/home/liu/actions-runner/_work/ccv/ccv/test/int/nnc/tensor.tests.c
Line | Count | Source |
1 | | #include "ccv.h" |
2 | | #include "case.h" |
3 | | #include "ccv_case.h" |
4 | | #include "ccv_nnc_case.h" |
5 | | #include "nnc/ccv_nnc.h" |
6 | | #include "nnc/ccv_nnc_easy.h" |
7 | | #include "3rdparty/sqlite3/sqlite3.h" |
8 | | #include "3rdparty/dsfmt/dSFMT.h" |
9 | | |
10 | | TEST_SETUP() |
11 | | { |
12 | | ccv_nnc_init(); |
13 | | } |
14 | | |
15 | | TEST_CASE("tensor persistence, to / from GPU") |
16 | | { |
17 | | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
18 | | sqlite3* handle; |
19 | | sqlite3_open("tensors_g.sqlite3", &handle); |
20 | | ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
21 | | ccv_nnc_tensor_t* const tensorG = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 20, 30), 0); |
22 | | int i; |
23 | | dsfmt_t dsfmt; |
24 | | dsfmt_init_gen_rand(&dsfmt, 1); |
25 | | for (i = 0; i < 10 * 20 * 30; i++) |
26 | | tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
27 | | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor), TENSOR_LIST(tensorG), 0); |
28 | | ccv_nnc_tensor_write(tensorG, handle, "x", 0); |
29 | | sqlite3_close(handle); |
30 | | handle = 0; |
31 | | sqlite3_open("tensors_g.sqlite3", &handle); |
32 | | ccv_nnc_tensor_t* tensor1 = 0; |
33 | | ccv_nnc_tensor_read(handle, "x", 0, 0, 0, &tensor1); |
34 | | ccv_nnc_tensor_t* tensor2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
35 | | ccv_nnc_tensor_read(handle, "x", 0, 0, 0, &tensor2); |
36 | | sqlite3_close(handle); |
37 | | ccv_nnc_tensor_t* const tensor1c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
38 | | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor1), TENSOR_LIST(tensor1c), 0); |
39 | | REQUIRE_TENSOR_EQ(tensor1c, tensor, "the first tensor should equal to the second"); |
40 | | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor2->data.f32, tensor->data.f32, 10, 1e-5, "the first 10 element should be equal"); |
41 | | REQUIRE(ccv_nnc_tensor_nd(tensor2->info.dim) == 1, "should be 1-d tensor"); |
42 | | REQUIRE_EQ(tensor2->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
43 | | ccv_nnc_tensor_free(tensor1); |
44 | | ccv_nnc_tensor_free(tensor1c); |
45 | | ccv_nnc_tensor_free(tensor2); |
46 | | ccv_nnc_tensor_free(tensor); |
47 | | ccv_nnc_tensor_free(tensorG); |
48 | | } |
49 | | |
50 | | TEST_CASE("tensor mapped from file") |
51 | 1 | { |
52 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
53 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_GPU_CUDNN) || ccv_nnc_cmd_ok(CCV_NNC_ADD_FORWARD, CCV_NNC_BACKEND_MPS)); |
54 | 1 | FILE* w = fopen("tensor.bin", "w+"); |
55 | 1 | float* w_a = (float*)ccmalloc(sizeof(float) * 4096 * 5); |
56 | 1 | int i; |
57 | 20.4k | for (i = 0; i < 4096 * 5; i++20.4k ) |
58 | 20.4k | w_a[i] = (float)(i + 1); |
59 | 1 | fwrite(w_a, 1, sizeof(float) * 4096 * 5, w); |
60 | 1 | fclose(w); |
61 | 1 | ccfree(w_a); |
62 | 1 | ccv_nnc_tensor_t* one = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 1), 0); |
63 | 1 | one->data.f32[0] = 1; |
64 | 1 | ccv_nnc_tensor_t* one_gpu = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 1), 0); |
65 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(one), TENSOR_LIST(one_gpu), 0); |
