You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

model_builder.cc 43 kB

5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #include "graph/build/model_builder.h"
  17. #include <securectype.h>
  18. #include <iostream>
  19. #include <set>
  20. #include <unordered_map>
  21. #include "common/ge/ge_util.h"
  22. #include "common/dump/dump_manager.h"
  23. #include "framework/common/debug/ge_log.h"
  24. #include "graph/anchor.h"
  25. #include "graph/attr_value.h"
  26. #include "graph/buffer.h"
  27. #include "graph/build/stream_allocator.h"
  28. #include "graph/common/omg_util.h"
  29. #include "graph/common/ge_call_wrapper.h"
  30. #include "graph/common/local_context.h"
  31. #include "graph/debug/ge_attr_define.h"
  32. #include "graph/ge_attr_value.h"
  33. #include "graph/ge_context.h"
  34. #include "graph/ge_error_codes.h"
  35. #include "graph/manager/graph_mem_allocator.h"
  36. #include "graph/manager/graph_var_manager.h"
  37. #include "graph/optimize/common/params.h"
  38. #include "graph/types.h"
  39. #include "graph/utils/attr_utils.h"
  40. #include "graph/utils/graph_utils.h"
  41. #include "graph/utils/node_utils.h"
  42. #include "graph/utils/op_desc_utils.h"
  43. #include "graph/utils/tensor_utils.h"
  44. #include "graph/utils/type_utils.h"
  45. #include "init/gelib.h"
  46. #include "memory/memory_assigner.h"
  47. #include "omg/version.h"
  48. #include "register/op_registry.h"
  49. #include "graph/passes/set_input_output_offset_pass.h"
  50. using std::map;
  51. using std::set;
  52. using std::string;
  53. using std::vector;
  54. namespace {
  55. const uint32_t kWeightsStartOffset = 512;
  56. const int32_t kWrongIndex = -2;
  57. const int kInvalidIndexNum = -1;
  58. const char *const kVectorCore = "VectorCore";
  59. const char *const kCoreType = "ge.engineType";
  60. const std::string kEnableL1Fusion = "ge.l1Fusion";
  61. const set<string> adjust_layer_type_ = {ge::CONVOLUTION};
  62. bool IsGeLocalOp(const ge::ConstOpDescPtr &op_desc) {
  63. auto type = op_desc->GetType();
  64. if (type == ge::CONSTANTOP) {
  65. // constant op just has one output
  66. ge::GeTensorDesc output_desc = op_desc->GetOutputDesc(0);
  67. return !(output_desc.GetDataType() == ge::DT_STRING);
  68. }
  69. const set<string> ge_local_set = {ge::STREAMMERGE, ge::MEMCPYASYNC, ge::STREAMACTIVE, ge::STREAMSWITCH,
  70. ge::VARIABLE, ge::NOOP, ge::CONSTANT, ge::ENTER,
  71. ge::REFENTER, ge::LOOPCOND, ge::NEXTITERATION, ge::REFNEXTITERATION,
  72. ge::EXIT, ge::REFEXIT, ge::MERGE, ge::MEMCPYADDRASYNC};
  73. return (ge_local_set.find(type) != ge_local_set.end());
  74. }
  75. } // namespace
  76. namespace ge {
  77. ModelBuilder::ModelBuilder(uint64_t session_id, ge::ComputeGraphPtr compute_graph,
  78. const Graph2SubGraphInfoList &subgraphs, const map<string, int> &stream_max_parallel_num,
  79. bool hcom_parallel, int mode)
  80. : session_id_(session_id),
  81. weight_offset_(kWeightsStartOffset),
  82. compute_graph_(std::move(compute_graph)),
  83. subgraphs_(subgraphs),
  84. stream_num_(0),
  85. event_num_(0),
  86. label_num_(0),
  87. stream_max_parallel_num_(stream_max_parallel_num),
  88. hcom_parallel_(hcom_parallel),
  89. build_mode_(mode),
  90. max_mem_offset_(0),
  91. p2p_mem_offset_(0),
  92. zero_copy_mem_size_(0),
  93. platform_type_(0),
  94. is_loop_graph_(false),
  95. is_l1_fusion_enable_(false) {}
  96. ModelBuilder::~ModelBuilder() {}
  97. Status ModelBuilder::CalcOutputSize(const ge::NodePtr &n) {
  98. GE_CHECK_NOTNULL(n);
  99. auto node_op_desc = n->GetOpDesc();
  100. GE_CHECK_NOTNULL(node_op_desc);
  101. uint32_t index = 0;
  102. for (const auto &output_desc_ptr : node_op_desc->GetAllOutputsDescPtr()) {
  103. GeTensorDesc &desc_temp = *output_desc_ptr;
  104. uint32_t dim_num = static_cast<uint32_t>(desc_temp.GetShape().GetDimNum());
  105. GE_IF_BOOL_EXEC(dim_num > DIM_DEFAULT_SIZE, TensorUtils::SetRealDimCnt(desc_temp, dim_num));
  106. // calculate tensor size
  107. int64_t size_temp = 0;
  108. graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(desc_temp, size_temp);
  109. if (graph_status != GRAPH_SUCCESS) {
  110. REPORT_CALL_ERROR("E19999", "Get tensor size in bytes failed for op:%s(%s) index:%u when CalcOutputSize",
  111. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str(), index);
  112. GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!");
  113. return FAILED;
  114. }
  115. TensorUtils::SetSize(desc_temp, size_temp);
  116. if (node_op_desc->UpdateOutputDesc(index, desc_temp) != SUCCESS) {
  117. REPORT_CALL_ERROR("E19999", "Update Output desc size failed for op:%s(%s) index:%u when CalcOutputSize",
  118. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str(), index);
