/home/liu/actions-runner/_work/ccv/ccv/test/int/nnc/mpsblas.tests.c
Line | Count | Source (jump to first uncovered line) |
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 | | TEST_CASE("gemm no transpose") |
15 | 1 | { |
16 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
17 | 0 | float ap[] = { |
18 | 0 | 1, 2, |
19 | 0 | 3, 4, |
20 | 0 | 5, 6, |
21 | 0 | 7, 8, |
22 | 0 | }; |
23 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
24 | 0 | float bp[] = { |
25 | 0 | 7, 8, 9, |
26 | 0 | 10, 11, 12, |
27 | 0 | }; |
28 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
29 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
30 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
31 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
32 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
33 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
34 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
35 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
36 | 0 | float ctp[] = { |
37 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
38 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
39 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
40 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
41 | 0 | }; |
42 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
43 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
44 | 0 | ccv_nnc_tensor_free(a); |
45 | 0 | ccv_nnc_tensor_free(b); |
46 | 0 | ccv_nnc_tensor_free(c); |
47 | 0 | ccv_nnc_tensor_free(ga); |
48 | 0 | ccv_nnc_tensor_free(gb); |
49 | 0 | ccv_nnc_tensor_free(gc); |
50 | 0 | } |
51 | | |
52 | | TEST_CASE("gemm transpose a") |
53 | 1 | { |
54 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
55 | 0 | float ap[] = { |
56 | 0 | 1, 3, 5, 7, |
57 | 0 | 2, 4, 6, 8, |
58 | 0 | }; |
59 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
60 | 0 | float bp[] = { |
61 | 0 | 7, 8, 9, |
62 | 0 | 10, 11, 12, |
63 | 0 | }; |
64 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
65 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
66 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4), 0); |
67 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
68 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
69 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
70 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
71 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
72 | 0 | float ctp[] = { |
73 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
74 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
75 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
76 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
77 | 0 | }; |
78 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
79 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
80 | 0 | ccv_nnc_tensor_free(a); |
81 | 0 | ccv_nnc_tensor_free(b); |
82 | 0 | ccv_nnc_tensor_free(c); |
83 | 0 | ccv_nnc_tensor_free(ga); |
84 | 0 | ccv_nnc_tensor_free(gb); |
85 | 0 | ccv_nnc_tensor_free(gc); |
86 | 0 | } |
87 | | |
88 | | TEST_CASE("gemm transpose b") |
89 | 1 | { |
90 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
91 | 0 | float ap[] = { |
92 | 0 | 1, 2, |
93 | 0 | 3, 4, |
94 | 0 | 5, 6, |
95 | 0 | 7, 8, |
96 | 0 | }; |
97 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
98 | 0 | float bp[] = { |
99 | 0 | 7, 10, |
100 | 0 | 8, 11, |
101 | 0 | 9, 12, |
102 | 0 | }; |
103 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
104 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
105 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
106 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
107 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
108 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
109 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
110 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
111 | 0 | float ctp[] = { |
112 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
113 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
114 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
115 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
116 | 0 | }; |
117 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
118 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
119 | 0 | ccv_nnc_tensor_free(a); |
120 | 0 | ccv_nnc_tensor_free(b); |
121 | 0 | ccv_nnc_tensor_free(c); |
122 | 0 | ccv_nnc_tensor_free(ga); |
123 | 0 | ccv_nnc_tensor_free(gb); |
124 | 0 | ccv_nnc_tensor_free(gc); |
125 | 0 | } |
126 | | |
127 | | TEST_CASE("gemm transpose a and b") |
128 | 1 | { |
129 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
130 | 0 | float ap[] = { |
131 | 0 | 1, 3, 5, 7, |
132 | 0 | 2, 4, 6, 8, |
133 | 0 | }; |
134 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
135 | 0 | float bp[] = { |
136 | 0 | 7, 10, |
137 | 0 | 8, 11, |
138 | 0 | 9, 12, |
139 | 0 | }; |
140 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
141 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
142 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4), 0); |
143 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
144 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
145 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
146 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(TRANSPOSE(0, 1), TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
147 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
148 | 0 | float ctp[] = { |
149 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
150 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
151 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
152 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
153 | 0 | }; |
154 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
155 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
156 | 0 | ccv_nnc_tensor_free(a); |
157 | 0 | ccv_nnc_tensor_free(b); |
158 | 0 | ccv_nnc_tensor_free(c); |
159 | 0 | ccv_nnc_tensor_free(ga); |
160 | 0 | ccv_nnc_tensor_free(gb); |
161 | 0 | ccv_nnc_tensor_free(gc); |
162 | 0 | } |
163 | | |
164 | | TEST_CASE("gemm no transpose with bias") |
165 | 1 | { |
166 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
167 | 0 | float ap[] = { |
168 | 0 | 1, 2, |
169 | 0 | 3, 4, |
170 | 0 | 5, 6, |
171 | 0 | 7, 8, |
172 | 0 | }; |
173 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
174 | 0 | float bp[] = { |
175 | 0 | 7, 8, 9, |
176 | 0 | 10, 11, 12, |
177 | 0 | }; |
178 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
179 | 0 | float dp[] = { |
180 | 0 | 1, -1, 1, |
181 | 0 | 1, -1, 1, |
182 | 0 | 1, -1, 1, |
183 | 0 | 1, -1, 1, |
184 | 0 | }; |
185 | 0 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(dp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
186 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
187 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
188 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
189 | 0 | ccv_nnc_tensor_t* gd = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
190 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
191 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, d), TENSOR_LIST(ga, gb, gd), 0); |
192 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb, gd), TENSOR_LIST(gc), 0); |
193 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
194 | 0 | float ctp[] = { |
195 | 0 | 1 * 7 + 2 * 10 + 1, 1 * 8 + 2 * 11 - 1, 1 * 9 + 2 * 12 + 1, |
196 | 0 | 3 * 7 + 4 * 10 + 1, 3 * 8 + 4 * 11 - 1, 3 * 9 + 4 * 12 + 1, |
197 | 0 | 5 * 7 + 6 * 10 + 1, 5 * 8 + 6 * 11 - 1, 5 * 9 + 6 * 12 + 1, |
198 | 0 | 7 * 7 + 8 * 10 + 1, 7 * 8 + 8 * 11 - 1, 7 * 9 + 8 * 12 + 1, |
199 | 0 | }; |
200 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
201 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
202 | 0 | ccv_nnc_tensor_free(a); |
203 | 0 | ccv_nnc_tensor_free(b); |
204 | 0 | ccv_nnc_tensor_free(c); |
205 | 0 | ccv_nnc_tensor_free(d); |
206 | 0 | ccv_nnc_tensor_free(ga); |
207 | 0 | ccv_nnc_tensor_free(gb); |
208 | 0 | ccv_nnc_tensor_free(gc); |
209 | 0 | ccv_nnc_tensor_free(gd); |
210 | 0 | } |
211 | | |
212 | | TEST_CASE("gemm no transpose batch 2, no batch b") |
213 | 1 | { |
214 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
215 | 0 | float ap[] = { |
216 | 0 | 1, 2, |
217 | 0 | 3, 4, |
218 | 0 | 5, 6, |
219 | 0 | 7, 8, |
220 | 0 | 2, 3, |
221 | 0 | 4, 5, |
222 | 0 | 6, 7, |
223 | 0 | 8, 9 |
224 | 0 | }; |
225 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
226 | 0 | float bp[] = { |
227 | 0 | 7, 8, 9, |
228 | 0 | 10, 11, 12, |
229 | 0 | }; |
230 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
231 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
232 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
233 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
234 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
235 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
236 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
237 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
238 | 0 | float ctp[] = { |
239 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
240 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
241 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
242 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
243 | 0 | 2 * 7 + 3 * 10, 2 * 8 + 3 * 11, 2 * 9 + 3 * 12, |
244 | 0 | 4 * 7 + 5 * 10, 4 * 8 + 5 * 11, 4 * 9 + 5 * 12, |
245 | 0 | 6 * 7 + 7 * 10, 6 * 8 + 7 * 11, 6 * 9 + 7 * 12, |
246 | 0 | 8 * 7 + 9 * 10, 8 * 8 + 9 * 11, 8 * 9 + 9 * 12, |
247 | 0 | }; |
248 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
249 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
250 | 0 | ccv_nnc_tensor_free(a); |
251 | 0 | ccv_nnc_tensor_free(b); |
252 | 0 | ccv_nnc_tensor_free(c); |
253 | 0 | ccv_nnc_tensor_free(ga); |
254 | 0 | ccv_nnc_tensor_free(gb); |
255 | 0 | ccv_nnc_tensor_free(gc); |
256 | 0 | } |
257 | | |
258 | | TEST_CASE("gemm no transpose batch 2") |
259 | 1 | { |
260 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
261 | 0 | float ap[] = { |
262 | 0 | 1, 2, |
263 | 0 | 3, 4, |
264 | 0 | 5, 6, |
265 | 0 | 7, 8, |
266 | 0 | 2, 3, |
267 | 0 | 4, 5, |
268 | 0 | 6, 7, |
269 | 0 | 8, 9 |
270 | 0 | }; |
271 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
272 | 0 | float bp[] = { |
273 | 0 | 7, 8, 9, |
274 | 0 | 10, 11, 12, |
275 | 0 | 8, 9, 10, |
276 | 0 | 11, 12, 13, |
277 | 0 | }; |
278 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
279 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
280 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
281 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 3), 0); |
282 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
283 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
284 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
285 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
286 | 0 | float ctp[] = { |
287 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
288 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
289 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
290 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
291 | 0 | 2 * 8 + 3 * 11, 2 * 9 + 3 * 12, 2 * 10 + 3 * 13, |
292 | 0 | 4 * 8 + 5 * 11, 4 * 9 + 5 * 12, 4 * 10 + 5 * 13, |
293 | 0 | 6 * 8 + 7 * 11, 6 * 9 + 7 * 12, 6 * 10 + 7 * 13, |
294 | 0 | 8 * 8 + 9 * 11, 8 * 9 + 9 * 12, 8 * 10 + 9 * 13, |
295 | 0 | }; |
296 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
297 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
298 | 0 | ccv_nnc_tensor_free(a); |
299 | 0 | ccv_nnc_tensor_free(b); |
300 | 0 | ccv_nnc_tensor_free(c); |
301 | 0 | ccv_nnc_tensor_free(ga); |
302 | 0 | ccv_nnc_tensor_free(gb); |
303 | 0 | ccv_nnc_tensor_free(gc); |
304 | 0 | } |
305 | | |
306 | | TEST_CASE("gemm transpose a batch 2, no batch b, with bias") |
307 | 1 | { |
308 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
309 | 0 | float ap[] = { |
310 | 0 | 1, 3, 5, 7, |
311 | 0 | 2, 4, 6, 8, |
312 | 0 | 2, 4, 6, 8, |
313 | 0 | 3, 5, 7, 9, |
314 | 0 | }; |
315 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
316 | 0 | float bp[] = { |
317 | 0 | 7, 8, 9, |
318 | 0 | 10, 11, 12, |
319 | 0 | }; |
320 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
321 | 0 | float dp[] = { |
322 | 0 | -1, 0, 1, |
323 | 0 | }; |
324 | 0 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(dp, CPU_TENSOR_NHWC(32F, 3), 0); |
325 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
326 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
327 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
328 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
329 | 0 | ccv_nnc_tensor_t* gd = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
330 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, d), TENSOR_LIST(ga, gb, gd), 0); |
331 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb, gd), TENSOR_LIST(gc), 0); |
332 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
333 | 0 | float ctp[] = { |
334 | 0 | 1 * 7 + 2 * 10 - 1, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12 + 1, |
335 | 0 | 3 * 7 + 4 * 10 - 1, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12 + 1, |
336 | 0 | 5 * 7 + 6 * 10 - 1, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12 + 1, |
337 | 0 | 7 * 7 + 8 * 10 - 1, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12 + 1, |
338 | 0 | 2 * 7 + 3 * 10 - 1, 2 * 8 + 3 * 11, 2 * 9 + 3 * 12 + 1, |
339 | 0 | 4 * 7 + 5 * 10 - 1, 4 * 8 + 5 * 11, 4 * 9 + 5 * 12 + 1, |
340 | 0 | 6 * 7 + 7 * 10 - 1, 6 * 8 + 7 * 11, 6 * 9 + 7 * 12 + 1, |
341 | 0 | 8 * 7 + 9 * 10 - 1, 8 * 8 + 9 * 11, 8 * 9 + 9 * 12 + 1, |
342 | 0 | }; |
343 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
344 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
345 | 0 | ccv_nnc_tensor_free(a); |
346 | 0 | ccv_nnc_tensor_free(b); |
347 | 0 | ccv_nnc_tensor_free(c); |
348 | 0 | ccv_nnc_tensor_free(d); |
349 | 0 | ccv_nnc_tensor_free(ga); |
350 | 0 | ccv_nnc_tensor_free(gb); |
351 | 0 | ccv_nnc_tensor_free(gc); |
352 | 0 | ccv_nnc_tensor_free(gd); |
353 | 0 | } |
354 | | |
355 | | TEST_CASE("gemm transpose a batch 2, with bias") |
356 | 1 | { |
357 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
358 | 0 | float ap[] = { |
359 | 0 | 1, 3, 5, 7, |
360 | 0 | 2, 4, 6, 8, |
361 | 0 | 2, 4, 6, 8, |
362 | 0 | 3, 5, 7, 9, |
363 | 0 | }; |
364 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
365 | 0 | float bp[] = { |
366 | 0 | 7, 8, 9, |
367 | 0 | 10, 11, 12, |
368 | 0 | 8, 9, 10, |
369 | 0 | 11, 12, 13, |
370 | 0 | }; |
371 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
372 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
373 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
374 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 3), 0); |
375 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
376 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
377 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
378 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
379 | 0 | float ctp[] = { |
380 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
381 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
382 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
383 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
384 | 0 | 2 * 8 + 3 * 11, 2 * 9 + 3 * 12, 2 * 10 + 3 * 13, |
385 | 0 | 4 * 8 + 5 * 11, 4 * 9 + 5 * 12, 4 * 10 + 5 * 13, |
386 | 0 | 6 * 8 + 7 * 11, 6 * 9 + 7 * 12, 6 * 10 + 7 * 13, |
387 | 0 | 8 * 8 + 9 * 11, 8 * 9 + 9 * 12, 8 * 10 + 9 * 13, |
388 | 0 | }; |
389 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
390 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
391 | 0 | ccv_nnc_tensor_free(a); |
392 | 0 | ccv_nnc_tensor_free(b); |
393 | 0 | ccv_nnc_tensor_free(c); |
394 | 0 | ccv_nnc_tensor_free(ga); |
395 | 0 | ccv_nnc_tensor_free(gb); |
396 | 0 | ccv_nnc_tensor_free(gc); |
397 | 0 | } |
398 | | |
399 | | TEST_CASE("gemm transpose b batch 2, with bias") |
400 | 1 | { |
401 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
402 | 0 | float ap[] = { |
403 | 0 | 1, 2, |
404 | 0 | 3, 4, |
405 | 0 | 5, 6, |
406 | 0 | 7, 8, |
407 | 0 | 2, 3, |
408 | 0 | 4, 5, |
409 | 0 | 6, 7, |
410 | 0 | 8, 9 |
411 | 0 | }; |
412 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
413 | 0 | float bp[] = { |
414 | 0 | 7, 10, |
415 | 0 | 8, 11, |
416 | 0 | 9, 12, |
417 | 0 | 80, 110, |
418 | 0 | 90, 120, |
419 | 0 | 10, 13, |
420 | 0 | }; |
421 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
422 | 0 | float dp[] = { |
423 | 0 | -1, 0, 1, |
424 | 0 | 2, 3, -4, |
425 | 0 | }; |
426 | 0 | ccv_nnc_tensor_t* const d = ccv_nnc_tensor_new(dp, CPU_TENSOR_NHWC(32F, 2, 1, 3), 0); |
427 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
428 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
429 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
430 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
431 | 0 | ccv_nnc_tensor_t* gd = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 1, 3), 0); |
432 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b, d), TENSOR_LIST(ga, gb, gd), 0); |
433 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb, gd), TENSOR_LIST(gc), 0); |
434 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
435 | 0 | float ctp[] = { |
436 | 0 | 1 * 7 + 2 * 10 - 1, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12 + 1, |
437 | 0 | 3 * 7 + 4 * 10 - 1, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12 + 1, |
438 | 0 | 5 * 7 + 6 * 10 - 1, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12 + 1, |
439 | 0 | 7 * 7 + 8 * 10 - 1, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12 + 1, |
440 | 0 | 2 * 80 + 3 * 110 + 2, 2 * 90 + 3 * 120 + 3, 2 * 10 + 3 * 13 - 4, |
441 | 0 | 4 * 80 + 5 * 110 + 2, 4 * 90 + 5 * 120 + 3, 4 * 10 + 5 * 13 - 4, |
442 | 0 | 6 * 80 + 7 * 110 + 2, 6 * 90 + 7 * 120 + 3, 6 * 10 + 7 * 13 - 4, |
443 | 0 | 8 * 80 + 9 * 110 + 2, 8 * 90 + 9 * 120 + 3, 8 * 10 + 9 * 13 - 4, |
444 | 0 | }; |
445 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
446 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
447 | 0 | ccv_nnc_tensor_free(a); |
448 | 0 | ccv_nnc_tensor_free(b); |
449 | 0 | ccv_nnc_tensor_free(c); |
450 | 0 | ccv_nnc_tensor_free(d); |
451 | 0 | ccv_nnc_tensor_free(ga); |
452 | 0 | ccv_nnc_tensor_free(gb); |
453 | 0 | ccv_nnc_tensor_free(gc); |
454 | 0 | ccv_nnc_tensor_free(gd); |
455 | 0 | } |
456 | | |
457 | | TEST_CASE("gemm transpose b batch 2") |
458 | 1 | { |
459 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
460 | 0 | float ap[] = { |
461 | 0 | 1, 2, |
462 | 0 | 3, 4, |
463 | 0 | 5, 6, |
464 | 0 | 7, 8, |
465 | 0 | 2, 3, |
466 | 0 | 4, 5, |
467 | 0 | 6, 7, |
468 | 0 | 8, 9 |
469 | 0 | }; |
470 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
471 | 0 | float bp[] = { |
472 | 0 | 7, 10, |
473 | 0 | 8, 11, |
474 | 0 | 9, 12, |
475 | 0 | 80, 110, |
476 | 0 | 90, 120, |
477 | 0 | 10, 13, |
478 | 0 | }; |
479 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