66 | 1 | ccv_nnc_tensor_t* tensor_a = ccv_nnc_tensor_new_from_file(GPU_TENSOR_NHWC(000, 32F, 5), "tensor.bin", 0, 0); |
67 | 1 | ccv_nnc_tensor_t* a_result = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 5), 0); |
68 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(0.5, 0.2), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor_a, one_gpu), TENSOR_LIST(a_result), 0); |
69 | 1 | float a[] = {1 * 0.5 + 0.2, 2 * 0.5 + 0.2, 3 * 0.5 + 0.2, 4 * 0.5 + 0.2, 5 * 0.5 + 0.2}; |
70 | 1 | ccv_nnc_tensor_t* tensor_b = ccv_nnc_tensor_new_from_file(GPU_TENSOR_NHWC(000, 32F, 4), "tensor.bin", (4096 * 4 * 4), 0); |
71 | 1 | ccv_nnc_tensor_t* b_result = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 4), 0); |
72 | 1 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(1, 1), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor_b, one_gpu), TENSOR_LIST(b_result), 0); |
73 | 1 | float b[] = {4096 * 4 + 1 + 1, 4096 * 4 + 2 + 1, 4096 * 4 + 3 + 1, 4096 * 4 + 4 + 1}; |
74 | 1 | ccv_nnc_tensor_t* at = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 5), 0); |
75 | 1 | ccv_nnc_tensor_t* bt = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 4), 0); |
76 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(a_result, b_result), TENSOR_LIST(at, bt), 0); |
77 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, at->data.f32, a, 5, 1e-5, "the first 5 element should be equal"); |
78 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, bt->data.f32, b, 4, 1e-5, "the first 4 element should be equal"); |
79 | 1 | ccv_nnc_tensor_free(tensor_a); |
80 | 1 | ccv_nnc_tensor_free(tensor_b); |
81 | 1 | ccv_nnc_tensor_free(one); |
82 | 1 | ccv_nnc_tensor_free(one_gpu); |
83 | 1 | ccv_nnc_tensor_free(a_result); |
84 | 1 | ccv_nnc_tensor_free(at); |
85 | 1 | ccv_nnc_tensor_free(b_result); |
86 | 1 | ccv_nnc_tensor_free(bt); |
87 | 1 | } |
88 | | |
89 | | static int _tensor_xor_encode(const void* const data, const size_t data_size, const int datatype, const int* const dimensions, const int dimension_count, void* const context, void* const encoded, size_t* const encoded_size, ccv_nnc_tensor_param_t* const params, unsigned int* const identifier) |
90 | | { |
91 | | unsigned char* const u8 = (unsigned char*)data; |
92 | | unsigned char* const u8enc = (unsigned char*)encoded; |
93 | | int i; |
94 | | for (i = 0; i < data_size; i++) |
95 | | u8enc[i] = u8[i] ^ 0x13; |
96 | | *encoded_size = data_size; |
97 | | *identifier = 1; |
98 | | return 1; |
99 | | } |
100 | | |
101 | | static int _tensor_xor_decode(const void* const data, const size_t data_size, const int datatype, const int* const dimensions, const int dimension_count, const unsigned int identifier, void* const context, const ccv_nnc_tensor_param_t tensor_params, ccv_nnc_tensor_t** const tensor_out, void* decoded, size_t* const decoded_size) |
102 | | { |
103 | | if (identifier != 1) |
104 | | return 0; |
105 | | if (!tensor_out[0]) |
106 | | { |
107 | | tensor_out[0] = ccv_nnc_tensor_new(0, tensor_params, 0); |
108 | | if (!decoded) |
109 | | decoded = tensor_out[0]->data.u8; |
110 | | } |
111 | | unsigned char* const u8 = (unsigned char*)data; |
112 | | unsigned char* const u8dec = (unsigned char*)decoded; |
113 | | const size_t expected_size = *decoded_size; |
114 | | int i; |
115 | | for (i = 0; i < ccv_min(expected_size, data_size); i++) |
