2, 2, 10), "scale"); |
2613 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
2614 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 1, 1, 2, 3), TENSOR_SYMBOL_LIST(cx, cscale), TENSOR_SYMBOL_LIST(cy, csaved_inv_std), "rmsnorm"); |
2615 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2616 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2617 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2618 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2619 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2620 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2621 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2622 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2623 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2624 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2625 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2626 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2627 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2628 | 1 | ccv_nnc_tensor_t* const cscale_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cscale); |
2629 | 1 | memcpy(cscale_tensor->data.f32, scaledata, sizeof(float) * 1 * 2 * 2 * 10); |
2630 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2631 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2632 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "rmsnorm gradient result from cudnn should match the one from reference implementation"); |
2633 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2634 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2635 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2636 | 1 | ccv_nnc_graph_free(cpu_graph); |
2637 | 1 | ccv_nnc_tensor_free(x_tensor); |
2638 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2639 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2640 | 1 | } |
2641 | | |
2642 | | TEST_CASE("compare rmsnorm with cudnn without scale") |
2643 | 1 | { |
2644 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2645 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2646 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2647 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2648 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host x"); |
2649 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
2650 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
2651 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "host y"); |
2652 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
2653 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(bx), "transfer x"); |
2654 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_inv_std), "rmsnorm"); |
2655 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(y), "transfer y"); |
2656 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2657 | 1 | ccv_nnc_graph_t* graph = 0; |
2658 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2659 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2660 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2661 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2662 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2663 | 1 | dsfmt_t dsfmt; |
2664 | 1 | float xdata[2 * 2 * 2 * 10]; |
2665 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
2666 | 1 | int i; |
2667 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2668 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2669 | 80 | x_tensor->data.f32[i] = xdata[i] = dsfmt_genrand_open_close(&dsfmt); |
2670 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2671 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
2672 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2673 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2674 | 1 | ccv_nnc_graph_free(graph); |
2675 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2676 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
2677 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
2678 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
2679 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_inv_std), "rmsnorm"); |
2680 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2681 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2682 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2683 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2684 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2685 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2686 | 1 | memcpy(cx_tensor->data.f32, xdata, sizeof(float) * 2 * 2 * 2 * 10); |
2687 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2688 | 1 | ccv_nnc_tensor_t* const cy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cy); |
2689 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cy_tensor->data.f32, 2 * 2 * 2 * 10, 1e-5, "rmsnorm result from cudnn should match the one from reference implementation"); |
2690 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2691 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2692 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2693 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2694 | 1 | ccv_nnc_graph_free(cpu_graph); |
2695 | 1 | } |
2696 | | |
2697 | | TEST_CASE("compare rmsnorm gradient with cudnn without scale") |
2698 | 1 | { |
2699 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2700 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2701 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2702 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2703 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
2704 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
2705 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
2706 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_inv_std), "rmsnorm"); |
2707 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2708 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2709 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2710 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2711 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2712 | 1 | ccv_nnc_graph_t* graph = 0; |
2713 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2714 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2715 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2716 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2717 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2718 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2719 | 1 | dsfmt_t dsfmt; |
2720 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2721 | 1 | int i; |
2722 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2723 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2724 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2725 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2726 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2727 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2728 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2729 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2730 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2731 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2732 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2733 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2734 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2735 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2736 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2737 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2738 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2739 | 1 | ccv_nnc_graph_free(graph); |
2740 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2741 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
2742 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
2743 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
2744 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_inv_std), "rmsnorm"); |
2745 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2746 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2747 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2748 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2749 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2750 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2751 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2752 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2753 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2754 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2755 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2756 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2757 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2758 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2759 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2760 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "rmsnorm gradient result from cudnn should match the one from reference implementation"); |
2761 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2762 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2763 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2764 | 1 | ccv_nnc_graph_free(cpu_graph); |
2765 | 1 | ccv_nnc_tensor_free(x_tensor); |
2766 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2767 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2768 | 1 | } |
2769 | | |
2770 | | TEST_CASE("compare rmsnorm only gradient with cudnn without scale") |
2771 | 1 | { |
2772 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2773 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RMSNORM_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
2774 | 1 | (ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN))); |
2775 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2776 | 1 | ccv_nnc_tensor_symbol_t bx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "x"); |
2777 | 1 | ccv_nnc_tensor_symbol_t by = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 2, 2, 10), "y"); |
2778 | 1 | ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 2, 1, 1, 1), "saved_inv_std"); |
2779 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(bx), TENSOR_SYMBOL_LIST(by, saved_inv_std), "rmsnorm"); |
2780 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2781 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(by), TENSOR_SYMBOL_LIST(bx), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
2782 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2783 | 1 | ccv_nnc_tensor_symbol_t dby = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, by); |
2784 | 1 | ccv_nnc_tensor_symbol_t dbx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, bx); |
2785 | 1 | ccv_nnc_graph_t* graph = 0; |
2786 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2787 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2788 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2789 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2790 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2791 | 1 | ccv_nnc_tensor_t* const bx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, bx); |
2792 | 1 | dsfmt_t dsfmt; |
2793 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2794 | 1 | int i; |
2795 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
2796 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2797 | 80 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 100; |
2798 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(bx_tensor), 0); |
2799 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2800 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2801 | 1 | ccv_nnc_tensor_t* const dby_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dby); |
2802 | 81 | for (i = 0; i < 2 * 2 * 2 * 10; i++80 ) |
2803 | 80 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
2804 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dby_tensor), 0); |
2805 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2806 | 1 | ccv_nnc_tensor_t* const dbx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dbx); |
2807 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), 0); |
2808 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dbx_tensor), TENSOR_LIST(dx_tensor), 0); |
2809 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2810 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2811 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2812 | 1 | ccv_nnc_graph_free(graph); |
2813 | 1 | ccv_nnc_symbolic_graph_t* const cpu_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2814 | 1 | ccv_nnc_tensor_symbol_t cx = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "x"); |
2815 | 1 | ccv_nnc_tensor_symbol_t cy = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 2, 2, 10), "y"); |
2816 | 1 | ccv_nnc_tensor_symbol_t csaved_inv_std = ccv_nnc_tensor_symbol_new(cpu_symbolic_graph, CPU_TENSOR_NHWC(32F, 2, 1, 1, 1), "saved_inv_std"); |
2817 | 1 | ccv_nnc_graph_exec_symbol_new(cpu_symbolic_graph, CMD_RMSNORM_FORWARD(1e-4, 0, 1, 2, 3), TENSOR_SYMBOL_LIST(cx), TENSOR_SYMBOL_LIST(cy, csaved_inv_std), "rmsnorm"); |
2818 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2819 | 1 | ccv_nnc_symbolic_graph_backward(cpu_symbolic_graph, TENSOR_SYMBOL_LIST(cy), TENSOR_SYMBOL_LIST(cx), SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph)); |
2820 | 1 | ccv_nnc_graph_exec_symbol_autogen(cpu_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2821 | 1 | ccv_nnc_tensor_symbol_t dcy = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cy); |
2822 | 1 | ccv_nnc_tensor_symbol_t dcx = ccv_nnc_tensor_symbol_for_backward(cpu_symbolic_graph, cx); |
2823 | 1 | ccv_nnc_graph_t* cpu_graph = 0; |
2824 | 1 | ccv_nnc_tensor_arena_t* cpu_tensor_arena = 0; |
2825 | 1 | ccv_nnc_graph_exec_arena_t* cpu_graph_exec_arena = 0; |
2826 | 1 | ccv_nnc_symbolic_graph_compile(cpu_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(cpu_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(cpu_symbolic_graph), &cpu_graph, &cpu_tensor_arena, &cpu_graph_exec_arena); |
2827 | 1 | ccv_nnc_tensor_t* const cx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, cx); |
2828 | 1 | memcpy(cx_tensor->data.f32, x_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2829 | 1 | ccv_nnc_tensor_t* const dcy_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcy); |
2830 | 1 | memcpy(dcy_tensor->data.f32, dy_tensor->data.f32, sizeof(float) * 2 * 2 * 2 * 10); |
2831 | 1 | ccv_nnc_graph_run(cpu_graph, 0, TRAVERSE_FULL, 0, 0); |
2832 | 1 | ccv_nnc_tensor_t* const dcx_tensor = ccv_nnc_tensor_from_symbol(cpu_tensor_arena, dcx); |
2833 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, dcx_tensor, "rmsnorm gradient result from cudnn should match the one from reference implementation"); |
2834 | 1 | ccv_nnc_symbolic_graph_free(cpu_symbolic_graph); |
2835 | 1 | ccv_nnc_tensor_arena_free(cpu_tensor_arena); |
2836 | 1 | ccv_nnc_graph_exec_arena_free(cpu_graph_exec_arena); |
2837 | 1 | ccv_nnc_graph_free(cpu_graph); |
2838 | 1 | ccv_nnc_tensor_free(x_tensor); |
2839 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2840 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2841 | 1 | } |
2842 | | |
2843 | | TEST_CASE("compare average pooling with cudnn") |
2844 | 1 | { |
2845 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_AVERAGE_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
2846 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2847 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 7, 7, 10), "x"); |
2848 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 3, 3, 10), "y"); |
2849 | 1 | ccv_nnc_graph_exec_symbol_t avg_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_AVERAGE_POOL_FORWARD(5, 5), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "avg_pool"); |
2850 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, avg_pool, HINT((2, 2), (1, 1))); |
2851 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2852 | 1 | ccv_nnc_graph_t* graph = 0; |
2853 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2854 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2855 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2856 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2857 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2858 | 1 | dsfmt_t dsfmt; |
2859 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
2860 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
2861 | 1 | int i; |
2862 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
2863 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2864 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
2865 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
2866 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2867 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
2868 | 1 | ccv_nnc_cmd_exec(CMD_AVERAGE_POOL_FORWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
2869 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
2870 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
2871 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y), 0); |
2872 | 1 | REQUIRE_TENSOR_EQ(y_tensor, cpu_y, "cudnn result should equal to cpu result"); |
2873 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2874 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2875 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2876 | 1 | ccv_nnc_graph_free(graph); |
2877 | 1 | ccv_nnc_tensor_free(x_tensor); |
2878 | 1 | ccv_nnc_tensor_free(y_tensor); |
2879 | 1 | ccv_nnc_tensor_free(cpu_y); |
2880 | 1 | } |
2881 | | |
2882 | | TEST_CASE("compare average pooling with cudnn in half precision") |
2883 | 1 | { |
2884 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_AVERAGE_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
2885 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2886 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 7, 7, 10), "x"); |
2887 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 3, 3, 10), "y"); |
2888 | 1 | ccv_nnc_graph_exec_symbol_t avg_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_AVERAGE_POOL_FORWARD(5, 5), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "avg_pool"); |
2889 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, avg_pool, HINT((2, 2), (1, 1))); |
2890 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2891 | 1 | ccv_nnc_graph_t* graph = 0; |
2892 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2893 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2894 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2895 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2896 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2897 | 1 | dsfmt_t dsfmt; |
2898 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
2899 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
2900 | 1 | int i; |
2901 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
2902 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2903 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
2904 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
2905 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
2906 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
2907 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2908 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
2909 | 1 | ccv_nnc_cmd_exec(CMD_AVERAGE_POOL_FORWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
2910 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
2911 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
2912 | 1 | ccv_nnc_tensor_t* const cpu_y16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 3, 3, 10), 0); |
2913 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y16), 0); |
2914 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_y16), TENSOR_LIST(cpu_y), 0); |
2915 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cpu_y->data.f32, 3 * 3 * 10, 1e-3, "cudnn result should equal to cpu result"); |
2916 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2917 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2918 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2919 | 1 | ccv_nnc_graph_free(graph); |
2920 | 1 | ccv_nnc_tensor_free(x_tensor); |
2921 | 1 | ccv_nnc_tensor_free(x16_tensor); |
2922 | 1 | ccv_nnc_tensor_free(y_tensor); |
2923 | 1 | ccv_nnc_tensor_free(cpu_y); |
2924 | 1 | ccv_nnc_tensor_free(cpu_y16); |
2925 | 1 | } |
2926 | | |
2927 | | TEST_CASE("compare average pooling gradient with cudnn") |
2928 | 1 | { |
2929 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_AVERAGE_POOL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
2930 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2931 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 7, 7, 10), "dx"); |
2932 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 3, 3, 10), "dy"); |
2933 | 1 | ccv_nnc_graph_exec_symbol_t avg_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_AVERAGE_POOL_BACKWARD(5, 5), TENSOR_SYMBOL_LIST(dy), TENSOR_SYMBOL_LIST(dx), "avg_pool"); |
2934 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, avg_pool, HINT((2, 2), (1, 1))); |
2935 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2936 | 1 | ccv_nnc_graph_t* graph = 0; |
2937 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2938 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2939 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2940 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2941 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2942 | 1 | dsfmt_t dsfmt; |
2943 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
2944 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
2945 | 1 | int i; |
2946 | 91 | for (i = 0; i < 3 * 3 * 10; i++90 ) |
2947 | 90 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2948 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_from_symbol(tensor_arena, dy); |
2949 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
2950 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2951 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
2952 | 1 | ccv_nnc_cmd_exec(CMD_AVERAGE_POOL_BACKWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dx_tensor), 0); |
2953 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
2954 | 1 | ccv_nnc_tensor_t* const cpu_dx = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
2955 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(cpu_dx), 0); |
2956 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, cpu_dx, "cudnn result should equal to cpu result"); |
2957 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
2958 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
2959 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
2960 | 1 | ccv_nnc_graph_free(graph); |
2961 | 1 | ccv_nnc_tensor_free(dy_tensor); |
2962 | 1 | ccv_nnc_tensor_free(dx_tensor); |
2963 | 1 | ccv_nnc_tensor_free(cpu_dx); |
2964 | 1 | } |
2965 | | |
2966 | | TEST_CASE("compare average pooling gradient with cudnn in half precision") |
2967 | 1 | { |
2968 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_AVERAGE_POOL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
2969 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
2970 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 7, 7, 10), "dx"); |
2971 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 3, 3, 10), "dy"); |
2972 | 1 | ccv_nnc_graph_exec_symbol_t avg_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_AVERAGE_POOL_BACKWARD(5, 5), TENSOR_SYMBOL_LIST(dy), TENSOR_SYMBOL_LIST(dx), "avg_pool"); |
2973 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, avg_pool, HINT((2, 2), (1, 1))); |
2974 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
2975 | 1 | ccv_nnc_graph_t* graph = 0; |
2976 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
2977 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
2978 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
2979 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
2980 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
2981 | 1 | dsfmt_t dsfmt; |
2982 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
2983 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
2984 | 1 | int i; |
2985 | 91 | for (i = 0; i < 3 * 3 * 10; i++90 ) |
2986 | 90 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2987 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 3, 3, 10), 0); |
2988 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_from_symbol(tensor_arena, dy); |
2989 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
2990 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
2991 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
2992 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
2993 | 1 | ccv_nnc_cmd_exec(CMD_AVERAGE_POOL_BACKWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dx_tensor), 0); |
2994 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
2995 | 1 | ccv_nnc_tensor_t* const cpu_dx16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
2996 | 1 | ccv_nnc_tensor_t* const cpu_dx = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
2997 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(cpu_dx16), 0); |
2998 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_dx16), TENSOR_LIST(cpu_dx), 0); |
2999 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dx_tensor->data.f32, cpu_dx->data.f32, 7 * 7 * 10, 1e-3, "cudnn result should equal to cpu result"); |