  119. GELOGE(FAILED, "UpdateOutputDesc failed.");
  120. return FAILED;
  121. }
  122. GELOGD("update output desc, dim_size: %u, mem_size: %ld, format: %s, type: %s, node name:%s", dim_num, size_temp,
  123. TypeUtils::FormatToSerialString(desc_temp.GetFormat()).c_str(),
  124. TypeUtils::DataTypeToSerialString(desc_temp.GetDataType()).c_str(), node_op_desc->GetName().c_str());
  125. index++;
  126. }
  127. return SUCCESS;
  128. }
  129. bool ModelBuilder::SetInputConst(const OpDescPtr &op_desc, const NodePtr &src_node, size_t index,
  130. vector<bool> &is_input_const) {
  131. GELOGI("SetIsInputConst const: %s, source node: %s", op_desc->GetName().c_str(), src_node->GetName().c_str());
  132. for (size_t i = is_input_const.size(); i <= index; ++i) {
  133. is_input_const.push_back(false);
  134. }
  135. is_input_const[index] = true;
  136. vector<GeTensorPtr> weights = OpDescUtils::MutableWeights(src_node);
  137. if (weights.empty()) {
  138. GELOGW("SetInputIsConst weights is empty, node: %s", src_node->GetName().c_str());
  139. return false;
  140. }
  141. GeTensorPtr weight = weights[0];
  142. GE_IF_BOOL_EXEC(weight == nullptr, return true);
  143. GeTensorDesc &tensor_desc = weight->MutableTensorDesc();
  144. int64_t data_offset = 0;
  145. if (TensorUtils::GetDataOffset(tensor_desc, data_offset) != GRAPH_SUCCESS) {
  146. GELOGW("Get Offset from weight failed");
  147. return false;
  148. }
  149. auto input_tensor = op_desc->MutableInputDesc(static_cast<uint32_t>(index));
  150. if (input_tensor == nullptr) {
  151. GELOGW("Get input_tensor failed");
  152. return false;
  153. }
  154. TensorUtils::SetDataOffset(*input_tensor, data_offset);
  155. return true;
  156. }
  157. void ModelBuilder::SetInputIsConst(const ge::NodePtr &n) {
  158. auto node_op_desc = n->GetOpDesc();
  159. GE_CHECK_NOTNULL_JUST_RETURN(node_op_desc);
  160. auto is_input_const = node_op_desc->GetIsInputConst();
  161. // must set all true input_const to false
  162. for (size_t i = 0; i < is_input_const.size(); i++) {
  163. is_input_const[i] = false;
  164. }
  165. std::string const_type;
  166. auto in_data_anchors = n->GetAllInDataAnchors();
  167. for (size_t index = 0; index < in_data_anchors.size(); index++) {
  168. auto in_data_anchor = in_data_anchors.at(index);
  169. const auto &peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  170. GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
  171. const auto &src_node = peer_out_anchor->GetOwnerNode();
  172. if (!NodeUtils::GetConstOpType(src_node, const_type)) {
  173. continue;
  174. }
  175. if (const_type == CONSTANT) {
  176. if (!SetInputConst(node_op_desc, src_node, index, is_input_const)) {
  177. return;
  178. }
  179. } else {
  180. if ((index < is_input_const.size()) && is_input_const[index]) {
  181. is_input_const[index] = false;
  182. }
  183. }
  184. }
  185. std::string input_const_info = ToString(is_input_const);
  186. GELOGD("update opdesc:%s InputConst:%s", node_op_desc->GetName().c_str(), input_const_info.c_str());
  187. node_op_desc->SetIsInputConst(is_input_const);
  188. }
  189. Status ModelBuilder::AdjustConstWeightSize(const ge::NodePtr &node, size_t &mem_offset) {
  190. GE_CHECK_NOTNULL(node);
  191. if (node->GetType() == CONSTANT) {
  192. vector<GeTensorPtr> weights = OpDescUtils::MutableWeights(node);
  193. if (weights.empty()) {
  194. REPORT_INNER_ERROR("E19999", "Check weights size of node %s(%s) is empty when AdjustConstWeightSize",
  195. node->GetName().c_str(), node->GetType().c_str());
  196. GELOGE(FAILED, "weights size of node %s is empty", node->GetName().c_str());
  197. return FAILED;
  198. }
  199. GeTensorPtr weight = weights[0];
  200. if (weight == nullptr) {
  201. REPORT_INNER_ERROR("E19999", "Check weight of node %s(%s) is nullptr when AdjustConstWeightSize",
  202. node->GetName().c_str(), node->GetType().c_str());
  203. GELOGE(FAILED, "weights[0] is null.");
  204. return FAILED;
  205. }
  206. GeTensorDesc &tensor_desc = weight->MutableTensorDesc();
  207. size_t output_size = weight->GetData().size();
  208. TensorUtils::SetDataOffset(tensor_desc, mem_offset);
  209. GELOGD("Node: %s, weight size: %zu.", node->GetName().c_str(), output_size);
  210. mem_offset += output_size;
  211. }
  212. return SUCCESS;
  213. }
  214. Status ModelBuilder::SetInputOutputDesc() {
  215. Status ret;
  216. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  217. auto node_op_desc = n->GetOpDesc();
  218. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  219. if (!is_loop_graph_ && node_op_desc->GetType() == LOOPCOND) {
  220. is_loop_graph_ = true;
  221. }
  222. // if user set input node format ND, the expected node for data and netoutput format is ND in
  223. // final graph.