116 | | u8dec[i] = u8[i] ^ 0x13; |
117 | | *decoded_size = ccv_min(expected_size, data_size); |
118 | | return 1; |
119 | | } |
120 | | |
121 | | static int _tensor_noop_encode(const void* const data, const size_t data_size, const int datatype, const int* const dimensions, const int dimension_count, void* const context, void* const encoded, size_t* const encoded_size, ccv_nnc_tensor_param_t* const params, unsigned int* const identifier) |
122 | | { |
123 | | return 0; |
124 | | } |
125 | | |
126 | | static int _tensor_noop_decode(const void* const data, const size_t data_size, const int datatype, const int* const dimensions, const int dimension_count, const unsigned int identifier, void* const context, const ccv_nnc_tensor_param_t tensor_params, ccv_nnc_tensor_t** const tensor_out, void* const decoded, size_t* const decoded_size) |
127 | | { |
128 | | return 0; |
129 | | } |
130 | | |
131 | | TEST_CASE("tensor persistence with encoder / decoder, to / from GPU") |
132 | 1 | { |
133 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
134 | 1 | sqlite3* handle; |
135 | 1 | sqlite3_open("tensors_de_g.sqlite3", &handle); |
136 | 1 | ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
137 | 1 | int i; |
138 | 1 | dsfmt_t dsfmt; |
139 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
140 | 6.00k | for (i = 0; i < 10 * 20 * 30; i++6.00k ) |
141 | 6.00k | tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
142 | 1 | ccv_nnc_tensor_t* const tensorG = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 20, 30), 0); |
143 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor), TENSOR_LIST(tensorG), 0); |
144 | 1 | ccv_nnc_tensor_io_option_t options = { |
145 | 1 | .encode = _tensor_xor_encode, |
146 | 1 | .decode = _tensor_xor_decode |
147 | 1 | }; |
148 | 1 | ccv_nnc_tensor_write(tensorG, handle, "y", &options); |
149 | 1 | sqlite3_close(handle); |
150 | 1 | handle = 0; |
151 | 1 | sqlite3_open("tensors_de_g.sqlite3", &handle); |
152 | 1 | ccv_nnc_tensor_t* tensor1 = 0; |
153 | 1 | ccv_nnc_tensor_read(handle, "y", &options, 0, 0, &tensor1); |
154 | 1 | ccv_nnc_tensor_t* tensor2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
155 | 1 | ccv_nnc_tensor_read(handle, "y", &options, 0, 0, &tensor2); |
156 | 1 | sqlite3_close(handle); |
157 | 1 | ccv_nnc_tensor_t* tensor1c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
158 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor1), TENSOR_LIST(tensor1c), 0); |
159 | 1 | REQUIRE_TENSOR_EQ(tensor1c, tensor, "the first tensor should equal to the second"); |
160 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor2->data.f32, tensor->data.f32, 10, 1e-5, "the first 10 element should be equal"); |
161 | 1 | REQUIRE(ccv_nnc_tensor_nd(tensor2->info.dim) == 1, "should be 1-d tensor"); |
162 | 1 | REQUIRE_EQ(tensor2->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
163 | 1 | ccv_nnc_tensor_free(tensor1); |
164 | 1 | ccv_nnc_tensor_free(tensor1c); |
165 | 1 | ccv_nnc_tensor_free(tensor2); |
166 | 1 | ccv_nnc_tensor_free(tensor); |
167 | 1 | ccv_nnc_tensor_free(tensorG); |
168 | 1 | } |
169 | | |
170 | | TEST_CASE("tensor persistence with noop encoder / decoder, to / from GPU") |
171 | 1 | { |
172 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
173 | 1 | sqlite3* handle; |