3000 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3001 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3002 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3003 | 1 | ccv_nnc_graph_free(graph); |
3004 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3005 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
3006 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3007 | 1 | ccv_nnc_tensor_free(cpu_dx); |
3008 | 1 | ccv_nnc_tensor_free(cpu_dx16); |
3009 | 1 | } |
3010 | | |
3011 | | TEST_CASE("compare max pooling with cudnn") |
3012 | 1 | { |
3013 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3014 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3015 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 7, 7, 10), "x"); |
3016 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 3, 3, 10), "y"); |
3017 | 1 | ccv_nnc_graph_exec_symbol_t max_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_MAX_POOL_FORWARD(5, 5), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "max_pool"); |
3018 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, max_pool, HINT((2, 2), (1, 1))); |
3019 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3020 | 1 | ccv_nnc_graph_t* graph = 0; |
3021 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3022 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3023 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3024 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3025 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3026 | 1 | dsfmt_t dsfmt; |
3027 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3028 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3029 | 1 | int i; |
3030 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
3031 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3032 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3033 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3034 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3035 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3036 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_FORWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3037 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3038 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3039 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y), 0); |
3040 | 1 | REQUIRE_TENSOR_EQ(y_tensor, cpu_y, "cudnn result should equal to cpu result"); |
3041 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3042 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3043 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3044 | 1 | ccv_nnc_graph_free(graph); |
3045 | 1 | ccv_nnc_tensor_free(x_tensor); |
3046 | 1 | ccv_nnc_tensor_free(y_tensor); |
3047 | 1 | ccv_nnc_tensor_free(cpu_y); |
3048 | 1 | } |
3049 | | |
3050 | | TEST_CASE("compare max pooling with cudnn in half precision") |
3051 | 1 | { |
3052 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3053 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3054 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 7, 7, 10), "x"); |
3055 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 3, 3, 10), "y"); |
3056 | 1 | ccv_nnc_graph_exec_symbol_t max_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_MAX_POOL_FORWARD(5, 5), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "max_pool"); |
3057 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, max_pool, HINT((2, 2), (1, 1))); |
3058 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3059 | 1 | ccv_nnc_graph_t* graph = 0; |
3060 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3061 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3062 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3063 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3064 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3065 | 1 | dsfmt_t dsfmt; |
3066 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3067 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3068 | 1 | int i; |
3069 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
3070 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3071 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3072 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
3073 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3074 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3075 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3076 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3077 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_FORWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3078 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3079 | 1 | ccv_nnc_tensor_t* const cpu_y16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 3, 3, 10), 0); |
3080 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3081 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y16), 0); |
3082 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_y16), TENSOR_LIST(cpu_y), 0); |
3083 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cpu_y->data.f32, 3 * 3 * 10, 1e-3, "cudnn result should equal to cpu result"); |
3084 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3085 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3086 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3087 | 1 | ccv_nnc_graph_free(graph); |
3088 | 1 | ccv_nnc_tensor_free(x_tensor); |
3089 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3090 | 1 | ccv_nnc_tensor_free(y_tensor); |
3091 | 1 | ccv_nnc_tensor_free(cpu_y); |
3092 | 1 | ccv_nnc_tensor_free(cpu_y16); |
3093 | 1 | } |
3094 | | |
3095 | | TEST_CASE("compare max pooling 2x2 with cudnn") |
3096 | 1 | { |
3097 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3098 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3099 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 10, 6, 6), "x"); |
3100 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 10, 3, 3), "y"); |
3101 | 1 | ccv_nnc_graph_exec_symbol_t max_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_MAX_POOL_FORWARD(2, 2), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "max_pool"); |
3102 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, max_pool, HINT((2, 2), (0, 0))); |
3103 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3104 | 1 | ccv_nnc_graph_t* graph = 0; |
3105 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3106 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3107 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3108 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3109 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3110 | 1 | dsfmt_t dsfmt; |
3111 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3112 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 6, 6), 0); |
3113 | 1 | int i, j; |
3114 | 361 | for (i = 0; i < 6 * 6 * 10; i++360 ) |
3115 | 360 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3116 | 1 | ccv_nnc_tensor_t* const gt_x = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 6, 6, 10), 0); |
3117 | 11 | for (i = 0; i < 10; i++10 ) |
3118 | 370 | for (j = 0; 10 j < 6 * 6; j++360 ) |
3119 | 360 | gt_x->data.f32[j * 10 + i] = x_tensor->data.f32[i * 6 * 6 + j]; |
3120 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3121 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3122 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3123 | 1 | ccv_nnc_tensor_t* const gt_y= ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3124 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_FORWARD(2, 2), HINT((2, 2), (0, 0)), 0, TENSOR_LIST(gt_x), TENSOR_LIST(gt_y), 0); |
3125 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3126 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 3, 3), 0); |
3127 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y), 0); |
3128 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 3, 3), 0); |
3129 | 11 | for (i = 0; i < 10; i++10 ) |
3130 | 100 | for (j = 0; 10 j < 3 * 3; j++90 ) |
3131 | 90 | y_tensor->data.f32[i * 3 * 3 + j] = gt_y->data.f32[j * 10 + i]; |
3132 | 1 | REQUIRE_TENSOR_EQ(y_tensor, cpu_y, "cudnn result should equal to cpu result"); |
3133 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3134 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3135 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3136 | 1 | ccv_nnc_graph_free(graph); |
3137 | 1 | ccv_nnc_tensor_free(x_tensor); |
3138 | 1 | ccv_nnc_tensor_free(y_tensor); |
3139 | 1 | ccv_nnc_tensor_free(cpu_y); |
3140 | 1 | } |
3141 | | |
3142 | | TEST_CASE("compare max pooling 2x2 with cudnn in half precision") |
3143 | 1 | { |
3144 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3145 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3146 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 10, 6, 6), "x"); |
3147 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 10, 3, 3), "y"); |
3148 | 1 | ccv_nnc_graph_exec_symbol_t max_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_MAX_POOL_FORWARD(2, 2), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "max_pool"); |
3149 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, max_pool, HINT((2, 2), (0, 0))); |
3150 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3151 | 1 | ccv_nnc_graph_t* graph = 0; |
3152 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3153 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3154 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3155 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3156 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3157 | 1 | dsfmt_t dsfmt; |
3158 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3159 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 6, 6), 0); |
3160 | 1 | int i, j; |
3161 | 361 | for (i = 0; i < 6 * 6 * 10; i++360 ) |
3162 | 360 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3163 | 1 | ccv_nnc_tensor_t* const gt_x = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 6, 6, 10), 0); |
3164 | 11 | for (i = 0; i < 10; i++10 ) |
3165 | 370 | for (j = 0; 10 j < 6 * 6; j++360 ) |
3166 | 360 | gt_x->data.f32[j * 10 + i] = x_tensor->data.f32[i * 6 * 6 + j]; |
3167 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3168 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 10, 6, 6), 0); |
3169 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3170 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3171 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3172 | 1 | ccv_nnc_tensor_t* const gt_y= ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3173 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_FORWARD(2, 2), HINT((2, 2), (0, 0)), 0, TENSOR_LIST(gt_x), TENSOR_LIST(gt_y), 0); |
3174 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3175 | 1 | ccv_nnc_tensor_t* const cpu_y16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 10, 3, 3), 0); |
3176 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 3, 3), 0); |
3177 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y16), 0); |
3178 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_y16), TENSOR_LIST(cpu_y), 0); |
3179 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 3, 3), 0); |
3180 | 11 | for (i = 0; i < 10; i++10 ) |
3181 | 100 | for (j = 0; 10 j < 3 * 3; j++90 ) |
3182 | 90 | y_tensor->data.f32[i * 3 * 3 + j] = gt_y->data.f32[j * 10 + i]; |
3183 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cpu_y->data.f32, 10 * 3 * 3, 1e-3, "cudnn result should equal to cpu result"); |
3184 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3185 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3186 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3187 | 1 | ccv_nnc_graph_free(graph); |
3188 | 1 | ccv_nnc_tensor_free(x_tensor); |
3189 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3190 | 1 | ccv_nnc_tensor_free(y_tensor); |
3191 | 1 | ccv_nnc_tensor_free(cpu_y); |
3192 | 1 | ccv_nnc_tensor_free(cpu_y16); |
3193 | 1 | } |
3194 | | |
3195 | | TEST_CASE("compare max pooling gradient with cudnn") |
3196 | 1 | { |
3197 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3198 | 1 | ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3199 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3200 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 7, 7, 10), "x"); |
3201 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 3, 3, 10), "y"); |
3202 | 1 | ccv_nnc_graph_exec_symbol_t max_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_MAX_POOL_FORWARD(5, 5), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "max_pool"); |
3203 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, max_pool, HINT((2, 2), (1, 1))); |
3204 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3205 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3206 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3207 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3208 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3209 | 1 | dsfmt_t dsfmt; |
3210 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3211 | 1 | int i; |
3212 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3213 | 91 | for (i = 0; i < 3 * 3 * 10; i++90 ) |
3214 | 90 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3215 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 3, 10), 0); |
3216 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
3217 | 1 | ccv_nnc_graph_t* graph = 0; |
3218 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3219 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3220 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3221 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3222 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3223 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3224 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
3225 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3226 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3227 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3228 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3229 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3230 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_FORWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3231 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3232 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_BACKWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(dy_tensor, x_tensor, y_tensor), TENSOR_LIST(dx_tensor), 0); |
3233 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3234 | 1 | ccv_nnc_tensor_t* const cpu_dx = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3235 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(cpu_dx), 0); |
3236 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, cpu_dx, "cudnn result should equal to cpu result"); |
3237 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3238 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3239 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3240 | 1 | ccv_nnc_graph_free(graph); |
3241 | 1 | ccv_nnc_tensor_free(x_tensor); |
3242 | 1 | ccv_nnc_tensor_free(y_tensor); |
3243 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3244 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3245 | 1 | ccv_nnc_tensor_free(cpu_dx); |
3246 | 1 | ccv_nnc_tensor_free(dyt); |
3247 | 1 | } |
3248 | | |
3249 | | TEST_CASE("compare max pooling gradient with cudnn in half precision") |
3250 | 1 | { |
3251 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3252 | 1 | ccv_nnc_cmd_ok(CCV_NNC_MAX_POOL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3253 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3254 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 7, 7, 10), "x"); |
3255 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 3, 3, 10), "y"); |
3256 | 1 | ccv_nnc_graph_exec_symbol_t max_pool = ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_MAX_POOL_FORWARD(5, 5), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "max_pool"); |
3257 | 1 | ccv_nnc_graph_exec_symbol_set_hint(symbolic_graph, max_pool, HINT((2, 2), (1, 1))); |
3258 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3259 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3260 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3261 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3262 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3263 | 1 | dsfmt_t dsfmt; |
3264 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3265 | 1 | int i; |
3266 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3267 | 91 | for (i = 0; i < 3 * 3 * 10; i++90 ) |
3268 | 90 | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3269 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 3, 3, 10), 0); |
3270 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 3, 3, 10), 0); |
3271 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
3272 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
3273 | 1 | ccv_nnc_graph_t* graph = 0; |
3274 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3275 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3276 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3277 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3278 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3279 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3280 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
3281 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3282 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
3283 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3284 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3285 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3286 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3287 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 3, 10), 0); |
3288 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_FORWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3289 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3290 | 1 | ccv_nnc_cmd_exec(CMD_MAX_POOL_BACKWARD(5, 5), HINT((2, 2), (1, 1)), 0, TENSOR_LIST(dy_tensor, x_tensor, y_tensor), TENSOR_LIST(dx_tensor), 0); |
3291 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3292 | 1 | ccv_nnc_tensor_t* const cpu_dx16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
3293 | 1 | ccv_nnc_tensor_t* const cpu_dx = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3294 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(cpu_dx16), 0); |
3295 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_dx16), TENSOR_LIST(cpu_dx), 0); |
3296 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dx_tensor->data.f32, cpu_dx->data.f32, 7 * 7 * 10, 5e-3, "cudnn result should equal to cpu result"); |
3297 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3298 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3299 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3300 | 1 | ccv_nnc_graph_free(graph); |
3301 | 1 | ccv_nnc_tensor_free(x_tensor); |
3302 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3303 | 1 | ccv_nnc_tensor_free(y_tensor); |
3304 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3305 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3306 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
3307 | 1 | ccv_nnc_tensor_free(cpu_dx); |
3308 | 1 | ccv_nnc_tensor_free(cpu_dx16); |
3309 | 1 | ccv_nnc_tensor_free(dyt); |
3310 | 1 | } |
3311 | | |
3312 | | TEST_CASE("compare relu with cudnn") |
3313 | 1 | { |
3314 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RELU_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3315 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3316 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 7, 7, 10), "x"); |
3317 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 7, 7, 10), "y"); |
3318 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RELU_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "relu"); |
3319 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3320 | 1 | ccv_nnc_graph_t* graph = 0; |
3321 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3322 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3323 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3324 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3325 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3326 | 1 | dsfmt_t dsfmt; |
3327 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3328 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3329 | 1 | int i; |
3330 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
3331 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
3332 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3333 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3334 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3335 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3336 | 1 | ccv_nnc_cmd_exec(CMD_RELU_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3337 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3338 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3339 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y), 0); |
3340 | 1 | REQUIRE_TENSOR_EQ(y_tensor, cpu_y, "cudnn result should equal to cpu result"); |
3341 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3342 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3343 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3344 | 1 | ccv_nnc_graph_free(graph); |
3345 | 1 | ccv_nnc_tensor_free(x_tensor); |
3346 | 1 | ccv_nnc_tensor_free(y_tensor); |
3347 | 1 | ccv_nnc_tensor_free(cpu_y); |
3348 | 1 | } |
3349 | | |
3350 | | TEST_CASE("compare relu with cudnn in half precision") |
3351 | 1 | { |
3352 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RELU_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3353 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3354 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 7, 7, 10), "x"); |
3355 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 7, 7, 10), "y"); |
3356 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RELU_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "relu"); |
3357 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3358 | 1 | ccv_nnc_graph_t* graph = 0; |
3359 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3360 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3361 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3362 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3363 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3364 | 1 | dsfmt_t dsfmt; |
3365 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3366 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3367 | 1 | int i; |
3368 | 491 | for (i = 0; i < 7 * 7 * 10; i++490 ) |
3369 | 490 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
3370 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
3371 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3372 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3373 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3374 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3375 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3376 | 1 | ccv_nnc_cmd_exec(CMD_RELU_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3377 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3378 | 1 | ccv_nnc_tensor_t* const cpu_y16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 7, 7, 10), 0); |
3379 | 1 | ccv_nnc_tensor_t* const cpu_y = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 7, 10), 0); |
3380 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(cpu_y16), 0); |
3381 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_y16), TENSOR_LIST(cpu_y), 0); |
3382 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, y_tensor->data.f32, cpu_y->data.f32, 7 * 7 * 10, 1e-3, "cudnn result should equal to cpu result"); |
3383 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3384 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3385 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3386 | 1 | ccv_nnc_graph_free(graph); |
3387 | 1 | ccv_nnc_tensor_free(x_tensor); |
3388 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3389 | 1 | ccv_nnc_tensor_free(y_tensor); |
3390 | 1 | ccv_nnc_tensor_free(cpu_y); |
3391 | 1 | ccv_nnc_tensor_free(cpu_y16); |
3392 | 1 | } |
3393 | | |
3394 | | TEST_CASE("compare relu gradient with cudnn") |
3395 | 1 | { |
3396 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RELU_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3397 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RELU_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3398 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3399 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 10, 10, 7, 7), "x"); |
3400 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 10, 10, 7, 7), "y"); |
3401 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RELU_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "relu"); |
3402 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3403 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3404 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3405 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3406 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3407 | 1 | dsfmt_t dsfmt; |
3408 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3409 | 1 | int i; |