  224. if ((GetLocalOmgContext().format == domi::DOMI_TENSOR_ND) && (!node_op_desc->HasAttr("_is_single_op")) &&
  225. ((node_op_desc->GetType() == DATA_TYPE) || (node_op_desc->GetType() == NETOUTPUT))) {
  226. auto inputDescsPtr = node_op_desc->GetAllInputsDescPtr();
  227. auto outputDescsPtr = node_op_desc->GetAllOutputsDescPtr();
  228. ge::Format format = ge::FORMAT_ND;
  229. for (auto &inputDescPtr : inputDescsPtr) {
  230. GE_CHECK_NOTNULL(inputDescPtr);
  231. inputDescPtr->SetFormat(format);
  232. inputDescPtr->SetOriginFormat(format);
  233. }
  234. for (auto &outputDescPtr : outputDescsPtr) {
  235. GE_CHECK_NOTNULL(outputDescPtr);
  236. outputDescPtr->SetFormat(format);
  237. outputDescPtr->SetOriginFormat(format);
  238. }
  239. }
  240. if (node_op_desc->GetType() == DATA_TYPE || node_op_desc->GetType() == AIPP_DATA_TYPE) {
  241. GELOGD("Data node: %s.", n->GetName().c_str());
  242. continue;
  243. }
  244. GE_IF_BOOL_EXEC(n->GetInAllNodes().empty() && n->GetOutAllNodes().empty(), continue;);
  245. SetInputIsConst(n);
  246. bool is_unknow = false;
  247. (void)NodeUtils::GetNodeUnknownShapeStatus(*n, is_unknow);
  248. if ((IsGeLocalOp(n->GetOpDesc())) && (!is_unknow)) {
  249. GE_CHK_STATUS_RET(CalcOutputSize(n), "Calculate output size failed");
  250. }
  251. ret = AdjustConstWeightSize(n, weight_offset_);
  252. GE_CHK_STATUS_RET(ret, "AdjustConstWeightSize failed");
  253. GE_IF_BOOL_EXEC(((weight_offset_ > 0) && (weight_offset_ % MEM_ALIGN_SIZE != 0)),
  254. weight_offset_ = (weight_offset_ + MEM_ALIGN_SIZE - 1) / MEM_ALIGN_SIZE * MEM_ALIGN_SIZE);
  255. }
  256. GE_CHK_STATUS_RET(compute_graph_->TopologicalSorting(), "TopologicalSorting failed");
  257. return SUCCESS;
  258. }
  259. void ModelBuilder::AddNodeInputProperty() {
  260. for (const ge::NodePtr &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  261. auto node_op_desc = node->GetOpDesc();
  262. GE_IF_BOOL_EXEC(node_op_desc == nullptr, GELOGW("node_op_desc is nullptr!"); return);
  263. vector<string> src_name_list;
  264. vector<int64_t> src_index_list;
  265. for (const auto &in_data_anchor : node->GetAllInDataAnchors()) {
  266. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  267. GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
  268. GE_IF_BOOL_EXEC(node_op_desc->HasAttr(MERGE_PRENODE_FLAG), continue);
  269. ge::NodePtr src_node = peer_out_anchor->GetOwnerNode();
  270. src_name_list.emplace_back(src_node->GetName());
  271. src_index_list.emplace_back(peer_out_anchor->GetIdx());
  272. }
  273. auto in_control_anchor = node->GetInControlAnchor();
  274. if (in_control_anchor != nullptr) {
  275. string src_name_temp;
  276. for (const auto &out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
  277. ge::NodePtr src_node = out_control_anchor->GetOwnerNode();
  278. src_name_temp = src_name_temp.empty() ? src_node->GetName() : src_name_temp + ":" + src_node->GetName();
  279. }
  280. GE_IF_BOOL_EXEC(!src_name_temp.empty(), src_name_list.emplace_back(src_name_temp);)
  281. }
  282. node_op_desc->SetSrcName(src_name_list);
  283. node_op_desc->SetSrcIndex(src_index_list);
  284. }
  285. for (const ge::NodePtr &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  286. auto node_op_desc = node->GetOpDesc();
  287. GE_IF_BOOL_EXEC(node_op_desc == nullptr, GELOGW("node_op_desc is nullptr!"); return);
  288. GE_IF_BOOL_EXEC(node_op_desc->GetType() == NETOUTPUT, continue);
  289. auto out_control_anchor = node->GetOutControlAnchor();
  290. GE_IF_BOOL_EXEC(out_control_anchor == nullptr, GELOGW("out_control_anchor is nullptr"); return);
  291. vector<string> dst_name_list;
  292. vector<int64_t> dst_index_list;
  293. string dst_name_temp;
  294. for (const auto &in_control_anchor : out_control_anchor->GetPeerInControlAnchors()) {
  295. ge::NodePtr dst_node = in_control_anchor->GetOwnerNode();
  296. dst_name_temp = dst_name_temp.empty() ? dst_node->GetName() : dst_name_temp + ":" + dst_node->GetName();
  297. }
  298. GE_IF_BOOL_EXEC(!dst_name_temp.empty(), dst_name_list.emplace_back(dst_name_temp));
  299. GE_IF_BOOL_EXEC(!out_control_anchor->GetPeerInControlAnchors().empty(),
  300. dst_index_list.emplace_back(kInvalidIndexNum));
  301. for (const auto &out_data_anchor : node->GetAllOutDataAnchors()) {
  302. GE_IF_BOOL_EXEC(node_op_desc->HasAttr(MERGE_PRENODE_FLAG), break);
  303. dst_name_temp = "";
  304. int64_t dst_index = kWrongIndex; // assign an impossible value to dst_index.