480 | 0 | ccv_nnc_tensor_t* const c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
481 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
482 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
483 | 0 | ccv_nnc_tensor_t* gc = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
484 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(ga, gb), 0); |
485 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(ga, gb), TENSOR_LIST(gc), 0); |
486 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0); |
487 | 0 | float ctp[] = { |
488 | 0 | 1 * 7 + 2 * 10, 1 * 8 + 2 * 11, 1 * 9 + 2 * 12, |
489 | 0 | 3 * 7 + 4 * 10, 3 * 8 + 4 * 11, 3 * 9 + 4 * 12, |
490 | 0 | 5 * 7 + 6 * 10, 5 * 8 + 6 * 11, 5 * 9 + 6 * 12, |
491 | 0 | 7 * 7 + 8 * 10, 7 * 8 + 8 * 11, 7 * 9 + 8 * 12, |
492 | 0 | 2 * 80 + 3 * 110, 2 * 90 + 3 * 120, 2 * 10 + 3 * 13, |
493 | 0 | 4 * 80 + 5 * 110, 4 * 90 + 5 * 120, 4 * 10 + 5 * 13, |
494 | 0 | 6 * 80 + 7 * 110, 6 * 90 + 7 * 120, 6 * 10 + 7 * 13, |
495 | 0 | 8 * 80 + 9 * 110, 8 * 90 + 9 * 120, 8 * 10 + 9 * 13, |
496 | 0 | }; |
497 | 0 | ccv_nnc_tensor_t ct = ccv_nnc_tensor(ctp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
498 | 0 | REQUIRE_TENSOR_EQ(c, &ct, "result should be equal"); |
499 | 0 | ccv_nnc_tensor_free(a); |
500 | 0 | ccv_nnc_tensor_free(b); |
501 | 0 | ccv_nnc_tensor_free(c); |
502 | 0 | ccv_nnc_tensor_free(ga); |
503 | 0 | ccv_nnc_tensor_free(gb); |
504 | 0 | ccv_nnc_tensor_free(gc); |
505 | 0 | } |
506 | | |
507 | | TEST_CASE("mps forward gemm") |
508 | 1 | { |
509 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
510 | 0 | dsfmt_t dsfmt; |
511 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
512 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
513 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
514 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
515 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
516 | |
|
517 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
518 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
519 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
520 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
521 | 0 | int i; |
522 | 0 | for (i = 0; i < 64 * 128; i++) |
523 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
524 | 0 | for (i = 0; i < 64; i++) |
525 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
526 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
527 | 0 | for (i = 0; i < 10 * 128; i++) |
528 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
529 | 0 | for (i = 0; i < 128; i++) |
530 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
531 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw, hbias), TENSOR_LIST(a, w, bias), 0); |
532 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(hb), 0); |
533 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
534 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
535 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
536 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
537 | 0 | for (i = 0; i < 64; i++) |
538 | 0 | tb1->data.f32[i] = tb->data.f32[i]; |
539 | 0 | REQUIRE_TENSOR_EQ(tb1, hb, "GPU computed output should be the same as CPU computed ones"); |
540 | 0 | ccv_nnc_tensor_free(a); |
541 | 0 | ccv_nnc_tensor_free(w); |
542 | 0 | ccv_nnc_tensor_free(bias); |
543 | 0 | ccv_nnc_tensor_free(tb); |
544 | 0 | ccv_nnc_tensor_free(b); |
545 | 0 | ccv_nnc_tensor_free(ha); |
546 | 0 | ccv_nnc_tensor_free(ha1); |
547 | 0 | ccv_nnc_tensor_free(tb1); |
548 | 0 | ccv_nnc_tensor_free(hw); |
549 | 0 | ccv_nnc_tensor_free(hbias); |
550 | 0 | ccv_nnc_tensor_free(hb); |
551 | 0 | } |
552 | | |
553 | | TEST_CASE("mps forward gemm in half precision") |
554 | 1 | { |
555 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
556 | 0 | dsfmt_t dsfmt; |
557 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
558 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 128), 0); |
559 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64, 128), 0); |
560 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64), 0); |
561 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 64), 0); |
562 | |
|
563 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
564 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
565 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
566 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
567 | 0 | int i; |
568 | 0 | for (i = 0; i < 64 * 128; i++) |
569 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
570 | 0 | for (i = 0; i < 64; i++) |
571 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
572 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
573 | 0 | for (i = 0; i < 10 * 128; i++) |
574 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
575 | 0 | for (i = 0; i < 128; i++) |
576 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
577 | 0 | ccv_nnc_tensor_t* ha2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 128), 0); |
578 | 0 | ccv_nnc_tensor_t* hw2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64, 128), 0); |
579 | 0 | ccv_nnc_tensor_t* hbias2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64), 0); |
580 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw, hbias), TENSOR_LIST(ha2, hw2, hbias2), 0); |
581 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha2, hw2, hbias2), TENSOR_LIST(a, w, bias), 0); |
582 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(hb), 0); |
583 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
584 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 64), 0); |
585 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
586 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
587 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tb), TENSOR_LIST(tb1), 0); |
588 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tb1->data.f32, hb->data.f32, 64, 5e-3, "GPU computed output should be the same as CPU computed ones"); |
589 | 0 | ccv_nnc_tensor_free(a); |
590 | 0 | ccv_nnc_tensor_free(w); |
591 | 0 | ccv_nnc_tensor_free(bias); |
592 | 0 | ccv_nnc_tensor_free(b); |
593 | 0 | ccv_nnc_tensor_free(tb); |
594 | 0 | ccv_nnc_tensor_free(ha); |
595 | 0 | ccv_nnc_tensor_free(ha1); |
596 | 0 | ccv_nnc_tensor_free(tb1); |
597 | 0 | ccv_nnc_tensor_free(hw); |
598 | 0 | ccv_nnc_tensor_free(hbias); |
599 | 0 | ccv_nnc_tensor_free(hb); |
600 | 0 | ccv_nnc_tensor_free(ha2); |
601 | 0 | ccv_nnc_tensor_free(hw2); |
602 | 0 | ccv_nnc_tensor_free(hbias2); |
603 | 0 | } |
604 | | |
605 | | TEST_CASE("mps forward gemv in half precision, variant 1") |
606 | 1 | { |
607 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
608 | 0 | dsfmt_t dsfmt; |
609 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
610 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 1, 128), 0); |
611 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64, 128), 0); |
612 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64), 0); |
613 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 1, 64), 0); |
614 | |
|
615 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
616 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
617 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
618 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
619 | 0 | int i; |
620 | 0 | for (i = 0; i < 64 * 128; i++) |
621 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
622 | 0 | for (i = 0; i < 64; i++) |
623 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
624 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
625 | 0 | for (i = 0; i < 128; i++) |
626 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
627 | 0 | for (i = 0; i < 128; i++) |
628 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
629 | 0 | ccv_nnc_tensor_t* ha2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 1, 128), 0); |
630 | 0 | ccv_nnc_tensor_t* hw2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64, 128), 0); |
631 | 0 | ccv_nnc_tensor_t* hbias2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64), 0); |
632 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw, hbias), TENSOR_LIST(ha2, hw2, hbias2), 0); |
633 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha2, hw2, hbias2), TENSOR_LIST(a, w, bias), 0); |
634 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(hb), 0); |
635 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
636 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 1, 64), 0); |
637 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
638 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
639 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tb), TENSOR_LIST(tb1), 0); |
640 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tb1->data.f32, hb->data.f32, 64, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
641 | 0 | ccv_nnc_tensor_free(a); |
642 | 0 | ccv_nnc_tensor_free(w); |
643 | 0 | ccv_nnc_tensor_free(bias); |
644 | 0 | ccv_nnc_tensor_free(b); |
645 | 0 | ccv_nnc_tensor_free(tb); |
646 | 0 | ccv_nnc_tensor_free(ha); |
647 | 0 | ccv_nnc_tensor_free(ha1); |
648 | 0 | ccv_nnc_tensor_free(tb1); |
649 | 0 | ccv_nnc_tensor_free(hw); |
650 | 0 | ccv_nnc_tensor_free(hbias); |
651 | 0 | ccv_nnc_tensor_free(hb); |
652 | 0 | ccv_nnc_tensor_free(ha2); |
653 | 0 | ccv_nnc_tensor_free(hw2); |
654 | 0 | ccv_nnc_tensor_free(hbias2); |
655 | 0 | } |
656 | | |
657 | | TEST_CASE("mps forward gemm no bias") |
658 | 1 | { |
659 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
660 | 0 | dsfmt_t dsfmt; |
661 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
662 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
663 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
664 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
665 | |
|
666 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
667 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
668 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
669 | 0 | int i; |
670 | 0 | for (i = 0; i < 64 * 128; i++) |
671 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
672 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
673 | 0 | for (i = 0; i < 10 * 128; i++) |
674 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
675 | 0 | for (i = 0; i < 128; i++) |
676 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
677 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw), TENSOR_LIST(a, w), 0); |
678 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(hb), 0); |
679 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w), TENSOR_LIST(b), 0); |
680 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
681 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
682 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
683 | 0 | for (i = 0; i < 64; i++) |
684 | 0 | tb1->data.f32[i] = tb->data.f32[i]; |
685 | 0 | REQUIRE_TENSOR_EQ(tb1, hb, "GPU computed output should be the same as CPU computed ones"); |
686 | 0 | ccv_nnc_tensor_free(a); |
687 | 0 | ccv_nnc_tensor_free(w); |
688 | 0 | ccv_nnc_tensor_free(b); |
689 | 0 | ccv_nnc_tensor_free(tb); |
690 | 0 | ccv_nnc_tensor_free(ha); |
691 | 0 | ccv_nnc_tensor_free(ha1); |
692 | 0 | ccv_nnc_tensor_free(tb1); |
693 | 0 | ccv_nnc_tensor_free(hw); |
694 | 0 | ccv_nnc_tensor_free(hb); |
695 | 0 | } |
696 | | |
697 | | TEST_CASE("mps forward gemm no bias in half precision") |
698 | 1 | { |
699 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
700 | 0 | dsfmt_t dsfmt; |
701 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
702 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 128), 0); |
703 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64, 128), 0); |
704 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 64), 0); |
705 | |
|
706 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
707 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
708 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
709 | 0 | int i; |
710 | 0 | for (i = 0; i < 64 * 128; i++) |
711 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
712 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
713 | 0 | for (i = 0; i < 10 * 128; i++) |
714 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
715 | 0 | for (i = 0; i < 128; i++) |
716 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
717 | 0 | ccv_nnc_tensor_t* ha2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 128), 0); |
718 | 0 | ccv_nnc_tensor_t* hw2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64, 128), 0); |
719 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw), TENSOR_LIST(ha2, hw2), 0); |
720 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha2, hw2), TENSOR_LIST(a, w), 0); |
721 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(hb), 0); |
722 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w), TENSOR_LIST(b), 0); |
723 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 64), 0); |
724 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
725 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
726 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tb), TENSOR_LIST(tb1), 0); |
727 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tb1->data.f32, hb->data.f32, 64, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
728 | 0 | ccv_nnc_tensor_free(a); |
729 | 0 | ccv_nnc_tensor_free(w); |
730 | 0 | ccv_nnc_tensor_free(b); |
731 | 0 | ccv_nnc_tensor_free(tb); |
732 | 0 | ccv_nnc_tensor_free(ha); |
733 | 0 | ccv_nnc_tensor_free(ha1); |
734 | 0 | ccv_nnc_tensor_free(tb1); |
735 | 0 | ccv_nnc_tensor_free(hw); |
736 | 0 | ccv_nnc_tensor_free(hb); |
737 | 0 | ccv_nnc_tensor_free(ha2); |
738 | 0 | ccv_nnc_tensor_free(hw2); |
739 | 0 | } |
740 | | |
741 | | TEST_CASE("mps forward gemv in half precision no bias, variant 1") |
742 | 1 | { |
743 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
744 | 0 | dsfmt_t dsfmt; |
745 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
746 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 1, 128), 0); |
747 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64, 128), 0); |
748 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 1, 64), 0); |
749 | |
|
750 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
751 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
752 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
753 | 0 | int i; |
754 | 0 | for (i = 0; i < 64 * 128; i++) |
755 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
756 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 128), 0); |
757 | 0 | for (i = 0; i < 128; i++) |
758 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
759 | 0 | for (i = 0; i < 128; i++) |
760 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
761 | 0 | ccv_nnc_tensor_t* ha2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 1, 128), 0); |
762 | 0 | ccv_nnc_tensor_t* hw2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64, 128), 0); |
763 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw), TENSOR_LIST(ha2, hw2), 0); |
764 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha2, hw2), TENSOR_LIST(a, w), 0); |
765 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(hb), 0); |
766 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w), TENSOR_LIST(b), 0); |
767 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 1, 64), 0); |
768 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
769 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 64), 0); |
770 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tb), TENSOR_LIST(tb1), 0); |
771 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tb1->data.f32, hb->data.f32, 64, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
772 | 0 | ccv_nnc_tensor_free(a); |
773 | 0 | ccv_nnc_tensor_free(w); |
774 | 0 | ccv_nnc_tensor_free(b); |
775 | 0 | ccv_nnc_tensor_free(tb); |
776 | 0 | ccv_nnc_tensor_free(ha); |
777 | 0 | ccv_nnc_tensor_free(ha1); |
778 | 0 | ccv_nnc_tensor_free(tb1); |
779 | 0 | ccv_nnc_tensor_free(hw); |
780 | 0 | ccv_nnc_tensor_free(hb); |
781 | 0 | ccv_nnc_tensor_free(ha2); |
782 | 0 | ccv_nnc_tensor_free(hw2); |
783 | 0 | } |
784 | | |
785 | | TEST_CASE("mps forward gemv in half precision no bias, variant 2") |
786 | 1 | { |
787 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
788 | 0 | dsfmt_t dsfmt; |
789 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
790 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64, 128), 0); |
791 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 128, 1), 0); |
792 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 64, 1), 0); |
793 | |
|
794 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
795 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 128, 1), 0); |
796 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 1), 0); |
797 | 0 | int i; |
798 | 0 | for (i = 0; i < 64 * 128; i++) |
799 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
800 | 0 | ccv_nnc_tensor_t* ha1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 128, 1), 0); |
801 | 0 | for (i = 0; i < 128; i++) |
802 | 0 | ha1->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
803 | 0 | for (i = 0; i < 128; i++) |
804 | 0 | ha->data.f32[i] = ha1->data.f32[i]; |
805 | 0 | ccv_nnc_tensor_t* hw2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64, 128), 0); |
806 | 0 | ccv_nnc_tensor_t* ha2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 128, 1), 0); |
807 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha1, hw), TENSOR_LIST(ha2, hw2), 0); |
808 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha2, hw2), TENSOR_LIST(a, w), 0); |
809 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, NO_TRANSPOSE), ccv_nnc_no_hint, 0, TENSOR_LIST(hw, ha), TENSOR_LIST(hb), 0); |
810 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, NO_TRANSPOSE), ccv_nnc_no_hint, 0, TENSOR_LIST(w, a), TENSOR_LIST(b), 0); |
811 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 64, 1), 0); |
812 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(tb), 0); |
813 | 0 | ccv_nnc_tensor_t* tb1 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 1), 0); |
814 | 0 | ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tb), TENSOR_LIST(tb1), 0); |
815 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tb1->data.f32, hb->data.f32, 64, 1e-3, "GPU computed output should be the same as CPU computed ones"); |
816 | 0 | ccv_nnc_tensor_free(a); |
817 | 0 | ccv_nnc_tensor_free(w); |
818 | 0 | ccv_nnc_tensor_free(b); |
819 | 0 | ccv_nnc_tensor_free(tb); |
820 | 0 | ccv_nnc_tensor_free(ha); |
821 | 0 | ccv_nnc_tensor_free(ha1); |
822 | 0 | ccv_nnc_tensor_free(tb1); |
823 | 0 | ccv_nnc_tensor_free(hw); |
824 | 0 | ccv_nnc_tensor_free(hb); |
825 | 0 | ccv_nnc_tensor_free(ha2); |
826 | 0 | ccv_nnc_tensor_free(hw2); |
827 | 0 | } |
828 | | |
829 | | TEST_CASE("mps handle permute") |
830 | 1 | { |
831 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
832 | 0 | dsfmt_t dsfmt; |
833 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
834 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 2, 128), 0); |