3410 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3411 | 4.90k | for (i = 0; i < 10 * 7 * 7 * 10; i++4.90k ) |
3412 | 4.90k | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3413 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 10, 7, 7), 0); |
3414 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
3415 | 1 | ccv_nnc_graph_t* graph = 0; |
3416 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3417 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3418 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3419 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3420 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3421 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3422 | 4.90k | for (i = 0; i < 10 * 7 * 7 * 10; i++4.90k ) |
3423 | 4.90k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
3424 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3425 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3426 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3427 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3428 | 1 | ccv_nnc_cmd_exec(CMD_RELU_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3429 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3430 | 1 | ccv_nnc_cmd_exec(CMD_RELU_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, y_tensor), TENSOR_LIST(dx_tensor), 0); |
3431 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3432 | 1 | ccv_nnc_tensor_t* const cpu_dx = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3433 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(cpu_dx), 0); |
3434 | 1 | REQUIRE_TENSOR_EQ(dx_tensor, cpu_dx, "cudnn result should equal to cpu result"); |
3435 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3436 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3437 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3438 | 1 | ccv_nnc_graph_free(graph); |
3439 | 1 | ccv_nnc_tensor_free(x_tensor); |
3440 | 1 | ccv_nnc_tensor_free(y_tensor); |
3441 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3442 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3443 | 1 | ccv_nnc_tensor_free(dyt); |
3444 | 1 | ccv_nnc_tensor_free(cpu_dx); |
3445 | 1 | } |
3446 | | |
3447 | | TEST_CASE("compare relu gradient with cudnn in half precision") |
3448 | 1 | { |
3449 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_RELU_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3450 | 1 | ccv_nnc_cmd_ok(CCV_NNC_RELU_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3451 | 1 | ccv_nnc_symbolic_graph_t* symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3452 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 10, 10, 7, 7), "x"); |
3453 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 10, 10, 7, 7), "y"); |
3454 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_RELU_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "relu"); |
3455 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3456 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3457 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3458 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3459 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3460 | 1 | dsfmt_t dsfmt; |
3461 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3462 | 1 | int i; |
3463 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3464 | 4.90k | for (i = 0; i < 10 * 7 * 7 * 10; i++4.90k ) |
3465 | 4.90k | dy_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3466 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 16F, 10, 10, 7, 7), 0); |
3467 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 10, 10, 7, 7), 0); |
3468 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
3469 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
3470 | 1 | ccv_nnc_graph_t* graph = 0; |
3471 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3472 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3473 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3474 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3475 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3476 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3477 | 4.90k | for (i = 0; i < 10 * 7 * 7 * 10; i++4.90k ) |
3478 | 4.90k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
3479 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3480 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 10, 10, 7, 7), 0); |
3481 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3482 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3483 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3484 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3485 | 1 | ccv_nnc_cmd_exec(CMD_RELU_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(y_tensor), 0); |
3486 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3487 | 1 | ccv_nnc_cmd_exec(CMD_RELU_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, y_tensor), TENSOR_LIST(dx_tensor), 0); |
3488 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3489 | 1 | ccv_nnc_tensor_t* const cpu_dx16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 10, 10, 7, 7), 0); |
3490 | 1 | ccv_nnc_tensor_t* const cpu_dx = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 10, 7, 7), 0); |
3491 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(cpu_dx16), 0); |
3492 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(cpu_dx16), TENSOR_LIST(cpu_dx), 0); |
3493 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dx_tensor->data.f32, cpu_dx->data.f32, 10 * 10 * 7 * 7, 1e-3, "cudnn result should equal to cpu result"); |
3494 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3495 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3496 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3497 | 1 | ccv_nnc_graph_free(graph); |
3498 | 1 | ccv_nnc_tensor_free(x_tensor); |
3499 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3500 | 1 | ccv_nnc_tensor_free(y_tensor); |
3501 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3502 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3503 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
3504 | 1 | ccv_nnc_tensor_free(dyt); |
3505 | 1 | ccv_nnc_tensor_free(cpu_dx); |
3506 | 1 | ccv_nnc_tensor_free(cpu_dx16); |
3507 | 1 | } |
3508 | | |
3509 | | TEST_CASE("compare dropout with cudnn") |
3510 | 1 | { |
3511 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3512 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3513 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 20 * 50), "x"); |
3514 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 20 * 50), "y"); |
3515 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, ccv_nnc_tensor_auto, "c"); |
3516 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DROPOUT_FORWARD(0.4), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y, c), "dropout"); |
3517 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3518 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3519 | 1 | ccv_nnc_graph_t* graph = 0; |
3520 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3521 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3522 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3523 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3524 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3525 | 1 | int i; |
3526 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3527 | 1.00k | x_tensor->data.f32[i] = (i + 1) * 0.01; |
3528 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3529 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3530 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3531 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3532 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3533 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0); |
3534 | 1 | int zero_count = 0; |
3535 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3536 | 1.00k | if (fabsf(y_tensor->data.f32[i]) < 1e-5) |
3537 | 419 | ++zero_count; |
3538 | 581 | else { |
3539 | 581 | REQUIRE_EQ_WITH_TOLERANCE(x_tensor->data.f32[i] / 0.6, y_tensor->data.f32[i], 1e-5, "should be scaled up by 1 / 0.6"); |
3540 | 581 | } |
3541 | 1 | REQUIRE_EQ_WITH_TOLERANCE((float)zero_count / (20 * 50), 0.4, 5 * 1e-2, "should be within 5%% of error"); |
3542 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3543 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3544 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3545 | 1 | ccv_nnc_graph_free(graph); |
3546 | 1 | ccv_nnc_tensor_free(x_tensor); |
3547 | 1 | ccv_nnc_tensor_free(y_tensor); |
3548 | 1 | } |
3549 | | |
3550 | | TEST_CASE("compare dropout with cudnn in half precision") |
3551 | 1 | { |
3552 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3553 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3554 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 20 * 50), "x"); |
3555 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 20 * 50), "y"); |
3556 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, ccv_nnc_tensor_auto, "c"); |
3557 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DROPOUT_FORWARD(0.4), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y, c), "dropout"); |
3558 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3559 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3560 | 1 | ccv_nnc_graph_t* graph = 0; |
3561 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3562 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3563 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3564 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3565 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3566 | 1 | int i; |
3567 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3568 | 1.00k | x_tensor->data.f32[i] = (i + 1) * 0.01; |
3569 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 20 * 50), 0); |
3570 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3571 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3572 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3573 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3574 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3575 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3576 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 20 * 50), 0); |
3577 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0); |
3578 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
3579 | 1 | int zero_count = 0; |
3580 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3581 | 1.00k | if (fabsf(y_tensor->data.f32[i]) < 1e-5) |
3582 | 385 | ++zero_count; |
3583 | 615 | else { |
3584 | 615 | REQUIRE_EQ_WITH_TOLERANCE(x_tensor->data.f32[i] / 0.6, y_tensor->data.f32[i], x_tensor->data.f32[i] * 2e-3, "should be scaled up by 1 / 0.6"); |
3585 | 615 | } |
3586 | 1 | REQUIRE_EQ_WITH_TOLERANCE((float)zero_count / (20 * 50), 0.4, 5 * 1e-2, "should be within 5%% of error"); |
3587 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3588 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3589 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3590 | 1 | ccv_nnc_graph_free(graph); |
3591 | 1 | ccv_nnc_tensor_free(x_tensor); |
3592 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3593 | 1 | ccv_nnc_tensor_free(y_tensor); |
3594 | 1 | ccv_nnc_tensor_free(y16_tensor); |
3595 | 1 | } |
3596 | | |
3597 | | TEST_CASE("compare dropout gradient with cudnn") |
3598 | 1 | { |
3599 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3600 | 1 | ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3601 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3602 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 20 * 50), "x"); |
3603 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 20 * 50), "y"); |
3604 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, ccv_nnc_tensor_auto, "c"); |
3605 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DROPOUT_FORWARD(0.4), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y, c), "dropout"); |
3606 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3607 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3608 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3609 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3610 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3611 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3612 | 1 | int i; |
3613 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3614 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3615 | 1.00k | dy_tensor->data.f32[i] = i + 1; |
3616 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20 * 50), 0); |
3617 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
3618 | 1 | ccv_nnc_graph_t* graph = 0; |
3619 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3620 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3621 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3622 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3623 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3624 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3625 | 1.00k | x_tensor->data.f32[i] = (i + 1) * 0.01; |
3626 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3627 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3628 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3629 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3630 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3631 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0); |
3632 | 1 | int zero_count = 0; |
3633 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3634 | 1.00k | if (fabsf(dx_tensor->data.f32[i]) < 1e-5) |
3635 | 370 | ++zero_count; |
3636 | 630 | else { |
3637 | 630 | REQUIRE_EQ_WITH_TOLERANCE(dx_tensor->data.f32[i], dy_tensor->data.f32[i] / 0.6, 1e-3, "should match the gradient"); |
3638 | 630 | } |
3639 | 1 | REQUIRE_EQ_WITH_TOLERANCE((float)zero_count / (20 * 50), 0.4, 5 * 1e-2, "should be within 5%% of error"); |
3640 | 1 | ccv_nnc_graph_free(graph); |
3641 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3642 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3643 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3644 | 1 | ccv_nnc_tensor_free(x_tensor); |
3645 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3646 | 1 | ccv_nnc_tensor_free(dyt); |
3647 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3648 | 1 | } |
3649 | | |
3650 | | TEST_CASE("compare dropout gradient with cudnn in half precision") |
3651 | 1 | { |
3652 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3653 | 1 | ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3654 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3655 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 20 * 50), "x"); |
3656 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 20 * 50), "y"); |
3657 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, ccv_nnc_tensor_auto, "c"); |
3658 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DROPOUT_FORWARD(0.4), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y, c), "dropout"); |
3659 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3660 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3661 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3662 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3663 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3664 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3665 | 1 | int i; |
3666 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3667 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3668 | 1.00k | dy_tensor->data.f32[i] = i + 1; |
3669 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 20 * 50), 0); |
3670 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 20 * 50), 0); |
3671 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
3672 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
3673 | 1 | ccv_nnc_graph_t* graph = 0; |
3674 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3675 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3676 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3677 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3678 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3679 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3680 | 1.00k | x_tensor->data.f32[i] = (i + 1) * 0.01; |
3681 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 20 * 50), 0); |
3682 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3683 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3684 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3685 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3686 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3687 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20 * 50), 0); |
3688 | 1 | ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 20 * 50), 0); |
3689 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0); |
3690 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0); |
3691 | 1 | int zero_count = 0; |
3692 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3693 | 1.00k | if (fabsf(dx_tensor->data.f32[i]) < 1e-5) |
3694 | 416 | ++zero_count; |
3695 | 584 | else { |
3696 | 584 | REQUIRE_EQ_WITH_TOLERANCE(dx_tensor->data.f32[i], dy_tensor->data.f32[i] / 0.6, dx_tensor->data.f32[i] * 1e-3, "should match the gradient"); |
3697 | 584 | } |
3698 | 1 | REQUIRE_EQ_WITH_TOLERANCE((float)zero_count / (20 * 50), 0.4, 5 * 1e-2, "should be within 5%% of error"); |
3699 | 1 | ccv_nnc_graph_free(graph); |
3700 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3701 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3702 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3703 | 1 | ccv_nnc_tensor_free(x_tensor); |
3704 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3705 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3706 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
3707 | 1 | ccv_nnc_tensor_free(dyt); |
3708 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3709 | 1 | ccv_nnc_tensor_free(dx16_tensor); |
3710 | 1 | } |
3711 | | |
3712 | | TEST_CASE("dropout entire matrix with 20% chance") |
3713 | 1 | { |
3714 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3715 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3716 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3717 | 1 | int i; |
3718 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3719 | 1.00k | ha->data.f32[i] = (i + 1) * 0.01; |
3720 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20, 50), 0); |
3721 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
3722 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20, 50), 0); |
3723 | 1 | ccv_nnc_tensor_param_t output_info[2]; |
3724 | 1 | ccv_nnc_hint_tensor_auto(CMD_DROPOUT_FORWARD(0.4), &a->info, 1, ccv_nnc_no_hint, output_info, 2); |
3725 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, output_info[1], 0); |
3726 | 1 | ccv_nnc_cmd_exec(CMD_DROPOUT_FORWARD(0.2, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b, c), 0); |
3727 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
3728 | 1 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3729 | 1 | if (hb->data.f32[0] == 0) |
3730 | 0 | for (i = 0; i < 20 * 50; i++) |
3731 | 0 | d->data.f32[i] = 0; |
3732 | 1 | else |
3733 | 1.00k | for (i = 0; 1 i < 20 * 50; i++1.00k ) |
3734 | 1.00k | d->data.f32[i] = ha->data.f32[i] / 0.8; |
3735 | 1 | REQUIRE_TENSOR_EQ(hb, d, "dropout chance should be equal"); |
3736 | 1 | ccv_nnc_tensor_free(ha); |
3737 | 1 | ccv_nnc_tensor_free(hb); |
3738 | 1 | ccv_nnc_tensor_free(a); |
3739 | 1 | ccv_nnc_tensor_free(b); |
3740 | 1 | ccv_nnc_tensor_free(c); |
3741 | 1 | ccv_nnc_tensor_free(d); |
3742 | 1 | } |
3743 | | |
3744 | | TEST_CASE("dropout gradient entire matrix with 20% chance") |
3745 | 1 | { |
3746 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3747 | 1 | ccv_nnc_cmd_ok(CCV_NNC_DROPOUT_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3748 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3749 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3750 | 1 | int i; |
3751 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3752 | 1.00k | ha->data.f32[i] = (i + 1) * 0.01; |
3753 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20, 50), 0); |
3754 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
3755 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20, 50), 0); |
3756 | 1 | ccv_nnc_tensor_param_t output_info[2]; |
3757 | 1 | ccv_nnc_hint_tensor_auto(CMD_DROPOUT_FORWARD(0.4), &a->info, 1, ccv_nnc_no_hint, output_info, 2); |
3758 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, output_info[1], 0); |
3759 | 1 | ccv_nnc_cmd_exec(CMD_DROPOUT_FORWARD(0.2, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b, c), 0); |
3760 | 1 | ccv_nnc_tensor_t* const hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3761 | 1.00k | for (i = 0; i < 20 * 50; i++1.00k ) |
3762 | 1.00k | hg->data.f32[i] = i + 1; |
3763 | 1 | ccv_nnc_tensor_t* const hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3764 | 1 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20, 50), 0); |
3765 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(hg), TENSOR_LIST(g), 0); |
3766 | 1 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 20, 50), 0); |
3767 | 1 | ccv_nnc_cmd_exec(CMD_DROPOUT_BACKWARD(0.2, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(g, 0, 0, 0, c), TENSOR_LIST(h), 0); |
3768 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, h), TENSOR_LIST(hb, hh), 0); |
3769 | 1 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 20, 50), 0); |
3770 | 1 | if (hb->data.f32[0] == 0) |
3771 | 0 | for (i = 0; i < 20 * 50; i++) |
3772 | 0 | d->data.f32[i] = 0; |
3773 | 1 | else |
3774 | 1.00k | for (i = 0; 1 i < 20 * 50; i++1.00k ) |
3775 | 1.00k | d->data.f32[i] = hg->data.f32[i] / 0.8; |
3776 | 1 | REQUIRE_TENSOR_EQ(hh, d, "dropout chance should be equal"); |
3777 | 1 | ccv_nnc_tensor_free(ha); |
3778 | 1 | ccv_nnc_tensor_free(hb); |
3779 | 1 | ccv_nnc_tensor_free(hg); |
3780 | 1 | ccv_nnc_tensor_free(hh); |
3781 | 1 | ccv_nnc_tensor_free(a); |
3782 | 1 | ccv_nnc_tensor_free(b); |
3783 | 1 | ccv_nnc_tensor_free(c); |
3784 | 1 | ccv_nnc_tensor_free(g); |
3785 | 1 | ccv_nnc_tensor_free(h); |
3786 | 1 | ccv_nnc_tensor_free(d); |
3787 | 1 | } |
3788 | | |
3789 | | TEST_CASE("compare softmax with cudnn") |
3790 | 1 | { |
3791 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3792 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3793 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "a"); |
3794 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "b"); |
3795 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SOFTMAX_FORWARD(), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "softmax"); |
3796 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3797 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3798 | 1 | ccv_nnc_graph_t* graph = 0; |
3799 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3800 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3801 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3802 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3803 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
3804 | 1 | dsfmt_t dsfmt; |
3805 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3806 | 1 | int i; |
3807 | 201 | for (i = 0; i < 20 * 10; i++200 ) |
3808 | 200 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3809 | 1 | ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a); |
3810 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(a_tensor), 0); |
3811 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3812 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
3813 | 1 | ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b); |
3814 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y_tensor), 0); |
3815 | 1 | ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
3816 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0); |
3817 | 1 | REQUIRE_TENSOR_EQ(ty, y_tensor, "softmax from cudnn should match from CPU"); |
3818 | 1 | ccv_nnc_tensor_free(x_tensor); |
3819 | 1 | ccv_nnc_tensor_free(y_tensor); |
3820 | 1 | ccv_nnc_tensor_free(ty); |
3821 | 1 | ccv_nnc_graph_free(graph); |
3822 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3823 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3824 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3825 | 1 | } |
3826 | | |
3827 | | TEST_CASE("compare softmax with cudnn in half precision") |
3828 | 1 | { |
3829 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3830 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3831 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "a"); |
3832 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "b"); |
3833 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SOFTMAX_FORWARD(), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "softmax"); |
3834 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3835 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3836 | 1 | ccv_nnc_graph_t* graph = 0; |
3837 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3838 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3839 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3840 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3841 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