  305. for (const auto &in_data_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  306. GE_IF_BOOL_EXEC(in_data_anchor == nullptr, GELOGW("in_data_anchor is nullptr"); return);
  307. ge::NodePtr dst_node = in_data_anchor->GetOwnerNode();
  308. dst_name_temp = dst_name_temp.empty() ? dst_node->GetName() : dst_name_temp + ":" + dst_node->GetName();
  309. dst_index = in_data_anchor->GetIdx();
  310. }
  311. GE_IF_BOOL_EXEC(dst_index != kWrongIndex, dst_index_list.emplace_back(dst_index)); // not found
  312. GE_IF_BOOL_EXEC(!dst_name_temp.empty(), dst_name_list.emplace_back(dst_name_temp));
  313. }
  314. node_op_desc->SetDstName(dst_name_list);
  315. node_op_desc->SetDstIndex(dst_index_list);
  316. }
  317. }
  318. Status ModelBuilder::AdjustInputTensorFlag() {
  319. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  320. if ((n->GetType() == DATA_TYPE) || (n->GetType() == AIPP_DATA_TYPE)) {
  321. GELOGD("Data node: %s.", n->GetName().c_str());
  322. for (const auto &anchor : n->GetAllOutDataAnchors()) {
  323. for (const auto &in_anchors : anchor->GetPeerInDataAnchors()) {
  324. GE_IF_BOOL_EXEC(in_anchors == nullptr, continue);
  325. auto owner_node = in_anchors->GetOwnerNode();
  326. auto owner_node_op_desc = owner_node->GetOpDesc();
  327. GE_IF_BOOL_EXEC(owner_node_op_desc == nullptr, continue);
  328. auto input_desc = owner_node_op_desc->GetInputDesc(in_anchors->GetIdx());
  329. ge::TensorUtils::SetInputTensor(input_desc, true);
  330. if (owner_node_op_desc->UpdateInputDesc(in_anchors->GetIdx(), input_desc) != SUCCESS) {
  331. REPORT_CALL_ERROR("E19999", "Update Input desc size failed for op:%s(%s) index:%u when %s",
  332. owner_node_op_desc->GetName().c_str(), owner_node_op_desc->GetType().c_str(),
  333. in_anchors->GetIdx(), __FUNCTION__);
  334. GELOGE(FAILED, "UpdateOutputDesc failed.");
  335. return FAILED;
  336. }
  337. }
  338. }
  339. }
  340. }
  341. return SUCCESS;
  342. }
  343. void ModelBuilder::InitL1FusionOption() {
  344. string buffer_optimize = "off_optimize";
  345. graphStatus ret = ge::GetContext().GetOption(BUFFER_OPTIMIZE, buffer_optimize);
  346. if (ret == GRAPH_SUCCESS) {
  347. bool off_superkernel = false;
  348. (void)AttrUtils::GetBool(compute_graph_, ATTR_NAME_OFF_SUPERKERNEL_ATTR, off_superkernel);
  349. is_l1_fusion_enable_ = ((buffer_optimize == "l1_optimize") && (!off_superkernel));
  350. GELOGI("Compute graph %s the value of %s is %s, superkernel flag %d.", compute_graph_->GetName().c_str(),
  351. BUFFER_OPTIMIZE.c_str(), buffer_optimize.c_str(), is_l1_fusion_enable_);
  352. } else {
  353. GELOGW("The value of %s is empty.", kEnableL1Fusion.c_str());
  354. }
  355. }
  356. Status ModelBuilder::BuildModelDef(ge::Model &model) {
  357. ClearOriginalFormat();
  358. max_mem_offset_ = mem_type_to_mem_offset_[RT_MEMORY_HBM];
  359. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_MEMORY_SIZE, max_mem_offset_),
  360. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  361. ATTR_MODEL_MEMORY_SIZE.c_str(), __FUNCTION__);
  362. GELOGE(FAILED, "SetInt of ATTR_MODEL_MEMORY_SIZE failed.");
  363. return FAILED);
  364. if (mem_type_to_mem_offset_.find(RT_MEMORY_P2P_DDR) != mem_type_to_mem_offset_.end()) {
  365. p2p_mem_offset_ = mem_type_to_mem_offset_[RT_MEMORY_P2P_DDR];
  366. }
  367. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_P2P_MEMORY_SIZE, p2p_mem_offset_),
  368. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  369. ATTR_MODEL_P2P_MEMORY_SIZE.c_str(), __FUNCTION__);
  370. GELOGE(FAILED, "SetInt of ATTR_MODEL_P2P_MEMORY_SIZE failed.");
  371. return FAILED);
  372. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_WEIGHT_SIZE, weight_offset_),
  373. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  374. ATTR_MODEL_WEIGHT_SIZE.c_str(), __FUNCTION__);
  375. GELOGE(FAILED, "SetInt of ATTR_MODEL_WEIGHT_SIZE failed.");
  376. return FAILED);
  377. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_STREAM_NUM, stream_num_),
  378. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  379. ATTR_MODEL_STREAM_NUM.c_str(), __FUNCTION__);