174 | 1 | sqlite3_open("tensors_noop_de_g.sqlite3", &handle); |
175 | 1 | ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
176 | 1 | int i; |
177 | 1 | dsfmt_t dsfmt; |
178 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
179 | 6.00k | for (i = 0; i < 10 * 20 * 30; i++6.00k ) |
180 | 6.00k | tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
181 | 1 | ccv_nnc_tensor_t* const tensorG = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 20, 30), 0); |
182 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor), TENSOR_LIST(tensorG), 0); |
183 | 1 | ccv_nnc_tensor_io_option_t options = { |
184 | 1 | .encode = _tensor_noop_encode, |
185 | 1 | .decode = _tensor_noop_decode |
186 | 1 | }; |
187 | 1 | ccv_nnc_tensor_write(tensorG, handle, "y", &options); |
188 | 1 | sqlite3_close(handle); |
189 | 1 | handle = 0; |
190 | 1 | sqlite3_open("tensors_noop_de_g.sqlite3", &handle); |
191 | 1 | ccv_nnc_tensor_t* tensor1 = 0; |
192 | 1 | ccv_nnc_tensor_read(handle, "y", &options, 0, 0, &tensor1); |
193 | 1 | ccv_nnc_tensor_t* tensor2 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
194 | 1 | ccv_nnc_tensor_read(handle, "y", &options, 0, 0, &tensor2); |
195 | 1 | sqlite3_close(handle); |
196 | 1 | ccv_nnc_tensor_t* tensor1c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
197 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor1), TENSOR_LIST(tensor1c), 0); |
198 | 1 | REQUIRE_TENSOR_EQ(tensor1c, tensor, "the first tensor should equal to the second"); |
199 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor2->data.f32, tensor->data.f32, 10, 1e-5, "the first 10 element should be equal"); |
200 | 1 | REQUIRE(ccv_nnc_tensor_nd(tensor2->info.dim) == 1, "should be 1-d tensor"); |
201 | 1 | REQUIRE_EQ(tensor2->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
202 | 1 | ccv_nnc_tensor_free(tensor1); |
203 | 1 | ccv_nnc_tensor_free(tensor1c); |
204 | 1 | ccv_nnc_tensor_free(tensor2); |
205 | 1 | ccv_nnc_tensor_free(tensor); |
206 | 1 | ccv_nnc_tensor_free(tensorG); |
207 | 1 | } |
208 | | |
209 | | TEST_CASE("tensor persistence with type coercion, to / from GPU") |
210 | 1 | { |
211 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
212 | 1 | sqlite3* handle; |
213 | 1 | sqlite3_open("tensors_tc_g.sqlite3", &handle); |
214 | 1 | ccv_nnc_tensor_t* const tensorf32 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
215 | 1 | int i; |
216 | 1 | dsfmt_t dsfmt; |
217 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
218 | 6.00k | for (i = 0; i < 10 * 20 * 30; i++6.00k ) |
219 | 6.00k | tensorf32->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
220 | 1 | ccv_nnc_tensor_t* const tensorf16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 20, 30), 0); |
221 | 1 | ccv_nnc_tensor_t* const tensorf16G = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 20, 30), 0); |
222 | 1 | ccv_float_to_half_precision(tensorf32->data.f32, (uint16_t*)tensorf16->data.f16, 10 * 20 * 30); |
223 | 6.00k | for (i = 0; i < 10 * 20 * 30; i++6.00k ) |
224 | 6.00k | tensorf32->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
225 | 1 | ccv_nnc_tensor_t* const tensorf32G = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 20, 30), 0); |
226 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensorf16, tensorf32), TENSOR_LIST(tensorf16G, tensorf32G), 0); |