3842 | 1 | dsfmt_t dsfmt; |
3843 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3844 | 1 | int i; |
3845 | 201 | for (i = 0; i < 20 * 10; i++200 ) |
3846 | 200 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3847 | 1 | ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a); |
3848 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0); |
3849 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3850 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(a_tensor), 0); |
3851 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3852 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0); |
3853 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
3854 | 1 | ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b); |
3855 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y16_tensor), 0); |
3856 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
3857 | 1 | ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
3858 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0); |
3859 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty->data.f32, y_tensor->data.f32, 20 * 10, 1e-3, "softmax from cudnn should match from CPU"); |
3860 | 1 | ccv_nnc_tensor_free(x_tensor); |
3861 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3862 | 1 | ccv_nnc_tensor_free(y16_tensor); |
3863 | 1 | ccv_nnc_tensor_free(y_tensor); |
3864 | 1 | ccv_nnc_tensor_free(ty); |
3865 | 1 | ccv_nnc_graph_free(graph); |
3866 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3867 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3868 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3869 | 1 | } |
3870 | | |
3871 | | TEST_CASE("compare softmax gradient with cudnn") |
3872 | 1 | { |
3873 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3874 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3875 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3876 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "x"); |
3877 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "y"); |
3878 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SOFTMAX_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "softmax"); |
3879 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3880 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3881 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3882 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3883 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3884 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3885 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3886 | 1 | dsfmt_t dsfmt; |
3887 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3888 | 1 | int i; |
3889 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
3890 | 1.00k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3891 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3892 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
3893 | 1.00k | dy_tensor->data.f32[i] = 0; |
3894 | 11 | for (i = 0; i < 10; i++10 ) |
3895 | 10 | dy_tensor->data.f32[i * 100 + i] = 1; |
3896 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
3897 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
3898 | 1 | ccv_nnc_graph_t* graph = 0; |
3899 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3900 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3901 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3902 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3903 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3904 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
3905 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3906 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3907 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3908 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3909 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3910 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0); |
3911 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0); |
3912 | 1 | ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3913 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0); |
3914 | 1 | REQUIRE_TENSOR_EQ(ty_tensor, y_tensor, "forward pass should match"); |
3915 | 1 | ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3916 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, 0, ty_tensor), TENSOR_LIST(tdx_tensor), 0); |
3917 | 1 | REQUIRE_TENSOR_EQ(tdx_tensor, dx_tensor, "backward pass should match"); |
3918 | 1 | ccv_nnc_tensor_free(x_tensor); |
3919 | 1 | ccv_nnc_tensor_free(y_tensor); |
3920 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3921 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3922 | 1 | ccv_nnc_tensor_free(ty_tensor); |
3923 | 1 | ccv_nnc_tensor_free(tdx_tensor); |
3924 | 1 | ccv_nnc_tensor_free(dyt); |
3925 | 1 | ccv_nnc_graph_free(graph); |
3926 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3927 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
3928 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
3929 | 1 | } |
3930 | | |
3931 | | TEST_CASE("compare softmax gradient with cudnn in half precision") |
3932 | 1 | { |
3933 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
3934 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
3935 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
3936 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "x"); |
3937 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "y"); |
3938 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SOFTMAX_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "softmax"); |
3939 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3940 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
3941 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
3942 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
3943 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
3944 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
3945 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3946 | 1 | dsfmt_t dsfmt; |
3947 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
3948 | 1 | int i; |
3949 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
3950 | 1.00k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
3951 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3952 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
3953 | 1.00k | dy_tensor->data.f32[i] = 0; |
3954 | 11 | for (i = 0; i < 10; i++10 ) |
3955 | 10 | dy_tensor->data.f32[i * 100 + i] = 1; |
3956 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
3957 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
3958 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
3959 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
3960 | 1 | ccv_nnc_graph_t* graph = 0; |
3961 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
3962 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
3963 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
3964 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
3965 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
3966 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
3967 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
3968 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
3969 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
3970 | 1 | ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
3971 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3972 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
3973 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
3974 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3975 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
3976 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0); |
3977 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0); |
3978 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0); |
3979 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
3980 | 1 | ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3981 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0); |
3982 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty_tensor->data.f32, y_tensor->data.f32, 10 * 100, 1e-3, "forward pass should match"); |
3983 | 1 | ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
3984 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, 0, ty_tensor), TENSOR_LIST(tdx_tensor), 0); |
3985 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tdx_tensor->data.f32, dx_tensor->data.f32, 10 * 100, 1e-3, "backward pass should match"); |
3986 | 1 | ccv_nnc_tensor_free(x_tensor); |
3987 | 1 | ccv_nnc_tensor_free(x16_tensor); |
3988 | 1 | ccv_nnc_tensor_free(y_tensor); |
3989 | 1 | ccv_nnc_tensor_free(y16_tensor); |
3990 | 1 | ccv_nnc_tensor_free(dx_tensor); |
3991 | 1 | ccv_nnc_tensor_free(dx16_tensor); |
3992 | 1 | ccv_nnc_tensor_free(dy_tensor); |
3993 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
3994 | 1 | ccv_nnc_tensor_free(ty_tensor); |
3995 | 1 | ccv_nnc_tensor_free(tdx_tensor); |
3996 | 1 | ccv_nnc_tensor_free(dyt); |
3997 | 1 | ccv_nnc_graph_free(graph); |
3998 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
3999 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4000 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4001 | 1 | } |
4002 | | |
4003 | | TEST_CASE("compare sigmoid with cudnn") |
4004 | 1 | { |
4005 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4006 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4007 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "a"); |
4008 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "b"); |
4009 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SIGMOID_FORWARD(), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "sigmoid"); |
4010 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4011 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4012 | 1 | ccv_nnc_graph_t* graph = 0; |
4013 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4014 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4015 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4016 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4017 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4018 | 1 | dsfmt_t dsfmt; |
4019 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4020 | 1 | int i; |
4021 | 201 | for (i = 0; i < 20 * 10; i++200 ) |
4022 | 200 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4023 | 1 | ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a); |
4024 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(a_tensor), 0); |
4025 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4026 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4027 | 1 | ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b); |
4028 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y_tensor), 0); |
4029 | 1 | ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4030 | 1 | ccv_nnc_cmd_exec(CMD_SIGMOID_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0); |
4031 | 1 | REQUIRE_TENSOR_EQ(ty, y_tensor, "sigmoid from cudnn should match from CPU"); |
4032 | 1 | ccv_nnc_tensor_free(x_tensor); |
4033 | 1 | ccv_nnc_tensor_free(y_tensor); |
4034 | 1 | ccv_nnc_tensor_free(ty); |
4035 | 1 | ccv_nnc_graph_free(graph); |
4036 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4037 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4038 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4039 | 1 | } |
4040 | | |
4041 | | TEST_CASE("compare sigmoid with cudnn in half precision") |
4042 | 1 | { |
4043 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4044 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4045 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "a"); |
4046 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "b"); |
4047 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SIGMOID_FORWARD(), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "sigmoid"); |
4048 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4049 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4050 | 1 | ccv_nnc_graph_t* graph = 0; |
4051 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4052 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4053 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4054 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4055 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4056 | 1 | dsfmt_t dsfmt; |
4057 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4058 | 1 | int i; |
4059 | 201 | for (i = 0; i < 20 * 10; i++200 ) |
4060 | 200 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4061 | 1 | ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a); |
4062 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0); |
4063 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
4064 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(a_tensor), 0); |
4065 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4066 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0); |
4067 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4068 | 1 | ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b); |
4069 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y16_tensor), 0); |
4070 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
4071 | 1 | ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4072 | 1 | ccv_nnc_cmd_exec(CMD_SIGMOID_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0); |
4073 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty->data.f32, y_tensor->data.f32, 20 * 10, 1e-3, "sigmoid from cudnn should match from CPU"); |
4074 | 1 | ccv_nnc_tensor_free(x_tensor); |
4075 | 1 | ccv_nnc_tensor_free(x16_tensor); |
4076 | 1 | ccv_nnc_tensor_free(y16_tensor); |
4077 | 1 | ccv_nnc_tensor_free(y_tensor); |
4078 | 1 | ccv_nnc_tensor_free(ty); |
4079 | 1 | ccv_nnc_graph_free(graph); |
4080 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4081 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4082 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4083 | 1 | } |
4084 | | |
4085 | | TEST_CASE("compare sigmoid gradient with cudnn") |
4086 | 1 | { |
4087 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4088 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4089 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4090 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "x"); |
4091 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "y"); |
4092 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SIGMOID_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "sigmoid"); |
4093 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4094 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
4095 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4096 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4097 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
4098 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
4099 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4100 | 1 | dsfmt_t dsfmt; |
4101 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4102 | 1 | int i; |
4103 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4104 | 1.00k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4105 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4106 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4107 | 1.00k | dy_tensor->data.f32[i] = 0; |
4108 | 11 | for (i = 0; i < 10; i++10 ) |
4109 | 10 | dy_tensor->data.f32[i * 100 + i] = 1; |
4110 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4111 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
4112 | 1 | ccv_nnc_graph_t* graph = 0; |
4113 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4114 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4115 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4116 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4117 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
4118 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
4119 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4120 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4121 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
4122 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4123 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
4124 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0); |
4125 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0); |
4126 | 1 | ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4127 | 1 | ccv_nnc_cmd_exec(CMD_SIGMOID_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0); |
4128 | 1 | REQUIRE_TENSOR_EQ(ty_tensor, y_tensor, "forward pass should match"); |
4129 | 1 | ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4130 | 1 | ccv_nnc_cmd_exec(CMD_SIGMOID_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, 0, ty_tensor), TENSOR_LIST(tdx_tensor), 0); |
4131 | 1 | REQUIRE_TENSOR_EQ(tdx_tensor, dx_tensor, "backward pass should match"); |
4132 | 1 | ccv_nnc_tensor_free(x_tensor); |
4133 | 1 | ccv_nnc_tensor_free(y_tensor); |
4134 | 1 | ccv_nnc_tensor_free(dx_tensor); |
4135 | 1 | ccv_nnc_tensor_free(dy_tensor); |
4136 | 1 | ccv_nnc_tensor_free(ty_tensor); |
4137 | 1 | ccv_nnc_tensor_free(tdx_tensor); |
4138 | 1 | ccv_nnc_tensor_free(dyt); |
4139 | 1 | ccv_nnc_graph_free(graph); |
4140 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4141 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4142 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4143 | 1 | } |
4144 | | |
4145 | | TEST_CASE("compare sigmoid gradient with cudnn in half precision") |
4146 | 1 | { |
4147 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4148 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4149 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4150 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "x"); |
4151 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "y"); |
4152 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_SIGMOID_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "sigmoid"); |
4153 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4154 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
4155 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4156 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4157 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
4158 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
4159 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4160 | 1 | dsfmt_t dsfmt; |
4161 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4162 | 1 | int i; |
4163 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4164 | 1.00k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4165 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4166 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4167 | 1.00k | dy_tensor->data.f32[i] = 0; |
4168 | 11 | for (i = 0; i < 10; i++10 ) |
4169 | 10 | dy_tensor->data.f32[i * 100 + i] = 1; |
4170 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4171 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4172 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
4173 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
4174 | 1 | ccv_nnc_graph_t* graph = 0; |
4175 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4176 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4177 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4178 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4179 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
4180 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4181 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
4182 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
4183 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4184 | 1 | ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4185 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4186 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
4187 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4188 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4189 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
4190 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0); |
4191 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0); |
4192 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0); |
4193 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
4194 | 1 | ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4195 | 1 | ccv_nnc_cmd_exec(CMD_SIGMOID_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0); |
4196 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty_tensor->data.f32, y_tensor->data.f32, 10 * 100, 1e-3, "forward pass should match"); |
4197 | 1 | ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4198 | 1 | ccv_nnc_cmd_exec(CMD_SIGMOID_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, 0, ty_tensor), TENSOR_LIST(tdx_tensor), 0); |
4199 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tdx_tensor->data.f32, dx_tensor->data.f32, 10 * 100, 1e-3, "backward pass should match"); |
4200 | 1 | ccv_nnc_tensor_free(x_tensor); |
4201 | 1 | ccv_nnc_tensor_free(x16_tensor); |
4202 | 1 | ccv_nnc_tensor_free(y_tensor); |
4203 | 1 | ccv_nnc_tensor_free(y16_tensor); |
4204 | 1 | ccv_nnc_tensor_free(dx_tensor); |
4205 | 1 | ccv_nnc_tensor_free(dx16_tensor); |
4206 | 1 | ccv_nnc_tensor_free(dy_tensor); |
4207 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
4208 | 1 | ccv_nnc_tensor_free(ty_tensor); |
4209 | 1 | ccv_nnc_tensor_free(tdx_tensor); |
4210 | 1 | ccv_nnc_tensor_free(dyt); |
4211 | 1 | ccv_nnc_graph_free(graph); |
4212 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4213 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4214 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4215 | 1 | } |
4216 | | |
4217 | | TEST_CASE("compare tanh with cudnn") |
4218 | 1 | { |
4219 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_TANH_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4220 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4221 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "a"); |
4222 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "b"); |
4223 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_TANH_FORWARD(), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "tanh"); |
4224 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4225 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4226 | 1 | ccv_nnc_graph_t* graph = 0; |
4227 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4228 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4229 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4230 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4231 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4232 | 1 | dsfmt_t dsfmt; |
4233 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4234 | 1 | int i; |
4235 | 201 | for (i = 0; i < 20 * 10; i++200 ) |
4236 | 200 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4237 | 1 | ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a); |
4238 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(a_tensor), 0); |
4239 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4240 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4241 | 1 | ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b); |
4242 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y_tensor), 0); |
4243 | 1 | ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4244 | 1 | ccv_nnc_cmd_exec(CMD_TANH_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0); |
4245 | 1 | REQUIRE_TENSOR_EQ(ty, y_tensor, "tanh from cudnn should match from CPU"); |
4246 | 1 | ccv_nnc_tensor_free(x_tensor); |
4247 | 1 | ccv_nnc_tensor_free(y_tensor); |
4248 | 1 | ccv_nnc_tensor_free(ty); |
4249 | 1 | ccv_nnc_graph_free(graph); |
4250 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4251 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4252 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4253 | 1 | } |
4254 | | |
4255 | | TEST_CASE("compare tanh with cudnn in half precision") |
4256 | 1 | { |
4257 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_TANH_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4258 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4259 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "a"); |
4260 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "b"); |
4261 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_TANH_FORWARD(), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "tanh"); |
4262 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4263 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4264 | 1 | ccv_nnc_graph_t* graph = 0; |
4265 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4266 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4267 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4268 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4269 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4270 | 1 | dsfmt_t dsfmt; |
4271 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4272 | 1 | int i; |
4273 | 201 | for (i = 0; i < 20 * 10; i++200 ) |
4274 | 200 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4275 | 1 | ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a); |
4276 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0); |
4277 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
4278 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(a_tensor), 0); |