  380. GELOGE(FAILED, "SetInt of ATTR_MODEL_STREAM_NUM failed.");
  381. return FAILED);
  382. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_EVENT_NUM, event_num_),
  383. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  384. ATTR_MODEL_EVENT_NUM.c_str(), __FUNCTION__);
  385. GELOGE(FAILED, "SetInt of ATTR_MODEL_EVENT_NUM failed.");
  386. return FAILED);
  387. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(&model, ATTR_MODEL_HUGE_STREAM_LIST, huge_streams_),
  388. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  389. ATTR_MODEL_HUGE_STREAM_LIST.c_str(), __FUNCTION__);
  390. GELOGE(FAILED, "SetInt of ATTR_MODEL_HUGE_STREAM_LIST failed.");
  391. return FAILED);
  392. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_LABEL_NUM, label_num_),
  393. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  394. ATTR_MODEL_LABEL_NUM.c_str(), __FUNCTION__);
  395. GELOGE(FAILED, "SetInt of ATTR_MODEL_LABEL_NUM failed.");
  396. return FAILED);
  397. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_ZERO_COPY_MEMORY_SIZE, zero_copy_mem_size_),
  398. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  399. ATTR_MODEL_ZERO_COPY_MEMORY_SIZE.c_str(), __FUNCTION__);
  400. GELOGE(FAILED, "SetInt of ATTR_MODEL_ZERO_COPY_MEMORY_SIZE failed.");
  401. return FAILED);
  402. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&model, ATTR_MODEL_OUT_NODES_NAME, GetLocalOmgContext().net_out_nodes),
  403. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  404. ATTR_MODEL_OUT_NODES_NAME.c_str(), __FUNCTION__);
  405. GELOGE(FAILED, "SetListStr of ATTR_MODEL_OUT_NODES_NAME failed.");
  406. return FAILED);
  407. GELOGI("For model, max_mem_offset_: %zu, p2p_mem_size: %zu, zero_copy_mem_size_: %zu", max_mem_offset_,
  408. p2p_mem_offset_, zero_copy_mem_size_);
  409. string fp_ceiling_mode;
  410. if (ge::GetContext().GetOption("ge.fpCeilingMode", fp_ceiling_mode) == SUCCESS) {
  411. if (!ge::AttrUtils::SetStr(&model, ATTR_FP_CEILING_MODE, fp_ceiling_mode)) {
  412. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  413. ATTR_FP_CEILING_MODE.c_str(), __FUNCTION__);
  414. GELOGE(FAILED, "Failed to set attr ATTR_FP_CEILING_MODE");
  415. return FAILED;
  416. }
  417. GELOGI("Set attr ATTR_FP_CEILING_MODE to model, value is %s.", fp_ceiling_mode.c_str());
  418. }
  419. string ge_core_type;
  420. Status ret = ge::GetContext().GetOption(kCoreType, ge_core_type);
  421. if (ret != SUCCESS) {
  422. GELOGW("get the option CORE_TYPE fail, set it to default value VECTOR_ENGINE");
  423. }
  424. int64_t core_type = (ge_core_type == kVectorCore) ? 1 : 0;
  425. GELOGI("core_type: %ld", core_type);
  426. if (!ge::AttrUtils::SetInt(&model, ATTR_MODEL_CORE_TYPE, core_type)) {
  427. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  428. ATTR_MODEL_CORE_TYPE.c_str(), __FUNCTION__);
  429. GELOGE(FAILED, "SetInt of ATTR_CORE_TYPE failed.");
  430. }
  431. InitL1FusionOption();
  432. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetBool(&model, ATTR_NAME_SWITCH_FOR_L1_FUSION, is_l1_fusion_enable_),
  433. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  434. ATTR_NAME_SWITCH_FOR_L1_FUSION.c_str(), __FUNCTION__);
  435. GELOGE(FAILED, "SetBool of ATTR_NAME_SWITCH_FOR_L1_FUSION failed.");
  436. return FAILED);
  437. const DumpProperties &dump_properties = DumpManager::GetInstance().GetDumpProperties(session_id_);
  438. bool is_op_debug = dump_properties.IsOpDebugOpen();
  439. if (is_op_debug) {
  440. if (!ge::AttrUtils::SetBool(&model, ATTR_OP_DEBUG_FLAG, is_op_debug)) {
  441. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  442. ATTR_OP_DEBUG_FLAG.c_str(), __FUNCTION__);
  443. GELOGE(FAILED, "SetBool of ATTR_OP_DEBUG_FLAG failed.");
  444. return FAILED;
  445. }
  446. uint32_t op_debug_mode = dump_properties.GetOpDebugMode();
  447. GELOGI("Get op debug mode:%d", op_debug_mode);
  448. if (!ge::AttrUtils::SetInt(&model, ATTR_OP_DEBUG_MODE, op_debug_mode)) {
  449. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed when %s",
  450. ATTR_OP_DEBUG_MODE.c_str(), __FUNCTION__);
