| File: | nnc/ccv_cnnp_model.c |
| Warning: | line 2507, column 1 Assigned value is garbage or undefined |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | #include "ccv_nnc.h" |
| 2 | #include "ccv_nnc_easy.h" |
| 3 | #include "ccv_nnc_internal.h" |
| 4 | #include "ccv_internal.h" |
| 5 | #include "_ccv_cnnp_model.h" |
| 6 | #include "_ccv_nnc_graph.h" |
| 7 | #include "_ccv_nnc_symbolic_graph.h" |
| 8 | #ifdef HAVE_CUDA1 |
| 9 | #include "gpu/ccv_nnc_compat.h" |
| 10 | #endif |
| 11 | |
| 12 | // MARK - Level-5 API |
| 13 | |
| 14 | ccv_cnnp_model_io_t ccv_cnnp_model_apply(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t* const inputs, const int input_size) |
| 15 | { |
| 16 | if (!model->io) |
| 17 | model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0); |
| 18 | ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s) + sizeof(ccv_nnc_tensor_symbol_t) * model->output_size); |
| 19 | model_io->param_ref = 0; |
| 20 | model_io->param_sel = 0; |
| 21 | model_io->visit = 0; |
| 22 | model_io->model = model; |
| 23 | model_io->dependencies = 0; |
| 24 | model_io->dependents = 0; |
| 25 | model_io->outgoings = 0; |
| 26 | model_io->outputs = (ccv_nnc_tensor_symbol_t*)(model_io + 1); |
| 27 | ccv_array_push(model->io, &model_io); |
| 28 | if (input_size > 0) |
| 29 | { |
| 30 | model_io->incomings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), input_size, 0); |
| 31 | ccv_array_resize(model_io->incomings, input_size); |
| 32 | int i; |
| 33 | memcpy(ccv_array_get(model_io->incomings, 0)((void*)(((char*)((model_io->incomings)->data)) + (size_t )(model_io->incomings)->rsize * (size_t)(0))), inputs, sizeof(ccv_cnnp_model_io_t) * input_size); |
| 34 | for (i = 0; i < input_size; i++) |
| 35 | { |
| 36 | if (!inputs[i]->outgoings) |
| 37 | inputs[i]->outgoings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0); |
| 38 | ccv_array_push(inputs[i]->outgoings, &model_io); |
| 39 | } |
| 40 | } else { |
| 41 | model_io->incomings = 0; |
| 42 | } |
| 43 | return model_io; |
| 44 | } |
| 45 | |
| 46 | void ccv_cnnp_model_add_dependencies(ccv_cnnp_model_io_t model_io, const ccv_cnnp_model_io_t* const dependencies, const int dependency_size) |
| 47 | { |
| 48 | assert(dependency_size > 0)((void) sizeof ((dependency_size > 0) ? 1 : 0), __extension__ ({ if (dependency_size > 0) ; else __assert_fail ("dependency_size > 0" , "ccv_cnnp_model.c", 48, __extension__ __PRETTY_FUNCTION__); })); |
| 49 | if (!model_io->dependencies) |
| 50 | model_io->dependencies = ccv_array_new(sizeof(ccv_cnnp_model_io_t), dependency_size, 0); |
| 51 | int i, j; |
| 52 | for (i = 0; i < dependency_size; i++) |
| 53 | { |
| 54 | int flag = 0; |
| 55 | // Check if it is already exist or not. |
| 56 | for (j = 0; !flag && j < model_io->dependencies->rnum; j++) |
| 57 | if (*(ccv_cnnp_model_io_t*)ccv_array_get(model_io->dependencies, j)((void*)(((char*)((model_io->dependencies)->data)) + (size_t )(model_io->dependencies)->rsize * (size_t)(j))) == dependencies[i]) |
| 58 | flag = 1; |
| 59 | if (flag) |
| 60 | continue; |
| 61 | ccv_array_push(model_io->dependencies, dependencies + i); |
| 62 | ++dependencies[i]->dependents; |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | int ccv_cnnp_model_output_size(const ccv_cnnp_model_t* const model) |
| 67 | { |
| 68 | return model->output_size; |
| 69 | } |
| 70 | |
| 71 | int ccv_cnnp_model_is_trainable(const ccv_cnnp_model_t* const model) |
| 72 | { |
| 73 | // If the model is compiled, it is default to 1 unless it is not. |
| 74 | if (model->compiled_data) |
| 75 | return model->is_trainable >= 0 ? model->is_trainable : 1; |
| 76 | return model->is_trainable; |
| 77 | } |
| 78 | |
| 79 | ccv_cnnp_model_io_t ccv_cnnp_model_parameters(ccv_cnnp_model_t* const model, const int selector, const int index) |
| 80 | { |
| 81 | if (!model->io) |
| 82 | model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0); |
| 83 | ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s)); |
| 84 | model_io->param_ref = index >= 0 ? index + 1 : ALL_PARAMETERS-1; |
| 85 | model_io->param_sel = selector >= 0 ? selector + 1 : ALL_PARAMETERS-1; |
| 86 | model_io->visit = 0; |
| 87 | model_io->model = model; |
| 88 | model_io->outputs = 0; |
| 89 | model_io->dependencies = 0; |
| 90 | model_io->dependents = 0; |
| 91 | model_io->incomings = 0; |
| 92 | model_io->outgoings = 0; |
| 93 | ccv_array_push(model->io, &model_io); |
| 94 | return model_io; |
| 95 | } |
| 96 | |
| 97 | void ccv_cnnp_model_notify_hook(ccv_cnnp_model_t* const model, ccv_cnnp_model_notify_f func, void* const context) |
| 98 | { |
| 99 | model->notify_hook.func = func; |
| 100 | model->notify_hook.context = context; |
| 101 | } |
| 102 | |
| 103 | void ccv_cnnp_model_notify(const ccv_cnnp_model_t* const model, const int tag, void* const payload) |
| 104 | { |
| 105 | if (model->notify_hook.func) |
| 106 | model->notify_hook.func(model, tag, payload, model->notify_hook.context); |
| 107 | if (model->isa->notify) |
| 108 | model->isa->notify(model, tag, payload); |
| 109 | } |
| 110 | |
| 111 | static int _ccv_nnc_array_dedup_graph_exec_symbols(ccv_nnc_graph_exec_symbol_t* const graph_exec_symbols, int graph_exec_symbol_size) |
| 112 | { |
| 113 | int i, j; |
| 114 | for (i = 0; i < graph_exec_symbol_size; i++) |
| 115 | { |
| 116 | ccv_nnc_graph_exec_symbol_t* const graph_exec_symbol = graph_exec_symbols + i; |
| 117 | // Check whether this tensor symbol has any duplicate. |
| 118 | for (j = i + 1; j < graph_exec_symbol_size;) |
| 119 | { |
| 120 | ccv_nnc_graph_exec_symbol_t* const other_symbol = graph_exec_symbols + j; |
| 121 | // If there is a same tensor symbol, remove it. |
| 122 | if (other_symbol->d == graph_exec_symbol->d && other_symbol->graph == graph_exec_symbol->graph) |
| 123 | { |
| 124 | if (j + 1 < graph_exec_symbol_size) |
| 125 | *other_symbol = graph_exec_symbols[graph_exec_symbol_size - 1]; |
| 126 | --graph_exec_symbol_size; |
| 127 | continue; |
| 128 | } |
| 129 | ++j; |
| 130 | } |
| 131 | } |
| 132 | return graph_exec_symbol_size; |
| 133 | } |
| 134 | |
| 135 | void ccv_cnnp_model_add_to_array(void* const context, const ccv_nnc_tensor_symbol_t symbol, const int is_trainable) |
| 136 | { |
| 137 | ccv_cnnp_model_add_to_array_context_t* const add_to_array_context = (ccv_cnnp_model_add_to_array_context_t*)context; |
| 138 | ccv_cnnp_model_t* const model = add_to_array_context->sequence->model; |
| 139 | int i; |
| 140 | if (add_to_array_context->add_parameter_indices && !model->parameter_indices) |
| 141 | model->parameter_indices = ccv_array_new(sizeof(int), 0, 0); |
| 142 | for (i = 0; i < add_to_array_context->symbols->rnum; i++) |
| 143 | { |
| 144 | const ccv_nnc_tensor_symbol_t other_symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(add_to_array_context->symbols, i)((void*)(((char*)((add_to_array_context->symbols)->data )) + (size_t)(add_to_array_context->symbols)->rsize * ( size_t)(i))); |
| 145 | if (other_symbol.d == symbol.d && other_symbol.graph == symbol.graph) |
| 146 | { |
| 147 | // Only add to parameter_indices if it is trainable. |
| 148 | if (add_to_array_context->add_parameter_indices) |
| 149 | ccv_array_add_unique_int(model->parameter_indices, i); |
| 150 | // Found it, return, don't add it. |
| 151 | return; |
| 152 | } |
| 153 | } |
| 154 | // Only add to parameter_indices if it is trainable. |
| 155 | if (add_to_array_context->add_parameter_indices) |
| 156 | ccv_array_push(model->parameter_indices, &add_to_array_context->symbols->rnum); |
| 157 | // This is a new one, no need to add_unique_int, it is unique. |
| 158 | ccv_array_push(add_to_array_context->symbols, &symbol); |
| 159 | if (add_to_array_context->trainables) |
| 160 | ccv_array_push(add_to_array_context->trainables, &is_trainable); |
| 161 | char id[2048]; |
| 162 | id[0] = add_to_array_context->prefix; |
| 163 | id[1] = '-'; |
| 164 | int total_len = 2; |
| 165 | for (i = 0; i < add_to_array_context->sequence->sequences->rnum; i++) |
| 166 | { |
| 167 | const ccv_cnnp_model_name_t* const name = (ccv_cnnp_model_name_t*)ccv_array_get(add_to_array_context->sequence->sequences, i)((void*)(((char*)((add_to_array_context->sequence->sequences )->data)) + (size_t)(add_to_array_context->sequence-> sequences)->rsize * (size_t)(i))); |
| 168 | int len; |
| 169 | if (name->name && name->name[0] != '\0') |
| 170 | len = snprintf(id + total_len, 2048 - total_len, "%s-%d-", name->name, name->sequence); |
| 171 | else |
| 172 | len = snprintf(id + total_len, 2048 - total_len, "%d-", name->sequence); |
| 173 | total_len += len; |
| 174 | if (total_len >= 2047) |
| 175 | break; |
| 176 | } |
| 177 | if (total_len < 2047) |
| 178 | total_len += snprintf(id + total_len, 2048 - total_len, "%d", add_to_array_context->sequence->it); |
| 179 | assert(total_len < 2048)((void) sizeof ((total_len < 2048) ? 1 : 0), __extension__ ({ if (total_len < 2048) ; else __assert_fail ("total_len < 2048" , "ccv_cnnp_model.c", 179, __extension__ __PRETTY_FUNCTION__) ; })); |
| 180 | char *heap_id = (char*)ccmallocmalloc(total_len + 1); |
| 181 | memcpy(heap_id, id, total_len + 1); |
| 182 | ccv_array_push(add_to_array_context->ids, &heap_id); |
| 183 | ++add_to_array_context->sequence->it; |
| 184 | } |
| 185 | |
| 186 | static void _ccv_cnnp_compiled_data_init(ccv_cnnp_compiled_data_t* const compiled_data, const int output_size, ccv_array_t* const gradient_checkpoints) |
| 187 | { |
| 188 | compiled_data->f = compiled_data->fits + output_size; |
| 189 | compiled_data->xpu_alloc.mp_hdr = -1; |
| 190 | compiled_data->xpu_alloc.freed = kh_init(dy_str)kh_init_dy_str(); |
| 191 | compiled_data->xpu_alloc.allocd = kh_init(dy_alloc)kh_init_dy_alloc(); |
| 192 | compiled_data->gradient_checkpoints = gradient_checkpoints; |
| 193 | } |
| 194 | |
| 195 | static int _ccv_cnnp_model_root_parallel_count(const ccv_cnnp_model_t* const model) |
| 196 | { |
| 197 | return ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 198 | } |
| 199 | |
| 200 | static int _ccv_cnnp_model_effective_parallel_count(const ccv_cnnp_model_t* const model) |
| 201 | { |
| 202 | int parallel_count = _ccv_cnnp_model_root_parallel_count(model); |
| 203 | if (model->graph && model->graph->data_parallel.count > parallel_count) |
| 204 | parallel_count = model->graph->data_parallel.count; |
| 205 | return parallel_count; |
| 206 | } |
| 207 | |
| 208 | static int _ccv_cnnp_compiled_data_parallel_count(const ccv_cnnp_model_t* const model, const ccv_cnnp_compiled_data_t* const compiled_data) |
| 209 | { |
| 210 | return compiled_data->parallel_count > 0 ? compiled_data->parallel_count : _ccv_cnnp_model_effective_parallel_count(model); |
| 211 | } |
| 212 | |
| 213 | ccv_nnc_tensor_symbol_t ccv_cnnp_model_get_symbol(ccv_cnnp_model_t* const self, const ccv_nnc_tensor_symbol_t symbol) |
| 214 | { |
| 215 | assert(self->data)((void) sizeof ((self->data) ? 1 : 0), __extension__ ({ if (self->data) ; else __assert_fail ("self->data", "ccv_cnnp_model.c" , 215, __extension__ __PRETTY_FUNCTION__); })); |
| 216 | ccv_cnnp_model_build_data_t* const build_data = (ccv_cnnp_model_build_data_t*)self->data; |
| 217 | if (build_data->parallel_count <= 1 || build_data->parallel_rank == 0) |
| 218 | return symbol; |
| 219 | const int rank = build_data->parallel_rank; |
| 220 | assert(rank > 0)((void) sizeof ((rank > 0) ? 1 : 0), __extension__ ({ if ( rank > 0) ; else __assert_fail ("rank > 0", "ccv_cnnp_model.c" , 220, __extension__ __PRETTY_FUNCTION__); })); |
| 221 | assert(rank < build_data->parallel_count)((void) sizeof ((rank < build_data->parallel_count) ? 1 : 0), __extension__ ({ if (rank < build_data->parallel_count ) ; else __assert_fail ("rank < build_data->parallel_count" , "ccv_cnnp_model.c", 221, __extension__ __PRETTY_FUNCTION__) ; })); |
| 222 | ccv_nnc_symbolic_graph_t* const graph = (ccv_nnc_symbolic_graph_t*)symbol.graph; |
| 223 | ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, symbol, rank); |
| 224 | if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 225 | return copy; |
| 226 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, symbol); |
| 227 | if (CCV_TENSOR_GET_MEMORY(params.type)((params.type) & 0x3) == CCV_TENSOR_GPU_MEMORY) |
| 228 | CCV_TENSOR_SET_DEVICE_ID(params.type, rank)(params.type) = (((params.type) & ~0xfff00) | (((rank) & 0xfff) << 8)); |
| 229 | copy = ccv_nnc_tensor_symbol_new(graph, params, 0); |
| 230 | ccv_nnc_tensor_symbol_set_copy(graph, symbol, rank, copy); |
| 231 | return copy; |
| 232 | } |
| 233 | |
| 234 | typedef struct { |
| 235 | void* old_graph_exec_symbol_new_hook_context; |
| 236 | ccv_nnc_graph_exec_symbol_new_hook_f old_graph_exec_symbol_new_hook; |
| 237 | ccv_nnc_symbolic_graph_t* graph; |
| 238 | ccv_cnnp_model_build_data_t* build_data; |
| 239 | } ccv_cnnp_model_set_exec_flags_context_t; |
| 240 | |
| 241 | static void _ccv_cnnp_model_set_exec_flags(void* context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, const ccv_nnc_tensor_symbol_t* const outputs, const int output_size, const char* const name) |
| 242 | { |
| 243 | ccv_cnnp_model_set_exec_flags_context_t* flags_context = (ccv_cnnp_model_set_exec_flags_context_t*)context; |
| 244 | if (flags_context->build_data->exec_flags) |
| 245 | ccv_nnc_graph_exec_symbol_set_flags(flags_context->graph, symbol, flags_context->build_data->exec_flags); |
| 246 | if (flags_context->old_graph_exec_symbol_new_hook) |
| 247 | flags_context->old_graph_exec_symbol_new_hook(flags_context->old_graph_exec_symbol_new_hook_context, symbol, cmd, inputs, input_size, outputs, output_size, name); |
| 248 | } |
| 249 | |
| 250 | static void _ccv_cnnp_model_compile(ccv_cnnp_model_t* const model, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_cmd_t loss) |
| 251 | { |
| 252 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 252, __extension__ __PRETTY_FUNCTION__); })); |
| 253 | model->inputs = ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * input_size); |
| 254 | int i; |
| 255 | for (i = 0; i < input_size; i++) |
| 256 | model->inputs[i] = ccv_nnc_tensor_symbol_new(model->graph, inputs[i], 0); |
| 257 | ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); |
| 258 | ccv_array_t* const parameter_ids = ccv_array_new(sizeof(char*), 0, 0); |
| 259 | ccv_array_t* const parameter_trainables = ccv_array_new(sizeof(int), 0, 0); |
| 260 | ccv_cnnp_model_sequence_t model_sequence = { |
| 261 | .bank = kh_init(ccv_cnnp_model_name_bank)kh_init_ccv_cnnp_model_name_bank() |
| 262 | }; |
| 263 | ccv_cnnp_model_add_to_array_context_t add_to_parameter_context = { |
| 264 | .add_parameter_indices = 1, |
| 265 | .prefix = 't', |
| 266 | .sequence = &model_sequence, |
| 267 | .symbols = parameters, |
| 268 | .ids = parameter_ids, |
| 269 | .trainables = parameter_trainables, |
| 270 | }; |
| 271 | ccv_array_t* const internals = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); |
| 272 | ccv_array_t* const internal_ids = ccv_array_new(sizeof(char*), 0, 0); |
| 273 | ccv_cnnp_model_add_to_array_context_t add_to_output_context = { |
| 274 | .add_parameter_indices = 0, |
| 275 | .prefix = 'r', |
| 276 | .sequence = &model_sequence, |
| 277 | .symbols = internals, |
| 278 | .ids = internal_ids, |
| 279 | .trainables = 0, |
| 280 | }; |
| 281 | ccv_cnnp_model_build_data_t build_data = { |
| 282 | .exec_flags = 0, |
| 283 | .is_trainable = model->is_trainable >= 0 ? model->is_trainable : 1, |
| 284 | .parallel_count = 1, |
| 285 | .parallel_rank = 0, |
| 286 | .model_sequence = &model_sequence, |
| 287 | .add_to_array = ccv_cnnp_model_add_to_array, |
| 288 | .parameters = parameters, |
| 289 | .context = { |
| 290 | .add_to_parameter = &add_to_parameter_context, |
| 291 | .add_to_output = &add_to_output_context, |
| 292 | }, |
| 293 | .gradient_checkpoints = 0, |
| 294 | }; |
| 295 | model->data = &build_data; |
| 296 | ccv_cnnp_model_set_exec_flags_context_t flags_context = { |
| 297 | .graph = model->graph, |
| 298 | .build_data = &build_data, |
| 299 | .old_graph_exec_symbol_new_hook = 0, |
| 300 | .old_graph_exec_symbol_new_hook_context = 0 |
| 301 | }; |
| 302 | flags_context.old_graph_exec_symbol_new_hook_context = ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_set_exec_flags, &flags_context, &flags_context.old_graph_exec_symbol_new_hook); |
| 303 | ccv_cnnp_model_build(model, model->graph, model->inputs, input_size, 0, 0); |
| 304 | // Reset back to previous hook. |
| 305 | ccv_nnc_graph_exec_symbol_new_hook(model->graph, flags_context.old_graph_exec_symbol_new_hook, flags_context.old_graph_exec_symbol_new_hook_context, 0); |
| 306 | for (i = 0; i < model->output_size; i++) |
| 307 | { |
| 308 | const ccv_nnc_tensor_symbol_t output = model->outputs[i]; |
| 309 | const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, output); |
| 310 | if (alias_to.d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 311 | continue; |
| 312 | // If output is an alias, insert data transform regardless for result correctness (we cannot bind an alias). You can check ccv_nnc_tensor_bind_symbol method |
| 313 | // to see that we can correctly bind a tensor which from it, has aliases, but we cannot bind an alias tensor correctly (this is expected, sort of, to be |
| 314 | // honest, because we cannot handle cases of alias is part of the original tensor but bind differently). |
| 315 | const ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(model->graph, output); |
| 316 | model->outputs[i] = ccv_nnc_tensor_symbol_new(model->graph, output_params, 0); |
| 317 | ccv_nnc_graph_exec_symbol_t make_contiguous = ccv_nnc_graph_exec_symbol_new(model->graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), &output, 1, model->outputs + i, 1, "contiguous"); |
| 318 | ccv_nnc_graph_exec_symbol_set_flags(model->graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT); |
| 319 | } |
| 320 | model->data = 0; |
| 321 | kh_destroy(ccv_cnnp_model_name_bank, model_sequence.bank)kh_destroy_ccv_cnnp_model_name_bank(model_sequence.bank); |
| 322 | if (model_sequence.sequences) |
| 323 | ccv_array_free(model_sequence.sequences); |
| 324 | // Check if there are parameters that are not trainables. If there are, we will allocate uint64 bitmap to record that. |
| 325 | int not_trainables = 0; |
| 326 | // Assert no parameter is alias. |
| 327 | for (i = 0; i < parameters->rnum; i++) |
| 328 | { |
| 329 | const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(parameters, i)((void*)(((char*)((parameters)->data)) + (size_t)(parameters )->rsize * (size_t)(i))); |
| 330 | const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(parameter.graph, parameter); |
| 331 | assert(alias_to.graph == 0)((void) sizeof ((alias_to.graph == 0) ? 1 : 0), __extension__ ({ if (alias_to.graph == 0) ; else __assert_fail ("alias_to.graph == 0" , "ccv_cnnp_model.c", 331, __extension__ __PRETTY_FUNCTION__) ; })); // Cannot find the one alias to. |
| 332 | if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t )(parameter_trainables)->rsize * (size_t)(i))) == 0) |
| 333 | not_trainables = 1; |
| 334 | } |
| 335 | assert(parameters->rnum == parameter_trainables->rnum)((void) sizeof ((parameters->rnum == parameter_trainables-> rnum) ? 1 : 0), __extension__ ({ if (parameters->rnum == parameter_trainables ->rnum) ; else __assert_fail ("parameters->rnum == parameter_trainables->rnum" , "ccv_cnnp_model.c", 335, __extension__ __PRETTY_FUNCTION__) ; })); |
| 336 | uint64_t* parameter_flags = 0; |
| 337 | if (not_trainables) |
| 338 | { |
| 339 | parameter_flags = (uint64_t*)cccalloccalloc(((parameters->rnum + 63) >> 6), sizeof(uint64_t)); |
| 340 | for (i = 0; i < parameter_trainables->rnum; i++) |
| 341 | if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t )(parameter_trainables)->rsize * (size_t)(i)))) |
| 342 | parameter_flags[i >> 6] |= ((uint64_t)1 << (i & 63)); |
| 343 | } |
| 344 | ccv_array_free(parameter_trainables); |
| 345 | // Assert no internal is alias. |
| 346 | for (i = 0; i < internals->rnum; i++) |
| 347 | { |
| 348 | const ccv_nnc_tensor_symbol_t internal = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(internals, i)((void*)(((char*)((internals)->data)) + (size_t)(internals )->rsize * (size_t)(i))); |
| 349 | const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(internal.graph, internal); |
| 350 | assert(alias_to.graph == 0)((void) sizeof ((alias_to.graph == 0) ? 1 : 0), __extension__ ({ if (alias_to.graph == 0) ; else __assert_fail ("alias_to.graph == 0" , "ccv_cnnp_model.c", 350, __extension__ __PRETTY_FUNCTION__) ; })); // Cannot find the one alias to. |
| 351 | } |
| 352 | const int output_size = model->output_size; |
| 353 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 354 | const int parameters_rnum = parameters->rnum; |
| 355 | if (input_size > 0) |
| 356 | { |
| 357 | ccv_array_resize(parameters, parameters_rnum + input_size); |
| 358 | memcpy(ccv_array_get(parameters, parameters_rnum)((void*)(((char*)((parameters)->data)) + (size_t)(parameters )->rsize * (size_t)(parameters_rnum))), model->inputs, input_size * sizeof(ccv_nnc_tensor_symbol_t)); |
| 359 | } |
| 360 | ccv_nnc_symbolic_graph_simplify(model->graph, |
| 361 | SYMBOLIC_GRAPH_PASSES(CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION , CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION , CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1) |
| 362 | CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION , CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION , CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1) |
| 363 | CCV_NNC_SIMPLIFY_OPS_FUSION,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION , CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION , CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1) |
| 364 | CCV_NNC_SIMPLIFY_GRAPH_PRUNING)(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION , CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION , CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), |
| 365 | ccv_array_get(parameters, 0)((void*)(((char*)((parameters)->data)) + (size_t)(parameters )->rsize * (size_t)(0))), parameters_rnum + input_size, |
| 366 | model->outputs, output_size, |
| 367 | SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph)); |
| 368 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 369 | // Size it down. |
| 370 | parameters->rnum = parameters_rnum; |
| 371 | ccv_cnnp_compiled_data_t* compiled_data = model->compiled_data = cccalloccalloc(1, sizeof(ccv_cnnp_compiled_data_t) + sizeof(ccv_nnc_tensor_symbol_t) * (output_size * 2 - 1)); |
| 372 | _ccv_cnnp_compiled_data_init(compiled_data, output_size, build_data.gradient_checkpoints); |
| 373 | const int evaluate_to_size = compiled_data->evaluate.to_size = ccv_nnc_symbolic_graph_destination_size(model->graph); |
| 374 | assert(evaluate_to_size > 0)((void) sizeof ((evaluate_to_size > 0) ? 1 : 0), __extension__ ({ if (evaluate_to_size > 0) ; else __assert_fail ("evaluate_to_size > 0" , "ccv_cnnp_model.c", 374, __extension__ __PRETTY_FUNCTION__) ; })); |
| 375 | compiled_data->evaluate.tos = ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size); |