4279 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4280 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0); |
4281 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4282 | 1 | ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b); |
4283 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y16_tensor), 0); |
4284 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
4285 | 1 | ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0); |
4286 | 1 | ccv_nnc_cmd_exec(CMD_TANH_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0); |
4287 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty->data.f32, y_tensor->data.f32, 20 * 10, 1e-3, "tanh from cudnn should match from CPU"); |
4288 | 1 | ccv_nnc_tensor_free(x_tensor); |
4289 | 1 | ccv_nnc_tensor_free(x16_tensor); |
4290 | 1 | ccv_nnc_tensor_free(y16_tensor); |
4291 | 1 | ccv_nnc_tensor_free(y_tensor); |
4292 | 1 | ccv_nnc_tensor_free(ty); |
4293 | 1 | ccv_nnc_graph_free(graph); |
4294 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4295 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4296 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4297 | 1 | } |
4298 | | |
4299 | | TEST_CASE("compare tanh gradient with cudnn") |
4300 | 1 | { |
4301 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_TANH_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4302 | 1 | ccv_nnc_cmd_ok(CCV_NNC_TANH_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4303 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4304 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "x"); |
4305 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "y"); |
4306 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_TANH_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "tanh"); |
4307 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4308 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
4309 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4310 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4311 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
4312 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
4313 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4314 | 1 | dsfmt_t dsfmt; |
4315 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4316 | 1 | int i; |
4317 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4318 | 1.00k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4319 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4320 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4321 | 1.00k | dy_tensor->data.f32[i] = 0; |
4322 | 11 | for (i = 0; i < 10; i++10 ) |
4323 | 10 | dy_tensor->data.f32[i * 100 + i] = 1; |
4324 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4325 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0); |
4326 | 1 | ccv_nnc_graph_t* graph = 0; |
4327 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4328 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4329 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4330 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4331 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
4332 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0); |
4333 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4334 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4335 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
4336 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4337 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
4338 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0); |
4339 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0); |
4340 | 1 | ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4341 | 1 | ccv_nnc_cmd_exec(CMD_TANH_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0); |
4342 | 1 | REQUIRE_TENSOR_EQ(ty_tensor, y_tensor, "forward pass should match"); |
4343 | 1 | ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4344 | 1 | ccv_nnc_cmd_exec(CMD_TANH_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, 0, ty_tensor), TENSOR_LIST(tdx_tensor), 0); |
4345 | 1 | REQUIRE_TENSOR_EQ(tdx_tensor, dx_tensor, "backward pass should match"); |
4346 | 1 | ccv_nnc_tensor_free(x_tensor); |
4347 | 1 | ccv_nnc_tensor_free(y_tensor); |
4348 | 1 | ccv_nnc_tensor_free(dx_tensor); |
4349 | 1 | ccv_nnc_tensor_free(dy_tensor); |
4350 | 1 | ccv_nnc_tensor_free(ty_tensor); |
4351 | 1 | ccv_nnc_tensor_free(tdx_tensor); |
4352 | 1 | ccv_nnc_tensor_free(dyt); |
4353 | 1 | ccv_nnc_graph_free(graph); |
4354 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4355 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4356 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4357 | 1 | } |
4358 | | |
4359 | | TEST_CASE("compare tanh gradient with cudnn in half precision") |
4360 | 1 | { |
4361 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_TANH_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4362 | 1 | ccv_nnc_cmd_ok(CCV_NNC_TANH_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4363 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4364 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "x"); |
4365 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "y"); |
4366 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_TANH_FORWARD(), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "tanh"); |
4367 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4368 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
4369 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4370 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4371 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
4372 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
4373 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4374 | 1 | dsfmt_t dsfmt; |
4375 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4376 | 1 | int i; |
4377 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4378 | 1.00k | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4379 | 1 | ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4380 | 1.00k | for (i = 0; i < 10 * 100; i++1.00k ) |
4381 | 1.00k | dy_tensor->data.f32[i] = 0; |
4382 | 11 | for (i = 0; i < 10; i++10 ) |
4383 | 10 | dy_tensor->data.f32[i * 100 + i] = 1; |
4384 | 1 | ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4385 | 1 | ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4386 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0); |
4387 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0); |
4388 | 1 | ccv_nnc_graph_t* graph = 0; |
4389 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4390 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4391 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4392 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4393 | 1 | ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x); |
4394 | 1 | ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4395 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0); |
4396 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0); |
4397 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4398 | 1 | ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4399 | 1 | ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4400 | 1 | ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
4401 | 1 | ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4402 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4403 | 1 | ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y); |
4404 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0); |
4405 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0); |
4406 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0); |
4407 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0); |
4408 | 1 | ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4409 | 1 | ccv_nnc_cmd_exec(CMD_TANH_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0); |
4410 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty_tensor->data.f32, y_tensor->data.f32, 10 * 100, 1e-3, "forward pass should match"); |
4411 | 1 | ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4412 | 1 | ccv_nnc_cmd_exec(CMD_TANH_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, 0, ty_tensor), TENSOR_LIST(tdx_tensor), 0); |
4413 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tdx_tensor->data.f32, dx_tensor->data.f32, 10 * 100, 1e-3, "backward pass should match"); |
4414 | 1 | ccv_nnc_tensor_free(x_tensor); |
4415 | 1 | ccv_nnc_tensor_free(x16_tensor); |
4416 | 1 | ccv_nnc_tensor_free(y_tensor); |
4417 | 1 | ccv_nnc_tensor_free(y16_tensor); |
4418 | 1 | ccv_nnc_tensor_free(dx_tensor); |
4419 | 1 | ccv_nnc_tensor_free(dx16_tensor); |
4420 | 1 | ccv_nnc_tensor_free(dy_tensor); |
4421 | 1 | ccv_nnc_tensor_free(dy16_tensor); |
4422 | 1 | ccv_nnc_tensor_free(ty_tensor); |
4423 | 1 | ccv_nnc_tensor_free(tdx_tensor); |
4424 | 1 | ccv_nnc_tensor_free(dyt); |
4425 | 1 | ccv_nnc_graph_free(graph); |
4426 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4427 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4428 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4429 | 1 | } |
4430 | | |
4431 | | TEST_CASE("compare add with cudnn") |
4432 | 1 | { |
4433 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4434 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4435 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "x"); |
4436 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), "y"); |
4437 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 5, 5, 3), "a"); |
4438 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 5, 1, 3), "b"); |
4439 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 5, 5, 3), "c"); |
4440 | 1 | ccv_nnc_tensor_symbol_t z = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "z"); |
4441 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x, y), TENSOR_SYMBOL_LIST(a, b), "transfer"); |
4442 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_ADD_FORWARD(0.5, 0.2), TENSOR_SYMBOL_LIST(a, b), TENSOR_SYMBOL_LIST(c), "add"); |
4443 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(c), TENSOR_SYMBOL_LIST(z), "transfer"); |
4444 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4445 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4446 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4447 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4448 | 1 | ccv_nnc_graph_t* graph = 0; |
4449 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4450 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4451 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(x, x_tensor), KV(y, y_tensor)), TENSOR_SYMBOL_LIST(z), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4452 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4453 | 1 | dsfmt_t dsfmt; |
4454 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4455 | 1 | int i; |
4456 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4457 | 750 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4458 | 151 | for (i = 0; i < 10 * 5 * 1 * 3; i++150 ) |
4459 | 150 | y_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4460 | 1 | ccv_nnc_tensor_t* zt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4461 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor, y_tensor), TENSOR_LIST(zt), 0); |
4462 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4463 | 1 | ccv_nnc_tensor_t* const z_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, z); |
4464 | 1 | REQUIRE_TENSOR_EQ(zt, z_tensor, "add should match"); |
4465 | 1 | ccv_nnc_tensor_free(x_tensor); |
4466 | 1 | ccv_nnc_tensor_free(y_tensor); |
4467 | 1 | ccv_nnc_tensor_free(zt); |
4468 | 1 | ccv_nnc_graph_free(graph); |
4469 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4470 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4471 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4472 | 1 | } |
4473 | | |
4474 | | TEST_CASE("compare add with cudnn in half precision") |
4475 | 1 | { |
4476 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4477 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4478 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "x"); |
4479 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), "y"); |
4480 | 1 | ccv_nnc_tensor_symbol_t x16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 10, 5, 5, 3), "x 16"); |
4481 | 1 | ccv_nnc_tensor_symbol_t y16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 10, 5, 1, 3), "y 16"); |
4482 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 5, 5, 3), "a"); |
4483 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 5, 1, 3), "b"); |
4484 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 5, 5, 3), "c"); |
4485 | 1 | ccv_nnc_tensor_symbol_t z = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "z"); |
4486 | 1 | ccv_nnc_tensor_symbol_t z16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 10, 5, 5, 3), "z 16"); |
4487 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(x, y), TENSOR_SYMBOL_LIST(x16, y16), "convert"); |
4488 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x16, y16), TENSOR_SYMBOL_LIST(a, b), "transfer"); |
4489 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_ADD_FORWARD(0.5, 0.2), TENSOR_SYMBOL_LIST(a, b), TENSOR_SYMBOL_LIST(c), "add"); |
4490 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(c), TENSOR_SYMBOL_LIST(z16), "transfer"); |
4491 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(z16), TENSOR_SYMBOL_LIST(z), "convert"); |
4492 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4493 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4494 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4495 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4496 | 1 | ccv_nnc_graph_t* graph = 0; |
4497 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4498 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4499 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(x, x_tensor), KV(y, y_tensor)), TENSOR_SYMBOL_LIST(z), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4500 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4501 | 1 | dsfmt_t dsfmt; |
4502 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4503 | 1 | int i; |
4504 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4505 | 750 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4506 | 151 | for (i = 0; i < 10 * 5 * 1 * 3; i++150 ) |
4507 | 150 | y_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4508 | 1 | ccv_nnc_tensor_t* zt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4509 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor, y_tensor), TENSOR_LIST(zt), 0); |
4510 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4511 | 1 | ccv_nnc_tensor_t* const z_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, z); |
4512 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, zt->data.f32, z_tensor->data.f32, 10 * 5 * 5 * 3, 1e-3, "add should match"); |
4513 | 1 | ccv_nnc_tensor_free(x_tensor); |
4514 | 1 | ccv_nnc_tensor_free(y_tensor); |
4515 | 1 | ccv_nnc_tensor_free(zt); |
4516 | 1 | ccv_nnc_graph_free(graph); |
4517 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4518 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4519 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4520 | 1 | } |
4521 | | |
4522 | | TEST_CASE("compare add with cudnn in bfloat precision") |
4523 | 1 | { |
4524 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4525 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4526 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "x"); |
4527 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), "y"); |
4528 | 1 | ccv_nnc_tensor_symbol_t x16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16BF, 10, 5, 5, 3), "x 16"); |
4529 | 1 | ccv_nnc_tensor_symbol_t y16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16BF, 10, 5, 1, 3), "y 16"); |
4530 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16BF, 10, 5, 5, 3), "a"); |
4531 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16BF, 10, 5, 1, 3), "b"); |
4532 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16BF, 10, 5, 5, 3), "c"); |
4533 | 1 | ccv_nnc_tensor_symbol_t z = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "z"); |
4534 | 1 | ccv_nnc_tensor_symbol_t z16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16BF, 10, 5, 5, 3), "z 16"); |
4535 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(x, y), TENSOR_SYMBOL_LIST(x16, y16), "convert"); |
4536 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x16, y16), TENSOR_SYMBOL_LIST(a, b), "transfer"); |
4537 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_ADD_FORWARD(0.5, 0.2), TENSOR_SYMBOL_LIST(a, b), TENSOR_SYMBOL_LIST(c), "add"); |
4538 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(c), TENSOR_SYMBOL_LIST(z16), "transfer"); |
4539 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(z16), TENSOR_SYMBOL_LIST(z), "convert"); |
4540 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4541 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4542 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4543 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4544 | 1 | ccv_nnc_graph_t* graph = 0; |
4545 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4546 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4547 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(x, x_tensor), KV(y, y_tensor)), TENSOR_SYMBOL_LIST(z), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4548 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4549 | 1 | dsfmt_t dsfmt; |
4550 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4551 | 1 | int i; |
4552 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4553 | 750 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4554 | 151 | for (i = 0; i < 10 * 5 * 1 * 3; i++150 ) |
4555 | 150 | y_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4556 | 1 | ccv_nnc_tensor_t* zt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4557 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor, y_tensor), TENSOR_LIST(zt), 0); |
4558 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4559 | 1 | ccv_nnc_tensor_t* const z_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, z); |
4560 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, zt->data.f32, z_tensor->data.f32, 10 * 5 * 5 * 3, 1e-2, "add should match"); |
4561 | 1 | ccv_nnc_tensor_free(x_tensor); |
4562 | 1 | ccv_nnc_tensor_free(y_tensor); |
4563 | 1 | ccv_nnc_tensor_free(zt); |
4564 | 1 | ccv_nnc_graph_free(graph); |
4565 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4566 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4567 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4568 | 1 | } |
4569 | | |
4570 | | TEST_CASE("compare add gradient with cudnn") |
4571 | 1 | { |
4572 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4573 | 1 | ccv_nnc_cmd_ok(CCV_NNC_ADD_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4574 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4575 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "x"); |
4576 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), "y"); |
4577 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 5, 5, 3), "a"); |
4578 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 5, 1, 3), "b"); |
4579 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 5, 5, 3), "c"); |
4580 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x, y), TENSOR_SYMBOL_LIST(a, b), "transfer"); |
4581 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_ADD_FORWARD(0.5, 0.2), TENSOR_SYMBOL_LIST(a, b), TENSOR_SYMBOL_LIST(c), "add"); |
4582 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4583 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(c), TENSOR_SYMBOL_LIST(x, y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
4584 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4585 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4586 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4587 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4588 | 1 | ccv_nnc_graph_t* graph = 0; |
4589 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4590 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4591 | 1 | ccv_nnc_tensor_symbol_t dc = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, c); |
4592 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
4593 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
4594 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(x, x_tensor), KV(y, y_tensor)), TENSOR_SYMBOL_LIST(dx, dy), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4595 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4596 | 1 | dsfmt_t dsfmt; |
4597 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4598 | 1 | int i; |
4599 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4600 | 750 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4601 | 151 | for (i = 0; i < 10 * 5 * 1 * 3; i++150 ) |
4602 | 150 | y_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4603 | 1 | ccv_nnc_tensor_t* dct = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4604 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4605 | 750 | dct->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4606 | 1 | ccv_nnc_tensor_t* const dc_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dc); |
4607 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dct), TENSOR_LIST(dc_tensor), 0); |
4608 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4609 | 1 | ccv_nnc_tensor_t* zt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4610 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor, y_tensor), TENSOR_LIST(zt), 0); |
4611 | 1 | ccv_nnc_tensor_t* dxt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4612 | 1 | ccv_nnc_tensor_t* dyt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4613 | 1 | ccv_nnc_cmd_exec(CMD_ADD_BACKWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(dct, x_tensor, y_tensor, zt), TENSOR_LIST(dxt, dyt), 0); |
4614 | 1 | ccv_nnc_tensor_t* dx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
4615 | 1 | ccv_nnc_tensor_t* dy_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dy); |
4616 | 1 | REQUIRE_TENSOR_EQ(dxt, dx_tensor, "backward pass should match"); |
4617 | 1 | REQUIRE_TENSOR_EQ(dyt, dy_tensor, "backward pass should match"); |
4618 | 1 | ccv_nnc_tensor_free(x_tensor); |
4619 | 1 | ccv_nnc_tensor_free(y_tensor); |
4620 | 1 | ccv_nnc_tensor_free(dct); |
4621 | 1 | ccv_nnc_tensor_free(zt); |
4622 | 1 | ccv_nnc_tensor_free(dxt); |
4623 | 1 | ccv_nnc_tensor_free(dyt); |
4624 | 1 | ccv_nnc_graph_free(graph); |
4625 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4626 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4627 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4628 | 1 | } |
4629 | | |
4630 | | TEST_CASE("compare add gradient with cudnn in half precision") |
4631 | 1 | { |
4632 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4633 | 1 | ccv_nnc_cmd_ok(CCV_NNC_ADD_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4634 | 1 | ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new(); |
4635 | 1 | ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), "x"); |
4636 | 1 | ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), "y"); |
4637 | 1 | ccv_nnc_tensor_symbol_t x16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 10, 5, 5, 3), "x 16"); |
4638 | 1 | ccv_nnc_tensor_symbol_t y16 = ccv_nnc_tensor_symbol_new(symbolic_graph, CPU_TENSOR_NHWC(16F, 10, 5, 1, 3), "y 16"); |
4639 | 1 | ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 5, 5, 3), "a"); |
4640 | 1 | ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 5, 1, 3), "b"); |
4641 | 1 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 5, 5, 3), "c"); |
4642 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATATYPE_CONVERSION_FORWARD(), TENSOR_SYMBOL_LIST(x, y), TENSOR_SYMBOL_LIST(x16, y16), "convert"); |
4643 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_DATA_TRANSFER_FORWARD(), TENSOR_SYMBOL_LIST(x16, y16), TENSOR_SYMBOL_LIST(a, b), "transfer"); |
4644 | 1 | ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_ADD_FORWARD(0.5, 0.2), TENSOR_SYMBOL_LIST(a, b), TENSOR_SYMBOL_LIST(c), "add"); |
4645 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4646 | 1 | ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(c), TENSOR_SYMBOL_LIST(x, y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph)); |
4647 | 1 | ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
4648 | 1 | SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH); |
4649 | 1 | ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4650 | 1 | ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4651 | 1 | ccv_nnc_graph_t* graph = 0; |
4652 | 1 | ccv_nnc_tensor_arena_t* tensor_arena = 0; |
4653 | 1 | ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0; |
4654 | 1 | ccv_nnc_tensor_symbol_t dc = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, c); |
4655 | 1 | ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x); |
4656 | 1 | ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y); |
4657 | 1 | ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(x, x_tensor), KV(y, y_tensor)), TENSOR_SYMBOL_LIST(dx, dy), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena); |
4658 | 1 | GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH); |