227 | 1 | ccv_nnc_tensor_write(tensorf16G, handle, "x", 0); |
228 | 1 | ccv_nnc_tensor_write(tensorf32G, handle, "y", 0); |
229 | 1 | sqlite3_close(handle); |
230 | 1 | handle = 0; |
231 | 1 | sqlite3_open("tensors_tc_g.sqlite3", &handle); |
232 | 1 | ccv_nnc_tensor_t* tensor1 = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
233 | 1 | ccv_nnc_tensor_t* tensor1c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
234 | 1 | ccv_nnc_tensor_read(handle, "x", 0, 0, 0, &tensor1); |
235 | 1 | ccv_nnc_tensor_t* tensor2 = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
236 | 1 | ccv_nnc_tensor_t* tensor2c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
237 | 1 | ccv_nnc_tensor_read(handle, "y", 0, 0, 0, &tensor2); |
238 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor1, tensor2), TENSOR_LIST(tensor1c, tensor2c), 0); |
239 | 1 | sqlite3_close(handle); |
240 | 1 | float* tensor1_ref = (float*)ccmalloc(sizeof(float) * 10); |
241 | 1 | ccv_half_precision_to_float((uint16_t*)tensorf16->data.f16, tensor1_ref, 10); |
242 | 1 | float* tensor2_ret = (float*)ccmalloc(sizeof(float) * 10); |
243 | 1 | ccv_half_precision_to_float((uint16_t*)tensor2c->data.f16, tensor2_ret, 10); |
244 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor1c->data.f32, tensor1_ref, 10, 1e-3, "the first 10 element should be equal"); |
245 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor2_ret, tensorf32->data.f32, 10, 1e-3, "the first 10 element should be equal"); |
246 | 1 | REQUIRE(ccv_nnc_tensor_nd(tensor1->info.dim) == 1, "should be 1-d tensor"); |
247 | 1 | REQUIRE(ccv_nnc_tensor_nd(tensor2->info.dim) == 1, "should be 1-d tensor"); |
248 | 1 | REQUIRE_EQ(tensor1->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
249 | 1 | REQUIRE_EQ(tensor2->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
250 | 1 | ccv_nnc_tensor_free(tensor1); |
251 | 1 | ccv_nnc_tensor_free(tensor2); |
252 | 1 | ccv_nnc_tensor_free(tensor1c); |
253 | 1 | ccv_nnc_tensor_free(tensor2c); |
254 | 1 | ccv_nnc_tensor_free(tensorf16); |
255 | 1 | ccv_nnc_tensor_free(tensorf32); |
256 | 1 | ccv_nnc_tensor_free(tensorf16G); |
257 | 1 | ccv_nnc_tensor_free(tensorf32G); |
258 | 1 | ccfree(tensor1_ref); |
259 | 1 | ccfree(tensor2_ret); |
260 | 1 | } |
261 | | |
262 | | TEST_CASE("tensor persistence with type coercion and encoder / decoder, to / from GPU") |
263 | 1 | { |
264 | 1 | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
265 | 1 | sqlite3* handle; |
266 | 1 | sqlite3_open("tensors_tc_de_g.sqlite3", &handle); |
267 | 1 | ccv_nnc_tensor_t* const tensorf32 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
268 | 1 | int i; |
269 | 1 | dsfmt_t dsfmt; |
270 | 1 | dsfmt_init_gen_rand(&dsfmt, 1); |
271 | 6.00k | for (i = 0; i < 10 * 20 * 30; i++6.00k ) |
272 | 6.00k | tensorf32->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
273 | 1 | ccv_nnc_tensor_t* const tensorf16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 20, 30), 0); |
274 | 1 | ccv_float_to_half_precision(tensorf32->data.f32, (uint16_t*)tensorf16->data.f16, 10 * 20 * 30); |
275 | 6.00k | for (i = 0; i < 10 * 20 * 30; i++6.00k ) |
276 | 6.00k | tensorf32->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