835 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 2, 128), 0); |
836 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 2, 128), 0); |
837 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 2, 128), 0); |
838 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 64), 0); |
839 | |
|
840 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 128), 0); |
841 | 0 | ccv_nnc_tensor_t* wt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 128), 0); |
842 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 64), 0); |
843 | 0 | int i; |
844 | 0 | for (i = 0; i < 2 * 64 * 128; i++) |
845 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
846 | 0 | for (i = 0; i < 2 * 10 * 128; i++) |
847 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
848 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(a, w), 0); |
849 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(0, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(at), 0); |
850 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(0, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(w), TENSOR_LIST(wt), 0); |
851 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(at, wt), TENSOR_LIST(bt), 0); |
852 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(128, 2 * 128, 1)); |
853 | 0 | ccv_nnc_tensor_view_t* wv = ccv_nnc_tensor_view_new(w, GPU_TENSOR_NHWC(000, 32F, 2, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(128, 2 * 128, 1)); |
854 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)av, (ccv_nnc_tensor_t*)wv), TENSOR_LIST(b), 0); |
855 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 64), 0); |
856 | 0 | ccv_nnc_tensor_t* hbt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 64), 0); |
857 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, bt), TENSOR_LIST(hb, hbt), 0); |
858 | 0 | REQUIRE_TENSOR_EQ(hb, hbt, "permute computed output should be the same as non-permute computed ones"); |
859 | 0 | ccv_nnc_tensor_free(ha); |
860 | 0 | ccv_nnc_tensor_free(hw); |
861 | 0 | ccv_nnc_tensor_free(a); |
862 | 0 | ccv_nnc_tensor_free(w); |
863 | 0 | ccv_nnc_tensor_free(b); |
864 | 0 | ccv_nnc_tensor_view_free(av); |
865 | 0 | ccv_nnc_tensor_view_free(wv); |
866 | 0 | ccv_nnc_tensor_free(at); |
867 | 0 | ccv_nnc_tensor_free(wt); |
868 | 0 | ccv_nnc_tensor_free(bt); |
869 | 0 | ccv_nnc_tensor_free(hb); |
870 | 0 | ccv_nnc_tensor_free(hbt); |
871 | 0 | } |
872 | | |
873 | | TEST_CASE("generalized batched gemm with batch (2, 4) compare mps") |
874 | 1 | { |
875 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
876 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
877 | 0 | dsfmt_t dsfmt; |
878 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
879 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
880 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
881 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
882 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
883 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 4, 128), 0); |
884 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
885 | |
|
886 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
887 | 0 | ccv_nnc_tensor_t* wt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 64, 128), 0); |
888 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
889 | 0 | int i; |
890 | 0 | for (i = 0; i < 8 * 64 * 128; i++) |
891 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
892 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
893 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
894 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
895 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(hw), TENSOR_LIST(wt), 0); |
896 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(a, w), 0); |
897 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
898 | 0 | ccv_nnc_tensor_view_t* wv = ccv_nnc_tensor_view_new(w, GPU_TENSOR_NHWC(000, 32F, 2, 4, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(64 * 4 * 128, 128, 4 * 128, 1)); |
899 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)av, (ccv_nnc_tensor_t*)wv), TENSOR_LIST(b), 0); |
900 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST(at, wt), TENSOR_LIST(bt), 0); |
901 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
902 | 0 | REQUIRE_TENSOR_EQ(hb, bt, "permute computed output should be the same as non-permute computed ones"); |
903 | 0 | ccv_nnc_tensor_free(ha); |
904 | 0 | ccv_nnc_tensor_free(hw); |
905 | 0 | ccv_nnc_tensor_free(hb); |
906 | 0 | ccv_nnc_tensor_free(a); |
907 | 0 | ccv_nnc_tensor_free(w); |
908 | 0 | ccv_nnc_tensor_free(b); |
909 | 0 | ccv_nnc_tensor_view_free(av); |
910 | 0 | ccv_nnc_tensor_view_free(wv); |
911 | 0 | ccv_nnc_tensor_free(at); |
912 | 0 | ccv_nnc_tensor_free(wt); |
913 | 0 | ccv_nnc_tensor_free(bt); |
914 | 0 | } |
915 | | |
916 | | TEST_CASE("generalized batched gemm with batch (2, 4) and broadcast compare mps") |
917 | 1 | { |
918 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
919 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
920 | 0 | dsfmt_t dsfmt; |
921 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
922 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
923 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
924 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
925 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
926 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
927 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
928 | |
|
929 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
930 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
931 | 0 | int i; |
932 | 0 | for (i = 0; i < 64 * 128; i++) |
933 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
934 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
935 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
936 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
937 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(a, w), 0); |
938 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
939 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)av, w), TENSOR_LIST(b), 0); |
940 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(at, hw), TENSOR_LIST(bt), 0); |
941 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
942 | 0 | REQUIRE_TENSOR_EQ(hb, bt, "permute computed output should be the same as non-permute computed ones"); |
943 | 0 | ccv_nnc_tensor_free(ha); |
944 | 0 | ccv_nnc_tensor_free(hw); |
945 | 0 | ccv_nnc_tensor_free(hb); |
946 | 0 | ccv_nnc_tensor_free(a); |
947 | 0 | ccv_nnc_tensor_free(w); |
948 | 0 | ccv_nnc_tensor_free(b); |
949 | 0 | ccv_nnc_tensor_view_free(av); |
950 | 0 | ccv_nnc_tensor_free(at); |
951 | 0 | ccv_nnc_tensor_free(bt); |
952 | 0 | } |
953 | | |
954 | | TEST_CASE("generalized batched gemm with batch (2, 4) with bias compare mps") |
955 | 1 | { |
956 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
957 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
958 | 0 | dsfmt_t dsfmt; |
959 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
960 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
961 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
962 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
963 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
964 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
965 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 4, 128), 0); |
966 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
967 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
968 | |
|
969 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
970 | 0 | ccv_nnc_tensor_t* wt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 64, 128), 0); |
971 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
972 | 0 | int i; |
973 | 0 | for (i = 0; i < 8 * 64 * 128; i++) |
974 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
975 | 0 | for (i = 0; i < 64; i++) |
976 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / 64; |
977 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
978 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
979 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
980 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(hw), TENSOR_LIST(wt), 0); |
981 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(a, w, bias), 0); |
982 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
983 | 0 | ccv_nnc_tensor_view_t* wv = ccv_nnc_tensor_view_new(w, GPU_TENSOR_NHWC(000, 32F, 2, 4, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(64 * 4 * 128, 128, 4 * 128, 1)); |
984 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)av, (ccv_nnc_tensor_t*)wv, bias), TENSOR_LIST(b), 0); |
985 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST(at, wt, hbias), TENSOR_LIST(bt), 0); |
986 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
987 | 0 | REQUIRE_TENSOR_EQ(hb, bt, "permute computed output should be the same as non-permute computed ones"); |
988 | 0 | ccv_nnc_tensor_free(ha); |
989 | 0 | ccv_nnc_tensor_free(hw); |
990 | 0 | ccv_nnc_tensor_free(hbias); |
991 | 0 | ccv_nnc_tensor_free(hb); |
992 | 0 | ccv_nnc_tensor_free(a); |
993 | 0 | ccv_nnc_tensor_free(w); |
994 | 0 | ccv_nnc_tensor_free(bias); |
995 | 0 | ccv_nnc_tensor_free(b); |
996 | 0 | ccv_nnc_tensor_view_free(av); |
997 | 0 | ccv_nnc_tensor_view_free(wv); |
998 | 0 | ccv_nnc_tensor_free(at); |
999 | 0 | ccv_nnc_tensor_free(wt); |
1000 | 0 | ccv_nnc_tensor_free(bt); |
1001 | 0 | } |
1002 | | |
1003 | | TEST_CASE("generalized batched gemm with batch (2, 4) with bias and broadcast compare mps") |
1004 | 1 | { |
1005 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS)); |
1006 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
1007 | 0 | dsfmt_t dsfmt; |
1008 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1009 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1010 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1011 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
1012 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
1013 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1014 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
1015 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
1016 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
1017 | |
|
1018 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1019 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
1020 | 0 | int i; |
1021 | 0 | for (i = 0; i < 64 * 128; i++) |
1022 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
1023 | 0 | for (i = 0; i < 64; i++) |
1024 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / 64; |
1025 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
1026 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1027 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
1028 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(a, w, bias), 0); |
1029 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1030 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)av, w, bias), TENSOR_LIST(b), 0); |
1031 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(at, hw, hbias), TENSOR_LIST(bt), 0); |
1032 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1033 | 0 | REQUIRE_TENSOR_EQ(hb, bt, "permute computed output should be the same as non-permute computed ones"); |
1034 | 0 | ccv_nnc_tensor_free(ha); |
1035 | 0 | ccv_nnc_tensor_free(hw); |
1036 | 0 | ccv_nnc_tensor_free(hbias); |
1037 | 0 | ccv_nnc_tensor_free(hb); |
1038 | 0 | ccv_nnc_tensor_free(a); |
1039 | 0 | ccv_nnc_tensor_free(w); |
1040 | 0 | ccv_nnc_tensor_free(bias); |
1041 | 0 | ccv_nnc_tensor_free(b); |
1042 | 0 | ccv_nnc_tensor_view_free(av); |
1043 | 0 | ccv_nnc_tensor_free(at); |
1044 | 0 | ccv_nnc_tensor_free(bt); |
1045 | 0 | } |
1046 | | |
1047 | | TEST_CASE("generalized batched backward gemm with batch (2, 4) compare mps") |
1048 | 1 | { |
1049 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1050 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
1051 | 0 | dsfmt_t dsfmt; |
1052 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1053 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1054 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
1055 | 0 | ccv_nnc_tensor_t* hda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1056 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
1057 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
1058 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1059 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 4, 128), 0); |
1060 | 0 | ccv_nnc_tensor_t* da = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1061 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 4, 128), 0); |
1062 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
1063 | |
|
1064 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1065 | 0 | ccv_nnc_tensor_t* wt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 64, 128), 0); |
1066 | 0 | ccv_nnc_tensor_t* dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1067 | 0 | ccv_nnc_tensor_t* dwt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 64, 128), 0); |
1068 | 0 | ccv_nnc_tensor_t* tda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1069 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
1070 | 0 | int i; |
1071 | 0 | for (i = 0; i < 8 * 64 * 128; i++) |
1072 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
1073 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
1074 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1075 | 0 | for (i = 0; i < 2 * 4 * 10 * 64; i++) |
1076 | 0 | hb->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1077 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
1078 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(hw), TENSOR_LIST(wt), 0); |
1079 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hb), TENSOR_LIST(a, w, b), 0); |
1080 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1081 | 0 | ccv_nnc_tensor_view_t* wv = ccv_nnc_tensor_view_new(w, GPU_TENSOR_NHWC(000, 32F, 2, 4, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(64 * 4 * 128, 128, 4 * 128, 1)); |
1082 | 0 | ccv_nnc_tensor_view_t* dav = ccv_nnc_tensor_view_new(da, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1083 | 0 | ccv_nnc_tensor_view_t* dwv = ccv_nnc_tensor_view_new(dw, GPU_TENSOR_NHWC(000, 32F, 2, 4, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(64 * 4 * 128, 128, 4 * 128, 1)); |
1084 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST(b, (ccv_nnc_tensor_t*)av, (ccv_nnc_tensor_t*)wv), TENSOR_LIST((ccv_nnc_tensor_t*)dav, (ccv_nnc_tensor_t*)dwv), 0); |
1085 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST(hb, at, wt), TENSOR_LIST(dat, dwt), 0); |
1086 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(dat), TENSOR_LIST(tda), 0); |
1087 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(dwt), TENSOR_LIST(tdw), 0); |
1088 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(da, dw), TENSOR_LIST(hda, hdw), 0); |
1089 | 0 | REQUIRE_TENSOR_EQ(hda, tda, "permute computed output should be the same as non-permute computed ones"); |
1090 | 0 | REQUIRE_TENSOR_EQ(hdw, tdw, "permute computed output should be the same as non-permute computed ones"); |
1091 | 0 | ccv_nnc_tensor_free(ha); |
1092 | 0 | ccv_nnc_tensor_free(hw); |
1093 | 0 | ccv_nnc_tensor_free(hda); |
1094 | 0 | ccv_nnc_tensor_free(hdw); |
1095 | 0 | ccv_nnc_tensor_free(hb); |
1096 | 0 | ccv_nnc_tensor_free(a); |
1097 | 0 | ccv_nnc_tensor_free(w); |
1098 | 0 | ccv_nnc_tensor_free(da); |
1099 | 0 | ccv_nnc_tensor_free(dw); |
1100 | 0 | ccv_nnc_tensor_free(b); |
1101 | 0 | ccv_nnc_tensor_view_free(av); |
1102 | 0 | ccv_nnc_tensor_view_free(wv); |
1103 | 0 | ccv_nnc_tensor_view_free(dav); |
1104 | 0 | ccv_nnc_tensor_view_free(dwv); |
1105 | 0 | ccv_nnc_tensor_free(at); |
1106 | 0 | ccv_nnc_tensor_free(wt); |
1107 | 0 | ccv_nnc_tensor_free(dat); |
1108 | 0 | ccv_nnc_tensor_free(tda); |
1109 | 0 | ccv_nnc_tensor_free(dwt); |
1110 | 0 | ccv_nnc_tensor_free(tdw); |
1111 | 0 | } |
1112 | | |
1113 | | TEST_CASE("generalized batched backward gemm with batch (2, 4) and broadcast compare mps") |
1114 | 1 | { |
1115 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1116 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
1117 | 0 | dsfmt_t dsfmt; |
1118 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1119 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1120 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1121 | 0 | ccv_nnc_tensor_t* hda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1122 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1123 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
1124 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1125 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
1126 | 0 | ccv_nnc_tensor_t* da = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1127 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
1128 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
1129 | |
|
1130 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1131 | 0 | ccv_nnc_tensor_t* dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1132 | 0 | ccv_nnc_tensor_t* tda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1133 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1134 | 0 | int i; |
1135 | 0 | for (i = 0; i < 64 * 128; i++) |
1136 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
1137 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
1138 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1139 | 0 | for (i = 0; i < 2 * 4 * 10 * 64; i++) |
1140 | 0 | hb->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1141 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
1142 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hb), TENSOR_LIST(a, w, b), 0); |
1143 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1144 | 0 | ccv_nnc_tensor_view_t* dav = ccv_nnc_tensor_view_new(da, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1145 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(b, (ccv_nnc_tensor_t*)av, w), TENSOR_LIST((ccv_nnc_tensor_t*)dav, dw), 0); |
1146 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(hb, at, hw), TENSOR_LIST(dat, tdw), 0); |
1147 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(dat), TENSOR_LIST(tda), 0); |
1148 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(da, dw), TENSOR_LIST(hda, hdw), 0); |
1149 | 0 | REQUIRE_TENSOR_EQ(hda, tda, "permute computed output should be the same as non-permute computed ones"); |
1150 | 0 | REQUIRE_TENSOR_EQ(hdw, tdw, "permute computed output should be the same as non-permute computed ones"); |
1151 | 0 | ccv_nnc_tensor_free(ha); |
1152 | 0 | ccv_nnc_tensor_free(hw); |
1153 | 0 | ccv_nnc_tensor_free(hda); |
1154 | 0 | ccv_nnc_tensor_free(hdw); |
1155 | 0 | ccv_nnc_tensor_free(hb); |
1156 | 0 | ccv_nnc_tensor_free(a); |
1157 | 0 | ccv_nnc_tensor_free(w); |
1158 | 0 | ccv_nnc_tensor_free(da); |
1159 | 0 | ccv_nnc_tensor_free(dw); |