4659 | 1 | dsfmt_t dsfmt; |
4660 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4661 | 1 | int i; |
4662 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4663 | 750 | x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4664 | 151 | for (i = 0; i < 10 * 5 * 1 * 3; i++150 ) |
4665 | 150 | y_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4666 | 1 | ccv_nnc_tensor_t* dct = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4667 | 1 | ccv_nnc_tensor_t* dct16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 5, 5, 3), 0); |
4668 | 751 | for (i = 0; i < 10 * 5 * 5 * 3; i++750 ) |
4669 | 750 | dct->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4670 | 1 | ccv_nnc_tensor_t* const dc_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dc); |
4671 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dct), TENSOR_LIST(dct16), 0); |
4672 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dct16), TENSOR_LIST(dc_tensor), 0); |
4673 | 1 | ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0); |
4674 | 1 | ccv_nnc_tensor_t* zt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4675 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor, y_tensor), TENSOR_LIST(zt), 0); |
4676 | 1 | ccv_nnc_tensor_t* dxt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 5, 3), 0); |
4677 | 1 | ccv_nnc_tensor_t* dyt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 5, 1, 3), 0); |
4678 | 1 | ccv_nnc_cmd_exec(CMD_ADD_BACKWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(dct, x_tensor, y_tensor, zt), TENSOR_LIST(dxt, dyt), 0); |
4679 | 1 | ccv_nnc_tensor_t* dx_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dx); |
4680 | 1 | ccv_nnc_tensor_t* dy_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, dy); |
4681 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dxt->data.f32, dx_tensor->data.f32, 10 * 5 * 5 * 3, 1e-3, "backward pass should match"); |
4682 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, dyt->data.f32, dy_tensor->data.f32, 10 * 5 * 1 * 3, 1e-3, "backward pass should match"); |
4683 | 1 | ccv_nnc_tensor_free(x_tensor); |
4684 | 1 | ccv_nnc_tensor_free(y_tensor); |
4685 | 1 | ccv_nnc_tensor_free(dct); |
4686 | 1 | ccv_nnc_tensor_free(dct16); |
4687 | 1 | ccv_nnc_tensor_free(zt); |
4688 | 1 | ccv_nnc_tensor_free(dxt); |
4689 | 1 | ccv_nnc_tensor_free(dyt); |
4690 | 1 | ccv_nnc_graph_free(graph); |
4691 | 1 | ccv_nnc_tensor_arena_free(tensor_arena); |
4692 | 1 | ccv_nnc_graph_exec_arena_free(graph_exec_arena); |
4693 | 1 | ccv_nnc_symbolic_graph_free(symbolic_graph); |
4694 | 1 | } |
4695 | | |
4696 | | TEST_CASE("compare softmax cross entropy forward") |
4697 | 1 | { |
4698 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4699 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4700 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4701 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4702 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4703 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4704 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4705 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4706 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4707 | 1 | dsfmt_t dsfmt; |
4708 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4709 | 1 | int i; |
4710 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4711 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4712 | 11 | for (i = 0; i < 10; i++10 ) |
4713 | 10 | hb->data.f32[i] = (i + 1) * 9; |
4714 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(a, b), 0); |
4715 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
4716 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
4717 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4718 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4719 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d), TENSOR_LIST(tc, td), 0); |
4720 | 1 | REQUIRE_TENSOR_EQ(tc, hc, "GPU computed output should be the same as CPU computed ones"); |
4721 | 1 | REQUIRE_TENSOR_EQ(td, hd, "GPU computed output should be the same as CPU computed ones"); |
4722 | 1 | ccv_nnc_tensor_free(a); |
4723 | 1 | ccv_nnc_tensor_free(b); |
4724 | 1 | ccv_nnc_tensor_free(c); |
4725 | 1 | ccv_nnc_tensor_free(d); |
4726 | 1 | ccv_nnc_tensor_free(ha); |
4727 | 1 | ccv_nnc_tensor_free(hb); |
4728 | 1 | ccv_nnc_tensor_free(hc); |
4729 | 1 | ccv_nnc_tensor_free(hd); |
4730 | 1 | ccv_nnc_tensor_free(tc); |
4731 | 1 | ccv_nnc_tensor_free(td); |
4732 | 1 | } |
4733 | | |
4734 | | TEST_CASE("compare softmax cross entropy forward in half precision") |
4735 | 1 | { |
4736 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4737 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4738 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4739 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4740 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4741 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4742 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4743 | 1 | ccv_nnc_tensor_t* ha16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4744 | 1 | ccv_nnc_tensor_t* hb16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
4745 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4746 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4747 | 1 | dsfmt_t dsfmt; |
4748 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4749 | 1 | int i; |
4750 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4751 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4752 | 11 | for (i = 0; i < 10; i++10 ) |
4753 | 10 | hb->data.f32[i] = (i + 1) * 9; |
4754 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(ha16, hb16), 0); |
4755 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha16, hb16), TENSOR_LIST(a, b), 0); |
4756 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
4757 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
4758 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4759 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4760 | 1 | ccv_nnc_tensor_t* tc16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
4761 | 1 | ccv_nnc_tensor_t* td16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4762 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d), TENSOR_LIST(tc16, td16), 0); |
4763 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tc16, td16), TENSOR_LIST(tc, td), 0); |
4764 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tc->data.f32, hc->data.f32, 10, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
4765 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, td->data.f32, hd->data.f32, 10 * 100, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
4766 | 1 | ccv_nnc_tensor_free(a); |
4767 | 1 | ccv_nnc_tensor_free(b); |
4768 | 1 | ccv_nnc_tensor_free(c); |
4769 | 1 | ccv_nnc_tensor_free(d); |
4770 | 1 | ccv_nnc_tensor_free(ha); |
4771 | 1 | ccv_nnc_tensor_free(hb); |
4772 | 1 | ccv_nnc_tensor_free(ha16); |
4773 | 1 | ccv_nnc_tensor_free(hb16); |
4774 | 1 | ccv_nnc_tensor_free(hc); |
4775 | 1 | ccv_nnc_tensor_free(hd); |
4776 | 1 | ccv_nnc_tensor_free(tc); |
4777 | 1 | ccv_nnc_tensor_free(td); |
4778 | 1 | ccv_nnc_tensor_free(tc16); |
4779 | 1 | ccv_nnc_tensor_free(td16); |
4780 | 1 | } |
4781 | | |
4782 | | TEST_CASE("compare softmax cross entropy forward with label smoothing") |
4783 | 1 | { |
4784 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4785 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4786 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4787 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4788 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4789 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4790 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4791 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4792 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4793 | 1 | dsfmt_t dsfmt; |
4794 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4795 | 1 | int i; |
4796 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4797 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4798 | 11 | for (i = 0; i < 10; i++10 ) |
4799 | 10 | hb->data.f32[i] = (i + 1) * 9; |
4800 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(a, b), 0); |
4801 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
4802 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
4803 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4804 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4805 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d), TENSOR_LIST(tc, td), 0); |
4806 | 1 | REQUIRE_TENSOR_EQ(tc, hc, "GPU computed output should be the same as CPU computed ones"); |
4807 | 1 | REQUIRE_TENSOR_EQ(td, hd, "GPU computed output should be the same as CPU computed ones"); |
4808 | 1 | ccv_nnc_tensor_free(a); |
4809 | 1 | ccv_nnc_tensor_free(b); |
4810 | 1 | ccv_nnc_tensor_free(c); |
4811 | 1 | ccv_nnc_tensor_free(d); |
4812 | 1 | ccv_nnc_tensor_free(ha); |
4813 | 1 | ccv_nnc_tensor_free(hb); |
4814 | 1 | ccv_nnc_tensor_free(hc); |
4815 | 1 | ccv_nnc_tensor_free(hd); |
4816 | 1 | ccv_nnc_tensor_free(tc); |
4817 | 1 | ccv_nnc_tensor_free(td); |
4818 | 1 | } |
4819 | | |
4820 | | TEST_CASE("compare softmax cross entropy forward in half precision with label smoothing") |
4821 | 1 | { |
4822 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4823 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4824 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4825 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4826 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4827 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4828 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4829 | 1 | ccv_nnc_tensor_t* ha16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4830 | 1 | ccv_nnc_tensor_t* hb16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
4831 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4832 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4833 | 1 | dsfmt_t dsfmt; |
4834 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4835 | 1 | int i; |
4836 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4837 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4838 | 11 | for (i = 0; i < 10; i++10 ) |
4839 | 10 | hb->data.f32[i] = (i + 1) * 9; |
4840 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(ha16, hb16), 0); |
4841 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha16, hb16), TENSOR_LIST(a, b), 0); |
4842 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
4843 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
4844 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4845 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4846 | 1 | ccv_nnc_tensor_t* tc16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
4847 | 1 | ccv_nnc_tensor_t* td16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4848 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d), TENSOR_LIST(tc16, td16), 0); |
4849 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tc16, td16), TENSOR_LIST(tc, td), 0); |
4850 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tc->data.f32, hc->data.f32, 10, 5e-2, "GPU computed output should be the same as CPU computed ones"); |
4851 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, td->data.f32, hd->data.f32, 10 * 100, 5e-2, "GPU computed output should be the same as CPU computed ones"); |
4852 | 1 | ccv_nnc_tensor_free(a); |
4853 | 1 | ccv_nnc_tensor_free(b); |
4854 | 1 | ccv_nnc_tensor_free(c); |
4855 | 1 | ccv_nnc_tensor_free(d); |
4856 | 1 | ccv_nnc_tensor_free(ha); |
4857 | 1 | ccv_nnc_tensor_free(hb); |
4858 | 1 | ccv_nnc_tensor_free(ha16); |
4859 | 1 | ccv_nnc_tensor_free(hb16); |
4860 | 1 | ccv_nnc_tensor_free(hc); |
4861 | 1 | ccv_nnc_tensor_free(hd); |
4862 | 1 | ccv_nnc_tensor_free(tc); |
4863 | 1 | ccv_nnc_tensor_free(td); |
4864 | 1 | ccv_nnc_tensor_free(tc16); |
4865 | 1 | ccv_nnc_tensor_free(td16); |
4866 | 1 | } |
4867 | | |
4868 | | TEST_CASE("compare softmax cross entropy backward") |
4869 | 1 | { |
4870 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4871 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4872 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4873 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4874 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4875 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4876 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4877 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4878 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4879 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4880 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4881 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4882 | 1 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4883 | 1 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4884 | 1 | dsfmt_t dsfmt; |
4885 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4886 | 1 | int i; |
4887 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4888 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4889 | 11 | for (i = 0; i < 10; i++10 ) |
4890 | 10 | hb->data.f32[i] = (i + 1) * 9; |
4891 | 11 | for (i = 0; i < 10; i++10 ) |
4892 | 10 | hg->data.f32[i] = i * 0.1; |
4893 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hg), TENSOR_LIST(a, b, g), 0); |
4894 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
4895 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, 0, 0, hb, hc, hd), TENSOR_LIST(hh, 0), 0); |
4896 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
4897 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, 0, 0, b, c, d), TENSOR_LIST(h, 0), 0); |
4898 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4899 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4900 | 1 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4901 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d, h), TENSOR_LIST(tc, td, th), 0); |
4902 | 1 | REQUIRE_TENSOR_EQ(tc, hc, "GPU computed output should be the same as CPU computed ones"); |
4903 | 1 | REQUIRE_TENSOR_EQ(td, hd, "GPU computed output should be the same as CPU computed ones"); |
4904 | 1 | REQUIRE_TENSOR_EQ(th, hh, "GPU computed output should be the same as CPU computed ones"); |
4905 | 1 | ccv_nnc_tensor_free(a); |
4906 | 1 | ccv_nnc_tensor_free(b); |
4907 | 1 | ccv_nnc_tensor_free(c); |
4908 | 1 | ccv_nnc_tensor_free(d); |
4909 | 1 | ccv_nnc_tensor_free(h); |
4910 | 1 | ccv_nnc_tensor_free(ha); |
4911 | 1 | ccv_nnc_tensor_free(hb); |
4912 | 1 | ccv_nnc_tensor_free(hc); |
4913 | 1 | ccv_nnc_tensor_free(hd); |
4914 | 1 | ccv_nnc_tensor_free(hg); |
4915 | 1 | ccv_nnc_tensor_free(hh); |
4916 | 1 | ccv_nnc_tensor_free(tc); |
4917 | 1 | ccv_nnc_tensor_free(td); |
4918 | 1 | ccv_nnc_tensor_free(th); |
4919 | 1 | } |
4920 | | |
4921 | | TEST_CASE("compare softmax cross entropy backward with label smoothing") |
4922 | 1 | { |
4923 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4924 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4925 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4926 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4927 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4928 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4929 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
4930 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0); |
4931 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4932 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4933 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4934 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4935 | 1 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4936 | 1 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4937 | 1 | dsfmt_t dsfmt; |
4938 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4939 | 1 | int i; |
4940 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4941 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4942 | 11 | for (i = 0; i < 10; i++10 ) |
4943 | 10 | hb->data.f32[i] = (i + 1) * 9; |
4944 | 11 | for (i = 0; i < 10; i++10 ) |
4945 | 10 | hg->data.f32[i] = i * 0.1; |
4946 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hg), TENSOR_LIST(a, b, g), 0); |
4947 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
4948 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, 0, 0, hb, hc, hd), TENSOR_LIST(hh, 0), 0); |
4949 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
4950 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(g, 0, 0, b, c, d), TENSOR_LIST(h, 0), 0); |
4951 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4952 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4953 | 1 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4954 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d, h), TENSOR_LIST(tc, td, th), 0); |
4955 | 1 | REQUIRE_TENSOR_EQ(tc, hc, "GPU computed output should be the same as CPU computed ones"); |
4956 | 1 | REQUIRE_TENSOR_EQ(td, hd, "GPU computed output should be the same as CPU computed ones"); |
4957 | 1 | REQUIRE_TENSOR_EQ(th, hh, "GPU computed output should be the same as CPU computed ones"); |
4958 | 1 | ccv_nnc_tensor_free(a); |
4959 | 1 | ccv_nnc_tensor_free(b); |
4960 | 1 | ccv_nnc_tensor_free(c); |
4961 | 1 | ccv_nnc_tensor_free(d); |
4962 | 1 | ccv_nnc_tensor_free(h); |
4963 | 1 | ccv_nnc_tensor_free(ha); |
4964 | 1 | ccv_nnc_tensor_free(hb); |
4965 | 1 | ccv_nnc_tensor_free(hc); |
4966 | 1 | ccv_nnc_tensor_free(hd); |
4967 | 1 | ccv_nnc_tensor_free(hg); |
4968 | 1 | ccv_nnc_tensor_free(hh); |
4969 | 1 | ccv_nnc_tensor_free(tc); |
4970 | 1 | ccv_nnc_tensor_free(td); |
4971 | 1 | ccv_nnc_tensor_free(th); |
4972 | 1 | } |
4973 | | |
4974 | | TEST_CASE("compare softmax cross entropy backward in half precision") |
4975 | 1 | { |
4976 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
4977 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
4978 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4979 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4980 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4981 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4982 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
4983 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
4984 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4985 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4986 | 1 | ccv_nnc_tensor_t* ha16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
4987 | 1 | ccv_nnc_tensor_t* hb16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
4988 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4989 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4990 | 1 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
4991 | 1 | ccv_nnc_tensor_t* hg16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
4992 | 1 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
4993 | 1 | dsfmt_t dsfmt; |
4994 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
4995 | 1 | int i; |
4996 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
4997 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
4998 | 11 | for (i = 0; i < 10; i++10 ) |
4999 | 10 | hb->data.f32[i] = (i + 1) * 9; |
5000 | 11 | for (i = 0; i < 10; i++10 ) |
5001 | 10 | hg->data.f32[i] = i * 0.1; |
5002 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hg), TENSOR_LIST(ha16, hb16, hg16), 0); |
5003 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha16, hb16, hg16), TENSOR_LIST(a, b, g), 0); |
5004 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
5005 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, 0, 0, hb, hc, hd), TENSOR_LIST(hh, 0), 0); |
5006 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
5007 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, 0, 0, b, c, d), TENSOR_LIST(h, 0), 0); |
5008 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
5009 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5010 | 1 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5011 | 1 | ccv_nnc_tensor_t* tc16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
5012 | 1 | ccv_nnc_tensor_t* td16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
5013 | 1 | ccv_nnc_tensor_t* th16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
5014 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d, h), TENSOR_LIST(tc16, td16, th16), 0); |
5015 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tc16, td16, th16), TENSOR_LIST(tc, td, th), 0); |
5016 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tc->data.f32, hc->data.f32, 10, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
5017 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, td->data.f32, hd->data.f32, 10 * 100, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
5018 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, th->data.f32, hh->data.f32, 10 * 100, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
5019 | 1 | ccv_nnc_tensor_free(a); |
5020 | 1 | ccv_nnc_tensor_free(b); |
5021 | 1 | ccv_nnc_tensor_free(c); |
5022 | 1 | ccv_nnc_tensor_free(d); |
5023 | 1 | ccv_nnc_tensor_free(h); |
5024 | 1 | ccv_nnc_tensor_free(ha); |
5025 | 1 | ccv_nnc_tensor_free(hb); |
5026 | 1 | ccv_nnc_tensor_free(ha16); |
5027 | 1 | ccv_nnc_tensor_free(hb16); |
5028 | 1 | ccv_nnc_tensor_free(hc); |
5029 | 1 | ccv_nnc_tensor_free(hd); |
5030 | 1 | ccv_nnc_tensor_free(hg); |
5031 | 1 | ccv_nnc_tensor_free(hg16); |
5032 | 1 | ccv_nnc_tensor_free(hh); |
5033 | 1 | ccv_nnc_tensor_free(tc); |
5034 | 1 | ccv_nnc_tensor_free(td); |
5035 | 1 | ccv_nnc_tensor_free(th); |
5036 | 1 | ccv_nnc_tensor_free(tc16); |
5037 | 1 | ccv_nnc_tensor_free(td16); |
5038 | 1 | ccv_nnc_tensor_free(th16); |
5039 | 1 | } |
5040 | | |
5041 | | TEST_CASE("compare softmax cross entropy backward in half precision with label smoothing") |
5042 | 1 | { |
5043 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
5044 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SOFTMAX_CROSSENTROPY_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5045 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
5046 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
5047 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
5048 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
5049 | 1 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
5050 | 1 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0); |
5051 | 1 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5052 | 1 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
5053 | 1 | ccv_nnc_tensor_t* ha16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
5054 | 1 | ccv_nnc_tensor_t* hb16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
5055 | 1 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
5056 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5057 | 1 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
5058 | 1 | ccv_nnc_tensor_t* hg16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
5059 | 1 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5060 | 1 | dsfmt_t dsfmt; |
5061 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
5062 | 1 | int i; |
5063 | 1.00k | for (i = 0; i < 1000; i++1.00k ) |
5064 | 1.00k | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
5065 | 11 | for (i = 0; i < 10; i++10 ) |
5066 | 10 | hb->data.f32[i] = (i + 1) * 9; |
5067 | 11 | for (i = 0; i < 10; i++10 ) |
5068 | 10 | hg->data.f32[i] = i * 0.1; |
5069 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hg), TENSOR_LIST(ha16, hb16, hg16), 0); |
5070 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha16, hb16, hg16), TENSOR_LIST(a, b, g), 0); |
5071 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(hc, hd), 0); |
5072 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, 0, 0, hb, hc, hd), TENSOR_LIST(hh, 0), 0); |
5073 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_FORWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c, d), 0); |
5074 | 1 | ccv_nnc_cmd_exec(CMD_SOFTMAX_CROSSENTROPY_BACKWARD(0.1, 0.9), ccv_nnc_no_hint, 0, TENSOR_LIST(g, 0, 0, b, c, d), TENSOR_LIST(h, 0), 0); |
5075 | 1 | ccv_nnc_tensor_t* tc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
5076 | 1 | ccv_nnc_tensor_t* td = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5077 | 1 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0); |
5078 | 1 | ccv_nnc_tensor_t* tc16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
5079 | 1 | ccv_nnc_tensor_t* td16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
5080 | 1 | ccv_nnc_tensor_t* th16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0); |
5081 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c, d, h), TENSOR_LIST(tc16, td16, th16), 0); |
5082 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tc16, td16, th16), TENSOR_LIST(tc, td, th), 0); |
5083 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tc->data.f32, hc->data.f32, 10, 5e-2, "GPU computed output should be the same as CPU computed ones"); |
5084 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, td->data.f32, hd->data.f32, 10 * 100, 5e-2, "GPU computed output should be the same as CPU computed ones"); |
5085 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, th->data.f32, hh->data.f32, 10 * 100, 5e-2, "GPU computed output should be the same as CPU computed ones"); |