| 376 | memcpy(compiled_data->evaluate.tos, ccv_nnc_symbolic_graph_destinations(model->graph), sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size); |
| 377 | compiled_data->loss = loss; |
| 378 | if (loss.cmd == CCV_NNC_NOOP) |
| 379 | { |
| 380 | // If no loss function provided, there is no fits. |
| 381 | for (i = 0; i < output_size; i++) |
| 382 | { |
| 383 | compiled_data->fits[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 384 | const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, model->outputs[i]); |
| 385 | if (alias_to.d < 0) |
| 386 | compiled_data->f[i] = model->outputs[i]; |
| 387 | else { // We cannot differentiate against an alias, therefore, we have to verify this output is full, and we can diff against the original. |
| 388 | int ofs[CCV_NNC_MAX_DIM_ALLOC(12)]; |
| 389 | int inc[CCV_NNC_MAX_DIM_ALLOC(12)]; |
| 390 | ccv_nnc_tensor_symbol_alias_params(model->graph, model->outputs[i], ofs, inc); |
| 391 | int j; |
| 392 | for (j = 0; j < CCV_NNC_MAX_DIM_ALLOC(12); j++) |
| 393 | { assert(ofs[j] == 0)((void) sizeof ((ofs[j] == 0) ? 1 : 0), __extension__ ({ if ( ofs[j] == 0) ; else __assert_fail ("ofs[j] == 0", "ccv_cnnp_model.c" , 393, __extension__ __PRETTY_FUNCTION__); })); } // There is no ofs. |
| 394 | compiled_data->f[i] = alias_to; // Unfortunately, I cannot assert the size yet. |
| 395 | } |
| 396 | } |
| 397 | } else { |
| 398 | for (i = 0; i < output_size; i++) |
| 399 | { |
| 400 | const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(model->graph, model->outputs[i]); |
| 401 | const ccv_nnc_tensor_symbol_t fit = compiled_data->fits[i] = ccv_nnc_tensor_symbol_new(model->graph, info, 0); |
| 402 | compiled_data->f[i] = ccv_nnc_tensor_symbol_new(model->graph, ccv_nnc_tensor_auto, 0); |
| 403 | ccv_nnc_graph_exec_symbol_new(model->graph, loss, TENSOR_SYMBOL_LIST(model->outputs[i], fit)(const ccv_nnc_tensor_symbol_t []){model->outputs[i], fit} , (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(compiled_data->f[i])(const ccv_nnc_tensor_symbol_t []){compiled_data->f[i]}, ( 1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0); |
| 404 | } |
| 405 | } |
| 406 | if (loss.cmd != CCV_NNC_NOOP) |
| 407 | { |
| 408 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 409 | ccv_nnc_symbolic_graph_simplify(model->graph, |
| 410 | SYMBOLIC_GRAPH_PASSES(CCV_NNC_SIMPLIFY_OPS_FUSION)(const int []){CCV_NNC_SIMPLIFY_OPS_FUSION}, (1 +1 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), // Only do Ops fusion, in this way, we can fuse the loss function. |
| 411 | 0, 0, // No need to provide binds at this point. |
| 412 | compiled_data->f, model->output_size, |
| 413 | SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph)); |
| 414 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 415 | } |
| 416 | // If inputs are from GPU, stream type is GPU. |
| 417 | compiled_data->parameters = parameters; |
| 418 | compiled_data->parameter_flags = parameter_flags; |
| 419 | compiled_data->internals = internals; |
| 420 | compiled_data->ids.parameters = parameter_ids; |
| 421 | compiled_data->ids.internals = internal_ids; |
| 422 | ccv_cnnp_model_gradient_checkpoints_cleanup_after_build(compiled_data, model->graph); |
| 423 | } |
| 424 | |
| 425 | static void _ccv_cnnp_graph_push_graph_exec_symbol(void* context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, const ccv_nnc_tensor_symbol_t* const outputs, const int output_size, const char* const name) |
| 426 | { |
| 427 | ccv_array_t* const stack = (ccv_array_t*)context; |
| 428 | ccv_array_push(stack, &symbol.d); |
| 429 | } |
| 430 | |
| 431 | static void _ccv_nnc_tensor_symbol_reinit(const ccv_nnc_symbolic_graph_t* const src_graph, ccv_nnc_symbolic_graph_t* const dest_graph, const int src_index, const int dest_index) |
| 432 | { |
| 433 | int src_alias_ref = src_index; |
| 434 | const ccv_nnc_tensor_symbol_info_t* src_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(src_graph->tensor_symbol_info, src_alias_ref)((void*)(((char*)((src_graph->tensor_symbol_info)->data )) + (size_t)(src_graph->tensor_symbol_info)->rsize * ( size_t)(src_alias_ref))); |
| 435 | while (src_info->alias_ref) |
| 436 | { |
| 437 | src_alias_ref = src_info->alias_ref - 1; |
| 438 | src_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(src_graph->tensor_symbol_info, src_alias_ref)((void*)(((char*)((src_graph->tensor_symbol_info)->data )) + (size_t)(src_graph->tensor_symbol_info)->rsize * ( size_t)(src_alias_ref))); |
| 439 | } |
| 440 | int dest_alias_ref = dest_index; |
| 441 | const ccv_nnc_tensor_symbol_info_t* dest_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(dest_graph->tensor_symbol_info, dest_alias_ref)((void*)(((char*)((dest_graph->tensor_symbol_info)->data )) + (size_t)(dest_graph->tensor_symbol_info)->rsize * ( size_t)(dest_alias_ref))); |
| 442 | while (dest_info->alias_ref) |
| 443 | { |
| 444 | dest_alias_ref = dest_info->alias_ref - 1; |
| 445 | dest_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(dest_graph->tensor_symbol_info, dest_alias_ref)((void*)(((char*)((dest_graph->tensor_symbol_info)->data )) + (size_t)(dest_graph->tensor_symbol_info)->rsize * ( size_t)(dest_alias_ref))); |
| 446 | } |
| 447 | assert((src_alias_ref != src_index) == (dest_alias_ref != dest_index))((void) sizeof (((src_alias_ref != src_index) == (dest_alias_ref != dest_index)) ? 1 : 0), __extension__ ({ if ((src_alias_ref != src_index) == (dest_alias_ref != dest_index)) ; else __assert_fail ("(src_alias_ref != src_index) == (dest_alias_ref != dest_index)" , "ccv_cnnp_model.c", 447, __extension__ __PRETTY_FUNCTION__) ; })); |
| 448 | if (src_alias_ref != src_index) |
| 449 | { |
| 450 | const ccv_nnc_tensor_symbol_t src_alias_symbol = { |
| 451 | .d = src_alias_ref, |
| 452 | .graph = src_graph |
| 453 | }; |
| 454 | const ccv_nnc_tensor_symbol_t dest_alias_symbol = { |
| 455 | .d = dest_alias_ref, |
| 456 | .graph = dest_graph |
| 457 | }; |
| 458 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_alias_symbol); |
| 459 | ccv_nnc_tensor_symbol_set(dest_graph, dest_alias_symbol, params); |
| 460 | } |
| 461 | const ccv_nnc_tensor_symbol_t src_symbol = { |
| 462 | .d = src_index, |
| 463 | .graph = src_graph |
| 464 | }; |
| 465 | const ccv_nnc_tensor_symbol_t dest_symbol = { |
| 466 | .d = dest_index, |
| 467 | .graph = dest_graph |
| 468 | }; |
| 469 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol); |
| 470 | ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params); |
| 471 | int ofs[CCV_NNC_MAX_DIM_ALLOC(12)]; |
| 472 | int inc[CCV_NNC_MAX_DIM_ALLOC(12)]; |
| 473 | if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc)) |
| 474 | ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc); |
| 475 | } |
| 476 | |
| 477 | static int _ccv_nnc_tensor_symbol_check_dim(const ccv_nnc_symbolic_graph_t* const src_graph, ccv_nnc_symbolic_graph_t* const dest_graph, const int src_index, const int dest_index) |
| 478 | { |
| 479 | const ccv_nnc_tensor_symbol_t src_symbol = { |
| 480 | .d = src_index, |
| 481 | .graph = src_graph |
| 482 | }; |
| 483 | const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol); |
| 484 | const ccv_nnc_tensor_symbol_t dest_symbol = { |
| 485 | .d = dest_index, |
| 486 | .graph = dest_graph |
| 487 | }; |
| 488 | const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol); |
| 489 | if (src_params.dim[0] == 0 || dest_params.dim[0] == 0) |
| 490 | return 1; |
| 491 | return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0; |
| 492 | } |
| 493 | |
| 494 | static void _ccv_cnnp_model_gradient_init(ccv_cnnp_model_t* const model, const int gradient_mode, const uint64_t disable_outgrad, ccv_nnc_tensor_t* const* const fits, const int fit_size); |
| 495 | static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data); |
| 496 | |
| 497 | typedef struct { |
| 498 | int parallel_count; |
| 499 | ccv_nnc_symbolic_graph_t* graph; |
| 500 | ccv_nnc_graph_exec_arena_t* graph_exec_arena; |
| 501 | } ccv_nnc_graph_exec_update_t; |
| 502 | |
| 503 | static void _ccv_cnnp_cmd_update_for_execs(void* const context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint) |
| 504 | { |
| 505 | ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context; |
| 506 | ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena; |
| 507 | ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol); |
| 508 | ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd); |
| 509 | ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint); |
| 510 | const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph; |
| 511 | const int parallel_count = graph_exec_update->parallel_count; |
| 512 | int i; |
| 513 | for (i = 1; i < parallel_count; i++) |
| 514 | { |
| 515 | const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, ccv_nnc_graph_exec_symbol_copy(graph, symbol, i)); |
| 516 | if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0)) |
| 517 | { |
| 518 | ccv_nnc_graph_exec_set(copy.graph, copy, cmd); |
| 519 | ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint); |
| 520 | } |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | void ccv_cnnp_model_absorb(ccv_cnnp_model_t* const model, ccv_cnnp_model_t* const init, const ccv_nnc_tensor_param_t* const inputs, const int input_size) |
| 525 | { |
| 526 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 526, __extension__ __PRETTY_FUNCTION__); })); |
| 527 | assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__ ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data" , "ccv_cnnp_model.c", 527, __extension__ __PRETTY_FUNCTION__) ; })); |
| 528 | assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c" , 528, __extension__ __PRETTY_FUNCTION__); })); |
| 529 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 530 | init->graph = ccv_nnc_symbolic_graph_new(); |
| 531 | ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0); |
| 532 | ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0); |
| 533 | _ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss); |
| 534 | init->parallel_count = model->parallel_count; |
| 535 | init->memory_compression = model->memory_compression; |
| 536 | init->memory_reduction = model->memory_reduction; |
| 537 | init->gradient_checkpointing = model->gradient_checkpointing; |
| 538 | init->compiled_data->stream_type = model->compiled_data->stream_type; |
| 539 | init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer; |
| 540 | init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size; |
| 541 | if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE) |
| 542 | _ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0); |
| 543 | ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0); |
| 544 | ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0); |
| 545 | int i, j; |
| 546 | // Verify parameters, internals and saved_aux in both graph has the same dimensionality. |
| 547 | for (i = 0; i < compiled_data->parameters->rnum; i++) |
| 548 | { |
| 549 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( i))))->d; |
| 550 | assert(_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d))((void) sizeof ((_ccv_nnc_tensor_symbol_check_dim(model->graph , init->graph, d, d)) ? 1 : 0), __extension__ ({ if (_ccv_nnc_tensor_symbol_check_dim (model->graph, init->graph, d, d)) ; else __assert_fail ("_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d)" , "ccv_cnnp_model.c", 550, __extension__ __PRETTY_FUNCTION__) ; })); |
| 551 | } |
| 552 | for (i = 0; i < compiled_data->internals->rnum; i++) |
| 553 | { |
| 554 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(i)) ))->d; |
| 555 | assert(_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d))((void) sizeof ((_ccv_nnc_tensor_symbol_check_dim(model->graph , init->graph, d, d)) ? 1 : 0), __extension__ ({ if (_ccv_nnc_tensor_symbol_check_dim (model->graph, init->graph, d, d)) ; else __assert_fail ("_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d)" , "ccv_cnnp_model.c", 555, __extension__ __PRETTY_FUNCTION__) ; })); |
| 556 | } |
| 557 | // Update inputs. |
| 558 | assert(model->input_size == init->input_size)((void) sizeof ((model->input_size == init->input_size) ? 1 : 0), __extension__ ({ if (model->input_size == init-> input_size) ; else __assert_fail ("model->input_size == init->input_size" , "ccv_cnnp_model.c", 558, __extension__ __PRETTY_FUNCTION__) ; })); |
| 559 | for (i = 0; i < model->input_size; i++) |
| 560 | if (model->inputs[i].d >= 0) |
| 561 | { |
| 562 | assert(init->inputs[i].d >= 0)((void) sizeof ((init->inputs[i].d >= 0) ? 1 : 0), __extension__ ({ if (init->inputs[i].d >= 0) ; else __assert_fail ("init->inputs[i].d >= 0" , "ccv_cnnp_model.c", 562, __extension__ __PRETTY_FUNCTION__) ; })); |
| 563 | _ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d); |
| 564 | } |
| 565 | // Update outputs. |
| 566 | assert(model->output_size == init->output_size)((void) sizeof ((model->output_size == init->output_size ) ? 1 : 0), __extension__ ({ if (model->output_size == init ->output_size) ; else __assert_fail ("model->output_size == init->output_size" , "ccv_cnnp_model.c", 566, __extension__ __PRETTY_FUNCTION__) ; })); |
| 567 | for (i = 0; i < model->output_size; i++) |
| 568 | { |
| 569 | if (model->outputs[i].d >= 0) |
| 570 | { |
| 571 | assert(init->outputs[i].d >= 0)((void) sizeof ((init->outputs[i].d >= 0) ? 1 : 0), __extension__ ({ if (init->outputs[i].d >= 0) ; else __assert_fail ( "init->outputs[i].d >= 0", "ccv_cnnp_model.c", 571, __extension__ __PRETTY_FUNCTION__); })); |
| 572 | _ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d); |
| 573 | } |
| 574 | if (model->outputs[i].d != model->compiled_data->f[i].d) |
| 575 | { |
| 576 | assert(init->outputs[i].d != init->compiled_data->f[i].d)((void) sizeof ((init->outputs[i].d != init->compiled_data ->f[i].d) ? 1 : 0), __extension__ ({ if (init->outputs[ i].d != init->compiled_data->f[i].d) ; else __assert_fail ("init->outputs[i].d != init->compiled_data->f[i].d" , "ccv_cnnp_model.c", 576, __extension__ __PRETTY_FUNCTION__) ; })); |
| 577 | if (model->compiled_data->f[i].d >= 0) |
| 578 | { |
| 579 | assert(init->compiled_data->f[i].d >= 0)((void) sizeof ((init->compiled_data->f[i].d >= 0) ? 1 : 0), __extension__ ({ if (init->compiled_data->f[i] .d >= 0) ; else __assert_fail ("init->compiled_data->f[i].d >= 0" , "ccv_cnnp_model.c", 579, __extension__ __PRETTY_FUNCTION__) ; })); |
| 580 | _ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d); |
| 581 | } |
| 582 | } |
| 583 | } |
| 584 | // Go through the graph to set tensor on matching symbols |
| 585 | for (i = 0; i < stack->rnum; i++) |
| 586 | { |
| 587 | const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize * (size_t)(i))); |
| 588 | // If exceed range, skip. |
| 589 | if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) || |
| 590 | d >= ccv_nnc_graph_exec_symbol_count(model->graph)) |
| 591 | continue; |
| 592 | const ccv_nnc_graph_exec_symbol_t src_symbol = { |
| 593 | .d = d, |
| 594 | .graph = init->graph |
| 595 | }; |
| 596 | const ccv_nnc_graph_exec_symbol_t dest_symbol = { |
| 597 | .d = d, |
| 598 | .graph = model->graph |
| 599 | }; |
| 600 | const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol); |
| 601 | const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol); |
| 602 | // If the name doesn't match, skip. |
| 603 | if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP) |
| 604 | continue; |
| 605 | // Now get all the inputs and outputs, if matches, set them. |
| 606 | const int* src_inputs; |
| 607 | int src_input_size; |
| 608 | const int* src_outputs; |
| 609 | int src_output_size; |
| 610 | ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size); |
| 611 | const int* dest_inputs; |
| 612 | int dest_input_size; |
| 613 | const int* dest_outputs; |
| 614 | int dest_output_size; |
| 615 | ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size); |
| 616 | // We may have unmatched input / output size because this is the minimizer and it has |
| 617 | // different saved_aux (for example, when we shrunk with CMD_NOOP). |
| 618 | if (src_input_size != dest_input_size) |
| 619 | continue; |
| 620 | if (src_output_size != dest_output_size) |
| 621 | continue; |
| 622 | ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd); |
| 623 | // There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because |
| 624 | // we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original |
| 625 | // graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That |
| 626 | // will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as |
| 627 | // well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec |
| 628 | // symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not |
| 629 | // a new exec symbol. |
| 630 | for (j = 0; j < src_input_size; j++) |
| 631 | if (src_inputs[j] >= 0) |
| 632 | _ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]); |
| 633 | for (j = 0; j < src_output_size; j++) |
| 634 | if (src_outputs[j] >= 0) |
| 635 | _ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]); |
| 636 | } |
| 637 | ccv_array_free(stack); |
| 638 | // After this, we get all tensors in the model graph resolved through tensor_auto. |
| 639 | ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0); |
| 640 | // Verify symbols we get matches. |
| 641 | const int parameter_size = compiled_data->parameters->rnum; |
| 642 | for (i = 0; i < parameter_size; i++) |
| 643 | { assert(((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->parameters, i))->d)((void) sizeof ((((ccv_nnc_tensor_symbol_t*)((void*)(((char*) ((compiled_data->parameters)->data)) + (size_t)(compiled_data ->parameters)->rsize * (size_t)(i))))->d == ((ccv_nnc_tensor_symbol_t *)((void*)(((char*)((init->compiled_data->parameters)-> data)) + (size_t)(init->compiled_data->parameters)-> rsize * (size_t)(i))))->d) ? 1 : 0), __extension__ ({ if ( ((ccv_nnc_tensor_symbol_t*)((void*)(((char*)((compiled_data-> parameters)->data)) + (size_t)(compiled_data->parameters )->rsize * (size_t)(i))))->d == ((ccv_nnc_tensor_symbol_t *)((void*)(((char*)((init->compiled_data->parameters)-> data)) + (size_t)(init->compiled_data->parameters)-> rsize * (size_t)(i))))->d) ; else __assert_fail ("((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->parameters, i))->d" , "ccv_cnnp_model.c", 643, __extension__ __PRETTY_FUNCTION__) ; })); } |
| 644 | const int internal_size = compiled_data->internals->rnum; |
| 645 | for (i = 0; i < internal_size; i++) |
| 646 | { assert(((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->internals, i))->d)((void) sizeof ((((ccv_nnc_tensor_symbol_t*)((void*)(((char*) ((compiled_data->internals)->data)) + (size_t)(compiled_data ->internals)->rsize * (size_t)(i))))->d == ((ccv_nnc_tensor_symbol_t *)((void*)(((char*)((init->compiled_data->internals)-> data)) + (size_t)(init->compiled_data->internals)->rsize * (size_t)(i))))->d) ? 1 : 0), __extension__ ({ if (((ccv_nnc_tensor_symbol_t *)((void*)(((char*)((compiled_data->internals)->data)) + (size_t)(compiled_data->internals)->rsize * (size_t)(i ))))->d == ((ccv_nnc_tensor_symbol_t*)((void*)(((char*)((init ->compiled_data->internals)->data)) + (size_t)(init-> compiled_data->internals)->rsize * (size_t)(i))))->d ) ; else __assert_fail ("((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->internals, i))->d" , "ccv_cnnp_model.c", 646, __extension__ __PRETTY_FUNCTION__) ; })); } |
| 647 | // Go through compiled data. |
| 648 | if (compiled_data->tensor_arena) |
| 649 | { |
| 650 | const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph); |
| 651 | if (flag == 0 && compiled_data->graph_exec_arena) |
| 652 | { |
| 653 | ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph); |
| 654 | // Since we will reinit, if we previously set is_test, we need to set it again. |
| 655 | if (compiled_data->is_test) |
| 656 | { |
| 657 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 658 | ccv_nnc_graph_exec_update_t update = { |
| 659 | .parallel_count = parallel_count, |
| 660 | .graph = model->graph, |
| 661 | .graph_exec_arena = compiled_data->graph_exec_arena, |
| 662 | }; |
| 663 | ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update); |
| 664 | } |
| 665 | } else |
| 666 | // Free-up tensor arena & graph exec arena. |
| 667 | _ccv_cnnp_compiled_data_graph_free(compiled_data); |
| 668 | } |
| 669 | // There are other compiled graphs, for accum and apply gradients. |
| 670 | // However, the main conclusion is, these absorb operations shouldn't impact parameters. |
| 671 | // Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we |
| 672 | // don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot |
| 673 | // be changed otherwise parameters' shape will be meaningless. The same goes to internals. |
| 674 | // That is why we don't update these compiled graphs at all this point. |
| 675 | // Free the model, we've already "absorbed" it. |
| 676 | ccv_cnnp_model_free(init); |
| 677 | } |
| 678 | |
| 679 | void ccv_cnnp_model_compile(ccv_cnnp_model_t* const model, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_cmd_t minimizer, const ccv_nnc_cmd_t loss) |
| 680 | { |
| 681 | assert(input_size == model->input_size || model->input_size == 0)((void) sizeof ((input_size == model->input_size || model-> input_size == 0) ? 1 : 0), __extension__ ({ if (input_size == model->input_size || model->input_size == 0) ; else __assert_fail ("input_size == model->input_size || model->input_size == 0" , "ccv_cnnp_model.c", 681, __extension__ __PRETTY_FUNCTION__) ; })); |
| 682 | if (model->input_size == 0) |
| 683 | model->input_size = input_size; |
| 684 | if (!model->graph) // The graph is not compiled yet. |
| 685 | { |
| 686 | model->graph = ccv_nnc_symbolic_graph_new(); |
| 687 | _ccv_cnnp_model_compile(model, inputs, input_size, loss); |
| 688 | assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__ ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data" , "ccv_cnnp_model.c", 688, __extension__ __PRETTY_FUNCTION__) ; })); |
| 689 | int i, flag = 0; |
| 690 | for (i = 0; !flag && i < input_size; i++) |
| 691 | flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY); |
| 692 | // If inputs are from GPU, stream type is GPU. |
| 693 | model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU; |
| 694 | model->compiled_data->minimize.minimizer = minimizer; |
| 695 | model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer); |
| 696 | } else { |
| 697 | // Now, finally fill in this part. If the graph is already compiled, we make a copy of the model. |
| 698 | // And then absorb the "new model" to the old one. |
| 699 | ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable); |
| 700 | ccv_cnnp_model_absorb(model, init, inputs, input_size); |
| 701 | // Reset minimizer. |
| 702 | ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0); |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable) |
| 707 | { |
| 708 | ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0); |
| 709 | new_model->is_trainable = is_trainable; |