1160 | 0 | ccv_nnc_tensor_free(b); |
1161 | 0 | ccv_nnc_tensor_view_free(av); |
1162 | 0 | ccv_nnc_tensor_view_free(dav); |
1163 | 0 | ccv_nnc_tensor_free(at); |
1164 | 0 | ccv_nnc_tensor_free(dat); |
1165 | 0 | ccv_nnc_tensor_free(tda); |
1166 | 0 | ccv_nnc_tensor_free(tdw); |
1167 | 0 | } |
1168 | | |
1169 | | TEST_CASE("generalized batched backward gemm with batch (2, 4) with bias compare mps") |
1170 | 1 | { |
1171 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1172 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
1173 | 0 | dsfmt_t dsfmt; |
1174 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1175 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1176 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
1177 | 0 | ccv_nnc_tensor_t* hda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1178 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
1179 | 0 | ccv_nnc_tensor_t* hdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
1180 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
1181 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1182 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 4, 128), 0); |
1183 | 0 | ccv_nnc_tensor_t* da = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1184 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 64, 4, 128), 0); |
1185 | 0 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
1186 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
1187 | |
|
1188 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1189 | 0 | ccv_nnc_tensor_t* wt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 64, 128), 0); |
1190 | 0 | ccv_nnc_tensor_t* dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1191 | 0 | ccv_nnc_tensor_t* dwt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 64, 128), 0); |
1192 | 0 | ccv_nnc_tensor_t* tda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1193 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 64, 4, 128), 0); |
1194 | 0 | ccv_nnc_tensor_t* tdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
1195 | 0 | int i; |
1196 | 0 | for (i = 0; i < 8 * 64 * 128; i++) |
1197 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
1198 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
1199 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1200 | 0 | for (i = 0; i < 2 * 4 * 10 * 64; i++) |
1201 | 0 | hb->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1202 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
1203 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(hw), TENSOR_LIST(wt), 0); |
1204 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hb), TENSOR_LIST(a, w, b), 0); |
1205 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1206 | 0 | ccv_nnc_tensor_view_t* wv = ccv_nnc_tensor_view_new(w, GPU_TENSOR_NHWC(000, 32F, 2, 4, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(64 * 4 * 128, 128, 4 * 128, 1)); |
1207 | 0 | ccv_nnc_tensor_view_t* dav = ccv_nnc_tensor_view_new(da, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1208 | 0 | ccv_nnc_tensor_view_t* dwv = ccv_nnc_tensor_view_new(dw, GPU_TENSOR_NHWC(000, 32F, 2, 4, 64, 128), ccv_nnc_no_ofs, DIM_ALLOC(64 * 4 * 128, 128, 4 * 128, 1)); |
1209 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST(b, (ccv_nnc_tensor_t*)av, (ccv_nnc_tensor_t*)wv), TENSOR_LIST((ccv_nnc_tensor_t*)dav, (ccv_nnc_tensor_t*)dwv, dbias), 0); |
1210 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(2, 3)), ccv_nnc_no_hint, 0, TENSOR_LIST(hb, at, wt), TENSOR_LIST(dat, dwt, tdbias), 0); |
1211 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(da, dw, dbias), TENSOR_LIST(hda, hdw, hdbias), 0); |
1212 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(dat), TENSOR_LIST(tda), 0); |
1213 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(dwt), TENSOR_LIST(tdw), 0); |
1214 | 0 | REQUIRE_TENSOR_EQ(hda, tda, "permute computed output should be the same as non-permute computed ones"); |
1215 | 0 | REQUIRE_TENSOR_EQ(hdw, tdw, "permute computed output should be the same as non-permute computed ones"); |
1216 | 0 | REQUIRE_TENSOR_EQ(hdbias, tdbias, "permute computed output should be the same as non-permute computed ones"); |
1217 | 0 | ccv_nnc_tensor_free(ha); |
1218 | 0 | ccv_nnc_tensor_free(hw); |
1219 | 0 | ccv_nnc_tensor_free(hda); |
1220 | 0 | ccv_nnc_tensor_free(hdw); |
1221 | 0 | ccv_nnc_tensor_free(hdbias); |
1222 | 0 | ccv_nnc_tensor_free(hb); |
1223 | 0 | ccv_nnc_tensor_free(a); |
1224 | 0 | ccv_nnc_tensor_free(w); |
1225 | 0 | ccv_nnc_tensor_free(da); |
1226 | 0 | ccv_nnc_tensor_free(dw); |
1227 | 0 | ccv_nnc_tensor_free(dbias); |
1228 | 0 | ccv_nnc_tensor_free(b); |
1229 | 0 | ccv_nnc_tensor_view_free(av); |
1230 | 0 | ccv_nnc_tensor_view_free(wv); |
1231 | 0 | ccv_nnc_tensor_view_free(dav); |
1232 | 0 | ccv_nnc_tensor_view_free(dwv); |
1233 | 0 | ccv_nnc_tensor_free(at); |
1234 | 0 | ccv_nnc_tensor_free(wt); |
1235 | 0 | ccv_nnc_tensor_free(dat); |
1236 | 0 | ccv_nnc_tensor_free(dwt); |
1237 | 0 | ccv_nnc_tensor_free(tda); |
1238 | 0 | ccv_nnc_tensor_free(tdw); |
1239 | 0 | ccv_nnc_tensor_free(tdbias); |
1240 | 0 | } |
1241 | | |
1242 | | TEST_CASE("generalized batched backward gemm with batch (2, 4) with bias and broadcast compare mps") |
1243 | 1 | { |
1244 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1245 | | // This is a particular batched gemm which treat every dimensions other than the last two as batching. |
1246 | 0 | dsfmt_t dsfmt; |
1247 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1248 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1249 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1250 | 0 | ccv_nnc_tensor_t* hda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1251 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1252 | 0 | ccv_nnc_tensor_t* hdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
1253 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 64), 0); |
1254 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1255 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
1256 | 0 | ccv_nnc_tensor_t* da = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 10, 4, 128), 0); |
1257 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
1258 | 0 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
1259 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 64), 0); |
1260 | |
|
1261 | 0 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1262 | 0 | ccv_nnc_tensor_t* dat = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 10, 128), 0); |
1263 | 0 | ccv_nnc_tensor_t* tda = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 10, 4, 128), 0); |
1264 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
1265 | 0 | ccv_nnc_tensor_t* tdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
1266 | 0 | int i; |
1267 | 0 | for (i = 0; i < 64 * 128; i++) |
1268 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
1269 | 0 | for (i = 0; i < 8 * 10 * 128; i++) |
1270 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1271 | 0 | for (i = 0; i < 2 * 4 * 10 * 64; i++) |
1272 | 0 | hb->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1273 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(at), 0); |
1274 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hb), TENSOR_LIST(a, w, b), 0); |
1275 | 0 | ccv_nnc_tensor_view_t* av = ccv_nnc_tensor_view_new(a, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1276 | 0 | ccv_nnc_tensor_view_t* dav = ccv_nnc_tensor_view_new(da, GPU_TENSOR_NHWC(000, 32F, 2, 4, 10, 128), ccv_nnc_no_ofs, DIM_ALLOC(10 * 4 * 128, 128, 4 * 128, 1)); |
1277 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(b, (ccv_nnc_tensor_t*)av, w, dbias), TENSOR_LIST((ccv_nnc_tensor_t*)dav, dw, dbias), 0); |
1278 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(hb, at, hw, hdbias), TENSOR_LIST(dat, tdw, tdbias), 0); |
1279 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(da, dw, dbias), TENSOR_LIST(hda, hdw, hdbias), 0); |
1280 | 0 | ccv_nnc_cmd_exec(CMD_TRANSPOSE_FORWARD(1, 2), ccv_nnc_no_hint, 0, TENSOR_LIST(dat), TENSOR_LIST(tda), 0); |
1281 | 0 | REQUIRE_TENSOR_EQ(hda, tda, "permute computed output should be the same as non-permute computed ones"); |
1282 | 0 | REQUIRE_TENSOR_EQ(hdw, tdw, "permute computed output should be the same as non-permute computed ones"); |
1283 | 0 | REQUIRE_TENSOR_EQ(hdbias, tdbias, "permute computed output should be the same as non-permute computed ones"); |
1284 | 0 | ccv_nnc_tensor_free(ha); |
1285 | 0 | ccv_nnc_tensor_free(hw); |
1286 | 0 | ccv_nnc_tensor_free(hda); |
1287 | 0 | ccv_nnc_tensor_free(hdw); |
1288 | 0 | ccv_nnc_tensor_free(hdbias); |
1289 | 0 | ccv_nnc_tensor_free(hb); |
1290 | 0 | ccv_nnc_tensor_free(a); |
1291 | 0 | ccv_nnc_tensor_free(w); |
1292 | 0 | ccv_nnc_tensor_free(da); |
1293 | 0 | ccv_nnc_tensor_free(dw); |
1294 | 0 | ccv_nnc_tensor_free(dbias); |
1295 | 0 | ccv_nnc_tensor_free(b); |
1296 | 0 | ccv_nnc_tensor_view_free(av); |
1297 | 0 | ccv_nnc_tensor_view_free(dav); |
1298 | 0 | ccv_nnc_tensor_free(at); |
1299 | 0 | ccv_nnc_tensor_free(dat); |
1300 | 0 | ccv_nnc_tensor_free(tdw); |
1301 | 0 | ccv_nnc_tensor_free(tdbias); |
1302 | 0 | } |
1303 | | |
1304 | | TEST_CASE("ewdiv forward with reciprocal") |
1305 | 1 | { |
1306 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWDIV_FORWARD, CCV_NNC_BACKEND_MPS)); |
1307 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1308 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1309 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1310 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1311 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1312 | 0 | dsfmt_t dsfmt; |
1313 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1314 | 0 | int i; |
1315 | 0 | for (i = 0; i < 1000; i++) |
1316 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 0.01; |
1317 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1318 | 0 | ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(0, a), TENSOR_LIST(b), 0); |
1319 | 0 | ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(0, ha), TENSOR_LIST(bt), 0); |
1320 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1321 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1322 | 0 | ccv_nnc_tensor_free(a); |
1323 | 0 | ccv_nnc_tensor_free(b); |
1324 | 0 | ccv_nnc_tensor_free(ha); |
1325 | 0 | ccv_nnc_tensor_free(hb); |
1326 | 0 | ccv_nnc_tensor_free(bt); |
1327 | 0 | } |
1328 | | |
1329 | | TEST_CASE("ewdiv forward") |
1330 | 1 | { |
1331 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWDIV_FORWARD, CCV_NNC_BACKEND_MPS)); |
1332 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1333 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1334 | 0 | ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1335 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1336 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1337 | 0 | ccv_nnc_tensor_t* hc = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1338 | 0 | ccv_nnc_tensor_t* ct = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1339 | 0 | dsfmt_t dsfmt; |
1340 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1341 | 0 | int i; |
1342 | 0 | for (i = 0; i < 1000; i++) |
1343 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 0.01; |
1344 | 0 | for (i = 0; i < 1000; i++) |
1345 | 0 | hb->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 0.01; |
1346 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(a, b), 0); |
1347 | 0 | ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a, b), TENSOR_LIST(c), 0); |
1348 | 0 | ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hb), TENSOR_LIST(ct), 0); |
1349 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(c), TENSOR_LIST(hc), 0); |
1350 | 0 | REQUIRE_TENSOR_EQ(ct, hc, "GPU computed output should be the same as CPU computed ones"); |
1351 | 0 | ccv_nnc_tensor_free(a); |
1352 | 0 | ccv_nnc_tensor_free(b); |
1353 | 0 | ccv_nnc_tensor_free(c); |
1354 | 0 | ccv_nnc_tensor_free(ha); |
1355 | 0 | ccv_nnc_tensor_free(hb); |
1356 | 0 | ccv_nnc_tensor_free(hc); |
1357 | 0 | ccv_nnc_tensor_free(ct); |
1358 | 0 | } |
1359 | | |
1360 | | TEST_CASE("exp forward") |
1361 | 1 | { |
1362 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWEXP_FORWARD, CCV_NNC_BACKEND_MPS)); |
1363 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1364 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1365 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1366 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1367 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1368 | 0 | dsfmt_t dsfmt; |
1369 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1370 | 0 | int i; |
1371 | 0 | for (i = 0; i < 1000; i++) |
1372 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 10 - 1; |
1373 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1374 | 0 | ccv_nnc_cmd_exec(CMD_EWEXP_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
1375 | 0 | ccv_nnc_cmd_exec(CMD_EWEXP_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(bt), 0); |
1376 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1377 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1378 | 0 | ccv_nnc_tensor_free(a); |
1379 | 0 | ccv_nnc_tensor_free(b); |
1380 | 0 | ccv_nnc_tensor_free(ha); |
1381 | 0 | ccv_nnc_tensor_free(hb); |
1382 | 0 | ccv_nnc_tensor_free(bt); |
1383 | 0 | } |
1384 | | |
1385 | | TEST_CASE("ewlog forward") |
1386 | 1 | { |
1387 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWLOG_FORWARD, CCV_NNC_BACKEND_MPS)); |
1388 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1389 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1390 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1391 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1392 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1393 | 0 | dsfmt_t dsfmt; |
1394 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1395 | 0 | int i; |
1396 | 0 | for (i = 0; i < 1000; i++) |
1397 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 10 + 0.0001; |
1398 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1399 | 0 | ccv_nnc_cmd_exec(CMD_EWLOG_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
1400 | 0 | ccv_nnc_cmd_exec(CMD_EWLOG_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(bt), 0); |
1401 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1402 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1403 | 0 | ccv_nnc_tensor_free(a); |
1404 | 0 | ccv_nnc_tensor_free(b); |
1405 | 0 | ccv_nnc_tensor_free(ha); |
1406 | 0 | ccv_nnc_tensor_free(hb); |
1407 | 0 | ccv_nnc_tensor_free(bt); |
1408 | 0 | } |
1409 | | |
1410 | | TEST_CASE("ewsqrt forward") |
1411 | 1 | { |
1412 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_EWSQRT_FORWARD, CCV_NNC_BACKEND_MPS)); |
1413 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1414 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1415 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1416 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1417 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1418 | 0 | dsfmt_t dsfmt; |
1419 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1420 | 0 | int i; |
1421 | 0 | for (i = 0; i < 1000; i++) |
1422 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 10 + 0.0001; |
1423 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1424 | 0 | ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
1425 | 0 | ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(bt), 0); |
1426 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1427 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1428 | 0 | ccv_nnc_tensor_free(a); |
1429 | 0 | ccv_nnc_tensor_free(b); |
1430 | 0 | ccv_nnc_tensor_free(ha); |
1431 | 0 | ccv_nnc_tensor_free(hb); |
1432 | 0 | ccv_nnc_tensor_free(bt); |
1433 | 0 | } |
1434 | | |
1435 | | TEST_CASE("clamp forward") |
1436 | 1 | { |
1437 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CLAMP_FORWARD, CCV_NNC_BACKEND_MPS)); |
1438 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1439 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1440 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1441 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1442 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1443 | 0 | dsfmt_t dsfmt; |
1444 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1445 | 0 | int i; |
1446 | 0 | for (i = 0; i < 1000; i++) |
1447 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 10 - 1; |
1448 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1449 | 0 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(0, 6), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
1450 | 0 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(0, 6), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(bt), 0); |
1451 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1452 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1453 | 0 | ccv_nnc_tensor_free(a); |
1454 | 0 | ccv_nnc_tensor_free(b); |
1455 | 0 | ccv_nnc_tensor_free(ha); |
1456 | 0 | ccv_nnc_tensor_free(hb); |
1457 | 0 | ccv_nnc_tensor_free(bt); |
1458 | 0 | } |
1459 | | |
1460 | | TEST_CASE("clamp forward with only max") |
1461 | 1 | { |
1462 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CLAMP_FORWARD, CCV_NNC_BACKEND_MPS)); |
1463 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1464 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1465 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1466 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1467 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1468 | 0 | dsfmt_t dsfmt; |
1469 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1470 | 0 | int i; |
1471 | 0 | for (i = 0; i < 1000; i++) |
1472 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 10 - 1; |
1473 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1474 | 0 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 6), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
1475 | 0 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 6), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(bt), 0); |
1476 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1477 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1478 | 0 | ccv_nnc_tensor_free(a); |
1479 | 0 | ccv_nnc_tensor_free(b); |
1480 | 0 | ccv_nnc_tensor_free(ha); |
1481 | 0 | ccv_nnc_tensor_free(hb); |
1482 | 0 | ccv_nnc_tensor_free(bt); |
1483 | 0 | } |
1484 | | |
1485 | | TEST_CASE("clamp forward with only min") |
1486 | 1 | { |
1487 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_CLAMP_FORWARD, CCV_NNC_BACKEND_MPS)); |
1488 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1489 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NCHW(000, 32F, 10, 100), 0); |
1490 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1491 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1492 | 0 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 10, 100), 0); |
1493 | 0 | dsfmt_t dsfmt; |