277 | 1 | ccv_nnc_tensor_t* const tensorf32G = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 20, 30), 0); |
278 | 1 | ccv_nnc_tensor_t* const tensorf16G = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 20, 30), 0); |
279 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensorf16, tensorf32), TENSOR_LIST(tensorf16G, tensorf32G), 0); |
280 | 1 | ccv_nnc_tensor_io_option_t options = { |
281 | 1 | .encode = _tensor_xor_encode, |
282 | 1 | .decode = _tensor_xor_decode |
283 | 1 | }; |
284 | 1 | ccv_nnc_tensor_write(tensorf16G, handle, "x", &options); |
285 | 1 | ccv_nnc_tensor_write(tensorf32G, handle, "y", &options); |
286 | 1 | sqlite3_close(handle); |
287 | 1 | handle = 0; |
288 | 1 | sqlite3_open("tensors_tc_de_g.sqlite3", &handle); |
289 | 1 | ccv_nnc_tensor_t* tensor1 = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
290 | 1 | ccv_nnc_tensor_t* tensor1c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
291 | 1 | ccv_nnc_tensor_read(handle, "x", &options, 0, 0, &tensor1); |
292 | 1 | ccv_nnc_tensor_t* tensor2 = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
293 | 1 | ccv_nnc_tensor_t* tensor2c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
294 | 1 | ccv_nnc_tensor_read(handle, "y", &options, 0, 0, &tensor2); |
295 | 1 | sqlite3_close(handle); |
296 | 1 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor1, tensor2), TENSOR_LIST(tensor1c, tensor2c), 0); |
297 | 1 | float* tensor1_ref = (float*)ccmalloc(sizeof(float) * 10); |
298 | 1 | ccv_half_precision_to_float((uint16_t*)tensorf16->data.f16, tensor1_ref, 10); |
299 | 1 | float* tensor2_ret = (float*)ccmalloc(sizeof(float) * 10); |
300 | 1 | ccv_half_precision_to_float((uint16_t*)tensor2c->data.f16, tensor2_ret, 10); |
301 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor1c->data.f32, tensor1_ref, 10, 1e-3, "the first 10 element should be equal"); |
302 | 1 | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor2_ret, tensorf32->data.f32, 10, 1e-3, "the first 10 element should be equal"); |
303 | 1 | REQUIRE(ccv_nnc_tensor_nd(tensor1->info.dim) == 1, "should be 1-d tensor"); |
304 | 1 | REQUIRE(ccv_nnc_tensor_nd(tensor2->info.dim) == 1, "should be 1-d tensor"); |
305 | 1 | REQUIRE_EQ(tensor1->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
306 | 1 | REQUIRE_EQ(tensor2->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
307 | 1 | ccv_nnc_tensor_free(tensor1); |
308 | 1 | ccv_nnc_tensor_free(tensor2); |
309 | 1 | ccv_nnc_tensor_free(tensor1c); |
310 | 1 | ccv_nnc_tensor_free(tensor2c); |
311 | 1 | ccv_nnc_tensor_free(tensorf16); |
312 | 1 | ccv_nnc_tensor_free(tensorf32); |
313 | 1 | ccv_nnc_tensor_free(tensorf16G); |
314 | 1 | ccv_nnc_tensor_free(tensorf32G); |
315 | 1 | ccfree(tensor1_ref); |
316 | 1 | ccfree(tensor2_ret); |
317 | 1 | } |
318 | | |
319 | | TEST_CASE("tensor persistence with type coercion and noop encoder / decoder, to / from GPU") |
320 | | { |
321 | | GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_MPS) || ccv_nnc_cmd_ok(CCV_NNC_DATA_TRANSFER_FORWARD, CCV_NNC_BACKEND_GPU_REF)); |
322 | | sqlite3* handle; |
323 | | sqlite3_open("tensors_tc_noop_de_g.sqlite3", &handle); |
324 | | ccv_nnc_tensor_t* const tensorf32 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 20, 30), 0); |
325 | | int i; |
326 | | dsfmt_t dsfmt; |
327 | | dsfmt_init_gen_rand(&dsfmt, 1); |