| 710 | return new_model; |
| 711 | } |
| 712 | |
| 713 | void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size) |
| 714 | { |
| 715 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 715, __extension__ __PRETTY_FUNCTION__); })); |
| 716 | assert(output_size == model->output_size)((void) sizeof ((output_size == model->output_size) ? 1 : 0 ), __extension__ ({ if (output_size == model->output_size) ; else __assert_fail ("output_size == model->output_size" , "ccv_cnnp_model.c", 716, __extension__ __PRETTY_FUNCTION__) ; })); |
| 717 | ccv_nnc_symbolic_graph_t* const graph = model->graph; |
| 718 | ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0); |
| 719 | int i; |
| 720 | for (i = 0; i < output_size; i++) |
| 721 | { |
| 722 | assert(model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL)((void) sizeof ((model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL ) ? 1 : 0), __extension__ ({ if (model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL ) ; else __assert_fail ("model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL" , "ccv_cnnp_model.c", 722, __extension__ __PRETTY_FUNCTION__) ; })); |
| 723 | outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]); |
| 724 | } |
| 725 | } |
| 726 | |
| 727 | void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size) |
| 728 | { |
| 729 | if (workspace_size == model->workspace_size) |
| 730 | return; |
| 731 | model->workspace_size = workspace_size; |
| 732 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 733 | if (compiled_data && compiled_data->graph) |
| 734 | ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0); |
| 735 | } |
| 736 | |
| 737 | size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model) |
| 738 | { |
| 739 | return model->workspace_size; |
| 740 | } |
| 741 | |
| 742 | void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel) |
| 743 | { |
| 744 | if (parallel == 0) |
| 745 | model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU); |
| 746 | else |
| 747 | model->parallel_count = parallel; |
| 748 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 749 | if (compiled_data) |
| 750 | { assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__ ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph" , "ccv_cnnp_model.c", 750, __extension__ __PRETTY_FUNCTION__) ; })); } |
| 751 | } |
| 752 | |
| 753 | void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count) |
| 754 | { |
| 755 | model->max_stream_count = max_stream_count; |
| 756 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 757 | if (compiled_data) |
| 758 | { assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__ ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph" , "ccv_cnnp_model.c", 758, __extension__ __PRETTY_FUNCTION__) ; })); } |
| 759 | } |
| 760 | |
| 761 | void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression) |
| 762 | { |
| 763 | model->memory_compression = memory_compression; |
| 764 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 765 | if (compiled_data) |
| 766 | { assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__ ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph" , "ccv_cnnp_model.c", 766, __extension__ __PRETTY_FUNCTION__) ; })); } |
| 767 | } |
| 768 | |
| 769 | void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction) |
| 770 | { |
| 771 | model->memory_reduction = memory_reduction; |
| 772 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 773 | if (compiled_data) |
| 774 | { assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__ ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph" , "ccv_cnnp_model.c", 774, __extension__ __PRETTY_FUNCTION__) ; })); } |
| 775 | } |
| 776 | |
| 777 | void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing) |
| 778 | { |
| 779 | model->gradient_checkpointing = gradient_checkpointing; |
| 780 | } |
| 781 | |
| 782 | int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model) |
| 783 | { |
| 784 | return model->gradient_checkpointing; |
| 785 | } |
| 786 | |
| 787 | typedef struct { |
| 788 | int parallel_count; |
| 789 | ccv_nnc_symbolic_graph_t* graph; |
| 790 | ccv_cnnp_compiled_data_t* compiled_data; |
| 791 | ccv_nnc_tensor_arena_t* tensor_arena; |
| 792 | } ccv_nnc_tensor_init_states_t; |
| 793 | |
| 794 | static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data) |
| 795 | { |
| 796 | int i; |
| 797 | const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 798 | for (i = 0; i < compiled_data->parameters->rnum; i++) |
| 799 | { |
| 800 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( i))))->d; |
| 801 | if (!(init_v[d >> 5] & (1u << (d & 0x1f)))) |
| 802 | return 1; |
| 803 | } |
| 804 | for (i = 0; i < compiled_data->internals->rnum; i++) |
| 805 | { |
| 806 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(i)) ))->d; |
| 807 | if (!(init_v[d >> 5] & (1u << (d & 0x1f)))) |
| 808 | return 1; |
| 809 | } |
| 810 | return 0; |
| 811 | } |
| 812 | |
| 813 | static void _ccv_cnnp_init_states_for_tensors(void* const context, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const input, const ccv_nnc_tensor_symbol_t output_symbol) |
| 814 | { |
| 815 | ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context; |
| 816 | ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena; |
| 817 | ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol); |
| 818 | if (!output_tensor) |
| 819 | return; |
| 820 | const int d = output_symbol.d; |
| 821 | assert(d < tensor_init_states->compiled_data->tensors_init.size)((void) sizeof ((d < tensor_init_states->compiled_data-> tensors_init.size) ? 1 : 0), __extension__ ({ if (d < tensor_init_states ->compiled_data->tensors_init.size) ; else __assert_fail ("d < tensor_init_states->compiled_data->tensors_init.size" , "ccv_cnnp_model.c", 821, __extension__ __PRETTY_FUNCTION__) ; })); |
| 822 | uint32_t* const init_v = CCV_NNC_INIT_V(tensor_init_states->compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(tensor_init_states->compiled_data ->tensors_init.v) & ~(uintptr_t)1)); |
| 823 | if (init_v[d >> 5] & (1u << (d & 0x1f))) |
| 824 | return; |
| 825 | init_v[d >> 5] |= (1u << (d & 0x1f)); |
| 826 | ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0); |
| 827 | const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph; |
| 828 | const int parallel_count = tensor_init_states->parallel_count; |
| 829 | int i; |
| 830 | for (i = 1; i < parallel_count; i++) |
| 831 | { |
| 832 | ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i)); |
| 833 | if (copy) |
| 834 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, &output_tensor, 1, ©, 1, 0); |
| 835 | } |
| 836 | } |
| 837 | |
| 838 | // This method can only handle cases we added new tensors and exec, never delete. This invariant is true because |
| 839 | // we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup. |
| 840 | static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model) |
| 841 | { |
| 842 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 842, __extension__ __PRETTY_FUNCTION__); })); |
| 843 | assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__ ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data" , "ccv_cnnp_model.c", 843, __extension__ __PRETTY_FUNCTION__) ; })); |
| 844 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 845 | assert(compiled_data->rewindables)((void) sizeof ((compiled_data->rewindables) ? 1 : 0), __extension__ ({ if (compiled_data->rewindables) ; else __assert_fail ( "compiled_data->rewindables", "ccv_cnnp_model.c", 845, __extension__ __PRETTY_FUNCTION__); })); |
| 846 | int i; |
| 847 | for (i = 0; i < compiled_data->rewindables->rnum; i++) |
| 848 | { |
| 849 | const ccv_cnnp_rewind_symbol_t* const rewind_symbol = (ccv_cnnp_rewind_symbol_t*)ccv_array_get(compiled_data->rewindables, i)((void*)(((char*)((compiled_data->rewindables)->data)) + (size_t)(compiled_data->rewindables)->rsize * (size_t) (i))); |
| 850 | if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC) |
| 851 | ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec); |
| 852 | else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR) |
| 853 | ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor); |
| 854 | } |
| 855 | ccv_array_clear(compiled_data->rewindables); |
| 856 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 857 | } |
| 858 | |
| 859 | static void _ccv_cnnp_model_tensor_symbol_new_hook(void* context, const ccv_nnc_tensor_symbol_t symbol, const ccv_nnc_tensor_param_t info, const char* const name) |
| 860 | { |
| 861 | const ccv_cnnp_rewind_symbol_t rewind_symbol = { |
| 862 | .type = CCV_CNNP_REWIND_TENSOR, |
| 863 | .tensor = symbol |
| 864 | }; |
| 865 | ccv_array_t* const rewind_symbols = (ccv_array_t*)context; |
| 866 | ccv_array_push(rewind_symbols, &rewind_symbol); |
| 867 | } |
| 868 | |
| 869 | static void _ccv_cnnp_model_tensor_symbol_alias_new_hook(void* context, const ccv_nnc_tensor_symbol_t symbol, const ccv_nnc_tensor_symbol_t from_symbol, const int ofs[CCV_NNC_MAX_DIM_ALLOC(12)], const int inc[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_tensor_param_t info, const char* const name) |
| 870 | { |
| 871 | const ccv_cnnp_rewind_symbol_t rewind_symbol = { |
| 872 | .type = CCV_CNNP_REWIND_TENSOR, |
| 873 | .tensor = symbol |
| 874 | }; |
| 875 | ccv_array_t* const rewind_symbols = (ccv_array_t*)context; |
| 876 | ccv_array_push(rewind_symbols, &rewind_symbol); |
| 877 | } |
| 878 | |
| 879 | static void _ccv_cnnp_model_graph_exec_symbol_new_hook(void* context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, const ccv_nnc_tensor_symbol_t* const outputs, const int output_size, const char* const name) |
| 880 | { |
| 881 | const ccv_cnnp_rewind_symbol_t rewind_symbol = { |
| 882 | .type = CCV_CNNP_REWIND_GRAPH_EXEC, |
| 883 | .graph_exec = symbol |
| 884 | }; |
| 885 | ccv_array_t* const rewind_symbols = (ccv_array_t*)context; |
| 886 | ccv_array_push(rewind_symbols, &rewind_symbol); |
| 887 | } |
| 888 | |
| 889 | static void _ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(const ccv_nnc_graph_exec_arena_t* const graph_exec_arena, const int parallel_count, const ccv_nnc_graph_exec_symbol_t exec_symbol, const ccv_nnc_cmd_t cmd, ccv_nnc_symbolic_graph_t* const symbolic_graph) |
| 890 | { |
| 891 | ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol); |
| 892 | if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0)) |
| 893 | ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd); |
| 894 | int i; |
| 895 | for (i = 1; i < parallel_count; i++) |
| 896 | { |
| 897 | ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i); |
| 898 | const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol); |
| 899 | if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0)) |
| 900 | ccv_nnc_graph_exec_set(copy.graph, copy, cmd); |
| 901 | } |
| 902 | } |
| 903 | |
| 904 | static void _ccv_cnnp_model_graph_exec_symbol_set(ccv_nnc_symbolic_graph_t* const symbolic_graph, ccv_cnnp_compiled_data_t* const compiled_data, const int parallel_count, const ccv_nnc_graph_exec_symbol_t exec_symbol, const ccv_nnc_cmd_t cmd) |
| 905 | { |
| 906 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 906, __extension__ __PRETTY_FUNCTION__); })); |
| 907 | assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c" , 907, __extension__ __PRETTY_FUNCTION__); })); |
| 908 | ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd); |
| 909 | int i; |
| 910 | for (i = 1; i < parallel_count; i++) |
| 911 | { |
| 912 | ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i); |
| 913 | if (copy_symbol.graph) |
| 914 | ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd); |
| 915 | } |
| 916 | ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena; |
| 917 | if (graph_exec_arena) |
| 918 | _ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph); |
| 919 | // Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph) |
| 920 | ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena; |
| 921 | if (gradient_graph_exec_arena) |
| 922 | _ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph); |
| 923 | } |
| 924 | |
| 925 | static int _ccv_cnnp_set_minimizer_for_parameter(ccv_nnc_symbolic_graph_t* const graph, ccv_cnnp_compiled_data_t* const compiled_data, ccv_nnc_graph_exec_symbol_t* const update_nodes, ccv_nnc_tensor_symbol_t* const updated_parameters, ccv_nnc_tensor_symbol_map_t* const saved_aux, const int parallel_count, const ccv_nnc_cmd_t minimizer, const int saved_aux_size, const int max_saved_aux_size, const int parameter_indice) |
| 926 | { |
| 927 | int this_parameter_flag = 0; |
| 928 | if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 929 | return this_parameter_flag; |
| 930 | const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]); |
| 931 | int j, k; |
| 932 | // For no-op, we can preserve previous saved_aux_size. |
| 933 | if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP) |
| 934 | { |
| 935 | // If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous |
| 936 | // saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between |
| 937 | // noop and a minimizer. We don't want that because we do that in high-level frameworks to |
| 938 | // make sure some model parameters don't update if we don't want them to. |
| 939 | int old_saved_aux_size; |
| 940 | if (old_minimizer.cmd == CCV_NNC_NOOP) |
| 941 | { |
| 942 | int input_size; |
| 943 | ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0); |
| 944 | if (input_size < 2) // This is not legit. |
| 945 | old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer); |
| 946 | else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters. |
| 947 | old_saved_aux_size = input_size - 2; |
| 948 | } else |
| 949 | old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer); |
| 950 | if (old_saved_aux_size != saved_aux_size) |
| 951 | { |
| 952 | this_parameter_flag = 1; |
| 953 | if (saved_aux_size > old_saved_aux_size) |
| 954 | { |
| 955 | // Allocate new tensor symbols. |
| 956 | const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]); |
| 957 | for (j = old_saved_aux_size; j < saved_aux_size; j++) |
| 958 | { |
| 959 | saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0); |
| 960 | saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0); |
| 961 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8); |
| 962 | for (k = 1; k < parallel_count; k++) |
| 963 | { |
| 964 | ccv_nnc_tensor_param_t dev_info = info; |
| 965 | if (k != device_id) |
| 966 | CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) & 0xfff) << 8)); |
| 967 | else |
| 968 | CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) & 0xfff) << 8)); |
| 969 | const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0); |
| 970 | const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0); |
| 971 | ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy); |
| 972 | ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy); |
| 973 | } |
| 974 | } |
| 975 | } else { |
| 976 | for (j = saved_aux_size; j < old_saved_aux_size; j++) |
| 977 | { |
| 978 | for (k = 1; k < parallel_count; k++) |
| 979 | { |
| 980 | const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k); |
| 981 | if (src_copy.d >= 0) |
| 982 | { |
| 983 | ccv_nnc_tensor_symbol_free(graph, src_copy); |
| 984 | ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); |
| 985 | } |
| 986 | const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k); |
| 987 | if (dest_copy.d >= 0) |
| 988 | { |
| 989 | ccv_nnc_tensor_symbol_free(graph, dest_copy); |
| 990 | ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); |
| 991 | } |
| 992 | } |
| 993 | ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source); |
| 994 | ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination); |
| 995 | saved_aux[parameter_indice * max_saved_aux_size + j].source = saved_aux[parameter_indice * max_saved_aux_size + j].destination = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 996 | } |
| 997 | } |
| 998 | } |
| 999 | } |
| 1000 | _ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer); |
| 1001 | if (this_parameter_flag) |
| 1002 | { |
| 1003 | ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2]; |
| 1004 | ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1]; |
| 1005 | const int* inputs = 0; |
| 1006 | int input_size = 0; |
| 1007 | ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0); |
| 1008 | assert(input_size >= 1)((void) sizeof ((input_size >= 1) ? 1 : 0), __extension__ ( { if (input_size >= 1) ; else __assert_fail ("input_size >= 1" , "ccv_cnnp_model.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1009 | update_inputs[0].d = inputs[0]; |
| 1010 | update_inputs[0].graph = graph; |
| 1011 | update_inputs[1].d = inputs[1]; |
| 1012 | update_inputs[1].graph = graph; |
| 1013 | update_outputs[0] = updated_parameters[parameter_indice]; |
| 1014 | for (j = 0; j < saved_aux_size; j++) |
| 1015 | { |
| 1016 | update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source; |
| 1017 | update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination; |
| 1018 | } |
| 1019 | ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1); |
| 1020 | for (k = 1; k < parallel_count; k++) |
| 1021 | { |
| 1022 | const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k); |
| 1023 | assert(copy.d >= 0)((void) sizeof ((copy.d >= 0) ? 1 : 0), __extension__ ({ if (copy.d >= 0) ; else __assert_fail ("copy.d >= 0", "ccv_cnnp_model.c" , 1023, __extension__ __PRETTY_FUNCTION__); })); |
| 1024 | ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0); |
| 1025 | assert(input_size >= 1)((void) sizeof ((input_size >= 1) ? 1 : 0), __extension__ ( { if (input_size >= 1) ; else __assert_fail ("input_size >= 1" , "ccv_cnnp_model.c", 1025, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1026 | update_inputs[0].d = inputs[0]; |
| 1027 | update_inputs[0].graph = graph; |
| 1028 | update_inputs[1].d = inputs[1]; |
| 1029 | update_inputs[1].graph = graph; |
| 1030 | update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k); |
| 1031 | for (j = 0; j < saved_aux_size; j++) |
| 1032 | { |
| 1033 | update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k); |
| 1034 | update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k); |
| 1035 | } |
| 1036 | ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1); |
| 1037 | } |
| 1038 | } |
| 1039 | return this_parameter_flag; |
| 1040 | } |
| 1041 | |
| 1042 | typedef struct { |
| 1043 | int parameter_size; |
| 1044 | ccv_nnc_cmd_t minimizer; |
| 1045 | ccv_cnnp_model_io_t parameters[1]; |
| 1046 | } ccv_cnnp_set_minimizer_for_parameter_t; |
| 1047 | |
| 1048 | static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model) |
| 1049 | { |
| 1050 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1051 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 1051, __extension__ __PRETTY_FUNCTION__); })); |
| 1052 | const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size; |
| 1053 | // We update all parameters, at this point, we have one minimizer. |
| 1054 | const int parameter_size = compiled_data->parameters->rnum; |
| 1055 | ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes; |
| 1056 | ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph; |
| 1057 | assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c" , 1057, __extension__ __PRETTY_FUNCTION__); })); |
| 1058 | const int parallel_count = _ccv_cnnp_model_root_parallel_count(model); |
| 1059 | assert(_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now")((void) sizeof ((_ccv_cnnp_model_effective_parallel_count(model ) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now" ) ? 1 : 0), __extension__ ({ if (_ccv_cnnp_model_effective_parallel_count (model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now" ) ; else __assert_fail ("_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && \"local replicated stateful models only support forward / no-grad evaluation for now\"" , "ccv_cnnp_model.c", 1059, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1060 | ccv_array_t* const parameters = compiled_data->minimize.parameters; |
| 1061 | ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0); |
| 1062 | int i, j, flag = 0; |
| 1063 | for (i = 0; i < parameters->rnum; i++) |
| 1064 | { |
| 1065 | ccv_cnnp_set_minimizer_for_parameter_t* const set_minimizer_for_parameter = *(ccv_cnnp_set_minimizer_for_parameter_t**)ccv_array_get(parameters, i)((void*)(((char*)((parameters)->data)) + (size_t)(parameters )->rsize * (size_t)(i))); |
| 1066 | for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++) |
| 1067 | { |
| 1068 | const int param_sel = set_minimizer_for_parameter->parameters[j]->param_sel > 0 ? set_minimizer_for_parameter->parameters[j]->param_sel - 1 : set_minimizer_for_parameter->parameters[j]->param_sel; |
| 1069 | assert(set_minimizer_for_parameter->parameters[j]->param_sel != 0)((void) sizeof ((set_minimizer_for_parameter->parameters[j ]->param_sel != 0) ? 1 : 0), __extension__ ({ if (set_minimizer_for_parameter ->parameters[j]->param_sel != 0) ; else __assert_fail ( "set_minimizer_for_parameter->parameters[j]->param_sel != 0" , "ccv_cnnp_model.c", 1069, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1070 | const int old_rnum = parameter_indices->rnum; |
| 1071 | ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices); |
| 1072 | const int param_ref = set_minimizer_for_parameter->parameters[j]->param_ref > 0 ? set_minimizer_for_parameter->parameters[j]->param_ref - 1 : set_minimizer_for_parameter->parameters[j]->param_ref; |
| 1073 | assert(set_minimizer_for_parameter->parameters[j]->param_ref != 0)((void) sizeof ((set_minimizer_for_parameter->parameters[j ]->param_ref != 0) ? 1 : 0), __extension__ ({ if (set_minimizer_for_parameter ->parameters[j]->param_ref != 0) ; else __assert_fail ( "set_minimizer_for_parameter->parameters[j]->param_ref != 0" , "ccv_cnnp_model.c", 1073, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1074 | if (param_ref >= 0) |
| 1075 | { |
| 1076 | assert(param_ref + old_rnum < parameter_indices->rnum)((void) sizeof ((param_ref + old_rnum < parameter_indices-> rnum) ? 1 : 0), __extension__ ({ if (param_ref + old_rnum < parameter_indices->rnum) ; else __assert_fail ("param_ref + old_rnum < parameter_indices->rnum" , "ccv_cnnp_model.c", 1076, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1077 | *(int*)ccv_array_get(parameter_indices, old_rnum)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(old_rnum))) = *(int*)ccv_array_get(parameter_indices, param_ref + old_rnum)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(param_ref + old_rnum))); |
| 1078 | parameter_indices->rnum = old_rnum + 1; |
| 1079 | } |
| 1080 | } |
| 1081 | const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer); |
| 1082 | // We may have duplicated indices, but that is OK, we will set it twice. |
| 1083 | for (j = 0; j < parameter_indices->rnum; j++) |
| 1084 | { |
| 1085 | const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(j))); |