5086 | 1 | ccv_nnc_tensor_free(a); |
5087 | 1 | ccv_nnc_tensor_free(b); |
5088 | 1 | ccv_nnc_tensor_free(c); |
5089 | 1 | ccv_nnc_tensor_free(d); |
5090 | 1 | ccv_nnc_tensor_free(h); |
5091 | 1 | ccv_nnc_tensor_free(ha); |
5092 | 1 | ccv_nnc_tensor_free(hb); |
5093 | 1 | ccv_nnc_tensor_free(ha16); |
5094 | 1 | ccv_nnc_tensor_free(hb16); |
5095 | 1 | ccv_nnc_tensor_free(hc); |
5096 | 1 | ccv_nnc_tensor_free(hd); |
5097 | 1 | ccv_nnc_tensor_free(hg); |
5098 | 1 | ccv_nnc_tensor_free(hg16); |
5099 | 1 | ccv_nnc_tensor_free(hh); |
5100 | 1 | ccv_nnc_tensor_free(tc); |
5101 | 1 | ccv_nnc_tensor_free(td); |
5102 | 1 | ccv_nnc_tensor_free(th); |
5103 | 1 | ccv_nnc_tensor_free(tc16); |
5104 | 1 | ccv_nnc_tensor_free(td16); |
5105 | 1 | ccv_nnc_tensor_free(th16); |
5106 | 1 | } |
5107 | | |
5108 | | TEST_CASE("compare ewsum with cudnn") |
5109 | 1 | { |
5110 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWSUM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5111 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 100), 0); |
5112 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 100), 0); |
5113 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 100), 0); |
5114 | 1 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 100), 0); |
5115 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5116 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5117 | 1 | ccv_nnc_tensor_t* const hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5118 | 1 | ccv_nnc_tensor_t* const hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5119 | 1 | ccv_nnc_tensor_t* const gd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5120 | 1 | int i; |
5121 | 101 | for (i = 0; i < 100; i++100 ) |
5122 | 100 | { |
5123 | 100 | ha->data.f32[i] = 1; |
5124 | 100 | hb->data.f32[i] = 0.5; |
5125 | 100 | hc->data.f32[i] = 0.25; |
5126 | 100 | gd->data.f32[i] = 1.75; |
5127 | 100 | } |
5128 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hc), TENSOR_LIST(a, b, c), 0); |
5129 | 1 | ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, c), TENSOR_LIST(d), 0); |
5130 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(d), TENSOR_LIST(hd), 0); |
5131 | 1 | REQUIRE_TENSOR_EQ(hd, gd, "ewsum result should be the same"); |
5132 | 1 | ccv_nnc_tensor_free(a); |
5133 | 1 | ccv_nnc_tensor_free(b); |
5134 | 1 | ccv_nnc_tensor_free(c); |
5135 | 1 | ccv_nnc_tensor_free(d); |
5136 | 1 | ccv_nnc_tensor_free(ha); |
5137 | 1 | ccv_nnc_tensor_free(hb); |
5138 | 1 | ccv_nnc_tensor_free(hc); |
5139 | 1 | ccv_nnc_tensor_free(hd); |
5140 | 1 | ccv_nnc_tensor_free(gd); |
5141 | 1 | } |
5142 | | |
5143 | | TEST_CASE("compare ewsum with cudnn in half precision") |
5144 | 1 | { |
5145 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWSUM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5146 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 100), 0); |
5147 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 100), 0); |
5148 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 100), 0); |
5149 | 1 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 100), 0); |
5150 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5151 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5152 | 1 | ccv_nnc_tensor_t* const hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5153 | 1 | ccv_nnc_tensor_t* const hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5154 | 1 | ccv_nnc_tensor_t* const ha16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 100), 0); |
5155 | 1 | ccv_nnc_tensor_t* const hb16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 100), 0); |
5156 | 1 | ccv_nnc_tensor_t* const hc16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 100), 0); |
5157 | 1 | ccv_nnc_tensor_t* const hd16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 100), 0); |
5158 | 1 | ccv_nnc_tensor_t* const gd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5159 | 1 | int i; |
5160 | 101 | for (i = 0; i < 100; i++100 ) |
5161 | 100 | { |
5162 | 100 | ha->data.f32[i] = 1; |
5163 | 100 | hb->data.f32[i] = 0.5; |
5164 | 100 | hc->data.f32[i] = 0.25; |
5165 | 100 | gd->data.f32[i] = 1.75; |
5166 | 100 | } |
5167 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hc), TENSOR_LIST(ha16, hb16, hc16), 0); |
5168 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha16, hb16, hc16), TENSOR_LIST(a, b, c), 0); |
5169 | 1 | ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, c), TENSOR_LIST(d), 0); |
5170 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(d), TENSOR_LIST(hd16), 0); |
5171 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(hd16), TENSOR_LIST(hd), 0); |
5172 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, hd->data.f32, gd->data.f32, 100, 1e-3, "ewsum result should be the same"); |
5173 | 1 | ccv_nnc_tensor_free(a); |
5174 | 1 | ccv_nnc_tensor_free(b); |
5175 | 1 | ccv_nnc_tensor_free(c); |
5176 | 1 | ccv_nnc_tensor_free(d); |
5177 | 1 | ccv_nnc_tensor_free(ha); |
5178 | 1 | ccv_nnc_tensor_free(hb); |
5179 | 1 | ccv_nnc_tensor_free(hc); |
5180 | 1 | ccv_nnc_tensor_free(hd); |
5181 | 1 | ccv_nnc_tensor_free(ha16); |
5182 | 1 | ccv_nnc_tensor_free(hb16); |
5183 | 1 | ccv_nnc_tensor_free(hc16); |
5184 | 1 | ccv_nnc_tensor_free(hd16); |
5185 | 1 | ccv_nnc_tensor_free(gd); |
5186 | 1 | } |
5187 | | |
5188 | | TEST_CASE("compare ewsum with cudnn in bfloat precision") |
5189 | 1 | { |
5190 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWSUM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5191 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16BF, 100), 0); |
5192 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16BF, 100), 0); |
5193 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16BF, 100), 0); |
5194 | 1 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16BF, 100), 0); |
5195 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5196 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5197 | 1 | ccv_nnc_tensor_t* const hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5198 | 1 | ccv_nnc_tensor_t* const hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5199 | 1 | ccv_nnc_tensor_t* const ha16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16BF, 100), 0); |
5200 | 1 | ccv_nnc_tensor_t* const hb16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16BF, 100), 0); |
5201 | 1 | ccv_nnc_tensor_t* const hc16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16BF, 100), 0); |
5202 | 1 | ccv_nnc_tensor_t* const hd16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16BF, 100), 0); |
5203 | 1 | ccv_nnc_tensor_t* const gd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 100), 0); |
5204 | 1 | int i; |
5205 | 101 | for (i = 0; i < 100; i++100 ) |
5206 | 100 | { |
5207 | 100 | ha->data.f32[i] = 1; |
5208 | 100 | hb->data.f32[i] = 0.5; |
5209 | 100 | hc->data.f32[i] = 0.25; |
5210 | 100 | gd->data.f32[i] = 1.75; |
5211 | 100 | } |
5212 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hc), TENSOR_LIST(ha16, hb16, hc16), 0); |
5213 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha16, hb16, hc16), TENSOR_LIST(a, b, c), 0); |
5214 | 1 | ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, c), TENSOR_LIST(d), 0); |
5215 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(d), TENSOR_LIST(hd16), 0); |
5216 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(hd16), TENSOR_LIST(hd), 0); |
5217 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, hd->data.f32, gd->data.f32, 100, 1e-3, "ewsum result should be the same"); |
5218 | 1 | ccv_nnc_tensor_free(a); |
5219 | 1 | ccv_nnc_tensor_free(b); |
5220 | 1 | ccv_nnc_tensor_free(c); |
5221 | 1 | ccv_nnc_tensor_free(d); |
5222 | 1 | ccv_nnc_tensor_free(ha); |
5223 | 1 | ccv_nnc_tensor_free(hb); |
5224 | 1 | ccv_nnc_tensor_free(hc); |
5225 | 1 | ccv_nnc_tensor_free(hd); |
5226 | 1 | ccv_nnc_tensor_free(ha16); |
5227 | 1 | ccv_nnc_tensor_free(hb16); |
5228 | 1 | ccv_nnc_tensor_free(hc16); |
5229 | 1 | ccv_nnc_tensor_free(hd16); |
5230 | 1 | ccv_nnc_tensor_free(gd); |
5231 | 1 | } |
5232 | | |
5233 | | TEST_CASE("compare ewsum with cudnn in int32") |
5234 | 1 | { |
5235 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWSUM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5236 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 100), 0); |
5237 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 100), 0); |
5238 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 100), 0); |
5239 | 1 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 100), 0); |
5240 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 100), 0); |
5241 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 100), 0); |
5242 | 1 | ccv_nnc_tensor_t* const hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 100), 0); |
5243 | 1 | ccv_nnc_tensor_t* const hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 100), 0); |
5244 | 1 | ccv_nnc_tensor_t* const gd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 100), 0); |
5245 | 1 | int i; |
5246 | 101 | for (i = 0; i < 100; i++100 ) |
5247 | 100 | { |
5248 | 100 | ha->data.i32[i] = 2; |
5249 | 100 | hb->data.i32[i] = 5; |
5250 | 100 | hc->data.i32[i] = 8; |
5251 | 100 | gd->data.i32[i] = 2 + 5 + 8; |
5252 | 100 | } |
5253 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb, hc), TENSOR_LIST(a, b, c), 0); |
5254 | 1 | ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, c), TENSOR_LIST(d), 0); |
5255 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(d), TENSOR_LIST(hd), 0); |
5256 | 1 | REQUIRE_TENSOR_EQ(hd, gd, "ewsum result should be the same"); |
5257 | 1 | ccv_nnc_tensor_free(a); |
5258 | 1 | ccv_nnc_tensor_free(b); |
5259 | 1 | ccv_nnc_tensor_free(c); |
5260 | 1 | ccv_nnc_tensor_free(d); |
5261 | 1 | ccv_nnc_tensor_free(ha); |
5262 | 1 | ccv_nnc_tensor_free(hb); |
5263 | 1 | ccv_nnc_tensor_free(hc); |
5264 | 1 | ccv_nnc_tensor_free(hd); |
5265 | 1 | ccv_nnc_tensor_free(gd); |
5266 | 1 | } |
5267 | | |
5268 | | TEST_CASE("compare transpose two tensor views") |
5269 | 1 | { |
5270 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_TRANSPOSE_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5271 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 6, 5, 4), 0); |
5272 | 1 | memset(ha->data.f32, 0, sizeof(float) * 7 * 6 * 5 * 4); |
5273 | 1 | ccv_nnc_tensor_view_t ha_view = ccv_nnc_tensor_view(ha, CPU_TENSOR_NHWC(32F, 4, 3, 2, 2), DIM_ALLOC(3, 2, 1, 0), DIM_ALLOC(6 * 5 * 4, 5 * 4, 4, 1)); |
5274 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 8, 7, 6, 5), 0); |
5275 | 1 | memset(hb->data.f32, 0, sizeof(float) * 8 * 7 * 6 * 5); |
5276 | 1 | ccv_nnc_tensor_view_t hb_view = ccv_nnc_tensor_view(hb, CPU_TENSOR_NHWC(32F, 4, 2, 2, 3), DIM_ALLOC(3, 2, 1, 0), DIM_ALLOC(7 * 6 * 5, 6 * 5, 5, 1)); |
5277 | 1 | int i, j, k, l; |
5278 | 5 | for (i = 0; i < 4; i++4 ) |
5279 | 16 | for (j = 0; 4 j < 3; j++12 ) |
5280 | 36 | for (k = 0; 12 k < 2; k++24 ) |
5281 | 72 | for (l = 0; 24 l < 2; l++48 ) |
5282 | 48 | ha->data.f32[(i + 3) * 6 * 5 * 4 + (j + 2) * 5 * 4 + (k + 1) * 4 + l] = i * 3 * 2 * 2 + j * 2 * 2 + k * 2 + l; |
5283 | 1 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 3), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)&ha_view), TENSOR_LIST((ccv_nnc_tensor_t*)&hb_view), 0); |
5284 | 1 | ccv_nnc_tensor_t* hd = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 6, 5, 4), 0); |
5285 | 1 | memset(hd->data.f32, 0, sizeof(float) * 7 * 6 * 5 * 4); |
5286 | 1 | ccv_nnc_tensor_view_t hd_view = ccv_nnc_tensor_view(hd, CPU_TENSOR_NHWC(32F, 4, 3, 2, 2), DIM_ALLOC(3, 2, 1, 0), DIM_ALLOC(6 * 5 * 4, 5 * 4, 4, 1)); |
5287 | 1 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 3), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)&hb_view), TENSOR_LIST((ccv_nnc_tensor_t*)&hd_view), 0); |
5288 | 1 | REQUIRE_TENSOR_EQ(hd, ha, "4x3x2x2 tensor should be exactly the same."); |
5289 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 7, 6, 5, 4), 0); |
5290 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
5291 | 1 | ccv_nnc_tensor_view_t a_view = ccv_nnc_tensor_view(a, GPU_TENSOR_NHWC(000, 32F, 4, 3, 2, 2), DIM_ALLOC(3, 2, 1, 0), DIM_ALLOC(6 * 5 * 4, 5 * 4, 4, 1)); |
5292 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 8, 7, 6, 5), 0); |
5293 | 1 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(b), 0); |
5294 | 1 | ccv_nnc_tensor_view_t b_view = ccv_nnc_tensor_view(b, GPU_TENSOR_NHWC(000, 32F, 4, 2, 2, 3), DIM_ALLOC(3, 2, 1, 0), DIM_ALLOC(7 * 6 * 5, 6 * 5, 5, 1)); |
5295 | 1 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 3), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)&a_view), TENSOR_LIST((ccv_nnc_tensor_t*)&b_view), 0); |
5296 | 1 | ccv_nnc_tensor_t* d = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 7, 6, 5, 4), 0); |
5297 | 1 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(d), 0); |
5298 | 1 | ccv_nnc_tensor_view_t d_view = ccv_nnc_tensor_view(d, GPU_TENSOR_NHWC(000, 32F, 4, 3, 2, 2), DIM_ALLOC(3, 2, 1, 0), DIM_ALLOC(6 * 5 * 4, 5 * 4, 4, 1)); |
5299 | 1 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 3), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)&b_view), TENSOR_LIST((ccv_nnc_tensor_t*)&d_view), 0); |
5300 | 1 | ccv_nnc_tensor_t* const hbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 8, 7, 6, 5), 0); |
5301 | 1 | ccv_nnc_tensor_t* const hdt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 7, 6, 5, 4), 0); |
5302 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, d), TENSOR_LIST(hbt, hdt), 0); |
5303 | 1 | REQUIRE_TENSOR_EQ(hbt, hb, "4x2x2x3 tensor should be exactly the same."); |
5304 | 1 | REQUIRE_TENSOR_EQ(hdt, hd, "4x3x2x2 tensor should be exactly the same."); |
5305 | 1 | ccv_nnc_tensor_free(ha); |
5306 | 1 | ccv_nnc_tensor_free(hb); |
5307 | 1 | ccv_nnc_tensor_free(hd); |
5308 | 1 | ccv_nnc_tensor_free(hbt); |
5309 | 1 | ccv_nnc_tensor_free(hdt); |
5310 | 1 | ccv_nnc_tensor_free(a); |
5311 | 1 | ccv_nnc_tensor_free(b); |
5312 | 1 | ccv_nnc_tensor_free(d); |
5313 | 1 | } |
5314 | | |
5315 | | TEST_CASE("compare format transform with cudnn in double precision") |
5316 | 1 | { |
5317 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_FORMAT_TRANSFORM_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5318 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 64F, 11, 10, 9, 8), 0); |
5319 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 64F, 11, 8, 10, 9), 0); |
5320 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(64F, 11, 10, 9, 8), 0); |
5321 | 1 | ccv_nnc_tensor_t* const hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(64F, 11, 8, 10, 9), 0); |
5322 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(64F, 11, 8, 10, 9), 0); |
5323 | 1 | int i; |
5324 | 1 | dsfmt_t dsfmt; |
5325 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
5326 | 7.92k | for (i = 0; i < 11 * 10 * 9 * 8; i++7.92k ) |
5327 | 7.92k | ha->data.f64[i] = dsfmt_genrand_open_close(&dsfmt); |
5328 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
5329 | 1 | ccv_nnc_cmd_exec(CMD_FORMAT_TRANSFORM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
5330 | 1 | ccv_nnc_cmd_exec(CMD_FORMAT_TRANSFORM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(gb), 0); |
5331 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
5332 | 1 | REQUIRE_TENSOR_EQ(hb, gb, "format transform result should be the same"); |
5333 | 1 | ccv_nnc_tensor_free(a); |
5334 | 1 | ccv_nnc_tensor_free(b); |
5335 | 1 | ccv_nnc_tensor_free(ha); |
5336 | 1 | ccv_nnc_tensor_free(hb); |
5337 | 1 | ccv_nnc_tensor_free(gb); |
5338 | 1 | } |
5339 | | |
5340 | | TEST_CASE("compare set with cudnn in double precision") |
5341 | 1 | { |
5342 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5343 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 64F, 11, 10, 9, 8), 0); |
5344 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(64F, 11, 10, 9, 8), 0); |
5345 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(64F, 11, 10, 9, 8), 0); |
5346 | 1 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(10), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(a), 0); |
5347 | 1 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(10), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(ga), 0); |
5348 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(ha), 0); |
5349 | 1 | REQUIRE_TENSOR_EQ(ha, ga, "format transform result should be the same"); |
5350 | 1 | ccv_nnc_tensor_free(a); |
5351 | 1 | ccv_nnc_tensor_free(ha); |
5352 | 1 | ccv_nnc_tensor_free(ga); |
5353 | 1 | } |
5354 | | |
5355 | | TEST_CASE("compare set with cudnn in integer") |
5356 | 1 | { |
5357 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5358 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 1), 0); |
5359 | 1 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 1), 0); |
5360 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32S, 1), 0); |
5361 | 1 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(10), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(a), 0); |
5362 | 1 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(10), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(ga), 0); |
5363 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(ha), 0); |
5364 | 1 | REQUIRE_TENSOR_EQ(ha, ga, "format transform result should be the same"); |
5365 | 1 | ccv_nnc_tensor_free(a); |
5366 | 1 | ccv_nnc_tensor_free(ha); |
5367 | 1 | ccv_nnc_tensor_free(ga); |
5368 | 1 | } |
5369 | | |
5370 | | TEST_CASE("broadcasting semantics for add [[1, 2, 3], [4, 5, 6]] + [7, 8, 9]") |
5371 | 1 | { |
5372 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5373 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5374 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5375 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5376 | 1 | a->data.f32[0] = 1; |
5377 | 1 | a->data.f32[1] = 2; |
5378 | 1 | a->data.f32[2] = 3; |
5379 | 1 | a->data.f32[3] = 4; |
5380 | 1 | a->data.f32[4] = 5; |
5381 | 1 | a->data.f32[5] = 6; |
5382 | 1 | b->data.f32[0] = 7; |
5383 | 1 | b->data.f32[1] = 8; |
5384 | 1 | b->data.f32[2] = 9; |
5385 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5386 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
5387 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5388 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5389 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(1, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
5390 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5391 | 1 | float ctp[] = { |
5392 | 1 | 8, 10, 12, |
5393 | 1 | 11, 13, 15 |
5394 | 1 | }; |
5395 | 1 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5396 | 1 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
5397 | 1 | ccv_nnc_tensor_free(a); |
5398 | 1 | ccv_nnc_tensor_free(b); |
5399 | 1 | ccv_nnc_tensor_free(c); |
5400 | 1 | ccv_nnc_tensor_free(ga); |
5401 | 1 | ccv_nnc_tensor_free(gb); |
5402 | 1 | ccv_nnc_tensor_free(gc); |
5403 | 1 | } |
5404 | | |
5405 | | TEST_CASE("broadcasting semantics for add [[1], [2], [3], [4]] + [5, 6]") |
5406 | 1 | { |
5407 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5408 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5409 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5410 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
5411 | 1 | a->data.f32[0] = 1; |
5412 | 1 | a->data.f32[1] = 2; |
5413 | 1 | a->data.f32[2] = 3; |
5414 | 1 | a->data.f32[3] = 4; |
5415 | 1 | b->data.f32[0] = 5; |
5416 | 1 | b->data.f32[1] = 6; |
5417 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5418 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5419 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
5420 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5421 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(1, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
5422 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5423 | 1 | float ctp[] = { |
5424 | 1 | 6, 7, |
5425 | 1 | 7, 8, |
5426 | 1 | 8, 9, |
5427 | 1 | 9, 10 |
5428 | 1 | }; |
5429 | 1 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
5430 | 1 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
5431 | 1 | ccv_nnc_tensor_free(a); |
5432 | 1 | ccv_nnc_tensor_free(b); |
5433 | 1 | ccv_nnc_tensor_free(c); |
5434 | 1 | ccv_nnc_tensor_free(ga); |
5435 | 1 | ccv_nnc_tensor_free(gb); |
5436 | 1 | ccv_nnc_tensor_free(gc); |
5437 | 1 | } |
5438 | | |
5439 | | TEST_CASE("broadcasting semantics for mul [[1, 2, 3], [4, 5, 6]] * [7, 8, 9]") |
5440 | 1 | { |
5441 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5442 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5443 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5444 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5445 | 1 | a->data.f32[0] = 1; |
5446 | 1 | a->data.f32[1] = 2; |
5447 | 1 | a->data.f32[2] = 3; |
5448 | 1 | a->data.f32[3] = 4; |
5449 | 1 | a->data.f32[4] = 5; |
5450 | 1 | a->data.f32[5] = 6; |
5451 | 1 | b->data.f32[0] = 7; |
5452 | 1 | b->data.f32[1] = 8; |
5453 | 1 | b->data.f32[2] = 9; |
5454 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5455 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
5456 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5457 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5458 | 1 | ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
5459 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5460 | 1 | float ctp[] = { |
5461 | 1 | 7, 16, 27, |
5462 | 1 | 28, 40, 54 |
5463 | 1 | }; |
5464 | 1 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5465 | 1 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
5466 | 1 | ccv_nnc_tensor_free(a); |
5467 | 1 | ccv_nnc_tensor_free(b); |
5468 | 1 | ccv_nnc_tensor_free(c); |
5469 | 1 | ccv_nnc_tensor_free(ga); |
5470 | 1 | ccv_nnc_tensor_free(gb); |
5471 | 1 | ccv_nnc_tensor_free(gc); |
5472 | 1 | } |
5473 | | |
5474 | | TEST_CASE("broadcasting semantics for mul [[1], [2], [3], [4]] * [5, 6]") |
5475 | 1 | { |
5476 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5477 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5478 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5479 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
5480 | 1 | a->data.f32[0] = 1; |
5481 | 1 | a->data.f32[1] = 2; |
5482 | 1 | a->data.f32[2] = 3; |
5483 | 1 | a->data.f32[3] = 4; |
5484 | 1 | b->data.f32[0] = 5; |
5485 | 1 | b->data.f32[1] = 6; |
5486 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5487 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5488 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
5489 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5490 | 1 | ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
5491 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5492 | 1 | float ctp[] = { |
5493 | 1 | 5, 6, |
5494 | 1 | 10, 12, |
5495 | 1 | 15, 18, |
5496 | 1 | 20, 24 |
5497 | 1 | }; |
5498 | 1 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
5499 | 1 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
5500 | 1 | ccv_nnc_tensor_free(a); |
5501 | 1 | ccv_nnc_tensor_free(b); |
5502 | 1 | ccv_nnc_tensor_free(c); |
5503 | 1 | ccv_nnc_tensor_free(ga); |
5504 | 1 | ccv_nnc_tensor_free(gb); |
5505 | 1 | ccv_nnc_tensor_free(gc); |
5506 | 1 | } |
5507 | | |
5508 | | TEST_CASE("scalar mul [[1, 2, 3], [4, 5, 6]] * 0.3") |
5509 | 1 | { |
5510 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5511 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5512 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5513 | 1 | a->data.f32[0] = 1; |
5514 | 1 | a->data.f32[1] = 2; |
5515 | 1 | a->data.f32[2] = 3; |
5516 | 1 | a->data.f32[3] = 4; |
5517 | 1 | a->data.f32[4] = 5; |
5518 | 1 | a->data.f32[5] = 6; |
5519 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5520 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5521 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(ga), 0); |
5522 | 1 | ccv_nnc_cmd_exec(CMD_SCALAR_MUL_FORWARD(0.3), ccv_nnc_no_hint, 0, TENSOR_LIST(ga), TENSOR_LIST(gc), 0); |
5523 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5524 | 1 | float ctp[] = { |
5525 | 1 | 0.3, 0.6, 0.9, |
5526 | 1 | 1.2, 1.5, 1.8, |
5527 | 1 | }; |
5528 | 1 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5529 | 1 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
5530 | 1 | ccv_nnc_tensor_free(a); |
5531 | 1 | ccv_nnc_tensor_free(c); |
5532 | 1 | ccv_nnc_tensor_free(ga); |
5533 | 1 | ccv_nnc_tensor_free(gc); |
5534 | 1 | } |
5535 | | |
5536 | | TEST_CASE("scalar mul [[1, 2, 3], [4, 5, 6]] * 0.5, int") |
5537 | 1 | { |
5538 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5539 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 2, 3), 0); |
5540 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 2, 3), 0); |
5541 | 1 | ccv_nnc_tensor_t* const ct = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32S, 2, 3), 0); |
5542 | 1 | a->data.i32[0] = 1; |
5543 | 1 | a->data.i32[1] = 2; |
5544 | 1 | a->data.i32[2] = 3; |
5545 | 1 | a->data.i32[3] = 4; |
5546 | 1 | a->data.i32[4] = 5; |
5547 | 1 | a->data.i32[5] = 6; |
5548 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 2, 3), 0); |
5549 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32S, 2, 3), 0); |
5550 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(ga), 0); |