1494 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
1495 | 0 | int i; |
1496 | 0 | for (i = 0; i < 1000; i++) |
1497 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 10 - 1; |
1498 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(a), 0); |
1499 | 0 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(0, NAN), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(b), 0); |
1500 | 0 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(0, NAN), ccv_nnc_no_hint, 0, TENSOR_LIST(ha), TENSOR_LIST(bt), 0); |
1501 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b), TENSOR_LIST(hb), 0); |
1502 | 0 | REQUIRE_TENSOR_EQ(bt, hb, "GPU computed output should be the same as CPU computed ones"); |
1503 | 0 | ccv_nnc_tensor_free(a); |
1504 | 0 | ccv_nnc_tensor_free(b); |
1505 | 0 | ccv_nnc_tensor_free(ha); |
1506 | 0 | ccv_nnc_tensor_free(hb); |
1507 | 0 | ccv_nnc_tensor_free(bt); |
1508 | 0 | } |
1509 | | |
1510 | | TEST_CASE("compare set with mps") |
1511 | 1 | { |
1512 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SET_FORWARD, CCV_NNC_BACKEND_MPS)); |
1513 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 11, 10, 9, 8), 0); |
1514 | 0 | ccv_nnc_tensor_t* const ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 11, 10, 9, 8), 0); |
1515 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 11, 10, 9, 8), 0); |
1516 | 0 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(10), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(a), 0); |
1517 | 0 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(10), ccv_nnc_no_hint, 0, TENSOR_LIST(), TENSOR_LIST(ga), 0); |
1518 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a), TENSOR_LIST(ha), 0); |
1519 | 0 | REQUIRE_TENSOR_EQ(ha, ga, "format transform result should be the same"); |
1520 | 0 | ccv_nnc_tensor_free(a); |
1521 | 0 | ccv_nnc_tensor_free(ha); |
1522 | 0 | ccv_nnc_tensor_free(ga); |
1523 | 0 | } |
1524 | | |
1525 | | TEST_CASE("scaled dot product attention with mps") |
1526 | 1 | { |
1527 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD, CCV_NNC_BACKEND_MPS)); |
1528 | | // Bypass error: variable-sized object may not be initialized |
1529 | 0 | #define num_long_trials 3 |
1530 | 0 | #define num_short_trials 2 |
1531 | 0 | #define num_trials (num_long_trials + num_short_trials) |
1532 | |
|
1533 | 0 | for (int trial = 0; trial < num_trials; ++trial) { |
1534 | 0 | int B_candidates[num_trials] = { 32, 32, 3, 2, 1 }; |
1535 | 0 | int R_candidates[num_trials] = { 128, 128, 61, 6, 2 }; |
1536 | 0 | int C_candidates[num_trials] = { 128, 128, 49, 2, 1 }; |
1537 | 0 | int Hq_candidates[num_trials] = { 8, 32, 13, 3, 1 }; |
1538 | 0 | int Hk_candidates[num_trials] = { 8, 8, 13, 3, 1 }; |
1539 | 0 | int D_candidates[num_trials] = { 64, 128, 191, 4, 8 }; |
1540 | 0 | int is_causal_candidates[num_trials] = { 0, 1, 0, 1, 0 }; |
1541 | |
|
1542 | 0 | int B = B_candidates[trial]; |
1543 | 0 | int R = R_candidates[trial]; |
1544 | 0 | int C = C_candidates[trial]; |
1545 | 0 | int Hq = Hq_candidates[trial]; |
1546 | 0 | int Hk = Hk_candidates[trial]; |
1547 | 0 | int D = D_candidates[trial]; |
1548 | 0 | int is_causal = is_causal_candidates[trial]; |
1549 | 0 | float scale = 1.0 / sqrt((float)D); |
1550 | |
|
1551 | 0 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD, CCV_NNC_BACKEND_MPS)); |
1552 | 0 | ccv_nnc_tensor_t* const q_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, Hq, D), 0); |
1553 | 0 | ccv_nnc_tensor_t* const k_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, Hk, D), 0); |
1554 | 0 | ccv_nnc_tensor_t* const v_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, Hk, D), 0); |
1555 | |
|
1556 | 0 | for (int i = 0; i < B * R * Hq * D; ++i) { |
1557 | 0 | q_tensor->data.f32[i] = (float)(i) / (float)(B * R * Hq * D); |
1558 | 0 | } |
1559 | 0 | for (int i = 0; i < B * C * Hk * D; ++i) { |
1560 | 0 | k_tensor->data.f32[i] = (float)(i) / (float)(B * C * Hk * D); |
1561 | 0 | } |
1562 | 0 | for (int i = 0; i < B * C * Hk * D; ++i) { |
1563 | 0 | v_tensor->data.f32[i] = (float)(i) / (float)(B * C * Hk * D); |
1564 | 0 | } |
1565 | |
|
1566 | 0 | ccv_nnc_tensor_t* const o_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, Hq, D), 0); |
1567 | 0 | ccv_nnc_cmd_exec(CMD_SCALED_DOT_PRODUCT_ATTENTION_FORWARD(scale, is_causal), ccv_nnc_no_hint, 0, TENSOR_LIST(q_tensor, k_tensor, v_tensor), TENSOR_LIST(o_tensor), 0); |
1568 | | |
1569 | | // Why it there 000 in the beginning of the argument list for GPU_TENSOR_NHWC? |
1570 | 0 | ccv_nnc_tensor_t* const gpu_q_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, R, Hq, D), 0); |
1571 | 0 | ccv_nnc_tensor_t* const gpu_k_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, C, Hk, D), 0); |
1572 | 0 | ccv_nnc_tensor_t* const gpu_v_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, C, Hk, D), 0); |
1573 | 0 | ccv_nnc_tensor_t* const gpu_o_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, R, Hq, D), 0); |
1574 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(q_tensor, k_tensor, v_tensor), TENSOR_LIST(gpu_q_tensor, gpu_k_tensor, gpu_v_tensor), 0); |
1575 | |
|
1576 | 0 | if (is_causal) |
1577 | 0 | { |
1578 | 0 | ccv_nnc_tensor_t* const causal_mask = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1, 1, R, C), 0); |
1579 | 0 | ccv_nnc_tensor_t* const gpu_causal_mask = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 1, 1, R, C), 0); |
1580 | 0 | for (int i = 0; i < R; i++) |
1581 | 0 | for (int j = 0; j < C; j++) |
1582 | 0 | causal_mask->data.f32[i * C + j] = 0; |
1583 | 0 | for (int i = 0; i < R - 1; i++) |
1584 | 0 | for (int j = i - R + C + 1; j < C; j++) |
1585 | 0 | causal_mask->data.f32[i * C + j] = -FLT_MAX; |
1586 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(causal_mask), TENSOR_LIST(gpu_causal_mask), 0); |
1587 | 0 | ccv_nnc_cmd_exec(CMD_SCALED_DOT_PRODUCT_ATTENTION_FORWARD(scale, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(gpu_q_tensor, gpu_k_tensor, gpu_v_tensor, gpu_causal_mask), TENSOR_LIST(gpu_o_tensor), 0); |
1588 | 0 | ccv_nnc_tensor_free(gpu_causal_mask); |
1589 | 0 | ccv_nnc_tensor_free(causal_mask); |
1590 | 0 | } else { |
1591 | 0 | ccv_nnc_cmd_exec(CMD_SCALED_DOT_PRODUCT_ATTENTION_FORWARD(scale, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(gpu_q_tensor, gpu_k_tensor, gpu_v_tensor), TENSOR_LIST(gpu_o_tensor), 0); |
1592 | 0 | } |
1593 | |
|
1594 | 0 | ccv_nnc_tensor_t* const copy_of_gpu_o_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, Hq, D), 0); |
1595 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gpu_o_tensor), TENSOR_LIST(copy_of_gpu_o_tensor), 0); |
1596 | |
|
1597 | 0 | REQUIRE_TENSOR_EQ(copy_of_gpu_o_tensor, o_tensor, "scaled dot product attention result should be the same"); |
1598 | |
|
1599 | 0 | ccv_nnc_tensor_free(o_tensor); |
1600 | 0 | ccv_nnc_tensor_free(gpu_o_tensor); |
1601 | 0 | ccv_nnc_tensor_free(copy_of_gpu_o_tensor); |
1602 | 0 | ccv_nnc_tensor_free(q_tensor); |
1603 | 0 | ccv_nnc_tensor_free(k_tensor); |
1604 | 0 | ccv_nnc_tensor_free(v_tensor); |
1605 | 0 | ccv_nnc_tensor_free(gpu_q_tensor); |
1606 | 0 | ccv_nnc_tensor_free(gpu_k_tensor); |
1607 | 0 | ccv_nnc_tensor_free(gpu_v_tensor); |
1608 | 0 | } |
1609 | 0 | #undef num_long_trials |
1610 | 0 | #undef num_short_trials |
1611 | 0 | #undef num_trials |
1612 | 0 | } |
1613 | | |
1614 | | TEST_CASE("scaled dot product attention + unify head with mps") |
1615 | 1 | { |
1616 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD, CCV_NNC_BACKEND_MPS)); |
1617 | 0 | ccv_nnc_symbolic_graph_t* const sdp_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1618 | 0 | ccv_nnc_tensor_symbol_t q = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 32, 128, 8, 64), "q"); |
1619 | 0 | ccv_nnc_tensor_symbol_t k = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 32, 128, 8, 64), "k"); |
1620 | 0 | ccv_nnc_tensor_symbol_t v = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 32, 128, 8, 64), "v"); |
1621 | 0 | ccv_nnc_tensor_symbol_t w = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 512, 512), "w"); |
1622 | 0 | ccv_nnc_tensor_symbol_t bias = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 512), "bias"); |
1623 | 0 | ccv_nnc_tensor_symbol_t c = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 32, 128, 8, 64), "c"); |
1624 | 0 | ccv_nnc_tensor_symbol_t r = ccv_nnc_tensor_symbol_new(sdp_symbolic_graph, CPU_TENSOR_NHWC(32F, 32, 128, 512), "r"); |
1625 | 0 | ccv_nnc_graph_exec_symbol_new(sdp_symbolic_graph, CMD_SCALED_DOT_PRODUCT_ATTENTION_FORWARD(1.0 / 8, 0), TENSOR_SYMBOL_LIST(q, k, v, NO_TENSOR_SYMBOL, w, bias), TENSOR_SYMBOL_LIST(r, NO_TENSOR_SYMBOL, c), "scaled_dot_product_attention"); |
1626 | 0 | ccv_nnc_graph_exec_symbol_autogen(sdp_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1627 | 0 | ccv_nnc_graph_t* sdp_graph = 0; |
1628 | 0 | ccv_nnc_tensor_arena_t* sdp_tensor_arena = 0; |
1629 | 0 | ccv_nnc_graph_exec_arena_t* sdp_graph_exec_arena = 0; |
1630 | 0 | ccv_nnc_symbolic_graph_compile(sdp_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(sdp_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(sdp_symbolic_graph), &sdp_graph, &sdp_tensor_arena, &sdp_graph_exec_arena); |
1631 | 0 | ccv_nnc_tensor_t* const q_tensor = ccv_nnc_tensor_from_symbol(sdp_tensor_arena, q); |
1632 | 0 | ccv_nnc_tensor_t* const k_tensor = ccv_nnc_tensor_from_symbol(sdp_tensor_arena, k); |
1633 | 0 | ccv_nnc_tensor_t* const v_tensor = ccv_nnc_tensor_from_symbol(sdp_tensor_arena, v); |
1634 | 0 | ccv_nnc_tensor_t* const w_tensor = ccv_nnc_tensor_from_symbol(sdp_tensor_arena, w); |
1635 | 0 | ccv_nnc_tensor_t* const bias_tensor = ccv_nnc_tensor_from_symbol(sdp_tensor_arena, bias); |
1636 | 0 | dsfmt_t dsfmt; |
1637 | 0 | int i; |
1638 | 0 | dsfmt_init_gen_rand(&dsfmt, 1); |
1639 | 0 | for (i = 0; i < 32 * 8 * 128 * 64; i++) |
1640 | 0 | q_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1641 | 0 | for (i = 0; i < 32 * 8 * 128 * 64; i++) |
1642 | 0 | k_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1643 | 0 | for (i = 0; i < 32 * 8 * 128 * 64; i++) |
1644 | 0 | v_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1645 | 0 | for (i = 0; i < 512 * 512; i++) |
1646 | 0 | w_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1647 | 0 | for (i = 0; i < 512; i++) |
1648 | 0 | bias_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1649 | 0 | ccv_nnc_symbolic_graph_t* const g_symbolic_graph = ccv_nnc_symbolic_graph_new(); |
1650 | 0 | ccv_nnc_tensor_symbol_t gq = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 32, 128, 8, 64), "q"); |
1651 | 0 | ccv_nnc_tensor_symbol_t gk = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 32, 128, 8, 64), "k"); |
1652 | 0 | ccv_nnc_tensor_symbol_t gv = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 32, 128, 8, 64), "v"); |
1653 | 0 | ccv_nnc_tensor_symbol_t gw = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 512, 512), "w"); |
1654 | 0 | ccv_nnc_tensor_symbol_t gbias = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 512), "bias"); |
1655 | 0 | ccv_nnc_tensor_symbol_t gc = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 32, 128, 8, 64), "c"); |
1656 | 0 | ccv_nnc_tensor_symbol_t gr = ccv_nnc_tensor_symbol_new(g_symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 32, 128, 512), "r"); |
1657 | 0 | ccv_nnc_graph_exec_symbol_new(g_symbolic_graph, CMD_SCALED_DOT_PRODUCT_ATTENTION_FORWARD(1.0 / 8, 0), TENSOR_SYMBOL_LIST(gq, gk, gv, NO_TENSOR_SYMBOL, gw, gbias), TENSOR_SYMBOL_LIST(gr, NO_TENSOR_SYMBOL, gc), "scaled_dot_product_attention"); |
1658 | 0 | ccv_nnc_graph_exec_symbol_autogen(g_symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
1659 | 0 | ccv_nnc_graph_t* g_graph = 0; |
1660 | 0 | ccv_nnc_tensor_arena_t* g_tensor_arena = 0; |
1661 | 0 | ccv_nnc_graph_exec_arena_t* g_graph_exec_arena = 0; |
1662 | 0 | ccv_nnc_symbolic_graph_compile(g_symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(g_symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(g_symbolic_graph), &g_graph, &g_tensor_arena, &g_graph_exec_arena); |
1663 | 0 | ccv_nnc_tensor_t* const gq_tensor = ccv_nnc_tensor_from_symbol(g_tensor_arena, gq); |
1664 | 0 | ccv_nnc_tensor_t* const gk_tensor = ccv_nnc_tensor_from_symbol(g_tensor_arena, gk); |
1665 | 0 | ccv_nnc_tensor_t* const gv_tensor = ccv_nnc_tensor_from_symbol(g_tensor_arena, gv); |
1666 | 0 | ccv_nnc_tensor_t* const gw_tensor = ccv_nnc_tensor_from_symbol(g_tensor_arena, gw); |
1667 | 0 | ccv_nnc_tensor_t* const gbias_tensor = ccv_nnc_tensor_from_symbol(g_tensor_arena, gbias); |
1668 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(q_tensor, k_tensor, v_tensor, w_tensor, bias_tensor), TENSOR_LIST(gq_tensor, gk_tensor, gv_tensor, gw_tensor, gbias_tensor), 0); |
1669 | 0 | ccv_nnc_graph_run(sdp_graph, 0, TRAVERSE_FULL, 0, 0); |
1670 | 0 | ccv_nnc_graph_run(g_graph, 0, TRAVERSE_FULL, 0, 0); |
1671 | 0 | ccv_nnc_tensor_t* const r_tensor = ccv_nnc_tensor_from_symbol(sdp_tensor_arena, r); |
1672 | 0 | ccv_nnc_tensor_t* const gr_tensor = ccv_nnc_tensor_from_symbol(g_tensor_arena, gr); |
1673 | 0 | ccv_nnc_tensor_t* const hr = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 32, 128, 512), 0); |
1674 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gr_tensor), TENSOR_LIST(hr), 0); |
1675 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, r_tensor->data.f32, hr->data.f32, 32 * 128 * 512, 1e-4, "graph computed result should match scaled dot product attention op result"); |
1676 | 0 | ccv_nnc_symbolic_graph_free(sdp_symbolic_graph); |
1677 | 0 | ccv_nnc_tensor_arena_free(sdp_tensor_arena); |
1678 | 0 | ccv_nnc_graph_exec_arena_free(sdp_graph_exec_arena); |
1679 | 0 | ccv_nnc_graph_free(sdp_graph); |
1680 | 0 | ccv_nnc_symbolic_graph_free(g_symbolic_graph); |
1681 | 0 | ccv_nnc_tensor_arena_free(g_tensor_arena); |
1682 | 0 | ccv_nnc_graph_exec_arena_free(g_graph_exec_arena); |
1683 | 0 | ccv_nnc_graph_free(g_graph); |
1684 | 0 | ccv_nnc_tensor_free(hr); |
1685 | 0 | } |
1686 | | |
1687 | | TEST_CASE("scaled dot product attention gradient with mps") |
1688 | 1 | { |
1689 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD, CCV_NNC_BACKEND_MPS) && |
1690 | 1 | ccv_nnc_cmd_ok(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1691 | 0 | #define num_long_trials 2 |
1692 | 0 | #define num_short_trials 2 |
1693 | 0 | #define num_trials (num_long_trials + num_short_trials) |
1694 | |
|
1695 | 0 | dsfmt_t dsfmt; |
1696 | 0 | dsfmt_init_gen_rand(&dsfmt, 10); |
1697 | 0 | for (int trial = 0; trial < num_trials; ++trial) { |
1698 | 0 | int B_candidates[num_trials] = { 32, 3, 2, 1 }; |
1699 | 0 | int R_candidates[num_trials] = { 128, 61, 6, 2 }; |
1700 | 0 | int C_candidates[num_trials] = { 128, 49, 2, 1 }; |
1701 | 0 | int H_candidates[num_trials] = { 8, 13, 3, 1 }; |
1702 | 0 | int D_candidates[num_trials] = { 64, 191, 4, 8 }; |
1703 | |
|
1704 | 0 | int B = B_candidates[trial]; |
1705 | 0 | int R = R_candidates[trial]; |
1706 | 0 | int C = C_candidates[trial]; |
1707 | 0 | int H = H_candidates[trial]; |
1708 | 0 | int D = D_candidates[trial]; |
1709 | 0 | float scale = 1.0 / sqrt((float)D); |
1710 | |
|
1711 | 0 | ccv_nnc_tensor_t* const q_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, H, D), 0); |
1712 | 0 | ccv_nnc_tensor_t* const k_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, H, D), 0); |
1713 | 0 | ccv_nnc_tensor_t* const v_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, H, D), 0); |
1714 | 0 | ccv_nnc_tensor_t* const dq_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, H, D), 0); |
1715 | 0 | ccv_nnc_tensor_t* const dk_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, H, D), 0); |
1716 | 0 | ccv_nnc_tensor_t* const dv_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, H, D), 0); |
1717 | |
|
1718 | 0 | for (int i = 0; i < B * R * H * D; ++i) { |
1719 | 0 | q_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1720 | 0 | } |
1721 | 0 | for (int i = 0; i < B * C * H * D; ++i) { |
1722 | 0 | k_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1723 | 0 | } |
1724 | 0 | for (int i = 0; i < B * C * H * D; ++i) { |
1725 | 0 | v_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1726 | 0 | } |
1727 | |
|
1728 | 0 | ccv_nnc_tensor_t* const o_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, H, D), 0); |
1729 | 0 | for (int i = 0; i < B * R * H * D; ++i) { |
1730 | 0 | o_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
1731 | 0 | } |
1732 | 0 | ccv_nnc_cmd_exec(CMD_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD(scale, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(o_tensor, 0, 0, q_tensor, k_tensor, v_tensor), TENSOR_LIST(dq_tensor, dk_tensor, dv_tensor), 0); |
1733 | |
|
1734 | 0 | ccv_nnc_tensor_t* const gpu_q_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, R, H, D), 0); |
1735 | 0 | ccv_nnc_tensor_t* const gpu_k_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, C, H, D), 0); |
1736 | 0 | ccv_nnc_tensor_t* const gpu_v_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, C, H, D), 0); |
1737 | 0 | ccv_nnc_tensor_t* const gpu_o_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, R, H, D), 0); |
1738 | 0 | ccv_nnc_tensor_t* const gpu_dq_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, R, H, D), 0); |
1739 | 0 | ccv_nnc_tensor_t* const gpu_dk_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, C, H, D), 0); |
1740 | 0 | ccv_nnc_tensor_t* const gpu_dv_tensor = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, B, C, H, D), 0); |
1741 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(q_tensor, k_tensor, v_tensor, o_tensor), TENSOR_LIST(gpu_q_tensor, gpu_k_tensor, gpu_v_tensor, gpu_o_tensor), 0); |
1742 | |
|
1743 | 0 | ccv_nnc_cmd_exec(CMD_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD(scale, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(gpu_o_tensor, 0, 0, gpu_q_tensor, gpu_k_tensor, gpu_v_tensor), TENSOR_LIST(gpu_dq_tensor, gpu_dk_tensor, gpu_dv_tensor), 0); |
1744 | |
|
1745 | 0 | ccv_nnc_tensor_t* const copy_of_gpu_dq_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, R, H, D), 0); |
1746 | 0 | ccv_nnc_tensor_t* const copy_of_gpu_dk_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, H, D), 0); |
1747 | 0 | ccv_nnc_tensor_t* const copy_of_gpu_dv_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, B, C, H, D), 0); |
1748 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gpu_dq_tensor, gpu_dk_tensor, gpu_dv_tensor), TENSOR_LIST(copy_of_gpu_dq_tensor, copy_of_gpu_dk_tensor, copy_of_gpu_dv_tensor), 0); |
1749 | |
|
1750 | 0 | REQUIRE_TENSOR_EQ(copy_of_gpu_dv_tensor, dv_tensor, "scaled dot product attention result should be the same"); |
1751 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, copy_of_gpu_dq_tensor->data.f32, dq_tensor->data.f32, B * R * H * D, 1e-6, "scaled dot product attention result should be the same"); |
1752 | 0 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, copy_of_gpu_dk_tensor->data.f32, dk_tensor->data.f32, B * C * H * D, 1e-6, "scaled dot product attention result should be the same"); |