| 1086 | assert(d <= parameter_size)((void) sizeof ((d <= parameter_size) ? 1 : 0), __extension__ ({ if (d <= parameter_size) ; else __assert_fail ("d <= parameter_size" , "ccv_cnnp_model.c", 1086, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1087 | if (_ccv_cnnp_set_minimizer_for_parameter(symbolic_graph, compiled_data, update_nodes, compiled_data->updated_parameters, compiled_data->saved_aux, parallel_count, set_minimizer_for_parameter->minimizer, saved_aux_size, max_saved_aux_size, d)) |
| 1088 | flag = 1; |
| 1089 | } |
| 1090 | ccv_array_clear(parameter_indices); |
| 1091 | } |
| 1092 | ccv_array_free(parameter_indices); |
| 1093 | return flag; |
| 1094 | } |
| 1095 | |
| 1096 | static void _ccv_cnnp_scatter_saved_aux(ccv_nnc_tensor_symbol_map_t* const saved_aux, const int parameter_size, const int old_saved_aux_size, const int new_saved_aux_size) |
| 1097 | { |
| 1098 | if (new_saved_aux_size == old_saved_aux_size) |
| 1099 | return; |
| 1100 | assert(new_saved_aux_size > old_saved_aux_size)((void) sizeof ((new_saved_aux_size > old_saved_aux_size) ? 1 : 0), __extension__ ({ if (new_saved_aux_size > old_saved_aux_size ) ; else __assert_fail ("new_saved_aux_size > old_saved_aux_size" , "ccv_cnnp_model.c", 1100, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1101 | int i, j; |
| 1102 | for (i = parameter_size - 1; i >= 0; i--) |
| 1103 | { |
| 1104 | for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--) |
| 1105 | saved_aux[i * new_saved_aux_size + j].source = saved_aux[i * new_saved_aux_size + j].destination = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 1106 | for (j = old_saved_aux_size - 1; j >= 0; j--) |
| 1107 | saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j]; |
| 1108 | } |
| 1109 | } |
| 1110 | |
| 1111 | static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model) |
| 1112 | { |
| 1113 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1114 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 1114, __extension__ __PRETTY_FUNCTION__); })); |
| 1115 | if (!compiled_data->rewindables) |
| 1116 | compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0); |
| 1117 | ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0); |
| 1118 | ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0); |
| 1119 | ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0); |
| 1120 | } |
| 1121 | |
| 1122 | static void _ccv_cnnp_model_gradient_init(ccv_cnnp_model_t* const model, const int gradient_mode, const uint64_t disable_outgrad, ccv_nnc_tensor_t* const* const fits, const int fit_size) |
| 1123 | { |
| 1124 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1125 | assert(compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE ) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE" , "ccv_cnnp_model.c", 1125, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1126 | assert(gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)((void) sizeof ((gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE ) ? 1 : 0), __extension__ ({ if (gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE ) ; else __assert_fail ("gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE" , "ccv_cnnp_model.c", 1126, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1127 | const int evaluate_to_size = compiled_data->evaluate.to_size; |
| 1128 | assert(evaluate_to_size > 0)((void) sizeof ((evaluate_to_size > 0) ? 1 : 0), __extension__ ({ if (evaluate_to_size > 0) ; else __assert_fail ("evaluate_to_size > 0" , "ccv_cnnp_model.c", 1128, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1129 | const int parallel_count = _ccv_cnnp_model_root_parallel_count(model); |
| 1130 | assert(_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now")((void) sizeof ((_ccv_cnnp_model_effective_parallel_count(model ) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now" ) ? 1 : 0), __extension__ ({ if (_ccv_cnnp_model_effective_parallel_count (model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now" ) ; else __assert_fail ("_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && \"local replicated stateful models only support forward / no-grad evaluation for now\"" , "ccv_cnnp_model.c", 1130, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1131 | compiled_data->evaluate.tos = ccreallocrealloc(compiled_data->evaluate.tos, sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size * parallel_count + sizeof(ccv_nnc_graph_exec_t) * evaluate_to_size * parallel_count); |
| 1132 | compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count); |
| 1133 | int i, j; |
| 1134 | const int output_size = model->output_size; |
| 1135 | assert(!fits || fit_size == output_size * parallel_count)((void) sizeof ((!fits || fit_size == output_size * parallel_count ) ? 1 : 0), __extension__ ({ if (!fits || fit_size == output_size * parallel_count) ; else __assert_fail ("!fits || fit_size == output_size * parallel_count" , "ccv_cnnp_model.c", 1135, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1136 | if (fits) |
| 1137 | for (i = 0; i < output_size; i++) |
| 1138 | ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info); |
| 1139 | const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size; |
| 1140 | const int parameter_size = compiled_data->parameters->rnum; |
| 1141 | compiled_data->updated_parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size + sizeof(ccv_nnc_graph_exec_symbol_t) * parameter_size + sizeof(ccv_nnc_tensor_symbol_map_t) * max_saved_aux_size * parameter_size); |
| 1142 | compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size); |
| 1143 | compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size); |
| 1144 | int parameter_size_maybe_more = parameter_size; |
| 1145 | compiled_data->disable_outgrad = disable_outgrad; |
| 1146 | int outgrad_size; |
| 1147 | if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0) |
| 1148 | outgrad_size = 0; |
| 1149 | else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs. |
| 1150 | outgrad_size = model->input_size; |
| 1151 | else { |
| 1152 | assert(disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL)((void) sizeof ((disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL ) ? 1 : 0), __extension__ ({ if (disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL ) ; else __assert_fail ("disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL" , "ccv_cnnp_model.c", 1152, __extension__ __PRETTY_FUNCTION__ ); })); // If it is disable all, gradient mode won't be this. |
| 1153 | outgrad_size = 0; |
| 1154 | for (i = 0; i < model->input_size; i++) |
| 1155 | if (!(disable_outgrad & ((uint64_t)1 << i))) |
| 1156 | ++outgrad_size; |
| 1157 | } |
| 1158 | compiled_data->outgrad_size = outgrad_size; |
| 1159 | parameter_size_maybe_more += outgrad_size; |
| 1160 | compiled_data->gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size_maybe_more + sizeof(ccv_nnc_graph_exec_symbol_t) * parameter_size_maybe_more * parallel_count); |
| 1161 | compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0; |
| 1162 | compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more); |
| 1163 | compiled_data->backward.to_size = parameter_size_maybe_more; |
| 1164 | ccv_nnc_tensor_symbol_t* parameters = (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))); |
| 1165 | if (compiled_data->parameter_flags) |
| 1166 | { |
| 1167 | parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size); |
| 1168 | for (i = 0; i < parameter_size; i++) |
| 1169 | if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))) |
| 1170 | parameters[i] = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( i))); |
| 1171 | else |
| 1172 | parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 1173 | } |
| 1174 | if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0) |
| 1175 | ccv_nnc_symbolic_graph_minimize(model->graph, compiled_data->minimize.minimizer, compiled_data->f, output_size, parameters, parameter_size, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph), compiled_data->gradients, compiled_data->updated_parameters, compiled_data->saved_aux, compiled_data->update_nodes); |
| 1176 | else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs. |
| 1177 | ccv_nnc_symbolic_graph_minimize(model->graph, compiled_data->minimize.minimizer, compiled_data->f, output_size, parameters, parameter_size, model->inputs, model->input_size, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph), compiled_data->gradients, compiled_data->updated_parameters, compiled_data->saved_aux, compiled_data->update_nodes); |
| 1178 | else { // Compute minimize with gradients including selected inputs. |
| 1179 | assert(model->input_size > 0)((void) sizeof ((model->input_size > 0) ? 1 : 0), __extension__ ({ if (model->input_size > 0) ; else __assert_fail ("model->input_size > 0" , "ccv_cnnp_model.c", 1179, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1180 | assert(disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL)((void) sizeof ((disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL ) ? 1 : 0), __extension__ ({ if (disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL ) ; else __assert_fail ("disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL" , "ccv_cnnp_model.c", 1180, __extension__ __PRETTY_FUNCTION__ ); })); // If it is disable all, gradient mode won't be this. |
| 1181 | assert(outgrad_size > 0)((void) sizeof ((outgrad_size > 0) ? 1 : 0), __extension__ ({ if (outgrad_size > 0) ; else __assert_fail ("outgrad_size > 0" , "ccv_cnnp_model.c", 1181, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1182 | ccv_nnc_tensor_symbol_t outgrads[outgrad_size]; |
| 1183 | j = 0; |
| 1184 | for (i = 0; i < model->input_size; i++) |
| 1185 | if (!(disable_outgrad & ((uint64_t)1 << i))) |
| 1186 | outgrads[j++] = model->inputs[i]; |
| 1187 | ccv_nnc_symbolic_graph_minimize(model->graph, compiled_data->minimize.minimizer, compiled_data->f, output_size, parameters, parameter_size, outgrads, outgrad_size, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph), compiled_data->gradients, compiled_data->updated_parameters, compiled_data->saved_aux, compiled_data->update_nodes); |
| 1188 | } |
| 1189 | if (compiled_data->parameter_flags) |
| 1190 | ccfreefree(parameters); |
| 1191 | _ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer), compiled_data->minimize.max_saved_aux_size); |
| 1192 | if (compiled_data->minimize.parameters) |
| 1193 | _ccv_cnnp_apply_parameters_with_minimizer(model); |
| 1194 | // Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass. |
| 1195 | ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph); |
| 1196 | for (i = 0; i < output_size; i++) |
| 1197 | { |
| 1198 | const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]); |
| 1199 | // Init this to 1 so we can backprop. |
| 1200 | ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES); |
| 1201 | } |
| 1202 | compiled_data->backward.to_size = 0; |
| 1203 | for (i = 0; i < parameter_size_maybe_more; i++) |
| 1204 | if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 1205 | compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]); |
| 1206 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS); |
| 1207 | ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size); |
| 1208 | for (i = 0; i < parameter_size_maybe_more - parameter_size; i++) |
| 1209 | { |
| 1210 | if (compiled_data->outgrads[i].d < 0) // When we go through input, we might find zero-length inputs, and for these, we cannot have any outgrads. |
| 1211 | continue; |
| 1212 | const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]); |
| 1213 | const int* tos; |
| 1214 | int to_size; |
| 1215 | ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size); |
| 1216 | if (to_size == 0) // If this is the end (no minimizers afterwards). We need to attach this as a destination. Otherwise this is covered in update_nodes. |
| 1217 | { |
| 1218 | const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph); |
| 1219 | const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph); |
| 1220 | int flag = 0; |
| 1221 | const int outgrad_destination_start = ccv_max(0, destination_count - i)({ typeof (0) _a = (0); typeof (destination_count - i) _b = ( destination_count - i); (_a > _b) ? _a : _b; }); |
| 1222 | for (j = i - 1; !flag && j >= 0; j--) |
| 1223 | if (j + outgrad_destination_start < destination_count) |
| 1224 | flag = (destinations[j + outgrad_destination_start].d == outgrad.d); |
| 1225 | if (!flag) // Only if we cannot find it, we add it. |
| 1226 | ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad); |
| 1227 | } |
| 1228 | } |
| 1229 | if (parallel_count > 1) |
| 1230 | { |
| 1231 | ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count, |
| 1232 | 0, 0, |
| 1233 | compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */, |
| 1234 | compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */, |
| 1235 | 0, 0, 0, |
| 1236 | CCV_NNC_PARALLEL_REDUCE_OP_SUM, |
| 1237 | SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph)); |
| 1238 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 1239 | for (i = 0; i < evaluate_to_size; i++) |
| 1240 | for (j = 1; j < parallel_count; j++) |
| 1241 | { |
| 1242 | const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j); |
| 1243 | if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL) |
| 1244 | compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy; |
| 1245 | } |
| 1246 | const int backward_to_size = compiled_data->backward.to_size; |
| 1247 | for (i = 0; i < backward_to_size; i++) |
| 1248 | for (j = 1; j < parallel_count; j++) |
| 1249 | { |
| 1250 | const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j); |
| 1251 | if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL) |
| 1252 | compiled_data->backward.tos[compiled_data->backward.to_size++] = copy; |
| 1253 | } |
| 1254 | } |
| 1255 | // Only use memory compression if we are in gradient parameter mode. |
| 1256 | if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS) |
| 1257 | { |
| 1258 | if (model->memory_compression) |
| 1259 | ccv_nnc_symbolic_graph_memory_compression(model->graph, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph)); |
| 1260 | if (model->memory_reduction) |
| 1261 | ccv_nnc_symbolic_graph_memory_reduction(model->graph, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph)); |
| 1262 | } |
| 1263 | compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size); |
| 1264 | compiled_data->gradient_mode = gradient_mode; |
| 1265 | } |
| 1266 | |
| 1267 | void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data) |
| 1268 | { |
| 1269 | assert(!compiled_data->tensors.parameters)((void) sizeof ((!compiled_data->tensors.parameters) ? 1 : 0), __extension__ ({ if (!compiled_data->tensors.parameters ) ; else __assert_fail ("!compiled_data->tensors.parameters" , "ccv_cnnp_model.c", 1269, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1270 | const int parameter_size = compiled_data->parameters->rnum; |
| 1271 | const int parallel_count = _ccv_cnnp_model_effective_parallel_count(model); |
| 1272 | compiled_data->parallel_count = parallel_count; |
| 1273 | const int internal_size = compiled_data->internals->rnum; |
| 1274 | compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph); |
| 1275 | compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t)); |
| 1276 | compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*)); |
| 1277 | compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count; |
| 1278 | } |
| 1279 | |
| 1280 | int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data) |
| 1281 | { |
| 1282 | int i, j; |
| 1283 | const int parameter_size = compiled_data->parameters->rnum; |
| 1284 | const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data); |
| 1285 | const int internal_size = compiled_data->internals->rnum; |
| 1286 | for (i = 0; i < parameter_size; i++) |
| 1287 | { |
| 1288 | // parameters has to be allocated all together. |
| 1289 | if (compiled_data->tensors.parameters[i]) |
| 1290 | { |
| 1291 | for (j = 1; j < parallel_count; j++) |
| 1292 | { assert(compiled_data->tensors.parameters[i + j * parameter_size])((void) sizeof ((compiled_data->tensors.parameters[i + j * parameter_size]) ? 1 : 0), __extension__ ({ if (compiled_data ->tensors.parameters[i + j * parameter_size]) ; else __assert_fail ("compiled_data->tensors.parameters[i + j * parameter_size]" , "ccv_cnnp_model.c", 1292, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 1293 | continue; |
| 1294 | } |
| 1295 | return 1; |
| 1296 | } |
| 1297 | for (i = 0; i < internal_size; i++) |
| 1298 | { |
| 1299 | if (!compiled_data->tensors.internals[i]) |
| 1300 | return 1; |
| 1301 | for (j = 1; j < parallel_count; j++) |
| 1302 | if (!compiled_data->tensors.internals[i + j * internal_size]) |
| 1303 | return 1; |
| 1304 | } |
| 1305 | return 0; |
| 1306 | } |
| 1307 | |
| 1308 | void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data) |
| 1309 | { |
| 1310 | int i, j; |
| 1311 | const int parameter_size = compiled_data->parameters->rnum; |
| 1312 | const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data); |
| 1313 | compiled_data->parallel_count = parallel_count; |
| 1314 | const int internal_size = compiled_data->internals->rnum; |
| 1315 | for (i = 0; i < parameter_size; i++) |
| 1316 | { |
| 1317 | // parameters has to be allocated all together. |
| 1318 | if (compiled_data->tensors.parameters[i]) |
| 1319 | { |
| 1320 | for (j = 1; j < parallel_count; j++) |
| 1321 | { assert(compiled_data->tensors.parameters[i + j * parameter_size])((void) sizeof ((compiled_data->tensors.parameters[i + j * parameter_size]) ? 1 : 0), __extension__ ({ if (compiled_data ->tensors.parameters[i + j * parameter_size]) ; else __assert_fail ("compiled_data->tensors.parameters[i + j * parameter_size]" , "ccv_cnnp_model.c", 1321, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 1322 | continue; |
| 1323 | } |
| 1324 | const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( i))); |
| 1325 | ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter); |
| 1326 | if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY) |
| 1327 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 1328 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8); |
| 1329 | compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0); |
| 1330 | for (j = 1; j < parallel_count; j++) |
| 1331 | { |
| 1332 | if (j != device_id) |
| 1333 | CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff ) << 8)); |
| 1334 | else |
| 1335 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 1336 | compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0); |
| 1337 | } |
| 1338 | } |
| 1339 | const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 1340 | for (i = 0; i < internal_size; i++) |
| 1341 | { |
| 1342 | const ccv_nnc_tensor_symbol_t retained = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(i)) ); |
| 1343 | const int d = retained.d; |
| 1344 | if (init_v[d >> 5] & (1u << (d & 0x1f))) |
| 1345 | continue; |
| 1346 | ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained); |
| 1347 | if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY) |
| 1348 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 1349 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8); |
| 1350 | if (!compiled_data->tensors.internals[i]) |
| 1351 | compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0); |
| 1352 | for (j = 1; j < parallel_count; j++) |
| 1353 | { |
| 1354 | if (j != device_id) |
| 1355 | CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff ) << 8)); |
| 1356 | else |
| 1357 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 1358 | if (!compiled_data->tensors.internals[i + j * internal_size]) |
| 1359 | compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0); |
| 1360 | } |
| 1361 | } |
| 1362 | compiled_data->tensors_init.v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); // Remove 1 if any. |
| 1363 | } |
| 1364 | |
| 1365 | static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data) |
| 1366 | { |
| 1367 | ccv_cnnp_model_tensors_init_0(model, compiled_data); |
| 1368 | ccv_cnnp_model_tensors_init_1(model, compiled_data); |
| 1369 | } |
| 1370 | |
| 1371 | static void _ccv_cnnp_model_copy_tensors(const uint32_t* const tensors_init, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t* const* const tensors, const int tensor_size, const int parallel_count) |
| 1372 | { |
| 1373 | assert(parallel_count > 0)((void) sizeof ((parallel_count > 0) ? 1 : 0), __extension__ ({ if (parallel_count > 0) ; else __assert_fail ("parallel_count > 0" , "ccv_cnnp_model.c", 1373, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1374 | int i, j; |
| 1375 | for (i = 0; i < tensor_size; i++) |
| 1376 | { |
| 1377 | if (!tensors[i]) |
| 1378 | continue; |
| 1379 | const int d = tensor_symbols[i].d; |
| 1380 | if (!(tensors_init[d >> 5] & (1u << (d & 0x1f)))) |
| 1381 | continue; |
| 1382 | for (j = 1; j < parallel_count; j++) |
| 1383 | if (tensors[i + j * tensor_size]) |
| 1384 | { |
| 1385 | ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t )1)); |
| 1386 | ccv_nnc_tensor_t* const output = CCV_NNC_TENSOR(tensors[i + j * tensor_size])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i + j * tensor_size] ) & ~(uintptr_t)1)); |
| 1387 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, &input, 1, &output, 1, 0); |
| 1388 | } |
| 1389 | } |
| 1390 | } |
| 1391 | |
| 1392 | static void _ccv_cnnp_model_remove_nocopies(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t** const tensors, const int tensor_size, const int parallel_count) |
| 1393 | { |
| 1394 | assert(parallel_count > 0)((void) sizeof ((parallel_count > 0) ? 1 : 0), __extension__ ({ if (parallel_count > 0) ; else __assert_fail ("parallel_count > 0" , "ccv_cnnp_model.c", 1394, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1395 | int i, j; |
| 1396 | for (i = 0; i < tensor_size; i++) |
| 1397 | { |
| 1398 | const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i]; |
| 1399 | for (j = 1; j < parallel_count; j++) |
| 1400 | { |
| 1401 | const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j); |
| 1402 | ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size]; |
| 1403 | if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 1404 | { // We shouldn't allocate this, free it up. |
| 1405 | ccv_nnc_tensor_free(tensors[i + j * tensor_size]); |
| 1406 | tensors[i + j * tensor_size] = 0; |
| 1407 | } |
| 1408 | } |
| 1409 | } |
| 1410 | } |
| 1411 | |
| 1412 | static void _ccv_cnnp_model_bind_tensors(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t* const* const tensors, const int tensor_size, const int parallel_count, ccv_array_t* const tensor_binds) |
| 1413 | { |
| 1414 | assert(parallel_count > 0)((void) sizeof ((parallel_count > 0) ? 1 : 0), __extension__ ({ if (parallel_count > 0) ; else __assert_fail ("parallel_count > 0" , "ccv_cnnp_model.c", 1414, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1415 | int i, j; |
| 1416 | for (i = 0; i < tensor_size; i++) |
| 1417 | { |
| 1418 | ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i]; |