5551 | 1 | ccv_nnc_cmd_exec(CMD_SCALAR_MUL_FORWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(ga), TENSOR_LIST(gc), 0); |
5552 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5553 | 1 | ccv_nnc_cmd_exec(CMD_SCALAR_MUL_FORWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(ct), 0); |
5554 | 1 | REQUIRE_TENSOR_EQ(c, ct, "result should be equal"); |
5555 | 1 | ccv_nnc_tensor_free(a); |
5556 | 1 | ccv_nnc_tensor_free(c); |
5557 | 1 | ccv_nnc_tensor_free(ct); |
5558 | 1 | ccv_nnc_tensor_free(ga); |
5559 | 1 | ccv_nnc_tensor_free(gc); |
5560 | 1 | } |
5561 | | |
5562 | | TEST_CASE("broadcasting semantics for add backward") |
5563 | 1 | { |
5564 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
5565 | 1 | ccv_nnc_cmd_ok(CCV_NNC_ADD_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5566 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5567 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5568 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
5569 | 1 | ccv_nnc_tensor_t* const da = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5570 | 1 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5571 | 1 | ccv_nnc_tensor_t* const dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5572 | 1 | ccv_nnc_tensor_t* const dbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5573 | 1 | a->data.f32[0] = 1; |
5574 | 1 | a->data.f32[1] = 2; |
5575 | 1 | a->data.f32[2] = 3; |
5576 | 1 | a->data.f32[3] = 4; |
5577 | 1 | b->data.f32[0] = 5; |
5578 | 1 | b->data.f32[1] = 6; |
5579 | 1 | float ctp[] = { |
5580 | 1 | 6, 7, |
5581 | 1 | 7, 8, |
5582 | 1 | 8, 9, |
5583 | 1 | 9, 10 |
5584 | 1 | }; |
5585 | 1 | memcpy(c->data.f32, ctp, sizeof(ctp)); |
5586 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5587 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5588 | 1 | ccv_nnc_tensor_t* const gda = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5589 | 1 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5590 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
5591 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, c), TENSOR_LIST(ga, gb, gc), 0); |
5592 | 1 | ccv_nnc_cmd_exec(CMD_ADD_BACKWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(gc, ga, gb), TENSOR_LIST(gda, gdb), 0); |
5593 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gda, gdb), TENSOR_LIST(da, db), 0); |
5594 | 1 | ccv_nnc_cmd_exec(CMD_ADD_BACKWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(c, a, b), TENSOR_LIST(dat, dbt), 0); |
5595 | 1 | REQUIRE_TENSOR_EQ(dat, da, "gradient of a should be equal"); |
5596 | 1 | REQUIRE_TENSOR_EQ(dbt, db, "gradient of b should be equal"); |
5597 | 1 | ccv_nnc_tensor_free(a); |
5598 | 1 | ccv_nnc_tensor_free(b); |
5599 | 1 | ccv_nnc_tensor_free(c); |
5600 | 1 | ccv_nnc_tensor_free(da); |
5601 | 1 | ccv_nnc_tensor_free(db); |
5602 | 1 | ccv_nnc_tensor_free(dat); |
5603 | 1 | ccv_nnc_tensor_free(dbt); |
5604 | 1 | ccv_nnc_tensor_free(ga); |
5605 | 1 | ccv_nnc_tensor_free(gb); |
5606 | 1 | ccv_nnc_tensor_free(gc); |
5607 | 1 | ccv_nnc_tensor_free(gda); |
5608 | 1 | ccv_nnc_tensor_free(gdb); |
5609 | 1 | } |
5610 | | |
5611 | | TEST_CASE("broadcasting semantics for mul backward") |
5612 | 1 | { |
5613 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
5614 | 1 | ccv_nnc_cmd_ok(CCV_NNC_MUL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5615 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5616 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5617 | 1 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
5618 | 1 | ccv_nnc_tensor_t* const da = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5619 | 1 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5620 | 1 | ccv_nnc_tensor_t* const dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5621 | 1 | ccv_nnc_tensor_t* const dbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5622 | 1 | a->data.f32[0] = 1; |
5623 | 1 | a->data.f32[1] = 2; |
5624 | 1 | a->data.f32[2] = 3; |
5625 | 1 | a->data.f32[3] = 4; |
5626 | 1 | b->data.f32[0] = 5; |
5627 | 1 | b->data.f32[1] = 6; |
5628 | 1 | float ctp[] = { |
5629 | 1 | 6, 7, |
5630 | 1 | 7, 8, |
5631 | 1 | 8, 9, |
5632 | 1 | 9, 10 |
5633 | 1 | }; |
5634 | 1 | memcpy(c->data.f32, ctp, sizeof(ctp)); |
5635 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5636 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5637 | 1 | ccv_nnc_tensor_t* const gda = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5638 | 1 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5639 | 1 | ccv_nnc_tensor_t* const gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
5640 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, c), TENSOR_LIST(ga, gb, gc), 0); |
5641 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(gc, ga, gb), TENSOR_LIST(gda, gdb), 0); |
5642 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gda, gdb), TENSOR_LIST(da, db), 0); |
5643 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(c, a, b), TENSOR_LIST(dat, dbt), 0); |
5644 | 1 | REQUIRE_TENSOR_EQ(dat, da, "gradient of a should be equal"); |
5645 | 1 | REQUIRE_TENSOR_EQ(dbt, db, "gradient of b should be equal"); |
5646 | 1 | ccv_nnc_tensor_free(a); |
5647 | 1 | ccv_nnc_tensor_free(b); |
5648 | 1 | ccv_nnc_tensor_free(c); |
5649 | 1 | ccv_nnc_tensor_free(da); |
5650 | 1 | ccv_nnc_tensor_free(db); |
5651 | 1 | ccv_nnc_tensor_free(dat); |
5652 | 1 | ccv_nnc_tensor_free(dbt); |
5653 | 1 | ccv_nnc_tensor_free(ga); |
5654 | 1 | ccv_nnc_tensor_free(gb); |
5655 | 1 | ccv_nnc_tensor_free(gc); |
5656 | 1 | ccv_nnc_tensor_free(gda); |
5657 | 1 | ccv_nnc_tensor_free(gdb); |
5658 | 1 | } |
5659 | | |
5660 | | TEST_CASE("broadcasting semantics for mul backward (no input grad)") |
5661 | 1 | { |
5662 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
5663 | 1 | ccv_nnc_cmd_ok(CCV_NNC_MUL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5664 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5665 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5666 | 1 | ccv_nnc_tensor_t* const da = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5667 | 1 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5668 | 1 | ccv_nnc_tensor_t* const dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 1), 0); |
5669 | 1 | ccv_nnc_tensor_t* const dbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2), 0); |
5670 | 1 | a->data.f32[0] = 1; |
5671 | 1 | a->data.f32[1] = 2; |
5672 | 1 | a->data.f32[2] = 3; |
5673 | 1 | a->data.f32[3] = 4; |
5674 | 1 | b->data.f32[0] = 5; |
5675 | 1 | b->data.f32[1] = 6; |
5676 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5677 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5678 | 1 | ccv_nnc_tensor_t* const gda = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 1), 0); |
5679 | 1 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2), 0); |
5680 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5681 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(0, ga, gb), TENSOR_LIST(gda, gdb), 0); |
5682 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gda, gdb), TENSOR_LIST(da, db), 0); |
5683 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(0, a, b), TENSOR_LIST(dat, dbt), 0); |
5684 | 1 | REQUIRE_TENSOR_EQ(dat, da, "gradient of a should be equal"); |
5685 | 1 | REQUIRE_TENSOR_EQ(dbt, db, "gradient of b should be equal"); |
5686 | 1 | ccv_nnc_tensor_free(a); |
5687 | 1 | ccv_nnc_tensor_free(b); |
5688 | 1 | ccv_nnc_tensor_free(da); |
5689 | 1 | ccv_nnc_tensor_free(db); |
5690 | 1 | ccv_nnc_tensor_free(dat); |
5691 | 1 | ccv_nnc_tensor_free(dbt); |
5692 | 1 | ccv_nnc_tensor_free(ga); |
5693 | 1 | ccv_nnc_tensor_free(gb); |
5694 | 1 | ccv_nnc_tensor_free(gda); |
5695 | 1 | ccv_nnc_tensor_free(gdb); |
5696 | 1 | } |
5697 | | |
5698 | | TEST_CASE("broadcasting semantics for mul backward (no input grad) for b") |
5699 | 1 | { |
5700 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
5701 | 1 | ccv_nnc_cmd_ok(CCV_NNC_MUL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5702 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5703 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5704 | 1 | ccv_nnc_tensor_t* const da = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5705 | 1 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5706 | 1 | ccv_nnc_tensor_t* const dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5707 | 1 | ccv_nnc_tensor_t* const dbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5708 | 1 | a->data.f32[0] = 1; |
5709 | 1 | a->data.f32[1] = 2; |
5710 | 1 | a->data.f32[2] = 3; |
5711 | 1 | a->data.f32[3] = 4; |
5712 | 1 | a->data.f32[4] = 5; |
5713 | 1 | a->data.f32[5] = 6; |
5714 | 1 | b->data.f32[0] = 7; |
5715 | 1 | b->data.f32[1] = 8; |
5716 | 1 | b->data.f32[2] = 9; |
5717 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5718 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
5719 | 1 | ccv_nnc_tensor_t* const gda = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5720 | 1 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
5721 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5722 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(0, ga, gb), TENSOR_LIST(gda, gdb), 0); |
5723 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gda, gdb), TENSOR_LIST(da, db), 0); |
5724 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(0, a, b), TENSOR_LIST(dat, dbt), 0); |
5725 | 1 | REQUIRE_TENSOR_EQ(dat, da, "gradient of a should be equal"); |
5726 | 1 | REQUIRE_TENSOR_EQ(dbt, db, "gradient of b should be equal"); |
5727 | 1 | ccv_nnc_tensor_free(a); |
5728 | 1 | ccv_nnc_tensor_free(b); |
5729 | 1 | ccv_nnc_tensor_free(da); |
5730 | 1 | ccv_nnc_tensor_free(db); |
5731 | 1 | ccv_nnc_tensor_free(dat); |
5732 | 1 | ccv_nnc_tensor_free(dbt); |
5733 | 1 | ccv_nnc_tensor_free(ga); |
5734 | 1 | ccv_nnc_tensor_free(gb); |
5735 | 1 | ccv_nnc_tensor_free(gda); |
5736 | 1 | ccv_nnc_tensor_free(gdb); |
5737 | 1 | } |
5738 | | |
5739 | | TEST_CASE("broadcasting semantics for mul backward (no input grad) for a") |
5740 | 1 | { |
5741 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_MUL_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) && |
5742 | 1 | ccv_nnc_cmd_ok(CCV_NNC_MUL_BACKWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5743 | 1 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5744 | 1 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5745 | 1 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5746 | 1 | ccv_nnc_tensor_t* const da = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5747 | 1 | ccv_nnc_tensor_t* const dbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
5748 | 1 | ccv_nnc_tensor_t* const dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
5749 | 1 | b->data.f32[0] = 1; |
5750 | 1 | b->data.f32[1] = 2; |
5751 | 1 | b->data.f32[2] = 3; |
5752 | 1 | b->data.f32[3] = 4; |
5753 | 1 | b->data.f32[4] = 5; |
5754 | 1 | b->data.f32[5] = 6; |
5755 | 1 | a->data.f32[0] = 7; |
5756 | 1 | a->data.f32[1] = 8; |
5757 | 1 | a->data.f32[2] = 9; |
5758 | 1 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5759 | 1 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
5760 | 1 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
5761 | 1 | ccv_nnc_tensor_t* const gda = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
5762 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
5763 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(0, ga, gb), TENSOR_LIST(gda, gdb), 0); |
5764 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gda, gdb), TENSOR_LIST(da, db), 0); |
5765 | 1 | ccv_nnc_cmd_exec(CMD_MUL_BACKWARD(0.5), ccv_nnc_no_hint, 0, TENSOR_LIST(0, a, b), TENSOR_LIST(dat, dbt), 0); |
5766 | 1 | REQUIRE_TENSOR_EQ(dat, da, "gradient of a should be equal"); |
5767 | 1 | REQUIRE_TENSOR_EQ(dbt, db, "gradient of b should be equal"); |
5768 | 1 | ccv_nnc_tensor_free(a); |
5769 | 1 | ccv_nnc_tensor_free(b); |
5770 | 1 | ccv_nnc_tensor_free(da); |
5771 | 1 | ccv_nnc_tensor_free(db); |
5772 | 1 | ccv_nnc_tensor_free(dat); |
5773 | 1 | ccv_nnc_tensor_free(dbt); |
5774 | 1 | ccv_nnc_tensor_free(ga); |
5775 | 1 | ccv_nnc_tensor_free(gb); |
5776 | 1 | ccv_nnc_tensor_free(gda); |
5777 | 1 | ccv_nnc_tensor_free(gdb); |
5778 | 1 | } |
5779 | | |
5780 | | TEST_CASE("cudnn forward convolution transpose") |
5781 | 1 | { |
5782 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5783 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5784 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5785 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(1, INPUT_DIM, 0, KERNEL_SIZE, KERNEL_SIZE, OUTPUT_DIM); |
5786 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
5787 | 1 | assert(cmd.backend >= 0); |
5788 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, b->info, a->info); |
5789 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5790 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, INPUT_DIM), 0); |
5791 | | // configure the inlets. |
5792 | 1 | dsfmt_t dsfmt; |
5793 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
5794 | 1 | int i; |
5795 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
5796 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
5797 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
5798 | 19.2M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
5799 | 4 | for (i = 0; i < INPUT_DIM; i++3 ) |
5800 | 3 | bias->data.f32[i] = (float)i / INPUT_DIM; |
5801 | | // Copy generated matrix values over to GPU. |
5802 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5803 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5804 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, INPUT_DIM), 0); |
5805 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
5806 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
5807 | 1 | assert(move.backend >= 0); |
5808 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
5809 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
5810 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5811 | | |
5812 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
5813 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
5814 | 1 | assert(cmd.backend >= 0); |
5815 | 1 | cmd.algorithm = -1; |
5816 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), stream_context); |
5817 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), stream_context)); |
5818 | 1 | ccv_nnc_stream_context_wait(stream_context); |
5819 | 1 | ccv_nnc_stream_context_free(stream_context); |
5820 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5821 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5822 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 2e-4, "output from cudnn should match from CPU"); |
5823 | 1 | ccv_nnc_tensor_free(c); |
5824 | 1 | ccv_nnc_tensor_free(gc); |
5825 | 1 | ccv_nnc_tensor_free(bias); |
5826 | 1 | ccv_nnc_tensor_free(w); |
5827 | 1 | ccv_nnc_tensor_free(b); |
5828 | 1 | ccv_nnc_tensor_free(a); |
5829 | 1 | ccv_nnc_tensor_free(gbias); |
5830 | 1 | ccv_nnc_tensor_free(gw); |
5831 | 1 | ccv_nnc_tensor_free(ga); |
5832 | 1 | } |
5833 | | |
5834 | | TEST_CASE("cudnn forward convolution transpose, w in nchw format") |
5835 | 1 | { |
5836 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5837 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5838 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5839 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(1, INPUT_DIM, 0, KERNEL_SIZE, KERNEL_SIZE, OUTPUT_DIM); |
5840 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
5841 | 1 | assert(cmd.backend >= 0); |
5842 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, b->info, a->info); |
5843 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5844 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, INPUT_DIM), 0); |
5845 | | // configure the inlets. |
5846 | 1 | dsfmt_t dsfmt; |
5847 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
5848 | 1 | int i; |
5849 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
5850 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
5851 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
5852 | 19.2M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
5853 | 4 | for (i = 0; i < INPUT_DIM; i++3 ) |
5854 | 3 | bias->data.f32[i] = (float)i / INPUT_DIM; |
5855 | | // Copy generated matrix values over to GPU. |
5856 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5857 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5858 | 1 | ccv_nnc_tensor_t* gwo = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
5859 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, INPUT_DIM), 0); |
5860 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
5861 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
5862 | 1 | assert(move.backend >= 0); |
5863 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
5864 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
5865 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5866 | 1 | ccv_nnc_cmd_exec(CMD_FORMAT_TRANSFORM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gw), TENSOR_LIST(gwo), 0); |
5867 | | |
5868 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
5869 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
5870 | 1 | assert(cmd.backend >= 0); |
5871 | 1 | cmd.algorithm = -1; |
5872 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context); |
5873 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gwo, gbias), TENSOR_LIST(gc), stream_context)); |
5874 | 1 | ccv_nnc_stream_context_wait(stream_context); |
5875 | 1 | ccv_nnc_stream_context_free(stream_context); |
5876 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5877 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5878 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 2e-4, "output from cudnn should match from CPU"); |
5879 | 1 | ccv_nnc_tensor_free(c); |
5880 | 1 | ccv_nnc_tensor_free(gc); |
5881 | 1 | ccv_nnc_tensor_free(bias); |
5882 | 1 | ccv_nnc_tensor_free(w); |
5883 | 1 | ccv_nnc_tensor_free(b); |
5884 | 1 | ccv_nnc_tensor_free(a); |
5885 | 1 | ccv_nnc_tensor_free(gbias); |
5886 | 1 | ccv_nnc_tensor_free(gw); |
5887 | 1 | ccv_nnc_tensor_free(gwo); |
5888 | 1 | ccv_nnc_tensor_free(ga); |
5889 | 1 | } |
5890 | | |
5891 | | TEST_CASE("cudnn forward convolution transpose in nchw format") |
5892 | 1 | { |
5893 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5894 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, OUTPUT_DIM, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
5895 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
5896 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(1, INPUT_DIM, 0, KERNEL_SIZE, KERNEL_SIZE, OUTPUT_DIM); |
5897 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
5898 | 1 | assert(cmd.backend >= 0); |
5899 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, b->info, a->info); |
5900 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
5901 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, INPUT_DIM), 0); |
5902 | | // configure the inlets. |
5903 | 1 | dsfmt_t dsfmt; |
5904 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
5905 | 1 | int i; |
5906 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
5907 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
5908 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
5909 | 19.2M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
5910 | 4 | for (i = 0; i < INPUT_DIM; i++3 ) |
5911 | 3 | bias->data.f32[i] = (float)i / INPUT_DIM; |
5912 | | // Copy generated matrix values over to GPU. |
5913 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, OUTPUT_DIM, OUTPUT_SIZE, OUTPUT_SIZE), 0); |
5914 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, OUTPUT_DIM, INPUT_DIM, KERNEL_SIZE, KERNEL_SIZE), 0); |
5915 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, INPUT_DIM), 0); |
5916 | 1 | ccv_nnc_cmd_t move = CMD_DATA_TRANSFER_FORWARD(); |
5917 | 1 | move.backend = CCV_NNC_BACKEND_GPU_REF; |
5918 | 1 | assert(move.backend >= 0); |
5919 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(ga, gw, gbias), 0); |
5920 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
5921 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
5922 | | |
5923 | 1 | ccv_nnc_cmd_t transform = CMD_FORMAT_TRANSFORM_FORWARD(); |
5924 | 1 | transform.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
5925 | 1 | assert(transform.backend >= 0); |
5926 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
5927 | 1 | assert(cmd.backend >= 0); |
5928 | 1 | cmd.algorithm = -1; |
5929 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 1 * 1024 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), 0); |
5930 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), 0)); |
5931 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, BATCH_SIZE, INPUT_DIM, INPUT_SIZE, INPUT_SIZE), 0); |
5932 | 1 | ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
5933 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 1e-5, "output from cudnn should match from CPU"); |
5934 | 1 | ccv_nnc_tensor_free(c); |
5935 | 1 | ccv_nnc_tensor_free(gc); |
5936 | 1 | ccv_nnc_tensor_free(bias); |
5937 | 1 | ccv_nnc_tensor_free(w); |
5938 | 1 | ccv_nnc_tensor_free(b); |
5939 | 1 | ccv_nnc_tensor_free(a); |
5940 | 1 | ccv_nnc_tensor_free(gbias); |
5941 | 1 | ccv_nnc_tensor_free(gw); |
5942 | 1 | ccv_nnc_tensor_free(ga); |
5943 | 1 | } |
5944 | | |
5945 | | TEST_CASE("cudnn forward convolution transpose in half precision") |
5946 | 1 | { |
5947 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN)); |
5948 | 1 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5949 | 1 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5950 | 1 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(1, INPUT_DIM, 0, KERNEL_SIZE, KERNEL_SIZE, OUTPUT_DIM); |
5951 | 1 | cmd.backend = CCV_NNC_BACKEND_CPU_REF; |
5952 | 1 | assert(cmd.backend >= 0); |
5953 | 1 | ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, b->info, a->info); |
5954 | 1 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5955 | 1 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, INPUT_DIM), 0); |
5956 | | // configure the inlets. |
5957 | 1 | dsfmt_t dsfmt; |
5958 | 1 | dsfmt_init_gen_rand(&dsfmt, 0); |
5959 | 1 | int i; |
5960 | 14.1k | for (i = 0; i < INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE * OUTPUT_DIM; i++14.1k ) |
5961 | 14.1k | w->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (INPUT_DIM * KERNEL_SIZE * KERNEL_SIZE); |
5962 | 19.2M | for (i = 0; i < OUTPUT_SIZE * OUTPUT_SIZE * OUTPUT_DIM * ccv_max(1, BATCH_SIZE); i++19.2M ) |
5963 | 19.2M | a->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
5964 | 4 | for (i = 0; i < INPUT_DIM; i++3 ) |
5965 | 3 | bias->data.f32[i] = (float)i / INPUT_DIM; |
5966 | 1 | ccv_nnc_tensor_t* a1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5967 | 1 | ccv_nnc_tensor_t* w1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5968 | 1 | ccv_nnc_tensor_t* bias1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, INPUT_DIM), 0); |
5969 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(a1, w1, bias1), 0); |
5970 | | // Copy generated matrix values over to GPU. |
5971 | 1 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, OUTPUT_SIZE, OUTPUT_SIZE, OUTPUT_DIM), 0); |
5972 | 1 | ccv_nnc_tensor_t* gw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, OUTPUT_DIM, KERNEL_SIZE, KERNEL_SIZE, INPUT_DIM), 0); |
5973 | 1 | ccv_nnc_tensor_t* gbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, INPUT_DIM), 0); |
5974 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a1, w1, bias1), TENSOR_LIST(ga, gw, gbias), 0); |
5975 | 1 | ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
5976 | 1 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5977 | | |
5978 | 1 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_context_new(CCV_STREAM_CONTEXT_GPU); |
5979 | | |
5980 | 1 | cmd.backend = CCV_NNC_BACKEND_GPU_CUDNN; |
5981 | 1 | assert(cmd.backend >= 0); |
5982 | 1 | cmd.algorithm = -1; |
5983 | 1 | cmd = ccv_nnc_cmd_autotune(cmd, 512 * 1024 * 1024, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), stream_context); |
5984 | 1 | assert(CCV_NNC_EXEC_SUCCESS == ccv_nnc_cmd_exec(cmd, hint, 0, TENSOR_LIST(ga, gw, gbias), TENSOR_LIST(gc), stream_context)); |
5985 | 1 | ccv_nnc_stream_context_wait(stream_context); |
5986 | 1 | ccv_nnc_stream_context_free(stream_context); |
5987 | 1 | ccv_nnc_tensor_t* c1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5988 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c1), 0); |
5989 | 1 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0); |
5990 | 1 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c1), TENSOR_LIST(c), 0); |
5991 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 5e-3, "output from cudnn should match from CPU"); |
5992 | 1 | ccv_nnc_tensor_free(c); |
5993 | 1 | ccv_nnc_tensor_free(gc); |
5994 | 1 | ccv_nnc_tensor_free(bias); |
5995 | 1 | ccv_nnc_tensor_free(w); |
5996 | 1 | ccv_nnc_tensor_free(b); |
5997 | 1 | ccv_nnc_tensor_free(a); |
5998 | 1 | ccv_nnc_tensor_free(c1); |
5999 | 1 | ccv_nnc_tensor_free(bias1); |
6000 | 1 | ccv_nnc_tensor_free(w1); |
6001 | 1 | ccv_nnc_tensor_free(a1); |
6002 | 1 | ccv_nnc_tensor_free(gbias); |
6003 | 1 | ccv_nnc_tensor_free(gw); |
6004 | 1 | ccv_nnc_tensor_free(ga); |
6005 | 1 | } |
6006 | | |
6007 | | #include "case_main.h" |