1753 | |
|
1754 | 0 | ccv_nnc_tensor_free(o_tensor); |
1755 | 0 | ccv_nnc_tensor_free(gpu_o_tensor); |
1756 | 0 | ccv_nnc_tensor_free(copy_of_gpu_dq_tensor); |
1757 | 0 | ccv_nnc_tensor_free(copy_of_gpu_dk_tensor); |
1758 | 0 | ccv_nnc_tensor_free(copy_of_gpu_dv_tensor); |
1759 | 0 | ccv_nnc_tensor_free(q_tensor); |
1760 | 0 | ccv_nnc_tensor_free(k_tensor); |
1761 | 0 | ccv_nnc_tensor_free(v_tensor); |
1762 | 0 | ccv_nnc_tensor_free(gpu_q_tensor); |
1763 | 0 | ccv_nnc_tensor_free(gpu_k_tensor); |
1764 | 0 | ccv_nnc_tensor_free(gpu_v_tensor); |
1765 | 0 | ccv_nnc_tensor_free(dq_tensor); |
1766 | 0 | ccv_nnc_tensor_free(dk_tensor); |
1767 | 0 | ccv_nnc_tensor_free(dv_tensor); |
1768 | 0 | ccv_nnc_tensor_free(gpu_dq_tensor); |
1769 | 0 | ccv_nnc_tensor_free(gpu_dk_tensor); |
1770 | 0 | ccv_nnc_tensor_free(gpu_dv_tensor); |
1771 | 0 | } |
1772 | 0 | #undef num_long_trials |
1773 | 0 | #undef num_short_trials |
1774 | 0 | #undef num_trials |
1775 | 0 | } |
1776 | | |
1777 | | TEST_CASE("backward gemm with no transpose") |
1778 | 1 | { |
1779 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
1780 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1781 | 0 | float gp[] = { |
1782 | 0 | 1, 2, 3, |
1783 | 0 | 4, 5, 6, |
1784 | 0 | 7, 8, 9, |
1785 | 0 | 10, 11, 12, |
1786 | 0 | }; |
1787 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
1788 | |
|
1789 | 0 | float ap[] = { |
1790 | 0 | 13, 14, |
1791 | 0 | 15, 16, |
1792 | 0 | 17, 18, |
1793 | 0 | 19, 20, |
1794 | 0 | }; |
1795 | |
|
1796 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
1797 | |
|
1798 | 0 | float bp[] = { |
1799 | 0 | 21, 22, 23, |
1800 | 0 | 24, 25, 26, |
1801 | 0 | }; |
1802 | |
|
1803 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
1804 | |
|
1805 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
1806 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
1807 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
1808 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
1809 | |
|
1810 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
1811 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
1812 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
1813 | 0 | ccv_nnc_cmd_t cmd = CMD_GEMM_BACKWARD(); |
1814 | 0 | cmd.backend = CCV_NNC_BACKEND_MPS; |
1815 | 0 | cmd.algorithm = 1; // This is cblas. |
1816 | |
|
1817 | 0 | ccv_nnc_cmd_exec(cmd, ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(h, db, dbias), 0); |
1818 | |
|
1819 | 0 | ccv_nnc_tensor_t* const ch = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC( 32F, 4, 2), 0); |
1820 | 0 | ccv_nnc_tensor_t* const cdb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC( 32F, 2, 3), 0); |
1821 | 0 | ccv_nnc_tensor_t* const cdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC( 32F, 3), 0); |
1822 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(h, db, dbias), TENSOR_LIST(ch, cdb, cdbias), 0); |
1823 | |
|
1824 | 0 | float dbiastp[] = { |
1825 | 0 | 22, 26, 30, |
1826 | 0 | }; |
1827 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
1828 | |
|
1829 | 0 | REQUIRE_TENSOR_EQ(cdbias, &dbiast, "bias should be equal"); |
1830 | 0 | float htp[] = { |
1831 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
1832 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
1833 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
1834 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
1835 | 0 | }; |
1836 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
1837 | |
|
1838 | 0 | REQUIRE_TENSOR_EQ(ch, &ht, "h should be equal"); |
1839 | 0 | float dbtp[] = { |
1840 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, |
1841 | 0 | 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
1842 | 0 | }; |
1843 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
1844 | 0 | REQUIRE_TENSOR_EQ(cdb, &dbt, "db should be equal"); |
1845 | 0 | ccv_nnc_tensor_free(g); |
1846 | 0 | ccv_nnc_tensor_free(a); |
1847 | 0 | ccv_nnc_tensor_free(b); |
1848 | 0 | ccv_nnc_tensor_free(h); |
1849 | 0 | ccv_nnc_tensor_free(db); |
1850 | 0 | ccv_nnc_tensor_free(dbias); |
1851 | 0 | } |
1852 | | |
1853 | | TEST_CASE("backward gemm with transpose a") |
1854 | 1 | { |
1855 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
1856 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1857 | 0 | float gp[] = { |
1858 | 0 | 1, 2, 3, |
1859 | 0 | 4, 5, 6, |
1860 | 0 | 7, 8, 9, |
1861 | 0 | 10, 11, 12, |
1862 | 0 | }; |
1863 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
1864 | 0 | float ap[] = { |
1865 | 0 | 13, 15, 17, 19, |
1866 | 0 | 14, 16, 18, 20, |
1867 | 0 | }; |
1868 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
1869 | 0 | float bp[] = { |
1870 | 0 | 21, 22, 23, |
1871 | 0 | 24, 25, 26, |
1872 | 0 | }; |
1873 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
1874 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
1875 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
1876 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
1877 | 0 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
1878 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4), 0); |
1879 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
1880 | 0 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4), 0); |
1881 | 0 | ccv_nnc_tensor_t* gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
1882 | 0 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
1883 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
1884 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
1885 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
1886 | 0 | float dbiastp[] = { |
1887 | 0 | 22, 26, 30, |
1888 | 0 | }; |
1889 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
1890 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
1891 | 0 | float htp[] = { |
1892 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 4 * 21 + 5 * 22 + 6 * 23, 7 * 21 + 8 * 22 + 9 * 23, 10 * 21 + 11 * 22 + 12 * 23, |
1893 | 0 | 1 * 24 + 2 * 25 + 3 * 26, 4 * 24 + 5 * 25 + 6 * 26, 7 * 24 + 8 * 25 + 9 * 26, 10 * 24 + 11 * 25 + 12 * 26, |
1894 | 0 | }; |
1895 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
1896 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
1897 | 0 | float dbtp[] = { |
1898 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, |
1899 | 0 | 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
1900 | 0 | }; |
1901 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
1902 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
1903 | 0 | ccv_nnc_tensor_free(g); |
1904 | 0 | ccv_nnc_tensor_free(a); |
1905 | 0 | ccv_nnc_tensor_free(b); |
1906 | 0 | ccv_nnc_tensor_free(h); |
1907 | 0 | ccv_nnc_tensor_free(db); |
1908 | 0 | ccv_nnc_tensor_free(dbias); |
1909 | 0 | ccv_nnc_tensor_free(gg); |
1910 | 0 | ccv_nnc_tensor_free(ga); |
1911 | 0 | ccv_nnc_tensor_free(gb); |
1912 | 0 | ccv_nnc_tensor_free(gh); |
1913 | 0 | ccv_nnc_tensor_free(gdb); |
1914 | 0 | ccv_nnc_tensor_free(gdbias); |
1915 | 0 | } |
1916 | | |
1917 | | TEST_CASE("backward gemm with transpose b") |
1918 | 1 | { |
1919 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
1920 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1921 | 0 | float gp[] = { |
1922 | 0 | 1, 2, 3, |
1923 | 0 | 4, 5, 6, |
1924 | 0 | 7, 8, 9, |
1925 | 0 | 10, 11, 12, |
1926 | 0 | }; |
1927 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
1928 | 0 | float ap[] = { |
1929 | 0 | 13, 14, |
1930 | 0 | 15, 16, |
1931 | 0 | 17, 18, |
1932 | 0 | 19, 20, |
1933 | 0 | }; |
1934 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
1935 | 0 | float bp[] = { |
1936 | 0 | 21, 24, |
1937 | 0 | 22, 25, |
1938 | 0 | 23, 26, |
1939 | 0 | }; |
1940 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
1941 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
1942 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
1943 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
1944 | 0 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
1945 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
1946 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
1947 | 0 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 2), 0); |
1948 | 0 | ccv_nnc_tensor_t* gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
1949 | 0 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
1950 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
1951 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
1952 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
1953 | 0 | float dbiastp[] = { |
1954 | 0 | 22, 26, 30, |
1955 | 0 | }; |
1956 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
1957 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
1958 | 0 | float htp[] = { |
1959 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
1960 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
1961 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
1962 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
1963 | 0 | }; |
1964 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 4, 2), 0); |
1965 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
1966 | 0 | float dbtp[] = { |
1967 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, |
1968 | 0 | 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, |
1969 | 0 | 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
1970 | 0 | }; |
1971 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
1972 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
1973 | 0 | ccv_nnc_tensor_free(g); |
1974 | 0 | ccv_nnc_tensor_free(a); |
1975 | 0 | ccv_nnc_tensor_free(b); |
1976 | 0 | ccv_nnc_tensor_free(h); |
1977 | 0 | ccv_nnc_tensor_free(db); |
1978 | 0 | ccv_nnc_tensor_free(dbias); |
1979 | 0 | ccv_nnc_tensor_free(gg); |
1980 | 0 | ccv_nnc_tensor_free(ga); |
1981 | 0 | ccv_nnc_tensor_free(gb); |
1982 | 0 | ccv_nnc_tensor_free(gh); |
1983 | 0 | ccv_nnc_tensor_free(gdb); |
1984 | 0 | ccv_nnc_tensor_free(gdbias); |
1985 | 0 | } |
1986 | | |
1987 | | TEST_CASE("backward gemm with transpose a and b") |
1988 | 1 | { |
1989 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
1990 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
1991 | 0 | float gp[] = { |
1992 | 0 | 1, 2, 3, |
1993 | 0 | 4, 5, 6, |
1994 | 0 | 7, 8, 9, |
1995 | 0 | 10, 11, 12, |
1996 | 0 | }; |
1997 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 4, 3), 0); |
1998 | 0 | float ap[] = { |
1999 | 0 | 13, 15, 17, 19, |
2000 | 0 | 14, 16, 18, 20, |
2001 | 0 | }; |
2002 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
2003 | 0 | float bp[] = { |
2004 | 0 | 21, 24, |
2005 | 0 | 22, 25, |
2006 | 0 | 23, 26, |
2007 | 0 | }; |
2008 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
2009 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
2010 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
2011 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
2012 | 0 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4, 3), 0); |
2013 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4), 0); |
2014 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
2015 | 0 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4), 0); |
2016 | 0 | ccv_nnc_tensor_t* gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
2017 | 0 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
2018 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2019 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(TRANSPOSE(0, 1), TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
2020 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
2021 | 0 | float dbiastp[] = { |
2022 | 0 | 22, 26, 30, |
2023 | 0 | }; |
2024 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
2025 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
2026 | 0 | float htp[] = { |
2027 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 4 * 21 + 5 * 22 + 6 * 23, 7 * 21 + 8 * 22 + 9 * 23, 10 * 21 + 11 * 22 + 12 * 23, |
2028 | 0 | 1 * 24 + 2 * 25 + 3 * 26, 4 * 24 + 5 * 25 + 6 * 26, 7 * 24 + 8 * 25 + 9 * 26, 10 * 24 + 11 * 25 + 12 * 26, |
2029 | 0 | }; |
2030 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4), 0); |
2031 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2032 | 0 | float dbtp[] = { |
2033 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, |
2034 | 0 | 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, |
2035 | 0 | 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
2036 | 0 | }; |
2037 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
2038 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2039 | 0 | ccv_nnc_tensor_free(g); |
2040 | 0 | ccv_nnc_tensor_free(a); |
2041 | 0 | ccv_nnc_tensor_free(b); |
2042 | 0 | ccv_nnc_tensor_free(h); |
2043 | 0 | ccv_nnc_tensor_free(db); |
2044 | 0 | ccv_nnc_tensor_free(dbias); |
2045 | 0 | ccv_nnc_tensor_free(gg); |
2046 | 0 | ccv_nnc_tensor_free(ga); |
2047 | 0 | ccv_nnc_tensor_free(gb); |
2048 | 0 | ccv_nnc_tensor_free(gh); |
2049 | 0 | ccv_nnc_tensor_free(gdb); |
2050 | 0 | ccv_nnc_tensor_free(gdbias); |
2051 | 0 | } |
2052 | | |
2053 | | |
2054 | | TEST_CASE("backward gemm large data set") |
2055 | 1 | { |
2056 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2057 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2058 | 0 | dsfmt_t dsfmt; |
2059 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
2060 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2061 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2062 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
2063 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2064 | 0 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2065 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2066 | 0 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
2067 | 0 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2068 | |
|
2069 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2070 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2071 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2072 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2073 | 0 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2074 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2075 | 0 | ccv_nnc_tensor_t* hdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2076 | 0 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2077 | 0 | int i; |
2078 | 0 | for (i = 0; i < 64 * 128; i++) |
2079 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
2080 | 0 | for (i = 0; i < 64; i++) |
2081 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2082 | 0 | for (i = 0; i < 10 * 128; i++) |
2083 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2084 | 0 | for (i = 0; i < 10 * 64; i++) |
2085 | 0 | hg->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2086 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias, hg), TENSOR_LIST(a, w, bias, g), 0); |
2087 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(hb), 0); |
2088 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, ha, hw, 0), TENSOR_LIST(hh, hdw, hdbias), 0); |
2089 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
2090 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, w, 0), TENSOR_LIST(h, dw, dbias), 0); |
2091 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2092 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2093 | 0 | ccv_nnc_tensor_t* tdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2094 | 0 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2095 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, dw, dbias, h), TENSOR_LIST(tb, tdw, tdbias, th), 0); |
2096 | 0 | REQUIRE_TENSOR_EQ(tb, hb, "GPU computed output should be the same as CPU computed ones"); |
2097 | 0 | REQUIRE_TENSOR_EQ(tdw, hdw, "GPU computed output should be the same as CPU computed ones"); |
2098 | 0 | REQUIRE_TENSOR_EQ(tdbias, hdbias, "GPU computed output should be the same as CPU computed ones"); |
2099 | 0 | REQUIRE_TENSOR_EQ(th, hh, "GPU computed output should be the same as CPU computed ones"); |
2100 | 0 | ccv_nnc_tensor_free(a); |
2101 | 0 | ccv_nnc_tensor_free(w); |
2102 | 0 | ccv_nnc_tensor_free(bias); |
2103 | 0 | ccv_nnc_tensor_free(b); |
2104 | 0 | ccv_nnc_tensor_free(g); |
2105 | 0 | ccv_nnc_tensor_free(dw); |
2106 | 0 | ccv_nnc_tensor_free(dbias); |
2107 | 0 | ccv_nnc_tensor_free(h); |
2108 | 0 | ccv_nnc_tensor_free(ha); |
2109 | 0 | ccv_nnc_tensor_free(hw); |
2110 | 0 | ccv_nnc_tensor_free(hbias); |
2111 | 0 | ccv_nnc_tensor_free(hb); |
2112 | 0 | ccv_nnc_tensor_free(hg); |
2113 | 0 | ccv_nnc_tensor_free(hdw); |
2114 | 0 | ccv_nnc_tensor_free(hdbias); |
2115 | 0 | ccv_nnc_tensor_free(hh); |
2116 | 0 | ccv_nnc_tensor_free(tb); |
2117 | 0 | ccv_nnc_tensor_free(th); |
2118 | 0 | ccv_nnc_tensor_free(tdw); |
2119 | 0 | ccv_nnc_tensor_free(tdbias); |
2120 | 0 | } |
2121 | | |
2122 | | TEST_CASE("backward gemm no bias") |
2123 | 1 | { |
2124 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2125 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2126 | 0 | dsfmt_t dsfmt; |
2127 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
2128 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2129 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2130 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2131 | 0 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2132 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2133 | 0 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2134 | |
|
2135 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2136 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2137 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2138 | 0 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2139 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2140 | 0 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2141 | 0 | int i; |
2142 | 0 | for (i = 0; i < 64 * 128; i++) |
2143 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
2144 | 0 | for (i = 0; i < 10 * 128; i++) |
2145 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2146 | 0 | for (i = 0; i < 10 * 64; i++) |
2147 | 0 | hg->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2148 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hg), TENSOR_LIST(a, w, g), 0); |
2149 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw), TENSOR_LIST(hb), 0); |
2150 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, ha, hw, 0), TENSOR_LIST(hh, hdw, 0), 0); |
2151 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w), TENSOR_LIST(b), 0); |
2152 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, w, 0), TENSOR_LIST(h, dw, 0), 0); |
2153 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2154 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2155 | 0 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2156 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, dw, h), TENSOR_LIST(tb, tdw, th), 0); |
2157 | 0 | REQUIRE_TENSOR_EQ(tb, hb, "GPU computed output should be the same as CPU computed ones"); |