  451. GELOGE(FAILED, "SetBool of ATTR_OP_DEBUG_MODE failed.");
  452. return FAILED;
  453. }
  454. }
  455. model.SetName(compute_graph_->GetName());
  456. model.SetGraph(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_));
  457. GELOGI("weight_offset_: %zu", weight_offset_);
  458. GELOGI("Set event num: %ld.", event_num_);
  459. if (Params::Instance() == nullptr) {
  460. return FAILED;
  461. }
  462. platform_type_ = Params::Instance()->GetTarget_8bit();
  463. return SUCCESS;
  464. }
  465. void ModelBuilder::ClearOriginalFormat() {
  466. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  467. auto node_op_desc = n->GetOpDesc();
  468. if (node_op_desc != nullptr) {
  469. if (node_op_desc->HasAttr(ATTR_NAME_FORMAT)) {
  470. if (node_op_desc->DelAttr(ATTR_NAME_FORMAT) != SUCCESS) {
  471. GELOGW("DelAttr ATTR_NAME_FORMAT failed.");
  472. }
  473. }
  474. GE_IF_BOOL_EXEC(
  475. node_op_desc->HasAttr(ATTR_NAME_INFERRED_FORMAT),
  476. if (node_op_desc->DelAttr(ATTR_NAME_INFERRED_FORMAT) != SUCCESS) {
  477. GELOGW("DelAttr ATTR_NAME_INFERRED_FORMAT failed.");
  478. });
  479. GE_IF_BOOL_EXEC(
  480. node_op_desc->HasAttr(ATTR_NAME_PRED_PERMUTE_DELETED),
  481. if (node_op_desc->DelAttr(ATTR_NAME_PRED_PERMUTE_DELETED) != SUCCESS) {
  482. GELOGW("DelAttr ATTR_NAME_PRED_PERMUTE_DELETED failed.");
  483. });
  484. GE_IF_BOOL_EXEC(
  485. node_op_desc->HasAttr(ATTR_NAME_IGNORE_PRED_FORMAT),
  486. if (node_op_desc->DelAttr(ATTR_NAME_IGNORE_PRED_FORMAT) != SUCCESS) {
  487. GELOGW("DelAttr ATTR_NAME_IGNORE_PRED_FORMAT failed.");
  488. });
  489. }
  490. }
  491. }
  492. Status ModelBuilder::MergeWeights() {
  493. if (weight_offset_ == 0) {
  494. return SUCCESS;
  495. }
  496. ge::Buffer buffer(weight_offset_);
  497. weight_buffer_ = buffer;
  498. auto base_addr = weight_buffer_.GetData();
  499. for (const ge::NodePtr &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  500. auto op_desc = node->GetOpDesc();
  501. GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
  502. if (node->GetType() != CONSTANT) {
  503. continue;
  504. }
  505. // Get const op weight pointer
  506. ge::GeTensorPtr weight = nullptr;
  507. // If MutableTensor failed, weight is nullptr.
  508. (void)ge::AttrUtils::MutableTensor(op_desc, ATTR_NAME_WEIGHTS, weight);
  509. if (weight == nullptr) {
  510. REPORT_INNER_ERROR("E19999", "Can't get const weight in op:%s(%s) when %s",
  511. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  512. GELOGE(FAILED, "Can't get const op weight, name: %s", node->GetName().c_str());
  513. return FAILED;
  514. }
  515. // Get const op weight offset
  516. int64_t offset = 0;
  517. if (ge::TensorUtils::GetDataOffset(weight->GetTensorDesc(), offset) != SUCCESS) {
  518. GELOGW("Can't get const op offset, name: %s", node->GetName().c_str());
  519. continue; // continue to merge if can not get offset
  520. }
  521. // Get const op weight data
  522. auto weight_data = weight->MutableData();
  523. // copy const op weight data to buffer
  524. GELOGI("Move to buffer, name: %s offset: %ld size: %zu", node->GetName().c_str(), offset, weight_data.size());
  525. ge::TensorUtils::SetWeightSize(weight->MutableTensorDesc(), static_cast<uint32_t>(weight_data.size()));
  526. if ((offset == 0) || (weight_data.size() == 0)) {
  527. GELOGI("Size or offset is 0. size: %lu offset: %ld", weight_data.size(), offset);
  528. continue;
  529. }
  530. if (weight_data.data() != nullptr) {
  531. GE_IF_BOOL_EXEC(base_addr == nullptr,
  532. REPORT_INNER_ERROR("E19999", "Check weight in op:%s(%s) is nullptr when %s",
  533. op_desc->GetName().c_str(), op_desc->GetType().c_str(), __FUNCTION__);
  534. GELOGE(FAILED, "Base addr is nullptr.");
  535. return FAILED);
  536. if (weight_offset_ - offset < weight_data.size()) {
  537. REPORT_INNER_ERROR("E19999", "left weight size not enough for op:%s(%s) left_size:%zu, weight_size:%zu when %s",
  538. op_desc->GetName().c_str(), op_desc->GetType().c_str(),
  539. weight_offset_ - offset, weight_data.size(), __FUNCTION__);