| 1419 | if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 1420 | continue; |
| 1421 | if (graph) |
| 1422 | { |
| 1423 | const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol); |
| 1424 | if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 1425 | tensor_symbol = alias_to; |
| 1426 | } |
| 1427 | ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t )1)); |
| 1428 | if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 1429 | { |
| 1430 | const ccv_nnc_tensor_bind_t retained_bind = { |
| 1431 | .symbol = tensor_symbol, |
| 1432 | .tensor = tensor |
| 1433 | }; |
| 1434 | ccv_array_push(tensor_binds, &retained_bind); |
| 1435 | } |
| 1436 | for (j = 1; j < parallel_count; j++) |
| 1437 | { |
| 1438 | const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j); |
| 1439 | ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size]; |
| 1440 | if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 1441 | { |
| 1442 | const ccv_nnc_tensor_bind_t bind = { |
| 1443 | .symbol = copy, |
| 1444 | .tensor = tensors[i + j * tensor_size] |
| 1445 | }; |
| 1446 | ccv_array_push(tensor_binds, &bind); |
| 1447 | } |
| 1448 | } |
| 1449 | } |
| 1450 | } |
| 1451 | |
| 1452 | static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data) |
| 1453 | { |
| 1454 | if (compiled_data->graph) |
| 1455 | ccv_nnc_graph_free(compiled_data->graph); |
| 1456 | compiled_data->graph = 0; |
| 1457 | compiled_data->is_test = 0; |
| 1458 | if (compiled_data->tensor_arena) |
| 1459 | ccv_nnc_tensor_arena_free(compiled_data->tensor_arena); |
| 1460 | compiled_data->tensor_arena = 0; |
| 1461 | if (compiled_data->graph_exec_arena) |
| 1462 | ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena); |
| 1463 | compiled_data->graph_exec_arena = 0; |
| 1464 | if (compiled_data->backward.from_ops) |
| 1465 | ccfreefree(compiled_data->backward.from_ops); |
| 1466 | compiled_data->backward.from_ops = 0; |
| 1467 | if (compiled_data->evaluate.schedule) |
| 1468 | ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule); |
| 1469 | compiled_data->evaluate.schedule = 0; |
| 1470 | if (compiled_data->backward.schedule) |
| 1471 | ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule); |
| 1472 | compiled_data->backward.schedule = 0; |
| 1473 | } |
| 1474 | |
| 1475 | static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data) |
| 1476 | { |
| 1477 | if (compiled_data->gradients) |
| 1478 | ccfreefree(compiled_data->gradients); |
| 1479 | compiled_data->gradients = 0; |
| 1480 | if (compiled_data->updated_parameters) |
| 1481 | ccfreefree(compiled_data->updated_parameters); |
| 1482 | compiled_data->updated_parameters = 0; |
| 1483 | compiled_data->update_nodes = 0; |
| 1484 | compiled_data->saved_aux = 0; |
| 1485 | } |
| 1486 | |
| 1487 | static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data) |
| 1488 | { |
| 1489 | if (compiled_data->backward.gradients) |
| 1490 | ccfreefree(compiled_data->backward.gradients); |
| 1491 | compiled_data->backward.gradients = 0; |
| 1492 | if (compiled_data->backward.accum) |
| 1493 | ccv_nnc_graph_free(compiled_data->backward.accum); |
| 1494 | compiled_data->backward.accum = 0; |
| 1495 | if (compiled_data->backward.tensor_arena) |
| 1496 | ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena); |
| 1497 | compiled_data->backward.tensor_arena = 0; |
| 1498 | if (compiled_data->backward.graph_exec_arena) |
| 1499 | ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena); |
| 1500 | compiled_data->backward.graph_exec_arena = 0; |
| 1501 | } |
| 1502 | |
| 1503 | static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data) |
| 1504 | { |
| 1505 | if (compiled_data->apply_gradients.graph) |
| 1506 | ccv_nnc_graph_free(compiled_data->apply_gradients.graph); |
| 1507 | compiled_data->apply_gradients.graph = 0; |
| 1508 | if (compiled_data->apply_gradients.tensor_arena) |
| 1509 | ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena); |
| 1510 | compiled_data->apply_gradients.tensor_arena = 0; |
| 1511 | if (compiled_data->apply_gradients.graph_exec_arena) |
| 1512 | ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena); |
| 1513 | compiled_data->apply_gradients.graph_exec_arena = 0; |
| 1514 | } |
| 1515 | |
| 1516 | // Compile the graph to run ccv_cnnp_model_fit |
| 1517 | static void _ccv_cnnp_model_fit_jit(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const fits, const int fit_size, ccv_nnc_tensor_t* const* const outputs, const int output_size) |
| 1518 | { |
| 1519 | int i, j; |
| 1520 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1521 | assert(!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)((void) sizeof ((!compiled_data->graph || compiled_data-> graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE) ? 1 : 0), __extension__ ({ if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE) ; else __assert_fail ("!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE" , "ccv_cnnp_model.c", 1521, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1522 | compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE; |
| 1523 | const int parallel_count = _ccv_cnnp_model_root_parallel_count(model); |
| 1524 | assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count ) ? 1 : 0), __extension__ ({ if (output_size == model->output_size * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count" , "ccv_cnnp_model.c", 1524, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1525 | assert(!fits || output_size == fit_size)((void) sizeof ((!fits || output_size == fit_size) ? 1 : 0), __extension__ ({ if (!fits || output_size == fit_size) ; else __assert_fail ("!fits || output_size == fit_size", "ccv_cnnp_model.c", 1525 , __extension__ __PRETTY_FUNCTION__); })); |
| 1526 | assert(output_size > 0)((void) sizeof ((output_size > 0) ? 1 : 0), __extension__ ( { if (output_size > 0) ; else __assert_fail ("output_size > 0" , "ccv_cnnp_model.c", 1526, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1527 | if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE) |
| 1528 | { |
| 1529 | _ccv_cnnp_model_set_rewindables(model); |
| 1530 | _ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size); |
| 1531 | } else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) { |
| 1532 | _ccv_cnnp_model_rewind_graph(model); |
| 1533 | _ccv_cnnp_compiled_data_gradient_free(compiled_data); |
| 1534 | compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE; |
| 1535 | _ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size); |
| 1536 | } |
| 1537 | const int tensors_init = !!compiled_data->tensors_init.v; |
| 1538 | if (!tensors_init) |
| 1539 | _ccv_cnnp_model_tensors_init(model, compiled_data); |
| 1540 | else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1) |
| 1541 | // Check if it is not fully allocated, if it is not, init_1. |
| 1542 | ccv_cnnp_model_tensors_init_1(model, compiled_data); |
| 1543 | ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0); |
| 1544 | assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0) , __extension__ ({ if ((input_size % parallel_count) == 0) ; else __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1544, __extension__ __PRETTY_FUNCTION__); })); |
| 1545 | assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0 ), __extension__ ({ if ((output_size % parallel_count) == 0) ; else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1545, __extension__ __PRETTY_FUNCTION__); })); |
| 1546 | assert((fit_size % parallel_count) == 0)((void) sizeof (((fit_size % parallel_count) == 0) ? 1 : 0), __extension__ ({ if ((fit_size % parallel_count) == 0) ; else __assert_fail ("(fit_size % parallel_count) == 0", "ccv_cnnp_model.c", 1546 , __extension__ __PRETTY_FUNCTION__); })); |
| 1547 | const int input_size_per_p = input_size / parallel_count; |
| 1548 | _ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds); |
| 1549 | const int output_size_per_p = output_size / parallel_count; |
| 1550 | _ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds); |
| 1551 | const int fit_size_per_p = fit_size / parallel_count; |
| 1552 | _ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds); |
| 1553 | const int parameter_size = compiled_data->parameters->rnum; |
| 1554 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds); |
| 1555 | _ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds); |
| 1556 | const int internal_size = compiled_data->internals->rnum; |
| 1557 | _ccv_cnnp_model_remove_nocopies(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(0)) ), compiled_data->tensors.internals, internal_size, parallel_count); |
| 1558 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(0)) ), compiled_data->tensors.internals, internal_size, parallel_count, tensor_binds); |
| 1559 | ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds )->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph), &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena); |
| 1560 | ccv_array_free(tensor_binds); |
| 1561 | const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 1562 | if (tensors_init && parallel_count > 1) |
| 1563 | _ccv_cnnp_model_copy_tensors(init_v, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, compiled_data->parameters->rnum, parallel_count); |
| 1564 | // If tensor is not init'ed, we need to init states first. |
| 1565 | if (_ccv_cnnp_any_to_init(compiled_data)) |
| 1566 | { |
| 1567 | ccv_nnc_tensor_init_states_t tensor_init_states = { |
| 1568 | .parallel_count = parallel_count, |
| 1569 | .graph = model->graph, |
| 1570 | .compiled_data = compiled_data, |
| 1571 | .tensor_arena = compiled_data->tensor_arena |
| 1572 | }; |
| 1573 | ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states); |
| 1574 | } |
| 1575 | compiled_data->is_test = 0; |
| 1576 | const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer); |
| 1577 | // No need to set because it is default to training mode. |
| 1578 | // ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update); |
| 1579 | for (i = 0; i < saved_aux_size * parameter_size; i++) |
| 1580 | { |
| 1581 | if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 1582 | continue; |
| 1583 | ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source); |
| 1584 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, &tensor, 1, 0); |
| 1585 | for (j = 1; j < parallel_count; j++) |
| 1586 | { |
| 1587 | ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, compiled_data->saved_aux[i].source, j)); |
| 1588 | if (copy) |
| 1589 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, ©, 1, 0); |
| 1590 | } |
| 1591 | } |
| 1592 | const int evaluate_to_size = compiled_data->evaluate.to_size; |
| 1593 | compiled_data->evaluate.to_op_size = 0; |
| 1594 | for (i = 0; i < evaluate_to_size; i++) |
| 1595 | { |
| 1596 | ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]); |
| 1597 | if (to.graph) |
| 1598 | compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to; |
| 1599 | } |
| 1600 | ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count); |
| 1601 | ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0); |
| 1602 | } |
| 1603 | |
| 1604 | ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model) |
| 1605 | { |
| 1606 | const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1607 | if (!compiled_data || !compiled_data->graph) |
| 1608 | return 0; |
| 1609 | return ccv_nnc_graph_default_stream(compiled_data->graph); |
| 1610 | } |
| 1611 | |
| 1612 | uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model) |
| 1613 | { |
| 1614 | const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1615 | if (!compiled_data || !compiled_data->tensor_arena) |
| 1616 | return 0; |
| 1617 | return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena); |
| 1618 | } |
| 1619 | |
| 1620 | static void _ccv_cnnp_bind_tensors_to_arena(ccv_nnc_tensor_arena_t* const tensor_arena, const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t* const* const tensors, const int tensor_size, const int parallel_count) |
| 1621 | { |
| 1622 | int i, j; |
| 1623 | for (i = 0; i < tensor_size; i++) |
| 1624 | { |
| 1625 | ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i]; |
| 1626 | if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 1627 | continue; |
| 1628 | if (graph) |
| 1629 | { |
| 1630 | const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol); |
| 1631 | if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 1632 | tensor_symbol = alias_to; |
| 1633 | } |
| 1634 | ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]); |
| 1635 | for (j = 1; j < parallel_count; j++) |
| 1636 | { |
| 1637 | const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j); |
| 1638 | if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL) |
| 1639 | ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]); |
| 1640 | } |
| 1641 | } |
| 1642 | } |
| 1643 | |
| 1644 | void ccv_cnnp_model_fit(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const fits, const int fit_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) |
| 1645 | { |
| 1646 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1647 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 1647, __extension__ __PRETTY_FUNCTION__); })); |
| 1648 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 1649 | assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count ) ? 1 : 0), __extension__ ({ if (output_size == model->output_size * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count" , "ccv_cnnp_model.c", 1649, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1650 | assert(input_size == model->input_size * parallel_count)((void) sizeof ((input_size == model->input_size * parallel_count ) ? 1 : 0), __extension__ ({ if (input_size == model->input_size * parallel_count) ; else __assert_fail ("input_size == model->input_size * parallel_count" , "ccv_cnnp_model.c", 1650, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1651 | assert(!fits || fit_size == output_size)((void) sizeof ((!fits || fit_size == output_size) ? 1 : 0), __extension__ ({ if (!fits || fit_size == output_size) ; else __assert_fail ("!fits || fit_size == output_size", "ccv_cnnp_model.c", 1651 , __extension__ __PRETTY_FUNCTION__); })); |
| 1652 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 1652, __extension__ __PRETTY_FUNCTION__); })); |
| 1653 | if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE) |
| 1654 | { |
| 1655 | _ccv_cnnp_compiled_data_graph_free(compiled_data); |
| 1656 | _ccv_cnnp_compiled_data_backward_free(compiled_data); |
| 1657 | _ccv_cnnp_compiled_data_apply_gradients_free(compiled_data); |
| 1658 | // Compile the symbolic graph down only when needed. |
| 1659 | _ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size); |
| 1660 | } else { |
| 1661 | assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0) , __extension__ ({ if ((input_size % parallel_count) == 0) ; else __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1661, __extension__ __PRETTY_FUNCTION__); })); |
| 1662 | assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0 ), __extension__ ({ if ((output_size % parallel_count) == 0) ; else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1662, __extension__ __PRETTY_FUNCTION__); })); |
| 1663 | assert((fit_size % parallel_count) == 0)((void) sizeof (((fit_size % parallel_count) == 0) ? 1 : 0), __extension__ ({ if ((fit_size % parallel_count) == 0) ; else __assert_fail ("(fit_size % parallel_count) == 0", "ccv_cnnp_model.c", 1663 , __extension__ __PRETTY_FUNCTION__); })); |
| 1664 | const int input_size_per_p = input_size / parallel_count; |
| 1665 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count); |
| 1666 | const int output_size_per_p = output_size / parallel_count; |
| 1667 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count); |
| 1668 | const int fit_size_per_p = fit_size / parallel_count; |
| 1669 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count); |
| 1670 | } |
| 1671 | if (compiled_data->is_test) |
| 1672 | { |
| 1673 | compiled_data->is_test = 0; |
| 1674 | ccv_nnc_graph_exec_update_t update = { |
| 1675 | .parallel_count = parallel_count, |
| 1676 | .graph = model->graph, |
| 1677 | .graph_exec_arena = compiled_data->graph_exec_arena, |
| 1678 | }; |
| 1679 | ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update); |
| 1680 | } |
| 1681 | ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context); |
| 1682 | } |
| 1683 | |
| 1684 | // Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD). |
| 1685 | static void _ccv_cnnp_model_multistage_no_grad_jit(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size) |
| 1686 | { |
| 1687 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1688 | compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD; |
| 1689 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 1690 | assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count ) ? 1 : 0), __extension__ ({ if (output_size == model->output_size * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count" , "ccv_cnnp_model.c", 1690, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1691 | assert(output_size > 0)((void) sizeof ((output_size > 0) ? 1 : 0), __extension__ ( { if (output_size > 0) ; else __assert_fail ("output_size > 0" , "ccv_cnnp_model.c", 1691, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1692 | // If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather, |
| 1693 | // we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel. |
| 1694 | if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE) |
| 1695 | { |
| 1696 | const int evaluate_to_size = compiled_data->evaluate.to_size; |
| 1697 | compiled_data->evaluate.tos = ccreallocrealloc(compiled_data->evaluate.tos, sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size * parallel_count + sizeof(ccv_nnc_graph_exec_t) * evaluate_to_size * parallel_count); |
| 1698 | _ccv_cnnp_model_set_rewindables(model); |
| 1699 | ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count, |
| 1700 | 0, 0, |
| 1701 | 0, 0, 0, |
| 1702 | 0, 0, 0, |
| 1703 | CCV_NNC_PARALLEL_REDUCE_OP_SUM, |
| 1704 | SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size (model->graph)); |
| 1705 | ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 1706 | int i, j; |
| 1707 | for (i = 0; i < evaluate_to_size; i++) |
| 1708 | for (j = 1; j < parallel_count; j++) |
| 1709 | { |
| 1710 | const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j); |
| 1711 | if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL) |
| 1712 | compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy; |
| 1713 | } |
| 1714 | } |
| 1715 | const int tensors_init = !!compiled_data->tensors_init.v; |
| 1716 | if (!tensors_init) |
| 1717 | _ccv_cnnp_model_tensors_init(model, compiled_data); |
| 1718 | else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1) |
| 1719 | // Check if it is not fully allocated, if it is not, init_1. |
| 1720 | ccv_cnnp_model_tensors_init_1(model, compiled_data); |
| 1721 | const int tensor_parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data); |
| 1722 | ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0); |
| 1723 | assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0) , __extension__ ({ if ((input_size % parallel_count) == 0) ; else __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1723, __extension__ __PRETTY_FUNCTION__); })); |
| 1724 | assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0 ), __extension__ ({ if ((output_size % parallel_count) == 0) ; else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1724, __extension__ __PRETTY_FUNCTION__); })); |
| 1725 | const int input_size_per_p = input_size / parallel_count; |
| 1726 | _ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds); |
| 1727 | const int output_size_per_p = output_size / parallel_count; |
| 1728 | _ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds); |
| 1729 | const int parameter_size = compiled_data->parameters->rnum; |
| 1730 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, parameter_size, tensor_parallel_count, tensor_binds); |
| 1731 | const int internal_size = compiled_data->internals->rnum; |
| 1732 | _ccv_cnnp_model_remove_nocopies(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(0)) ), compiled_data->tensors.internals, internal_size, tensor_parallel_count); |
| 1733 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(0)) ), compiled_data->tensors.internals, internal_size, tensor_parallel_count, tensor_binds); |
| 1734 | // If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation. |
| 1735 | ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds )->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), compiled_data->evaluate.tos, compiled_data->evaluate.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena); |
| 1736 | ccv_array_free(tensor_binds); |
| 1737 | const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 1738 | // If tensor is not init'ed, we need to init states first. |
| 1739 | if (tensors_init && tensor_parallel_count > 1) |
| 1740 | _ccv_cnnp_model_copy_tensors(init_v, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, compiled_data->parameters->rnum, tensor_parallel_count); |
| 1741 | if (_ccv_cnnp_any_to_init(compiled_data)) |
| 1742 | { |
| 1743 | ccv_nnc_tensor_init_states_t tensor_init_states = { |
| 1744 | .parallel_count = tensor_parallel_count, |
| 1745 | .graph = model->graph, |
| 1746 | .compiled_data = compiled_data, |
| 1747 | .tensor_arena = compiled_data->tensor_arena |
| 1748 | }; |
| 1749 | ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states); |
| 1750 | } |
| 1751 | compiled_data->is_test = 1; |
| 1752 | ccv_nnc_graph_exec_update_t update = { |
| 1753 | .parallel_count = parallel_count, |
| 1754 | .graph = model->graph, |
| 1755 | .graph_exec_arena = compiled_data->graph_exec_arena, |
| 1756 | }; |
| 1757 | ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update); |
| 1758 | ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count); |
| 1759 | ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0); |
| 1760 | } |
| 1761 | |
| 1762 | static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data) |
| 1763 | { |
| 1764 | assert(!compiled_data->tensors.gradients)((void) sizeof ((!compiled_data->tensors.gradients) ? 1 : 0 ), __extension__ ({ if (!compiled_data->tensors.gradients) ; else __assert_fail ("!compiled_data->tensors.gradients" , "ccv_cnnp_model.c", 1764, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1765 | const int parameter_size = compiled_data->parameters->rnum; |
| 1766 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 1767 | compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count); |
| 1768 | compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count; |
| 1769 | int i, j; |
| 1770 | for (i = 0; i < parameter_size; i++) |
| 1771 | { |
| 1772 | if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))) |
| 1773 | { |
| 1774 | compiled_data->tensors.gradients[i] = 0; |
| 1775 | compiled_data->tensors.accum_gradients[i] = 0; |
| 1776 | for (j = 1; j < parallel_count; j++) |
| 1777 | { |