2158 | 0 | REQUIRE_TENSOR_EQ(tdw, hdw, "GPU computed output should be the same as CPU computed ones"); |
2159 | 0 | REQUIRE_TENSOR_EQ(th, hh, "GPU computed output should be the same as CPU computed ones"); |
2160 | 0 | ccv_nnc_tensor_free(a); |
2161 | 0 | ccv_nnc_tensor_free(w); |
2162 | 0 | ccv_nnc_tensor_free(b); |
2163 | 0 | ccv_nnc_tensor_free(g); |
2164 | 0 | ccv_nnc_tensor_free(dw); |
2165 | 0 | ccv_nnc_tensor_free(h); |
2166 | 0 | ccv_nnc_tensor_free(ha); |
2167 | 0 | ccv_nnc_tensor_free(hw); |
2168 | 0 | ccv_nnc_tensor_free(hb); |
2169 | 0 | ccv_nnc_tensor_free(hg); |
2170 | 0 | ccv_nnc_tensor_free(hdw); |
2171 | 0 | ccv_nnc_tensor_free(hh); |
2172 | 0 | ccv_nnc_tensor_free(tb); |
2173 | 0 | ccv_nnc_tensor_free(th); |
2174 | 0 | ccv_nnc_tensor_free(tdw); |
2175 | 0 | } |
2176 | | |
2177 | | TEST_CASE("backward gemm no h") |
2178 | 1 | { |
2179 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2180 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2181 | 0 | dsfmt_t dsfmt; |
2182 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
2183 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2184 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2185 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
2186 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2187 | 0 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2188 | 0 | ccv_nnc_tensor_t* dw = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2189 | 0 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
2190 | |
|
2191 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2192 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2193 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2194 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2195 | 0 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2196 | 0 | ccv_nnc_tensor_t* hdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2197 | 0 | ccv_nnc_tensor_t* hdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2198 | 0 | int i; |
2199 | 0 | for (i = 0; i < 64 * 128; i++) |
2200 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
2201 | 0 | for (i = 0; i < 64; i++) |
2202 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2203 | 0 | for (i = 0; i < 10 * 128; i++) |
2204 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2205 | 0 | for (i = 0; i < 10 * 64; i++) |
2206 | 0 | hg->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2207 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias, hg), TENSOR_LIST(a, w, bias, g), 0); |
2208 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(hb), 0); |
2209 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, ha, hw, 0), TENSOR_LIST(0, hdw, hdbias), 0); |
2210 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
2211 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, w, 0), TENSOR_LIST(0, dw, dbias), 0); |
2212 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2213 | 0 | ccv_nnc_tensor_t* tdw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2214 | 0 | ccv_nnc_tensor_t* tdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2215 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, dw, dbias, 0), TENSOR_LIST(tb, tdw, tdbias, 0), 0); |
2216 | 0 | REQUIRE_TENSOR_EQ(tb, hb, "GPU computed output should be the same as CPU computed ones"); |
2217 | 0 | REQUIRE_TENSOR_EQ(tdw, hdw, "GPU computed output should be the same as CPU computed ones"); |
2218 | 0 | REQUIRE_TENSOR_EQ(tdbias, hdbias, "GPU computed output should be the same as CPU computed ones"); |
2219 | 0 | ccv_nnc_tensor_free(a); |
2220 | 0 | ccv_nnc_tensor_free(w); |
2221 | 0 | ccv_nnc_tensor_free(bias); |
2222 | 0 | ccv_nnc_tensor_free(b); |
2223 | 0 | ccv_nnc_tensor_free(g); |
2224 | 0 | ccv_nnc_tensor_free(dw); |
2225 | 0 | ccv_nnc_tensor_free(dbias); |
2226 | 0 | ccv_nnc_tensor_free(ha); |
2227 | 0 | ccv_nnc_tensor_free(hw); |
2228 | 0 | ccv_nnc_tensor_free(hbias); |
2229 | 0 | ccv_nnc_tensor_free(hb); |
2230 | 0 | ccv_nnc_tensor_free(hg); |
2231 | 0 | ccv_nnc_tensor_free(hdw); |
2232 | 0 | ccv_nnc_tensor_free(hdbias); |
2233 | 0 | ccv_nnc_tensor_free(tb); |
2234 | 0 | ccv_nnc_tensor_free(tdw); |
2235 | 0 | ccv_nnc_tensor_free(tdbias); |
2236 | 0 | } |
2237 | | |
2238 | | TEST_CASE("backward gemm no dw") |
2239 | 1 | { |
2240 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2241 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2242 | 0 | dsfmt_t dsfmt; |
2243 | 0 | dsfmt_init_gen_rand(&dsfmt, 0); |
2244 | 0 | ccv_nnc_tensor_t* a = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2245 | 0 | ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64, 128), 0); |
2246 | 0 | ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
2247 | 0 | ccv_nnc_tensor_t* b = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2248 | 0 | ccv_nnc_tensor_t* g = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 64), 0); |
2249 | 0 | ccv_nnc_tensor_t* dbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 64), 0); |
2250 | 0 | ccv_nnc_tensor_t* h = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 128), 0); |
2251 | |
|
2252 | 0 | ccv_nnc_tensor_t* ha = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2253 | 0 | ccv_nnc_tensor_t* hw = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64, 128), 0); |
2254 | 0 | ccv_nnc_tensor_t* hbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2255 | 0 | ccv_nnc_tensor_t* hb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2256 | 0 | ccv_nnc_tensor_t* hg = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2257 | 0 | ccv_nnc_tensor_t* hdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2258 | 0 | ccv_nnc_tensor_t* hh = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2259 | 0 | int i; |
2260 | 0 | for (i = 0; i < 64 * 128; i++) |
2261 | 0 | hw->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) / (64 * 128); |
2262 | 0 | for (i = 0; i < 64; i++) |
2263 | 0 | hbias->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2264 | 0 | for (i = 0; i < 10 * 128; i++) |
2265 | 0 | ha->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2266 | 0 | for (i = 0; i < 10 * 64; i++) |
2267 | 0 | hg->data.f32[i] = dsfmt_genrand_open_close(&dsfmt); |
2268 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias, hg), TENSOR_LIST(a, w, bias, g), 0); |
2269 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(ha, hw, hbias), TENSOR_LIST(hb), 0); |
2270 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(hg, ha, hw, 0), TENSOR_LIST(hh, 0, hdbias), 0); |
2271 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_FORWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(a, w, bias), TENSOR_LIST(b), 0); |
2272 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, w, 0), TENSOR_LIST(h, 0, dbias), 0); |
2273 | 0 | ccv_nnc_tensor_t* tb = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 64), 0); |
2274 | 0 | ccv_nnc_tensor_t* tdbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 64), 0); |
2275 | 0 | ccv_nnc_tensor_t* th = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 128), 0); |
2276 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b, 0, dbias, h), TENSOR_LIST(tb, 0, tdbias, th), 0); |
2277 | 0 | REQUIRE_TENSOR_EQ(tb, hb, "GPU computed output should be the same as CPU computed ones"); |
2278 | 0 | REQUIRE_TENSOR_EQ(tdbias, hdbias, "GPU computed output should be the same as CPU computed ones"); |
2279 | 0 | REQUIRE_TENSOR_EQ(th, hh, "GPU computed output should be the same as CPU computed ones"); |
2280 | 0 | ccv_nnc_tensor_free(a); |
2281 | 0 | ccv_nnc_tensor_free(w); |
2282 | 0 | ccv_nnc_tensor_free(bias); |
2283 | 0 | ccv_nnc_tensor_free(b); |
2284 | 0 | ccv_nnc_tensor_free(g); |
2285 | 0 | ccv_nnc_tensor_free(dbias); |
2286 | 0 | ccv_nnc_tensor_free(h); |
2287 | 0 | ccv_nnc_tensor_free(ha); |
2288 | 0 | ccv_nnc_tensor_free(hw); |
2289 | 0 | ccv_nnc_tensor_free(hbias); |
2290 | 0 | ccv_nnc_tensor_free(hb); |
2291 | 0 | ccv_nnc_tensor_free(hg); |
2292 | 0 | ccv_nnc_tensor_free(hdbias); |
2293 | 0 | ccv_nnc_tensor_free(hh); |
2294 | 0 | ccv_nnc_tensor_free(tb); |
2295 | 0 | ccv_nnc_tensor_free(th); |
2296 | 0 | ccv_nnc_tensor_free(tdbias); |
2297 | 0 | } |
2298 | | |
2299 | | TEST_CASE("backwar gemm with no transpose batch 2, same b") |
2300 | 1 | { |
2301 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2302 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2303 | 0 | float gp[] = { |
2304 | 0 | 1, 2, 3, |
2305 | 0 | 4, 5, 6, |
2306 | 0 | 7, 8, 9, |
2307 | 0 | 10, 11, 12, |
2308 | 0 | 10, 20, 30, |
2309 | 0 | 40, 50, 60, |
2310 | 0 | 70, 80, 90, |
2311 | 0 | 100, 110, 120, |
2312 | 0 | }; |
2313 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2314 | 0 | float ap[] = { |
2315 | 0 | 13, 14, |
2316 | 0 | 15, 16, |
2317 | 0 | 17, 18, |
2318 | 0 | 19, 20, |
2319 | 0 | 131, 141, |
2320 | 0 | 151, 161, |
2321 | 0 | 171, 181, |
2322 | 0 | 191, 201, |
2323 | 0 | }; |
2324 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2325 | 0 | float bp[] = { |
2326 | 0 | 21, 22, 23, |
2327 | 0 | 24, 25, 26, |
2328 | 0 | }; |
2329 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
2330 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2331 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
2332 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
2333 | 0 | ccv_nnc_tensor_t* gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2334 | 0 | ccv_nnc_tensor_t* ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2335 | 0 | ccv_nnc_tensor_t* gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
2336 | 0 | ccv_nnc_tensor_t* gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2337 | 0 | ccv_nnc_tensor_t* gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
2338 | 0 | ccv_nnc_tensor_t* gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
2339 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2340 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
2341 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
2342 | 0 | float dbiastp[] = { |
2343 | 0 | 22 + 220, 26 + 260, 30 + 300, |
2344 | 0 | }; |
2345 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
2346 | | |
2347 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
2348 | 0 | float htp[] = { |
2349 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
2350 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
2351 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
2352 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
2353 | 0 | 10 * 21 + 20 * 22 + 30 * 23, 10 * 24 + 20 * 25 + 30 * 26, |
2354 | 0 | 40 * 21 + 50 * 22 + 60 * 23, 40 * 24 + 50 * 25 + 60 * 26, |
2355 | 0 | 70 * 21 + 80 * 22 + 90 * 23, 70 * 24 + 80 * 25 + 90 * 26, |
2356 | 0 | 100 * 21 + 110 * 22 + 120 * 23, 100 * 24 + 110 * 25 + 120 * 26, |
2357 | 0 | }; |
2358 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2359 | | |
2360 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2361 | 0 | float dbtp[] = { |
2362 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19 + 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19 + 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19 + 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, |
2363 | 0 | 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20 + 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20 + 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20 + 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2364 | 0 | }; |
2365 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
2366 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2367 | 0 | ccv_nnc_tensor_free(g); |
2368 | 0 | ccv_nnc_tensor_free(a); |
2369 | 0 | ccv_nnc_tensor_free(b); |
2370 | 0 | ccv_nnc_tensor_free(h); |
2371 | 0 | ccv_nnc_tensor_free(db); |
2372 | 0 | ccv_nnc_tensor_free(dbias); |
2373 | 0 | ccv_nnc_tensor_free(gg); |
2374 | 0 | ccv_nnc_tensor_free(ga); |
2375 | 0 | ccv_nnc_tensor_free(gb); |
2376 | 0 | ccv_nnc_tensor_free(gh); |
2377 | 0 | ccv_nnc_tensor_free(gdb); |
2378 | 0 | ccv_nnc_tensor_free(gdbias); |
2379 | 0 | } |
2380 | | |
2381 | | TEST_CASE("backward gemm with no transpose batch 2, batched b") |
2382 | 1 | { |
2383 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2384 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2385 | 0 | float gp[] = { |
2386 | 0 | 1, 2, 3, |
2387 | 0 | 4, 5, 6, |
2388 | 0 | 7, 8, 9, |
2389 | 0 | 10, 11, 12, |
2390 | 0 | 10, 20, 30, |
2391 | 0 | 40, 50, 60, |
2392 | 0 | 70, 80, 90, |
2393 | 0 | 100, 110, 120, |
2394 | 0 | }; |
2395 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2396 | 0 | float ap[] = { |
2397 | 0 | 13, 14, |
2398 | 0 | 15, 16, |
2399 | 0 | 17, 18, |
2400 | 0 | 19, 20, |
2401 | 0 | 131, 141, |
2402 | 0 | 151, 161, |
2403 | 0 | 171, 181, |
2404 | 0 | 191, 201, |
2405 | 0 | }; |
2406 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2407 | 0 | float bp[] = { |
2408 | 0 | 21, 22, 23, |
2409 | 0 | 24, 25, 26, |
2410 | 0 | 212, 222, 232, |
2411 | 0 | 242, 252, 262, |
2412 | 0 | }; |
2413 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
2414 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2415 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
2416 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 1, 3), 0); |
2417 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2418 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2419 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 3), 0); |
2420 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2421 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 3), 0); |
2422 | 0 | ccv_nnc_tensor_t* const gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 1, 3), 0); |
2423 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2424 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
2425 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
2426 | 0 | float dbiastp[] = { |
2427 | 0 | 22, 26, 30, |
2428 | 0 | 220, 260, 300, |
2429 | 0 | }; |
2430 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 2, 1, 3), 0); |
2431 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
2432 | 0 | float htp[] = { |
2433 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
2434 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
2435 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
2436 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
2437 | 0 | 10 * 212 + 20 * 222 + 30 * 232, 10 * 242 + 20 * 252 + 30 * 262, |
2438 | 0 | 40 * 212 + 50 * 222 + 60 * 232, 40 * 242 + 50 * 252 + 60 * 262, |
2439 | 0 | 70 * 212 + 80 * 222 + 90 * 232, 70 * 242 + 80 * 252 + 90 * 262, |
2440 | 0 | 100 * 212 + 110 * 222 + 120 * 232, 100 * 242 + 110 * 252 + 120 * 262, |
2441 | 0 | }; |
2442 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2443 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2444 | 0 | float dbtp[] = { |
2445 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, |
2446 | 0 | 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
2447 | 0 | 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, |
2448 | 0 | 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2449 | 0 | }; |
2450 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
2451 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2452 | 0 | ccv_nnc_tensor_free(g); |
2453 | 0 | ccv_nnc_tensor_free(a); |
2454 | 0 | ccv_nnc_tensor_free(b); |
2455 | 0 | ccv_nnc_tensor_free(h); |
2456 | 0 | ccv_nnc_tensor_free(db); |
2457 | 0 | ccv_nnc_tensor_free(dbias); |
2458 | 0 | ccv_nnc_tensor_free(gg); |
2459 | 0 | ccv_nnc_tensor_free(ga); |
2460 | 0 | ccv_nnc_tensor_free(gb); |
2461 | 0 | ccv_nnc_tensor_free(gh); |
2462 | 0 | ccv_nnc_tensor_free(gdb); |
2463 | 0 | ccv_nnc_tensor_free(gdbias); |
2464 | 0 | } |
2465 | | |
2466 | | TEST_CASE("backward gemm with transpose a batch 2, same b") |
2467 | 1 | { |
2468 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2469 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2470 | 0 | float gp[] = { |
2471 | 0 | 1, 2, 3, |
2472 | 0 | 4, 5, 6, |
2473 | 0 | 7, 8, 9, |
2474 | 0 | 10, 11, 12, |
2475 | 0 | 10, 20, 30, |
2476 | 0 | 40, 50, 60, |
2477 | 0 | 70, 80, 90, |
2478 | 0 | 100, 110, 120, |
2479 | 0 | }; |
2480 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2481 | 0 | float ap[] = { |
2482 | 0 | 13, 15, 17, 19, |
2483 | 0 | 14, 16, 18, 20, |
2484 | 0 | 131, 151, 171, 191, |
2485 | 0 | 141, 161, 181, 201, |
2486 | 0 | }; |
2487 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2488 | 0 | float bp[] = { |
2489 | 0 | 21, 22, 23, |
2490 | 0 | 24, 25, 26, |
2491 | 0 | }; |
2492 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
2493 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2494 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
2495 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
2496 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2497 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
2498 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
2499 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
2500 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3), 0); |
2501 | 0 | ccv_nnc_tensor_t* const gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
2502 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2503 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
2504 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
2505 | 0 | float dbiastp[] = { |
2506 | 0 | 22 + 220, 26 + 260, 30 + 300, |
2507 | 0 | }; |
2508 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
2509 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
2510 | 0 | float htp[] = { |
2511 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 4 * 21 + 5 * 22 + 6 * 23, 7 * 21 + 8 * 22 + 9 * 23, 10 * 21 + 11 * 22 + 12 * 23, |
2512 | 0 | 1 * 24 + 2 * 25 + 3 * 26, 4 * 24 + 5 * 25 + 6 * 26, 7 * 24 + 8 * 25 + 9 * 26, 10 * 24 + 11 * 25 + 12 * 26, |
2513 | 0 | 10 * 21 + 20 * 22 + 30 * 23, 40 * 21 + 50 * 22 + 60 * 23, 70 * 21 + 80 * 22 + 90 * 23, 100 * 21 + 110 * 22 + 120 * 23, |
2514 | 0 | 10 * 24 + 20 * 25 + 30 * 26, 40 * 24 + 50 * 25 + 60 * 26, 70 * 24 + 80 * 25 + 90 * 26, 100 * 24 + 110 * 25 + 120 * 26, |
2515 | 0 | }; |
2516 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2517 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2518 | 0 | float dbtp[] = { |
2519 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19 + 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19 + 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19 + 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, |
2520 | 0 | 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20 + 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20 + 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20 + 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2521 | 0 | }; |
2522 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3), 0); |
2523 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2524 | 0 | ccv_nnc_tensor_free(g); |
2525 | 0 | ccv_nnc_tensor_free(a); |
2526 | 0 | ccv_nnc_tensor_free(b); |
2527 | 0 | ccv_nnc_tensor_free(h); |