328 | | for (i = 0; i < 10 * 20 * 30; i++) |
329 | | tensorf32->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
330 | | ccv_nnc_tensor_t* const tensorf16 = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 20, 30), 0); |
331 | | ccv_float_to_half_precision(tensorf32->data.f32, (uint16_t*)tensorf16->data.f16, 10 * 20 * 30); |
332 | | for (i = 0; i < 10 * 20 * 30; i++) |
333 | | tensorf32->data.f32[i] = dsfmt_genrand_open_close(&dsfmt) * 2 - 1; |
334 | | ccv_nnc_tensor_t* const tensorf32G = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 20, 30), 0); |
335 | | ccv_nnc_tensor_t* const tensorf16G = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 20, 30), 0); |
336 | | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensorf16, tensorf32), TENSOR_LIST(tensorf16G, tensorf32G), 0); |
337 | | ccv_nnc_tensor_io_option_t options = { |
338 | | .encode = _tensor_noop_encode, |
339 | | .decode = _tensor_noop_decode |
340 | | }; |
341 | | ccv_nnc_tensor_write(tensorf16G, handle, "x", &options); |
342 | | ccv_nnc_tensor_write(tensorf32G, handle, "y", &options); |
343 | | sqlite3_close(handle); |
344 | | handle = 0; |
345 | | sqlite3_open("tensors_tc_noop_de_g.sqlite3", &handle); |
346 | | ccv_nnc_tensor_t* tensor1 = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10), 0); |
347 | | ccv_nnc_tensor_t* tensor1c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10), 0); |
348 | | ccv_nnc_tensor_read(handle, "x", &options, 0, 0, &tensor1); |
349 | | ccv_nnc_tensor_t* tensor2 = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10), 0); |
350 | | ccv_nnc_tensor_t* tensor2c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10), 0); |
351 | | ccv_nnc_tensor_read(handle, "y", &options, 0, 0, &tensor2); |
352 | | sqlite3_close(handle); |
353 | | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor1, tensor2), TENSOR_LIST(tensor1c, tensor2c), 0); |
354 | | float* tensor1_ref = (float*)ccmalloc(sizeof(float) * 10); |
355 | | ccv_half_precision_to_float((uint16_t*)tensorf16->data.f16, tensor1_ref, 10); |
356 | | float* tensor2_ret = (float*)ccmalloc(sizeof(float) * 10); |
357 | | ccv_half_precision_to_float((uint16_t*)tensor2c->data.f16, tensor2_ret, 10); |
358 | | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor1c->data.f32, tensor1_ref, 10, 1e-3, "the first 10 element should be equal"); |
359 | | REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tensor2_ret, tensorf32->data.f32, 10, 1e-3, "the first 10 element should be equal"); |
360 | | REQUIRE(ccv_nnc_tensor_nd(tensor1->info.dim) == 1, "should be 1-d tensor"); |
361 | | REQUIRE(ccv_nnc_tensor_nd(tensor2->info.dim) == 1, "should be 1-d tensor"); |
362 | | REQUIRE_EQ(tensor1->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
363 | | REQUIRE_EQ(tensor2->info.dim[0], 10, "should be 1-d tensor with 10-element"); |
364 | | ccv_nnc_tensor_free(tensor1); |
365 | | ccv_nnc_tensor_free(tensor2); |
366 | | ccv_nnc_tensor_free(tensor1c); |
367 | | ccv_nnc_tensor_free(tensor2c); |
368 | | ccv_nnc_tensor_free(tensorf16); |
369 | | ccv_nnc_tensor_free(tensorf32); |
370 | | ccv_nnc_tensor_free(tensorf16G); |
371 | | ccv_nnc_tensor_free(tensorf32G); |
372 | | ccfree(tensor1_ref); |
373 | | ccfree(tensor2_ret); |
374 | | } |
375 | | |
376 | | #include "case_main.h" |