| 1778 | compiled_data->tensors.gradients[i + j * parameter_size] = 0; |
| 1779 | compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0; |
| 1780 | } |
| 1781 | continue; |
| 1782 | } |
| 1783 | const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( i))); |
| 1784 | ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter); |
| 1785 | if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY) |
| 1786 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 1787 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8); |
| 1788 | compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0); |
| 1789 | compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it. |
| 1790 | for (j = 1; j < parallel_count; j++) |
| 1791 | { |
| 1792 | if (j != device_id) |
| 1793 | CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff ) << 8)); |
| 1794 | else |
| 1795 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 1796 | compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0); |
| 1797 | compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0; |
| 1798 | } |
| 1799 | } |
| 1800 | } |
| 1801 | |
| 1802 | static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size) |
| 1803 | { |
| 1804 | if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL) |
| 1805 | return 1; |
| 1806 | if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) |
| 1807 | return 0; |
| 1808 | int i; |
| 1809 | for (i = 0; i < input_size; i++) |
| 1810 | if (!(disable_outgrad & ((uint64_t)1 << i))) |
| 1811 | return 0; |
| 1812 | return 1; |
| 1813 | } |
| 1814 | |
| 1815 | // Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE). |
| 1816 | // Particularly, this method compiles the evaluation and backprop graph (the main graph). |
| 1817 | static void _ccv_cnnp_model_multistage_jit_0(ccv_cnnp_model_t* const model, const uint64_t disable_outgrad, const int is_test, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size) |
| 1818 | { |
| 1819 | int i, j; |
| 1820 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1821 | const int target_gradient_mode = _ccv_cnnp_is_disable_outgrad_all(disable_outgrad, model->input_size) ? CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES : CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS; |
| 1822 | assert(!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->gradient_mode != target_gradient_mode)((void) sizeof ((!compiled_data->graph || compiled_data-> graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data ->gradient_mode != target_gradient_mode) ? 1 : 0), __extension__ ({ if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data-> gradient_mode != target_gradient_mode) ; else __assert_fail ( "!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->gradient_mode != target_gradient_mode" , "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1823 | compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE; |
| 1824 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 1825 | assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count ) ? 1 : 0), __extension__ ({ if (output_size == model->output_size * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count" , "ccv_cnnp_model.c", 1825, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1826 | assert(output_size > 0)((void) sizeof ((output_size > 0) ? 1 : 0), __extension__ ( { if (output_size > 0) ; else __assert_fail ("output_size > 0" , "ccv_cnnp_model.c", 1826, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1827 | // There shouldn't be a loss function if we evaluate with multistage jit. |
| 1828 | assert(compiled_data->loss.cmd == CCV_NNC_NOOP)((void) sizeof ((compiled_data->loss.cmd == CCV_NNC_NOOP) ? 1 : 0), __extension__ ({ if (compiled_data->loss.cmd == CCV_NNC_NOOP ) ; else __assert_fail ("compiled_data->loss.cmd == CCV_NNC_NOOP" , "ccv_cnnp_model.c", 1828, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1829 | if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE) |
| 1830 | { |
| 1831 | _ccv_cnnp_model_set_rewindables(model); |
| 1832 | _ccv_cnnp_model_gradient_init(model, target_gradient_mode, disable_outgrad, 0, 0); // The type of outputs and fits should be the same. We only use type here. |
| 1833 | } else if (compiled_data->gradient_mode != target_gradient_mode) { |
| 1834 | _ccv_cnnp_model_rewind_graph(model); |
| 1835 | _ccv_cnnp_compiled_data_gradient_free(compiled_data); |
| 1836 | compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE; |
| 1837 | _ccv_cnnp_model_gradient_init(model, target_gradient_mode, disable_outgrad, 0, 0); // The type of outputs and fits should be the same. We only use type here. |
| 1838 | } |
| 1839 | const int tensors_init = !!compiled_data->tensors_init.v; |
| 1840 | if (!tensors_init) |
| 1841 | _ccv_cnnp_model_tensors_init(model, compiled_data); |
| 1842 | else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1) |
| 1843 | // Check if it is not fully allocated, if it is not, init_1. |
| 1844 | ccv_cnnp_model_tensors_init_1(model, compiled_data); |
| 1845 | ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0); |
| 1846 | assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0) , __extension__ ({ if ((input_size % parallel_count) == 0) ; else __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1846, __extension__ __PRETTY_FUNCTION__); })); |
| 1847 | assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0 ), __extension__ ({ if ((output_size % parallel_count) == 0) ; else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1847, __extension__ __PRETTY_FUNCTION__); })); |
| 1848 | const int input_size_per_p = input_size / parallel_count; |
| 1849 | _ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds); |
| 1850 | const int output_size_per_p = output_size / parallel_count; |
| 1851 | _ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds); |
| 1852 | const int parameter_size = compiled_data->parameters->rnum; |
| 1853 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds); |
| 1854 | const int internal_size = compiled_data->internals->rnum; |
| 1855 | _ccv_cnnp_model_remove_nocopies(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(0)) ), compiled_data->tensors.internals, internal_size, parallel_count); |
| 1856 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + ( size_t)(compiled_data->internals)->rsize * (size_t)(0)) ), compiled_data->tensors.internals, internal_size, parallel_count, tensor_binds); |
| 1857 | if (!compiled_data->tensors.gradients) |
| 1858 | _ccv_cnnp_model_gradient_tensors_init(model, compiled_data); |
| 1859 | _ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds); |
| 1860 | if (compiled_data->backward.to_size > 0) |
| 1861 | ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds )->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), compiled_data->backward.tos, compiled_data->backward.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena); |
| 1862 | else |
| 1863 | ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds )->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size (model->graph), compiled_data->evaluate.tos, compiled_data->evaluate.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena); |
| 1864 | ccv_array_free(tensor_binds); |
| 1865 | const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 1866 | if (tensors_init && parallel_count > 1) |
| 1867 | _ccv_cnnp_model_copy_tensors(init_v, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, compiled_data->parameters->rnum, parallel_count); |
| 1868 | // If tensor is not init'ed, we need to init states first. |
| 1869 | if (_ccv_cnnp_any_to_init(compiled_data)) |
| 1870 | { |
| 1871 | ccv_nnc_tensor_init_states_t tensor_init_states = { |
| 1872 | .parallel_count = parallel_count, |
| 1873 | .graph = model->graph, |
| 1874 | .compiled_data = compiled_data, |
| 1875 | .tensor_arena = compiled_data->tensor_arena |
| 1876 | }; |
| 1877 | ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states); |
| 1878 | } |
| 1879 | compiled_data->is_test = is_test; |
| 1880 | ccv_nnc_graph_exec_update_t update = { |
| 1881 | .parallel_count = parallel_count, |
| 1882 | .graph = model->graph, |
| 1883 | .graph_exec_arena = compiled_data->graph_exec_arena, |
| 1884 | }; |
| 1885 | ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update); |
| 1886 | const int evaluate_to_size = compiled_data->evaluate.to_size; |
| 1887 | compiled_data->evaluate.to_op_size = 0; |
| 1888 | ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0); |
| 1889 | for (i = 0; i < evaluate_to_size; i++) |
| 1890 | { |
| 1891 | ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]); |
| 1892 | if (to_op.graph) |
| 1893 | compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op; |
| 1894 | const int* tos; |
| 1895 | int to_size; |
| 1896 | ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size); |
| 1897 | for (j = 0; j < to_size; j++) |
| 1898 | { |
| 1899 | ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){ |
| 1900 | .d = tos[j], |
| 1901 | .graph = model->graph |
| 1902 | }); |
| 1903 | if (to_op.graph) |
| 1904 | ccv_array_add_unique_int(backward_from, to_op.d); |
| 1905 | } |
| 1906 | } |
| 1907 | assert(backward_from->rnum > 0)((void) sizeof ((backward_from->rnum > 0) ? 1 : 0), __extension__ ({ if (backward_from->rnum > 0) ; else __assert_fail ( "backward_from->rnum > 0", "ccv_cnnp_model.c", 1907, __extension__ __PRETTY_FUNCTION__); })); |
| 1908 | compiled_data->backward.from_op_size = backward_from->rnum; |
| 1909 | compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum); |
| 1910 | for (i = 0; i < backward_from->rnum; i++) |
| 1911 | compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){ |
| 1912 | .d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from )->rsize * (size_t)(i))), |
| 1913 | .graph = compiled_data->graph, |
| 1914 | }; |
| 1915 | // If there are any set node (to set some tensors to 0) inserted through backward pass, these won't be executed if we just do sources -> evaluate.to_ops, backward.from_ops -> destinations. We need this logic to find out these nodes and explicitly adding them to backward.from_ops. |
| 1916 | ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(compiled_data->graph->exec_info, 0)((void*)(((char*)((compiled_data->graph->exec_info)-> data)) + (size_t)(compiled_data->graph->exec_info)-> rsize * (size_t)(0))); |
| 1917 | const int exec_info_size = compiled_data->graph->exec_info->rnum; |
| 1918 | uint32_t* const visited = cccalloccalloc((exec_info_size + 31) >> 5, sizeof(uint32_t)); |
| 1919 | const ccv_nnc_graph_exec_t* const sources = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->sources, 0)((void*)(((char*)((compiled_data->graph->sources)->data )) + (size_t)(compiled_data->graph->sources)->rsize * (size_t)(0))); |
| 1920 | const int source_size = compiled_data->graph->sources->rnum; |
| 1921 | ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, compiled_data->evaluate.to_ops, compiled_data->evaluate.to_op_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc (sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) * ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t ; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ < (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_ ].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024 ); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * ( exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t ) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_ , 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t* _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), ( int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size), }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (( (sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_] = (sources)[_i_].d; } int _exist_size_[2] = { (source_size), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_ [_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if ( _incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2 ; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info )[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)(( void*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + ( size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_ ))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (( (sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_] = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_ [1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[ _idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings )->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_ [d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d ].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d]. c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue ; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_ ++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_ ].graph == compiled_data->graph) ? 1 : 0), __extension__ ( { if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_]. d].r = 5; _exists_[0][_i_] = (compiled_data->evaluate.to_ops )[_i_].d; } _exist_size_[0] = (compiled_data->evaluate.to_op_size ); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[ _idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_ [_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_ [_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_ ++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_ ].graph == compiled_data->graph) ? 1 : 0), __extension__ ( { if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_]. d].d = 1; } for (_i_ = 0; _i_ < (source_size); _i_++) { (( void) sizeof (((sources)[_i_].graph == compiled_data->graph ) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_ ] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size ].index = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_ [_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info )[_idx_].outgoings) { if ((exec_info)[_idx_].outgoings->rnum == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[ _idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].outgoings )->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_ [d].c == 0 && _incomings_[d].r == 6 && _d_ < (compiled_data->evaluate.to_op_size)) { _exists_[_p_][_i_ ] = d; continue; } } else for (_j_ = 0; _j_ < (exec_info)[ _idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void *)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (size_t )((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); -- _incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_ [d].r == 6 && _d_ < (compiled_data->evaluate.to_op_size )) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size) ) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size )) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size ); _i_++) { ((void) sizeof (((compiled_data->evaluate.to_ops )[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(compiled_data->evaluate.to_ops)[_i_ ].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_ [(compiled_data->evaluate.to_ops)[_i_].d].c == 0) ? 1 : 0) , __extension__ ({ if (_incomings_[(compiled_data->evaluate .to_ops)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(compiled_data->evaluate.to_ops)[_i_].d].c == 0" , "ccv_cnnp_model.c", 1921, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(compiled_data->evaluate.to_ops )[_i_].d].c > 0) continue; _visit_->node[_visit_->size ].index = (((compiled_data->evaluate.to_ops)[_i_].d)); _visit_ ->node[_visit_->size].term = ((_incomings_[(compiled_data ->evaluate.to_ops)[_i_].d].d)); ++_visit_->size;; } if ( _heap_mem_) free(_incomings_); } while (0);; ((void) sizeof ( (_visit_->size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)", "ccv_cnnp_model.c" , 1921, __extension__ __PRETTY_FUNCTION__); })); _visit_; }); |
| 1922 | ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const int idx __attribute__((unused)) = (visit)->node[_i_].index ; const int _node_unused_ __attribute__((unused)) = (visit)-> node[_i_].term; typeof ((exec_info)) const node __attribute__ ((unused)) = (exec_info) + idx; { |
| 1923 | visited[(idx >> 5)] |= (1u << (idx & 31)); |
| 1924 | } ccv_nnc_graph_visit_endfor} } |
| 1925 | ccv_nnc_graph_visit_free(visit); |
| 1926 | const ccv_nnc_graph_exec_t* const destinations = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->destinations, 0)((void*)(((char*)((compiled_data->graph->destinations)-> data)) + (size_t)(compiled_data->graph->destinations)-> rsize * (size_t)(0))); |
| 1927 | const int destination_size = compiled_data->graph->destinations->rnum; |
| 1928 | visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, compiled_data->backward.from_ops, compiled_data->backward.from_op_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc (sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) * ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t ; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ < (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_ ].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024 ); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * ( exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t ) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_ , 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t* _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), ( int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size), }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for (_i_ = 0; _i_ < (compiled_data->backward.from_op_size) ; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops )[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(compiled_data->backward.from_ops)[_i_ ].d].r = 1; _exists_[0][_i_] = (compiled_data->backward.from_ops )[_i_].d; } int _exist_size_[2] = { (compiled_data->backward .from_op_size), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 1) continue; _incomings_[ _idx_].r = 2; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings )->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_ [_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (compiled_data->backward.from_op_size) ; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops )[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(compiled_data->backward.from_ops)[_i_ ].d].r = 3; _exists_[0][_i_] = (compiled_data->backward.from_ops )[_i_].d; } _exist_size_[0] = (compiled_data->backward.from_op_size ); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 3) continue ; _incomings_[_idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum ; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info )[_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_]. outgoings)->rsize * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_[d].edges = _bump_; _bump_ += _incomings_ [d].c; _incomings_[d].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].c] = _idx_; ++_incomings_[d].c; if (_incomings_ [d].r != 2) continue; _incomings_[d].r = 3; ((void) sizeof (( _exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)", "ccv_cnnp_model.c" , 1928, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_ ][_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = ( _p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations) [_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(destinations)[_i_].d].r = 5; _exists_[0] [_i_] = (destinations)[_i_].d; } _exist_size_[0] = (destination_size ); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[ _idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_ [_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_ [_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (compiled_data->backward.from_op_size); _i_++ ) { ((void) sizeof (((compiled_data->backward.from_ops)[_i_ ].graph == compiled_data->graph) ? 1 : 0), __extension__ ( { if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[0][_i_] = (compiled_data->backward.from_ops )[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (compiled_data ->backward.from_op_size); _exist_size_[1] = 0; int _d_ = 0 ; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_ [_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info)[_idx_].outgoings ) { if ((exec_info)[_idx_].outgoings->rnum == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)-> data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * ( size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (destination_size) ) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)-> data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * ( size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (destination_size )) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size) ) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size )) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ( (void) sizeof (((destinations)[_i_].graph == compiled_data-> graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(destinations)[_i_].d].r == 7) continue ; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_ ].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(destinations )[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(destinations)[_i_].d].c > 0 ) continue; _visit_->node[_visit_->size].index = (((destinations )[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_ [(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_ ) free(_incomings_); } while (0);; ((void) sizeof ((_visit_-> size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_ ->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)" , "ccv_cnnp_model.c", 1928, __extension__ __PRETTY_FUNCTION__ ); })); _visit_; }); |
| 1929 | ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const int idx __attribute__((unused)) = (visit)->node[_i_].index ; const int _node_unused_ __attribute__((unused)) = (visit)-> node[_i_].term; typeof ((exec_info)) const node __attribute__ ((unused)) = (exec_info) + idx; { |
| 1930 | visited[(idx >> 5)] |= (1u << (idx & 31)); |
| 1931 | } ccv_nnc_graph_visit_endfor} } |
| 1932 | ccv_nnc_graph_visit_free(visit); |