  540. GELOGE(FAILED, "left weight size not enough. left_size:%lu, weight_size:%lu",
  541. weight_offset_ - offset, weight_data.size());
  542. return FAILED;
  543. }
  544. uintptr_t dst_ptr = reinterpret_cast<uintptr_t>(base_addr) + offset;
  545. uintptr_t src_ptr = reinterpret_cast<uintptr_t>(weight_data.data());
  546. size_t left_size = weight_data.size();
  547. while (left_size > SECUREC_MEM_MAX_LEN) {
  548. auto err = memcpy_s(reinterpret_cast<void *>(dst_ptr), SECUREC_MEM_MAX_LEN, reinterpret_cast<void *>(src_ptr),
  549. SECUREC_MEM_MAX_LEN);
  550. if (err != EOK) {
  551. REPORT_CALL_ERROR("E19999", "mem copy failed. errret:%u, "
  552. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu, when %s",
  553. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN, __FUNCTION__);
  554. GELOGE(FAILED, "mem copy failed. errret:%u, "
  555. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu",
  556. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN);
  557. return FAILED;
  558. }
  559. left_size -= SECUREC_MEM_MAX_LEN;
  560. dst_ptr = dst_ptr + SECUREC_MEM_MAX_LEN;
  561. src_ptr = src_ptr + SECUREC_MEM_MAX_LEN;
  562. }
  563. auto err = memcpy_s(reinterpret_cast<void *>(dst_ptr), left_size, reinterpret_cast<void *>(src_ptr), left_size);
  564. if (err != EOK) {
  565. REPORT_CALL_ERROR("E19999", "mem copy failed. errret:%u, "
  566. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu, when %s",
  567. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN, __FUNCTION__);
  568. GELOGE(FAILED, "mem copy failed. errret:%u, "
  569. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu",
  570. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN);
  571. return FAILED;
  572. }
  573. }
  574. weight->ClearData();
  575. }
  576. return SUCCESS;
  577. }
  578. Status ModelBuilder::SaveDataToModel(ge::Model &model, ge::GeModel &ge_model) {
  579. // Add weight
  580. ge_model.SetWeight(weight_buffer_);
  581. // Add TBE Kernels and custom aicpu op bin
  582. std::set<std::string> tbe_name_set;
  583. std::set<std::string> aicpu_name_set;
  584. std::set<std::string> aicpu_op_types;
  585. std::set<std::string> aicpu_tf_op_types;
  586. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  587. auto node_op_desc = n->GetOpDesc();
  588. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  589. // check aicpu op type
  590. CollectCheckAicpuAttr(node_op_desc, aicpu_op_types, aicpu_tf_op_types);
  591. TBEKernelPtr tbe_kernel = node_op_desc->TryGetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
  592. if (tbe_kernel == nullptr) {
  593. std::string kernel_name;
  594. GeAttrValue::BYTES kernel_buffer;
  595. (void) AttrUtils::GetStr(node_op_desc, ATTR_NAME_TBE_KERNEL_NAME, kernel_name);
  596. (void) AttrUtils::GetBytes(node_op_desc, ATTR_NAME_TBE_KERNEL_BUFFER, kernel_buffer);
  597. if (!kernel_name.empty() && (kernel_buffer.GetSize() > 0)) {
  598. GE_CHECK_NOTNULL(kernel_buffer.GetData());
  599. std::vector<char> data(kernel_buffer.GetData(), kernel_buffer.GetData() + kernel_buffer.GetSize());
  600. tbe_kernel = std::make_shared<OpKernelBin>(kernel_name, std::move(data));
  601. }
  602. }
  603. GE_IF_BOOL_EXEC(tbe_kernel == nullptr, continue);
  604. if (tbe_name_set.count(tbe_kernel->GetName()) > 0) {
  605. REPORT_INNER_ERROR("E19999", "tbe_kernel name %s can't be the same, judge for op:%s(%s), when %s",
  606. tbe_kernel->GetName().c_str(), n->GetName().c_str(), n->GetType().c_str(), __FUNCTION__);
  607. GELOGE(FAILED, "tbe_kernel name %s can't be the same", tbe_kernel->GetName().c_str());
  608. return FAILED;
  609. }
  610. tbe_name_set.insert(tbe_kernel->GetName());
  611. tbe_kernel_store_.AddTBEKernel(tbe_kernel);
  612. }
  613. SetModelCheckAicpuAttr(model, aicpu_op_types, aicpu_tf_op_types);
  614. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  615. auto node_op_desc = n->GetOpDesc();
  616. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  617. CustAICPUKernelPtr cust_aicpu_kernel =
  618. node_op_desc->TryGetExtAttr(ge::OP_EXTATTR_CUSTAICPU_KERNEL, CustAICPUKernelPtr());