2528 | 0 | ccv_nnc_tensor_free(db); |
2529 | 0 | ccv_nnc_tensor_free(dbias); |
2530 | 0 | ccv_nnc_tensor_free(gg); |
2531 | 0 | ccv_nnc_tensor_free(ga); |
2532 | 0 | ccv_nnc_tensor_free(gb); |
2533 | 0 | ccv_nnc_tensor_free(gh); |
2534 | 0 | ccv_nnc_tensor_free(gdb); |
2535 | 0 | ccv_nnc_tensor_free(gdbias); |
2536 | 0 | } |
2537 | | |
2538 | | TEST_CASE("backward gemm with transpose b batch 2, batched b") |
2539 | 1 | { |
2540 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2541 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2542 | 0 | float gp[] = { |
2543 | 0 | 1, 2, 3, |
2544 | 0 | 4, 5, 6, |
2545 | 0 | 7, 8, 9, |
2546 | 0 | 10, 11, 12, |
2547 | 0 | 10, 20, 30, |
2548 | 0 | 40, 50, 60, |
2549 | 0 | 70, 80, 90, |
2550 | 0 | 100, 110, 120, |
2551 | 0 | }; |
2552 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2553 | 0 | float ap[] = { |
2554 | 0 | 13, 14, |
2555 | 0 | 15, 16, |
2556 | 0 | 17, 18, |
2557 | 0 | 19, 20, |
2558 | 0 | 131, 141, |
2559 | 0 | 151, 161, |
2560 | 0 | 171, 181, |
2561 | 0 | 191, 201, |
2562 | 0 | }; |
2563 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2564 | 0 | float bp[] = { |
2565 | 0 | 21, 24, |
2566 | 0 | 22, 25, |
2567 | 0 | 23, 26, |
2568 | 0 | 212, 242, |
2569 | 0 | 222, 252, |
2570 | 0 | 232, 262, |
2571 | 0 | }; |
2572 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2573 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2574 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2575 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 1, 3), 0); |
2576 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2577 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2578 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
2579 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2580 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
2581 | 0 | ccv_nnc_tensor_t* const gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 1, 3), 0); |
2582 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2583 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
2584 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
2585 | 0 | float dbiastp[] = { |
2586 | 0 | 22, 26, 30, |
2587 | 0 | 220, 260, 300, |
2588 | 0 | }; |
2589 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 2, 1, 3), 0); |
2590 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
2591 | 0 | float htp[] = { |
2592 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
2593 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
2594 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
2595 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
2596 | 0 | 10 * 212 + 20 * 222 + 30 * 232, 10 * 242 + 20 * 252 + 30 * 262, |
2597 | 0 | 40 * 212 + 50 * 222 + 60 * 232, 40 * 242 + 50 * 252 + 60 * 262, |
2598 | 0 | 70 * 212 + 80 * 222 + 90 * 232, 70 * 242 + 80 * 252 + 90 * 262, |
2599 | 0 | 100 * 212 + 110 * 222 + 120 * 232, 100 * 242 + 110 * 252 + 120 * 262, |
2600 | 0 | }; |
2601 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2602 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2603 | 0 | float dbtp[] = { |
2604 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, |
2605 | 0 | 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, |
2606 | 0 | 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
2607 | 0 | 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, |
2608 | 0 | 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, |
2609 | 0 | 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2610 | 0 | }; |
2611 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2612 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2613 | 0 | ccv_nnc_tensor_free(g); |
2614 | 0 | ccv_nnc_tensor_free(a); |
2615 | 0 | ccv_nnc_tensor_free(b); |
2616 | 0 | ccv_nnc_tensor_free(h); |
2617 | 0 | ccv_nnc_tensor_free(db); |
2618 | 0 | ccv_nnc_tensor_free(dbias); |
2619 | 0 | ccv_nnc_tensor_free(gg); |
2620 | 0 | ccv_nnc_tensor_free(ga); |
2621 | 0 | ccv_nnc_tensor_free(gb); |
2622 | 0 | ccv_nnc_tensor_free(gh); |
2623 | 0 | ccv_nnc_tensor_free(gdb); |
2624 | 0 | ccv_nnc_tensor_free(gdbias); |
2625 | 0 | } |
2626 | | |
2627 | | TEST_CASE("backward gemm with transpose a and b batch 2, same b") |
2628 | 1 | { |
2629 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2630 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2631 | 0 | float gp[] = { |
2632 | 0 | 1, 2, 3, |
2633 | 0 | 4, 5, 6, |
2634 | 0 | 7, 8, 9, |
2635 | 0 | 10, 11, 12, |
2636 | 0 | 10, 20, 30, |
2637 | 0 | 40, 50, 60, |
2638 | 0 | 70, 80, 90, |
2639 | 0 | 100, 110, 120, |
2640 | 0 | }; |
2641 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2642 | 0 | float ap[] = { |
2643 | 0 | 13, 15, 17, 19, |
2644 | 0 | 14, 16, 18, 20, |
2645 | 0 | 131, 151, 171, 191, |
2646 | 0 | 141, 161, 181, 201, |
2647 | 0 | }; |
2648 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2649 | 0 | float bp[] = { |
2650 | 0 | 21, 24, |
2651 | 0 | 22, 25, |
2652 | 0 | 23, 26, |
2653 | 0 | }; |
2654 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
2655 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2656 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
2657 | 0 | ccv_nnc_tensor_t* const dbias = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 3), 0); |
2658 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2659 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
2660 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
2661 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
2662 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3, 2), 0); |
2663 | 0 | ccv_nnc_tensor_t* const gdbias = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 3), 0); |
2664 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2665 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(TRANSPOSE(1, 2), TRANSPOSE(0, 1)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb, gdbias), 0); |
2666 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb, gdbias), TENSOR_LIST(h, db, dbias), 0); |
2667 | 0 | float dbiastp[] = { |
2668 | 0 | 22 + 220, 26 + 260, 30 + 300, |
2669 | 0 | }; |
2670 | 0 | ccv_nnc_tensor_t dbiast = ccv_nnc_tensor(dbiastp, CPU_TENSOR_NHWC(32F, 3), 0); |
2671 | 0 | REQUIRE_TENSOR_EQ(dbias, &dbiast, "bias should be equal"); |
2672 | 0 | float htp[] = { |
2673 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 4 * 21 + 5 * 22 + 6 * 23, 7 * 21 + 8 * 22 + 9 * 23, 10 * 21 + 11 * 22 + 12 * 23, |
2674 | 0 | 1 * 24 + 2 * 25 + 3 * 26, 4 * 24 + 5 * 25 + 6 * 26, 7 * 24 + 8 * 25 + 9 * 26, 10 * 24 + 11 * 25 + 12 * 26, |
2675 | 0 | 10 * 21 + 20 * 22 + 30 * 23, 40 * 21 + 50 * 22 + 60 * 23, 70 * 21 + 80 * 22 + 90 * 23, 100 * 21 + 110 * 22 + 120 * 23, |
2676 | 0 | 10 * 24 + 20 * 25 + 30 * 26, 40 * 24 + 50 * 25 + 60 * 26, 70 * 24 + 80 * 25 + 90 * 26, 100 * 24 + 110 * 25 + 120 * 26, |
2677 | 0 | }; |
2678 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2679 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2680 | 0 | float dbtp[] = { |
2681 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19 + 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20 + 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, |
2682 | 0 | 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19 + 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20 + 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, |
2683 | 0 | 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19 + 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20 + 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2684 | 0 | }; |
2685 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 3, 2), 0); |
2686 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2687 | 0 | ccv_nnc_tensor_free(g); |
2688 | 0 | ccv_nnc_tensor_free(a); |
2689 | 0 | ccv_nnc_tensor_free(b); |
2690 | 0 | ccv_nnc_tensor_free(h); |
2691 | 0 | ccv_nnc_tensor_free(db); |
2692 | 0 | ccv_nnc_tensor_free(dbias); |
2693 | 0 | ccv_nnc_tensor_free(gg); |
2694 | 0 | ccv_nnc_tensor_free(ga); |
2695 | 0 | ccv_nnc_tensor_free(gb); |
2696 | 0 | ccv_nnc_tensor_free(gh); |
2697 | 0 | ccv_nnc_tensor_free(gdb); |
2698 | 0 | ccv_nnc_tensor_free(gdbias); |
2699 | 0 | } |
2700 | | |
2701 | | TEST_CASE("backward gemm with no transpose batch 2, batched b, no bias") |
2702 | 1 | { |
2703 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2704 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2705 | 0 | float gp[] = { |
2706 | 0 | 1, 2, 3, |
2707 | 0 | 4, 5, 6, |
2708 | 0 | 7, 8, 9, |
2709 | 0 | 10, 11, 12, |
2710 | 0 | 10, 20, 30, |
2711 | 0 | 40, 50, 60, |
2712 | 0 | 70, 80, 90, |
2713 | 0 | 100, 110, 120, |
2714 | 0 | }; |
2715 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2716 | 0 | float ap[] = { |
2717 | 0 | 13, 14, |
2718 | 0 | 15, 16, |
2719 | 0 | 17, 18, |
2720 | 0 | 19, 20, |
2721 | 0 | 131, 141, |
2722 | 0 | 151, 161, |
2723 | 0 | 171, 181, |
2724 | 0 | 191, 201, |
2725 | 0 | }; |
2726 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2727 | 0 | float bp[] = { |
2728 | 0 | 21, 22, 23, |
2729 | 0 | 24, 25, 26, |
2730 | 0 | 212, 222, 232, |
2731 | 0 | 242, 252, 262, |
2732 | 0 | }; |
2733 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
2734 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2735 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
2736 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2737 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2738 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 3), 0); |
2739 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2740 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 3), 0); |
2741 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2742 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb), 0); |
2743 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb), TENSOR_LIST(h, db), 0); |
2744 | 0 | float htp[] = { |
2745 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
2746 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
2747 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
2748 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
2749 | 0 | 10 * 212 + 20 * 222 + 30 * 232, 10 * 242 + 20 * 252 + 30 * 262, |
2750 | 0 | 40 * 212 + 50 * 222 + 60 * 232, 40 * 242 + 50 * 252 + 60 * 262, |
2751 | 0 | 70 * 212 + 80 * 222 + 90 * 232, 70 * 242 + 80 * 252 + 90 * 262, |
2752 | 0 | 100 * 212 + 110 * 222 + 120 * 232, 100 * 242 + 110 * 252 + 120 * 262, |
2753 | 0 | }; |
2754 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2755 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2756 | 0 | float dbtp[] = { |
2757 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, |
2758 | 0 | 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
2759 | 0 | 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, |
2760 | 0 | 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2761 | 0 | }; |
2762 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 2, 3), 0); |
2763 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2764 | 0 | ccv_nnc_tensor_free(g); |
2765 | 0 | ccv_nnc_tensor_free(a); |
2766 | 0 | ccv_nnc_tensor_free(b); |
2767 | 0 | ccv_nnc_tensor_free(h); |
2768 | 0 | ccv_nnc_tensor_free(db); |
2769 | 0 | ccv_nnc_tensor_free(gg); |
2770 | 0 | ccv_nnc_tensor_free(ga); |
2771 | 0 | ccv_nnc_tensor_free(gb); |
2772 | 0 | ccv_nnc_tensor_free(gh); |
2773 | 0 | ccv_nnc_tensor_free(gdb); |
2774 | 0 | } |
2775 | | |
2776 | | TEST_CASE("backward gemm with transpose b batch 2, batched b, no bias") |
2777 | 1 | { |
2778 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2779 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2780 | 0 | float gp[] = { |
2781 | 0 | 1, 2, 3, |
2782 | 0 | 4, 5, 6, |
2783 | 0 | 7, 8, 9, |
2784 | 0 | 10, 11, 12, |
2785 | 0 | 10, 20, 30, |
2786 | 0 | 40, 50, 60, |
2787 | 0 | 70, 80, 90, |
2788 | 0 | 100, 110, 120, |
2789 | 0 | }; |
2790 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2791 | 0 | float ap[] = { |
2792 | 0 | 13, 14, |
2793 | 0 | 15, 16, |
2794 | 0 | 17, 18, |
2795 | 0 | 19, 20, |
2796 | 0 | 131, 141, |
2797 | 0 | 151, 161, |
2798 | 0 | 171, 181, |
2799 | 0 | 191, 201, |
2800 | 0 | }; |
2801 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2802 | 0 | float bp[] = { |
2803 | 0 | 21, 24, |
2804 | 0 | 22, 25, |
2805 | 0 | 23, 26, |
2806 | 0 | 212, 242, |
2807 | 0 | 222, 252, |
2808 | 0 | 232, 262, |
2809 | 0 | }; |
2810 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2811 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2812 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2813 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2814 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2815 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
2816 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 2), 0); |
2817 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
2818 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2819 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(NO_TRANSPOSE, TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb), 0); |
2820 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb), TENSOR_LIST(h, db), 0); |
2821 | 0 | float htp[] = { |
2822 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 1 * 24 + 2 * 25 + 3 * 26, |
2823 | 0 | 4 * 21 + 5 * 22 + 6 * 23, 4 * 24 + 5 * 25 + 6 * 26, |
2824 | 0 | 7 * 21 + 8 * 22 + 9 * 23, 7 * 24 + 8 * 25 + 9 * 26, |
2825 | 0 | 10 * 21 + 11 * 22 + 12 * 23, 10 * 24 + 11 * 25 + 12 * 26, |
2826 | 0 | 10 * 212 + 20 * 222 + 30 * 232, 10 * 242 + 20 * 252 + 30 * 262, |
2827 | 0 | 40 * 212 + 50 * 222 + 60 * 232, 40 * 242 + 50 * 252 + 60 * 262, |
2828 | 0 | 70 * 212 + 80 * 222 + 90 * 232, 70 * 242 + 80 * 252 + 90 * 262, |
2829 | 0 | 100 * 212 + 110 * 222 + 120 * 232, 100 * 242 + 110 * 252 + 120 * 262, |
2830 | 0 | }; |
2831 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 4, 2), 0); |
2832 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2833 | 0 | float dbtp[] = { |
2834 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, |
2835 | 0 | 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, |
2836 | 0 | 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
2837 | 0 | 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, |
2838 | 0 | 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, |
2839 | 0 | 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2840 | 0 | }; |
2841 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2842 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2843 | 0 | ccv_nnc_tensor_free(g); |
2844 | 0 | ccv_nnc_tensor_free(a); |
2845 | 0 | ccv_nnc_tensor_free(b); |
2846 | 0 | ccv_nnc_tensor_free(h); |
2847 | 0 | ccv_nnc_tensor_free(db); |
2848 | 0 | ccv_nnc_tensor_free(gg); |
2849 | 0 | ccv_nnc_tensor_free(ga); |
2850 | 0 | ccv_nnc_tensor_free(gb); |
2851 | 0 | ccv_nnc_tensor_free(gh); |
2852 | 0 | ccv_nnc_tensor_free(gdb); |
2853 | 0 | } |
2854 | | |
2855 | | TEST_CASE("backward gemm with transpose a and b batch 2, batch b, no bias") |
2856 | 1 | { |
2857 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GEMM_FORWARD, CCV_NNC_BACKEND_MPS) && |
2858 | 1 | ccv_nnc_cmd_ok(CCV_NNC_GEMM_BACKWARD, CCV_NNC_BACKEND_MPS)); |
2859 | 0 | float gp[] = { |
2860 | 0 | 1, 2, 3, |
2861 | 0 | 4, 5, 6, |
2862 | 0 | 7, 8, 9, |
2863 | 0 | 10, 11, 12, |
2864 | 0 | 10, 20, 30, |
2865 | 0 | 40, 50, 60, |
2866 | 0 | 70, 80, 90, |
2867 | 0 | 100, 110, 120, |
2868 | 0 | }; |
2869 | 0 | ccv_nnc_tensor_t* const g = ccv_nnc_tensor_new(gp, CPU_TENSOR_NHWC(32F, 2, 4, 3), 0); |
2870 | 0 | float ap[] = { |
2871 | 0 | 13, 15, 17, 19, |
2872 | 0 | 14, 16, 18, 20, |
2873 | 0 | 131, 151, 171, 191, |
2874 | 0 | 141, 161, 181, 201, |
2875 | 0 | }; |
2876 | 0 | ccv_nnc_tensor_t* const a = ccv_nnc_tensor_new(ap, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2877 | 0 | float bp[] = { |
2878 | 0 | 21, 24, |
2879 | 0 | 22, 25, |
2880 | 0 | 23, 26, |
2881 | 0 | 212, 242, |
2882 | 0 | 222, 252, |
2883 | 0 | 232, 262, |
2884 | 0 | }; |
2885 | 0 | ccv_nnc_tensor_t* const b = ccv_nnc_tensor_new(bp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2886 | 0 | ccv_nnc_tensor_t* const h = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2887 | 0 | ccv_nnc_tensor_t* const db = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2888 | 0 | ccv_nnc_tensor_t* const gg = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 4, 3), 0); |
2889 | 0 | ccv_nnc_tensor_t* const ga = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
2890 | 0 | ccv_nnc_tensor_t* const gb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
2891 | 0 | ccv_nnc_tensor_t* const gh = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 2, 4), 0); |
2892 | 0 | ccv_nnc_tensor_t* const gdb = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 2, 3, 2), 0); |
2893 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(g, a, b), TENSOR_LIST(gg, ga, gb), 0); |
2894 | 0 | ccv_nnc_cmd_exec(CMD_GEMM_BACKWARD(TRANSPOSE(1, 2), TRANSPOSE(1, 2)), ccv_nnc_no_hint, 0, TENSOR_LIST(gg, ga, gb), TENSOR_LIST(gh, gdb), 0); |
2895 | 0 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(gh, gdb), TENSOR_LIST(h, db), 0); |
2896 | 0 | float htp[] = { |
2897 | 0 | 1 * 21 + 2 * 22 + 3 * 23, 4 * 21 + 5 * 22 + 6 * 23, 7 * 21 + 8 * 22 + 9 * 23, 10 * 21 + 11 * 22 + 12 * 23, |
2898 | 0 | 1 * 24 + 2 * 25 + 3 * 26, 4 * 24 + 5 * 25 + 6 * 26, 7 * 24 + 8 * 25 + 9 * 26, 10 * 24 + 11 * 25 + 12 * 26, |
2899 | 0 | 10 * 212 + 20 * 222 + 30 * 232, 40 * 212 + 50 * 222 + 60 * 232, 70 * 212 + 80 * 222 + 90 * 232, 100 * 212 + 110 * 222 + 120 * 232, |
2900 | 0 | 10 * 242 + 20 * 252 + 30 * 262, 40 * 242 + 50 * 252 + 60 * 262, 70 * 242 + 80 * 252 + 90 * 262, 100 * 242 + 110 * 252 + 120 * 262, |
2901 | 0 | }; |
2902 | 0 | ccv_nnc_tensor_t ht = ccv_nnc_tensor(htp, CPU_TENSOR_NHWC(32F, 2, 2, 4), 0); |
2903 | 0 | REQUIRE_TENSOR_EQ(h, &ht, "h should be equal"); |
2904 | 0 | float dbtp[] = { |
2905 | 0 | 1 * 13 + 4 * 15 + 7 * 17 + 10 * 19, 1 * 14 + 4 * 16 + 7 * 18 + 10 * 20, |
2906 | 0 | 2 * 13 + 5 * 15 + 8 * 17 + 11 * 19, 2 * 14 + 5 * 16 + 8 * 18 + 11 * 20, |
2907 | 0 | 3 * 13 + 6 * 15 + 9 * 17 + 12 * 19, 3 * 14 + 6 * 16 + 9 * 18 + 12 * 20, |
2908 | 0 | 10 * 131 + 40 * 151 + 70 * 171 + 100 * 191, 10 * 141 + 40 * 161 + 70 * 181 + 100 * 201, |
2909 | 0 | 20 * 131 + 50 * 151 + 80 * 171 + 110 * 191, 20 * 141 + 50 * 161 + 80 * 181 + 110 * 201, |
2910 | 0 | 30 * 131 + 60 * 151 + 90 * 171 + 120 * 191, 30 * 141 + 60 * 161 + 90 * 181 + 120 * 201, |
2911 | 0 | }; |
2912 | 0 | ccv_nnc_tensor_t dbt = ccv_nnc_tensor(dbtp, CPU_TENSOR_NHWC(32F, 2, 3, 2), 0); |
2913 | 0 | REQUIRE_TENSOR_EQ(db, &dbt, "db should be equal"); |
2914 | 0 | ccv_nnc_tensor_free(g); |
2915 | 0 | ccv_nnc_tensor_free(a); |
2916 | 0 | ccv_nnc_tensor_free(b); |
2917 | 0 | ccv_nnc_tensor_free(h); |
2918 | 0 | ccv_nnc_tensor_free(db); |
2919 | 0 | ccv_nnc_tensor_free(gg); |
2920 | 0 | ccv_nnc_tensor_free(ga); |
2921 | 0 | ccv_nnc_tensor_free(gb); |
2922 | 0 | ccv_nnc_tensor_free(gh); |
2923 | 0 | ccv_nnc_tensor_free(gdb); |
2924 | 0 | } |
2925 | | |
2926 | | #include "case_main.h" |