| 1933 | visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc (sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) * ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t ; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ < (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_ ].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024 ); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * ( exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t ) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_ , 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t* _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), ( int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size), }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (( (sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_] = (sources)[_i_].d; } int _exist_size_[2] = { (source_size), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_ [_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if ( _incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2 ; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info )[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)(( void*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + ( size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_ ))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (( (sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_] = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_ [1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[ _idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings )->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_ [d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d ].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d]. c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue ; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(destinations)[_i_].d].r = 5; _exists_[0] [_i_] = (destinations)[_i_].d; } _exist_size_[0] = (destination_size ); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[ _idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_ [_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_ [_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data ->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources )[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_ ] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size ].index = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_ [_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info )[_idx_].outgoings) { if ((exec_info)[_idx_].outgoings->rnum == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[ _idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].outgoings )->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_ [d].c == 0 && _incomings_[d].r == 6 && _d_ < (destination_size)) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum ; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info )[_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_]. outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (destination_size)) { ((void) sizeof ((_exist_size_ [_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_ [_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ( (void) sizeof (((destinations)[_i_].graph == compiled_data-> graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(destinations)[_i_].d].r == 7) continue ; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_ ].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(destinations )[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(destinations)[_i_].d].c > 0 ) continue; _visit_->node[_visit_->size].index = (((destinations )[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_ [(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_ ) free(_incomings_); } while (0);; ((void) sizeof ((_visit_-> size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_ ->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)" , "ccv_cnnp_model.c", 1933, __extension__ __PRETTY_FUNCTION__ ); })); _visit_; }); |
| 1934 | // Find any missing nodes to be added as source. Right now, these are only set nodes. |
| 1935 | ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const int idx __attribute__((unused)) = (visit)->node[_i_].index ; const int _node_unused_ __attribute__((unused)) = (visit)-> node[_i_].term; typeof ((exec_info)) const node __attribute__ ((unused)) = (exec_info) + idx; { |
| 1936 | if (!(visited[(idx >> 5)] & (1u << (idx & 31)))) |
| 1937 | { |
| 1938 | assert(exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD)((void) sizeof ((exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD ) ? 1 : 0), __extension__ ({ if (exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD ) ; else __assert_fail ("exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD" , "ccv_cnnp_model.c", 1938, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1939 | if (exec_info[idx].cmd.info.blas.a[0] == 0) // Special-casing for empty out the tensor set function, not for the set grad to 1 one. |
| 1940 | ccv_array_add_unique_int(backward_from, idx); |
| 1941 | } |
| 1942 | } ccv_nnc_graph_visit_endfor} } |
| 1943 | ccv_nnc_graph_visit_free(visit); |
| 1944 | ccfreefree(visited); |
| 1945 | if (backward_from->rnum != compiled_data->backward.from_op_size) // If it doesn't match, need to redo this. |
| 1946 | { |
| 1947 | compiled_data->backward.from_op_size = backward_from->rnum; |
| 1948 | compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccreallocrealloc(compiled_data->backward.from_ops, sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum); |
| 1949 | for (i = 0; i < backward_from->rnum; i++) |
| 1950 | compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){ |
| 1951 | .d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from )->rsize * (size_t)(i))), |
| 1952 | .graph = compiled_data->graph, |
| 1953 | }; |
| 1954 | } |
| 1955 | ccv_array_free(backward_from); |
| 1956 | ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count); |
| 1957 | ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0); |
| 1958 | } |
| 1959 | |
| 1960 | void ccv_cnnp_model_dry_run(ccv_cnnp_model_t* const model, const ccv_cnnp_evaluate_param_t params, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size) |
| 1961 | { |
| 1962 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 1963 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 1963, __extension__ __PRETTY_FUNCTION__); })); |
| 1964 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 1965 | assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count ) ? 1 : 0), __extension__ ({ if (output_size == model->output_size * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count" , "ccv_cnnp_model.c", 1965, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1966 | assert(input_size == model->input_size * parallel_count)((void) sizeof ((input_size == model->input_size * parallel_count ) ? 1 : 0), __extension__ ({ if (input_size == model->input_size * parallel_count) ; else __assert_fail ("input_size == model->input_size * parallel_count" , "ccv_cnnp_model.c", 1966, __extension__ __PRETTY_FUNCTION__ ); })); |
| 1967 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 1967, __extension__ __PRETTY_FUNCTION__); })); |
| 1968 | const int target_gradient_mode = _ccv_cnnp_is_disable_outgrad_all(params.disable_outgrad, model->input_size) ? CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES : CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS; |
| 1969 | const int mode_mismatch = (params.requires_grad && (compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->gradient_mode != target_gradient_mode || compiled_data->disable_outgrad != params.disable_outgrad)); |
| 1970 | if (!compiled_data->graph || mode_mismatch) |
| 1971 | { |
| 1972 | _ccv_cnnp_compiled_data_graph_free(compiled_data); |
| 1973 | if (mode_mismatch) // If mode mismatch, we need to redo the backward as well (no need to redo apply_gradients, it doesn't require target_gradient_mode or disable_outgrad. |
| 1974 | _ccv_cnnp_compiled_data_backward_free(compiled_data); |
| 1975 | if (params.requires_grad) |
| 1976 | _ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size); |
| 1977 | else |
| 1978 | _ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size); |
| 1979 | } else { |
| 1980 | ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena); |
| 1981 | assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0) , __extension__ ({ if ((input_size % parallel_count) == 0) ; else __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1981, __extension__ __PRETTY_FUNCTION__); })); |
| 1982 | const int input_size_per_p = input_size / parallel_count; |
| 1983 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count); |
| 1984 | assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0 ), __extension__ ({ if ((output_size % parallel_count) == 0) ; else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c" , 1984, __extension__ __PRETTY_FUNCTION__); })); |
| 1985 | const int output_size_per_p = output_size / parallel_count; |
| 1986 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count); |
| 1987 | } |
| 1988 | if (compiled_data->is_test != params.is_test) |
| 1989 | { |
| 1990 | compiled_data->is_test = params.is_test; |
| 1991 | ccv_nnc_graph_exec_update_t update = { |
| 1992 | .parallel_count = parallel_count, |
| 1993 | .graph = model->graph, |
| 1994 | .graph_exec_arena = compiled_data->graph_exec_arena, |
| 1995 | }; |
| 1996 | ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update); |
| 1997 | } |
| 1998 | } |
| 1999 | |
| 2000 | void ccv_cnnp_model_evaluate(ccv_cnnp_model_t* const model, const ccv_cnnp_evaluate_param_t params, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) |
| 2001 | { |
| 2002 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2003 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 2003, __extension__ __PRETTY_FUNCTION__); })); |
| 2004 | ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size); |
| 2005 | if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD) |
| 2006 | ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context); |
| 2007 | else { |
| 2008 | if (!compiled_data->evaluate.schedule) |
| 2009 | compiled_data->evaluate.schedule = ccv_nnc_graph_static_schedule_new(compiled_data->graph, compiled_data->stream_type, model->max_stream_count, 0, 0, compiled_data->evaluate.to_ops, compiled_data->evaluate.to_op_size); |
| 2010 | ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context); |
| 2011 | } |
| 2012 | } |
| 2013 | |
| 2014 | // Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE). |
| 2015 | // Particularly, this method compiles the accumulator graph. |
| 2016 | static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model) |
| 2017 | { |
| 2018 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2019 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 2019, __extension__ __PRETTY_FUNCTION__); })); |
| 2020 | assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE ) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE" , "ccv_cnnp_model.c", 2020, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2021 | ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new(); |
| 2022 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 2023 | const int parameter_size = compiled_data->parameters->rnum; |
| 2024 | int i, j; |
| 2025 | compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3); |
| 2026 | compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count; |
| 2027 | compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count; |
| 2028 | for (i = 0; i < parameter_size; i++) |
| 2029 | for (j = 0; j < parallel_count; j++) |
| 2030 | if (compiled_data->tensors.gradients[i + j * parameter_size]) |
| 2031 | { |
| 2032 | const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info; |
| 2033 | // Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them. |
| 2034 | compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size]; |
| 2035 | compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0); |
| 2036 | ccv_nnc_tensor_symbol_t inputs[2]; |
| 2037 | inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0); |
| 2038 | inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0); |
| 2039 | ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0); |
| 2040 | ccv_nnc_graph_exec_symbol_new(accum, CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), inputs, 2, &output, 1, 0); |
| 2041 | } else { |
| 2042 | compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 2043 | compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 2044 | compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; |
| 2045 | } |
| 2046 | ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); |
| 2047 | if (ccv_nnc_symbolic_graph_source_size(accum) == 0) |
| 2048 | { |
| 2049 | ccv_nnc_symbolic_graph_free(accum); |
| 2050 | // Create empty graph. |
| 2051 | compiled_data->backward.accum = ccv_nnc_graph_new(); |
| 2052 | ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0); |
| 2053 | return; |
| 2054 | } |
| 2055 | ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0); |
| 2056 | _ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds); |
| 2057 | _ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds); |
| 2058 | _ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds); |
| 2059 | ccv_nnc_symbolic_graph_compile(accum, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds )->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(accum)ccv_nnc_symbolic_graph_sources(accum), ccv_nnc_symbolic_graph_source_size (accum), SYMBOLIC_GRAPH_DESTINATIONS(accum)ccv_nnc_symbolic_graph_destinations(accum), ccv_nnc_symbolic_graph_destination_size (accum), &compiled_data->backward.accum, &compiled_data->backward.tensor_arena, &compiled_data->backward.graph_exec_arena); |
| 2060 | ccv_nnc_symbolic_graph_free(accum); |
| 2061 | ccv_array_free(tensor_binds); |
| 2062 | ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count); |
| 2063 | } |
| 2064 | |
| 2065 | void ccv_cnnp_model_backward(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const ingrads, const int ingrad_size, ccv_nnc_tensor_t* const* const outgrads, const int outgrad_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) |
| 2066 | { |
| 2067 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2068 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 2068, __extension__ __PRETTY_FUNCTION__); })); |
| 2069 | assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE ) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE" , "ccv_cnnp_model.c", 2069, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2070 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 2071 | assert(ingrad_size == 0 || ingrad_size == model->output_size * parallel_count)((void) sizeof ((ingrad_size == 0 || ingrad_size == model-> output_size * parallel_count) ? 1 : 0), __extension__ ({ if ( ingrad_size == 0 || ingrad_size == model->output_size * parallel_count ) ; else __assert_fail ("ingrad_size == 0 || ingrad_size == model->output_size * parallel_count" , "ccv_cnnp_model.c", 2071, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2072 | if (outgrad_size > 0) |
| 2073 | { assert(outgrad_size == compiled_data->outgrad_size * parallel_count)((void) sizeof ((outgrad_size == compiled_data->outgrad_size * parallel_count) ? 1 : 0), __extension__ ({ if (outgrad_size == compiled_data->outgrad_size * parallel_count) ; else __assert_fail ("outgrad_size == compiled_data->outgrad_size * parallel_count" , "ccv_cnnp_model.c", 2073, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 2074 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 2074, __extension__ __PRETTY_FUNCTION__); })); |
| 2075 | assert(compiled_data->graph)((void) sizeof ((compiled_data->graph) ? 1 : 0), __extension__ ({ if (compiled_data->graph) ; else __assert_fail ("compiled_data->graph" , "ccv_cnnp_model.c", 2075, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2076 | const int parameter_size = compiled_data->parameters->rnum; |
| 2077 | // If we need to accumulate the gradients now, do jit on accumulator. |
| 2078 | if (compiled_data->backward.count > 0) |
| 2079 | { |
| 2080 | if (!compiled_data->backward.accum) |
| 2081 | _ccv_cnnp_model_multistage_jit_1(model); |
| 2082 | else if (compiled_data->backward.count == 1) { |
| 2083 | // On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly). |
| 2084 | int i; |
| 2085 | for (i = 0; i < parameter_size * parallel_count; i++) |
| 2086 | { |
| 2087 | ccv_nnc_tensor_t* tensor; |
| 2088 | CCV_SWAP(compiled_data->tensors.accum_gradients[i], compiled_data->tensors.gradients[i], tensor)((tensor) = (compiled_data->tensors.accum_gradients[i]), ( compiled_data->tensors.accum_gradients[i]) = (compiled_data ->tensors.gradients[i]), (compiled_data->tensors.gradients [i]) = (tensor)); |
| 2089 | } |
| 2090 | if (compiled_data->backward.tensor_arena) |
| 2091 | { |
| 2092 | ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena); |
| 2093 | // Do rebind in case we messed up the binding (we switch accum_gradients and gradients). |
| 2094 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->backward.tensor_arena, 0, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1); |
| 2095 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->backward.tensor_arena, 0, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1); |
| 2096 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->backward.tensor_arena, 0, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1); |
| 2097 | } |
| 2098 | } |
| 2099 | } |
| 2100 | const int ingrad_size_per_p = model->output_size; |
| 2101 | const int outgrad_size_per_p = compiled_data->outgrad_size; |
| 2102 | int i, j; |
| 2103 | for (i = 0; i < ingrad_size_per_p; i++) |
| 2104 | { |
| 2105 | const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]); |
| 2106 | if (!ingrad_size || !ingrads || ingrads[i] == 0) |
| 2107 | { |
| 2108 | // Set it to 1 if it is not specified. |
| 2109 | ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad); |
| 2110 | if (ingrad_tensor) |
| 2111 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(ingrad_tensor)(ccv_nnc_tensor_t* []){ingrad_tensor}, (1 +1 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); |
| 2112 | for (j = 1; j < parallel_count; j++) |
| 2113 | { |
| 2114 | ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, ingrad, j)); |
| 2115 | if (ingrad_tensor) |
| 2116 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(ingrad_tensor)(ccv_nnc_tensor_t* []){ingrad_tensor}, (1 +1 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); |
| 2117 | } |
| 2118 | } else { |
| 2119 | // Make sure the length matches, in case it is an alias. |
| 2120 | assert(ccv_nnc_tensor_count(ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params(model->graph, ingrad)))((void) sizeof ((ccv_nnc_tensor_count(ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params(model-> graph, ingrad))) ? 1 : 0), __extension__ ({ if (ccv_nnc_tensor_count (ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params (model->graph, ingrad))) ; else __assert_fail ("ccv_nnc_tensor_count(ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params(model->graph, ingrad))" , "ccv_cnnp_model.c", 2120, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2121 | ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]); |
| 2122 | for (j = 1; j < parallel_count; j++) |
| 2123 | ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, ingrad, j), ingrads[i + ingrad_size_per_p * j]); |
| 2124 | } |
| 2125 | } |
| 2126 | if (outgrad_size > 0) |
| 2127 | { |
| 2128 | assert(compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS && "shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad")((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS && "shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad" ) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS && "shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad" ) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS && \"shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad\"" , "ccv_cnnp_model.c", 2128, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2129 | for (i = 0; i < outgrad_size_per_p; i++) |
| 2130 | if (outgrads[i]) |
| 2131 | { |
| 2132 | const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i]; |
| 2133 | ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]); |
| 2134 | for (j = 1; j < parallel_count; j++) |
| 2135 | ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, outgrad, j), outgrads[i + outgrad_size_per_p * j]); |
| 2136 | } |
| 2137 | } else { |
| 2138 | assert(compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES ||((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS ) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data ->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS ) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS" , "ccv_cnnp_model.c", 2139, __extension__ __PRETTY_FUNCTION__ ); })) |
| 2139 | compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS ) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data ->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS ) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS" , "ccv_cnnp_model.c", 2139, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2140 | } |
| 2141 | // We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients. |
| 2142 | // For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these |
| 2143 | // parameters and internals. The same cannot be said for gradients due to the accum_gradients switching. |
| 2144 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count); |
| 2145 | if (!compiled_data->backward.schedule) |
| 2146 | compiled_data->backward.schedule = ccv_nnc_graph_static_schedule_new(compiled_data->graph, compiled_data->stream_type, model->max_stream_count, compiled_data->backward.from_ops, compiled_data->backward.from_op_size, 0, 0); |
| 2147 | // Run the backward pass. |
| 2148 | ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context); |
| 2149 | // If we need to run accumulation round, do that now. |
| 2150 | if (compiled_data->backward.count > 0) |
| 2151 | ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context); |
| 2152 | // Update the count, this determines whether we need to accumulate or not. |
| 2153 | ++compiled_data->backward.count; |
| 2154 | } |
| 2155 | |
| 2156 | // Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE). |
| 2157 | // Particularly, this method compiles the parameter update graph. |
| 2158 | static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model) |
| 2159 | { |
| 2160 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2161 | assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE ) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE" , "ccv_cnnp_model.c", 2161, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2162 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 2163 | const int parameter_size = compiled_data->parameters->rnum; |
| 2164 | ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0); |
| 2165 | _ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( 0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds); |
| 2166 | _ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds); |
| 2167 | // Bind accumulated gradients. |
| 2168 | if (compiled_data->backward.count > 1) |
| 2169 | _ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds); |
| 2170 | else |
| 2171 | _ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds); |
| 2172 | ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0); |
| 2173 | int i, j; |
| 2174 | for (i = 0; i < compiled_data->backward.to_size; i++) |
| 2175 | { |
| 2176 | const int* tos; |
| 2177 | int to_size; |
| 2178 | ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size); |
| 2179 | for (j = 0; j < to_size; j++) |
| 2180 | { |
| 2181 | // Check if this is already show up in the backward graph, if that is the case, it won't be in the apply |
| 2182 | // gradients graph. |
| 2183 | const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){ |
| 2184 | .d = tos[j], |
| 2185 | .graph = model->graph, |
| 2186 | }); |
| 2187 | if (!exec.graph) |
| 2188 | ccv_array_add_unique_int(apply_gradients_from, tos[j]); |
| 2189 | } |
| 2190 | } |
| 2191 | const int from_size = apply_gradients_from->rnum; |
| 2192 | if (from_size == 0) |
| 2193 | { |
| 2194 | ccv_array_free(apply_gradients_from); |
| 2195 | ccv_array_free(tensor_binds); |
| 2196 | return; |
| 2197 | } |
| 2198 | ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size); |
| 2199 | for (i = 0; i < from_size; i++) |
| 2200 | froms[i] = (ccv_nnc_graph_exec_symbol_t){ |
| 2201 | .d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t )(apply_gradients_from)->rsize * (size_t)(i))), |
| 2202 | .graph = model->graph |
| 2203 | }; |
| 2204 | ccv_array_free(apply_gradients_from); |
| 2205 | // It can only ends with updates on the parameters. |
| 2206 | ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0); |
| 2207 | for (i = 0; i < parameter_size; i++) |
| 2208 | { |
| 2209 | if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 2210 | continue; |
| 2211 | ccv_array_push(tos, &compiled_data->update_nodes[i]); |
| 2212 | for (j = 1; j < parallel_count; j++) |
| 2213 | { |
| 2214 | const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j); |
| 2215 | ccv_array_push(tos, ©); |
| 2216 | } |
| 2217 | } |
| 2218 | ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds )->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, froms, from_size, (ccv_nnc_graph_exec_symbol_t*)ccv_array_get(tos, 0)((void*)(((char*)((tos)->data)) + (size_t)(tos)->rsize * (size_t)(0))), tos->rnum, &compiled_data->apply_gradients.graph, &compiled_data->apply_gradients.tensor_arena, &compiled_data->apply_gradients.graph_exec_arena); |
| 2219 | ccv_array_free(tos); |
| 2220 | ccv_array_free(tensor_binds); |
| 2221 | ccfreefree(froms); |
| 2222 | const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size; |
| 2223 | for (i = 0; i < max_saved_aux_size * parameter_size; i++) |
| 2224 | { |
| 2225 | // Skip on no tensor. |
| 2226 | if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL) |
| 2227 | continue; |
| 2228 | ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source); |
| 2229 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, &tensor, 1, 0); |
| 2230 | for (j = 1; j < parallel_count; j++) |
| 2231 | { |
| 2232 | ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, compiled_data->saved_aux[i].source, j)); |
| 2233 | if (copy) |
| 2234 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, ©, 1, 0); |
| 2235 | } |
| 2236 | } |
| 2237 | ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count); |
| 2238 | } |
| 2239 | |
| 2240 | void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context) |
| 2241 | { |
| 2242 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2243 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c" , 2243, __extension__ __PRETTY_FUNCTION__); })); |
| 2244 | assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE ) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE" , "ccv_cnnp_model.c", 2244, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2245 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); |
| 2246 | assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c" , 2246, __extension__ __PRETTY_FUNCTION__); })); |
| 2247 | assert(compiled_data->graph)((void) sizeof ((compiled_data->graph) ? 1 : 0), __extension__ ({ if (compiled_data->graph) ; else __assert_fail ("compiled_data->graph" , "ccv_cnnp_model.c", 2247, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2248 | // Skip if there is no backward pass. |
| 2249 | if (compiled_data->backward.count <= 0) |
| 2250 | return; |
| 2251 | // Skip if there is no parameters. |
| 2252 | if (compiled_data->parameters->rnum == 0) |
| 2253 | { |
| 2254 | compiled_data->backward.count = 0; |
| 2255 | return; |
| 2256 | } |
| 2257 | if (!compiled_data->apply_gradients.graph) |
| 2258 | _ccv_cnnp_model_multistage_jit_2(model); |
| 2259 | else { |
| 2260 | const int parameter_size = compiled_data->parameters->rnum; |
| 2261 | ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena); |
| 2262 | // Change to bind accum_gradients if we do gradient accumulation (run backward more than once). |
| 2263 | if (compiled_data->backward.count > 1) |
| 2264 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->apply_gradients.tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count); |
| 2265 | else |
| 2266 | _ccv_cnnp_bind_tensors_to_arena(compiled_data->apply_gradients.tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count); |
| 2267 | } |
| 2268 | if (compiled_data->apply_gradients.graph) |
| 2269 | ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context); |
| 2270 | // Reset backward count to 0. |
| 2271 | compiled_data->backward.count = 0; |
| 2272 | } |
| 2273 | |
| 2274 | void ccv_cnnp_model_set_parameter(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, const ccv_nnc_tensor_t* const tensor) |
| 2275 | { |
| 2276 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2277 | const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel; |
| 2278 | assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__ ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0" , "ccv_cnnp_model.c", 2278, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2279 | const int tensors_init = !!compiled_data->tensors_init.v; |
| 2280 | int this_tensor_init = tensors_init; |
| 2281 | if (!tensors_init) |
| 2282 | ccv_cnnp_model_tensors_init_0(model, compiled_data); |
| 2283 | else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1) |
| 2284 | // Check if it is not fully allocated, if it is not, init_1. |
| 2285 | this_tensor_init = 0; |
| 2286 | ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0); |
| 2287 | ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices); |
| 2288 | const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref; |
| 2289 | if (param_ref < 0) |
| 2290 | { assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__ ({ if (parameter_indices->rnum == 1) ; else __assert_fail ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2290 , __extension__ __PRETTY_FUNCTION__); })); } |
| 2291 | else |
| 2292 | { assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ? 1 : 0), __extension__ ({ if (param_ref < parameter_indices ->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum" , "ccv_cnnp_model.c", 2292, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 2293 | const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(param_ref >= 0 ? param_ref : 0))); |
| 2294 | ccv_array_free(parameter_indices); |
| 2295 | const int parameter_size = compiled_data->parameters->rnum; |
| 2296 | assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >= 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2296 , __extension__ __PRETTY_FUNCTION__); })); |
| 2297 | assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__ ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size" , "ccv_cnnp_model.c", 2297, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2298 | const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data); |
| 2299 | int i; |
| 2300 | if (!this_tensor_init) |
| 2301 | { |
| 2302 | if (compiled_data->tensors.parameters[d]) |
| 2303 | { |
| 2304 | for (i = 1; i < parallel_count; i++) |
| 2305 | { assert(compiled_data->tensors.parameters[d + i * parameter_size])((void) sizeof ((compiled_data->tensors.parameters[d + i * parameter_size]) ? 1 : 0), __extension__ ({ if (compiled_data ->tensors.parameters[d + i * parameter_size]) ; else __assert_fail ("compiled_data->tensors.parameters[d + i * parameter_size]" , "ccv_cnnp_model.c", 2305, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 2306 | this_tensor_init = 1; |
| 2307 | } else { |
| 2308 | const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, d)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( d))); |
| 2309 | ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter); |
| 2310 | if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY) |
| 2311 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 2312 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8); |
| 2313 | compiled_data->tensors.parameters[d] = ccv_nnc_tensor_new(0, info, 0); |
| 2314 | for (i = 1; i < parallel_count; i++) |
| 2315 | { |
| 2316 | if (i != device_id) |
| 2317 | CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff ) << 8)); |
| 2318 | else |
| 2319 | CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff ) << 8)); |
| 2320 | compiled_data->tensors.parameters[d + i * parameter_size] = ccv_nnc_tensor_new(0, info, 0); |
| 2321 | } |
| 2322 | } |
| 2323 | } |
| 2324 | ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters [d]) & ~(uintptr_t)1)); |
| 2325 | assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ; else __assert_fail ("dest", "ccv_cnnp_model.c", 2325, __extension__ __PRETTY_FUNCTION__); })); |
| 2326 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)tensor)(ccv_nnc_tensor_t* []){(ccv_nnc_tensor_t*)tensor}, (1 +1 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1 ), TENSOR_LIST(dest)(ccv_nnc_tensor_t* []){dest}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0); |
| 2327 | for (i = 1; i < parallel_count; i++) |
| 2328 | { |
| 2329 | ccv_nnc_tensor_t* const copy_tensor = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d + i * parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters [d + i * parameter_size]) & ~(uintptr_t)1)); |
| 2330 | if (copy_tensor) |
| 2331 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, TENSOR_LIST(dest)(ccv_nnc_tensor_t* []){dest}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(copy_tensor)(ccv_nnc_tensor_t* []){copy_tensor}, (1 +1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0); |
| 2332 | } |
| 2333 | // Mark this symbol as init'ed. |
| 2334 | const int s = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, d)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( d))))->d; |
| 2335 | uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 2336 | init_v[s >> 5] |= (1u << (s & 0x1f)); |
| 2337 | // If we just allocated this tensor, now it is time to check if we need to mark it as fully allocated. |
| 2338 | if (!this_tensor_init) |
| 2339 | { |
| 2340 | if (ccv_cnnp_model_tensors_any_to_alloc(model, compiled_data)) |
| 2341 | compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)compiled_data->tensors_init.v | (uintptr_t)1); |
| 2342 | else // Remove the flag. |
| 2343 | compiled_data->tensors_init.v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)); |
| 2344 | } |
| 2345 | } |
| 2346 | |
| 2347 | void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor) |
| 2348 | { |
| 2349 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2350 | const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel; |
| 2351 | assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__ ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0" , "ccv_cnnp_model.c", 2351, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2352 | assert(compiled_data->tensors.parameters)((void) sizeof ((compiled_data->tensors.parameters) ? 1 : 0 ), __extension__ ({ if (compiled_data->tensors.parameters) ; else __assert_fail ("compiled_data->tensors.parameters" , "ccv_cnnp_model.c", 2352, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2353 | ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0); |
| 2354 | ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices); |
| 2355 | const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref; |
| 2356 | if (param_ref < 0) |
| 2357 | { assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__ ({ if (parameter_indices->rnum == 1) ; else __assert_fail ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2357 , __extension__ __PRETTY_FUNCTION__); })); } |
| 2358 | else |
| 2359 | { assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ? 1 : 0), __extension__ ({ if (param_ref < parameter_indices ->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum" , "ccv_cnnp_model.c", 2359, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 2360 | const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(param_ref >= 0 ? param_ref : 0))); |
| 2361 | ccv_array_free(parameter_indices); |
| 2362 | const int parameter_size = compiled_data->parameters->rnum; |
| 2363 | assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >= 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2363 , __extension__ __PRETTY_FUNCTION__); })); |
| 2364 | assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__ ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size" , "ccv_cnnp_model.c", 2364, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2365 | // We don't need to consider parallel_count, every parameter on each device is identical. |
| 2366 | ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters [d]) & ~(uintptr_t)1)); |
| 2367 | assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else __assert_fail ("src", "ccv_cnnp_model.c", 2367, __extension__ __PRETTY_FUNCTION__); })); |
| 2368 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, TENSOR_LIST(src)(ccv_nnc_tensor_t* []){src}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(tensor)(ccv_nnc_tensor_t* []){tensor}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0); |
| 2369 | } |
| 2370 | |
| 2371 | ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter) |
| 2372 | { |
| 2373 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2374 | const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel; |
| 2375 | assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__ ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0" , "ccv_cnnp_model.c", 2375, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2376 | assert(compiled_data->tensors.parameters)((void) sizeof ((compiled_data->tensors.parameters) ? 1 : 0 ), __extension__ ({ if (compiled_data->tensors.parameters) ; else __assert_fail ("compiled_data->tensors.parameters" , "ccv_cnnp_model.c", 2376, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2377 | ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0); |
| 2378 | ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices); |
| 2379 | const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref; |
| 2380 | if (param_ref < 0) |
| 2381 | { assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__ ({ if (parameter_indices->rnum == 1) ; else __assert_fail ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2381 , __extension__ __PRETTY_FUNCTION__); })); } |
| 2382 | else |
| 2383 | { assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ? 1 : 0), __extension__ ({ if (param_ref < parameter_indices ->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum" , "ccv_cnnp_model.c", 2383, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 2384 | const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(param_ref >= 0 ? param_ref : 0))); |
| 2385 | ccv_array_free(parameter_indices); |
| 2386 | const int parameter_size = compiled_data->parameters->rnum; |
| 2387 | assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >= 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2387 , __extension__ __PRETTY_FUNCTION__); })); |
| 2388 | assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__ ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size" , "ccv_cnnp_model.c", 2388, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2389 | // We don't need to consider parallel_count, every parameter on each device is identical. |
| 2390 | ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters [d]) & ~(uintptr_t)1)); |
| 2391 | assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor ) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2391, __extension__ __PRETTY_FUNCTION__); })); |
| 2392 | return tensor->info; |
| 2393 | } |
| 2394 | |
| 2395 | const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter) |
| 2396 | { |
| 2397 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2398 | const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel; |
| 2399 | assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__ ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0" , "ccv_cnnp_model.c", 2399, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2400 | ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0); |
| 2401 | ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices); |
| 2402 | const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref; |
| 2403 | if (param_ref < 0) |
| 2404 | { assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__ ({ if (parameter_indices->rnum == 1) ; else __assert_fail ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2404 , __extension__ __PRETTY_FUNCTION__); })); } |
| 2405 | else |
| 2406 | { assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ? 1 : 0), __extension__ ({ if (param_ref < parameter_indices ->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum" , "ccv_cnnp_model.c", 2406, __extension__ __PRETTY_FUNCTION__ ); })); } |
| 2407 | const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices )->rsize * (size_t)(param_ref >= 0 ? param_ref : 0))); |
| 2408 | ccv_array_free(parameter_indices); |
| 2409 | const int parameter_size = compiled_data->parameters->rnum; |
| 2410 | assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >= 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2410 , __extension__ __PRETTY_FUNCTION__); })); |
| 2411 | assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__ ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size" , "ccv_cnnp_model.c", 2411, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2412 | return *(char**)ccv_array_get(compiled_data->ids.parameters, d)((void*)(((char*)((compiled_data->ids.parameters)->data )) + (size_t)(compiled_data->ids.parameters)->rsize * ( size_t)(d))); |
| 2413 | } |
| 2414 | |
| 2415 | int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model) |
| 2416 | { |
| 2417 | assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__ ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data" , "ccv_cnnp_model.c", 2417, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2418 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2419 | return compiled_data->parameters->rnum; |
| 2420 | } |
| 2421 | |
| 2422 | uint64_t ccv_cnnp_model_parameters_size(ccv_cnnp_model_t* const model) |
| 2423 | { |
| 2424 | assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__ ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data" , "ccv_cnnp_model.c", 2424, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2425 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2426 | const int parameter_size = compiled_data->parameters->rnum; |
| 2427 | int i; |
| 2428 | const ccv_nnc_symbolic_graph_t* const graph = model->graph; |
| 2429 | uint64_t size = 0; |
| 2430 | const int tensors_init = !!compiled_data->tensors_init.v; |
| 2431 | uint32_t* const init_v = tensors_init ? CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) & ~(uintptr_t)1)) : 0; |
| 2432 | for (i = 0; i < parameter_size; i++) |
| 2433 | { |
| 2434 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) + (size_t)(compiled_data->parameters)->rsize * (size_t)( i))))->d; |
| 2435 | if (tensors_init && compiled_data->tensors.parameters && (init_v[d >> 5] | (1u << (d & 0x1f))) && compiled_data->tensors.parameters[i]) |
| 2436 | { |
| 2437 | ccv_nnc_tensor_param_t params = compiled_data->tensors.parameters[i]->info; |
| 2438 | size += ccv_nnc_tensor_data_size(params); |
| 2439 | continue; |
| 2440 | } |
| 2441 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, (ccv_nnc_tensor_symbol_t){ |
| 2442 | .graph = graph, |
| 2443 | .d = d |
| 2444 | }); |
| 2445 | size += ccv_nnc_tensor_data_size(params); |
| 2446 | } |
| 2447 | return size; |
| 2448 | } |
| 2449 | |
| 2450 | int ccv_cnnp_model_parameters_move(ccv_cnnp_model_t* const model, char** const names, ccv_nnc_tensor_t** const tensors, const int count, int type) |
| 2451 | { |
| 2452 | assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__ ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data" , "ccv_cnnp_model.c", 2452, __extension__ __PRETTY_FUNCTION__ ); })); |
| 2453 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; |
| 2454 | if (count != compiled_data->parameters->rnum) |
| 2455 | return 0; |
| 2456 | if (CCV_TENSOR_GET_DEVICE(type)((type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY) |
| 2457 | CCV_TENSOR_SET_DEVICE_ID(type, 0)(type) = (((type) & ~0xfff00) | (((0) & 0xfff) << 8)); |
| 2458 | int i; |
| 2459 | // We don't need to consider parallel_count, every parameter on each device is identical. |
| 2460 | for (i = 0; i < count; i++) |
| 2461 | { |
| 2462 | ccv_nnc_tensor_t* tensor = compiled_data->tensors.parameters[i]; |
| 2463 | if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything. |
| 2464 | { |
| 2465 | tensors[i] = 0; |
| 2466 | continue; |
| 2467 | } |
| 2468 | tensor = CCV_NNC_TENSOR(tensor)((ccv_nnc_tensor_t*)((uintptr_t)(tensor) & ~(uintptr_t)1) ); |
| 2469 | if (tensor->info.type == type) |
| 2470 | tensors[i] = tensor; |
| 2471 | else { |
| 2472 | ccv_nnc_tensor_param_t info = tensor->info; |
| 2473 | info.type = type; |
| 2474 | tensors[i] = ccv_nnc_tensor_new(0, info, 0); // Create this tensor, don't initiate copy yet. |
| 2475 | } |
| 2476 | } |
| 2477 | for (i = 0; i < count; i++) |
| 2478 | { |
| 2479 | ccv_nnc_tensor_t* tensor = compiled_data->tensors.parameters[i]; |
| 2480 | if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything. |
| 2481 | continue; |
| 2482 | tensor = CCV_NNC_TENSOR(tensor)((ccv_nnc_tensor_t*)((uintptr_t)(tensor) & ~(uintptr_t)1) ); |
| 2483 | // Now initiate transfer. We should do this one on a stream. |
| 2484 | if (tensor->info.type != type) |
| 2485 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, TENSOR_LIST(tensor)(ccv_nnc_tensor_t* []){tensor}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(tensors[i])(ccv_nnc_tensor_t* []){tensors[i]}, (1 +1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0); |
| 2486 | } |
| 2487 | // Copy names and remove parameters. |
| 2488 | for (i = 0; i < count; i++) |
| 2489 | { |
| 2490 | ccv_nnc_tensor_t* const tensor = compiled_data->tensors.parameters[i]; |
| 2491 | if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything. |
| 2492 | { |
| 2493 | names[i] = 0; |
| 2494 | continue; |
| 2495 | } |
| 2496 | const char* const name = *(char**)ccv_array_get(compiled_data->ids.parameters, i)((void*)(((char*)((compiled_data->ids.parameters)->data )) + (size_t)(compiled_data->ids.parameters)->rsize * ( size_t)(i))); |
| 2497 | const size_t name_len = ccv_min(strnlen(name, 1023), 1023)({ typeof (strnlen(name, 1023)) _a = (strnlen(name, 1023)); typeof (1023) _b = (1023); (_a < _b) ? _a : _b; }); |
| 2498 | names[i] = ccmallocmalloc(name_len + 1); |
| 2499 | names[i][name_len] = 0; |
| 2500 | memcpy(names[i], name, name_len); |
| 2501 | if (tensor->info.type == type) |
| 2502 | compiled_data->tensors.parameters[i] = 0; // Only move when it is moved. |
| 2503 | } |
| 2504 | return 1; |
| 2505 | } |
| 2506 | |
| 2507 |