  619. GE_IF_BOOL_EXEC(cust_aicpu_kernel == nullptr, continue);
  620. if (aicpu_name_set.count(cust_aicpu_kernel->GetName()) > 0) {
  621. REPORT_INNER_ERROR("E19999", "aicpu_kernel name %s can't be the same, judge for op:%s(%s), when %s",
  622. cust_aicpu_kernel->GetName().c_str(), n->GetName().c_str(), n->GetType().c_str(), __FUNCTION__);
  623. GELOGE(FAILED, "aicpu_kernel name %s can't be the same", cust_aicpu_kernel->GetName().c_str());
  624. return FAILED;
  625. }
  626. aicpu_name_set.insert(cust_aicpu_kernel->GetName());
  627. cust_aicpu_kernel_store_.AddCustAICPUKernel(cust_aicpu_kernel);
  628. GELOGI("Add cust aicpu kernel bin %s", cust_aicpu_kernel->GetName().c_str());
  629. }
  630. if (!tbe_kernel_store_.Build()) {
  631. GELOGE(FAILED, "TBE Kernels store build failed!");
  632. return FAILED;
  633. }
  634. if (!cust_aicpu_kernel_store_.Build()) {
  635. GELOGE(FAILED, "custom AICPU kernels store build failed!");
  636. return FAILED;
  637. }
  638. ge_model.SetTBEKernelStore(tbe_kernel_store_);
  639. ge_model.SetCustAICPUKernelStore(cust_aicpu_kernel_store_);
  640. // Add task
  641. GeAttrValue::BYTES task_def_bytes;
  642. if (!AttrUtils::GetZeroCopyBytes(model, MODEL_ATTR_TASKS, task_def_bytes)) {
  643. REPORT_CALL_ERROR("E19999", "Get attr:%s in model fail when %s", MODEL_ATTR_TASKS.c_str(), __FUNCTION__);
  644. GELOGE(INTERNAL_ERROR, "Get zero copy bytes fail.");
  645. return INTERNAL_ERROR;
  646. }
  647. int byte_size = static_cast<int>(task_def_bytes.GetSize());
  648. std::shared_ptr<domi::ModelTaskDef> task = ge::MakeShared<domi::ModelTaskDef>();
  649. GE_CHECK_NOTNULL(task);
  650. GE_CHK_BOOL_EXEC(ReadProtoFromArray(task_def_bytes.GetData(), byte_size, task.get()), return INTERNAL_ERROR,
  651. "ReadProtoFromArray failed.");
  652. ge_model.SetModelTaskDef(task);
  653. // Add graph
  654. ge_model.SetName(model.GetName());
  655. ge_model.SetGraph(model.GetGraph());
  656. ge_model.SetVersion(model.GetVersion());
  657. ge_model.SetPlatformVersion(model.GetPlatformVersion());
  658. ge_model.SetPlatformType(platform_type_);
  659. ge_model.SetAttr(model.MutableAttrMap());
  660. return SUCCESS;
  661. }
  662. void ModelBuilder::SetModelVersion(ge::Model &model) {
  663. // set framework_version TO model
  664. string framework_version;
  665. uint32_t counter = 0;
  666. Status frame_rt = PlatformVersionManager::GetPlatformVersion(framework_version);
  667. GE_IF_BOOL_EXEC((frame_rt == SUCCESS),
  668. string model_framework_version = framework_version + "." + std::to_string(counter);
  669. model.SetPlatformVersion(model_framework_version););
  670. // set IR Version TO model
  671. model.SetVersion(static_cast<uint32_t>(OM_PROTO_VERSION));
  672. }
  673. Status ModelBuilder::PreBuildModel() {
  674. if ((compute_graph_ == nullptr) || !(compute_graph_->IsValid())) {
  675. REPORT_INNER_ERROR("E19999", "Check compute_graph no valid when %s", __FUNCTION__);
  676. GELOGE(FAILED, "Graph_ is not valid.");
  677. return FAILED;
  678. }
  679. GE_CHK_STATUS_RET(SetInputOutputDesc(), "SetInputOutputDesc Failed!");
  680. AddNodeInputProperty();
  681. return SUCCESS;
  682. }
  683. Status ModelBuilder::BuildModelForGetTask(ge::Model &model) {
  684. GE_CHK_STATUS_RET(AdjustInputTensorFlag(), "AdjustInputTensorFlag failed!");
  685. ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kStreamAlloc);
  686. // Assign logical streams.
  687. StreamAllocator stream_allocator(compute_graph_, subgraphs_);
  688. GE_TIMESTAMP_START(AssignLogicalStreams);
  689. GE_CHK_STATUS_RET(stream_allocator.AssignLogicalStreams(stream_max_parallel_num_, hcom_parallel_),
  690. "Assign logical streams failed.");
  691. GE_TIMESTAMP_END(AssignLogicalStreams, "GraphBuilder::AssignLogicalStreams");
  692. ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kMemoryAlloc);
  693. // Assign functional op labels.
  694. auto root_graph = GraphUtils::FindRootGraph(compute_graph_);
  695. (void)AttrUtils::GetInt(*root_graph, ATTR_MODEL_LABEL_NUM, label_num_);
  696. GE_TIMESTAMP_START(AssignMemory);
  697. MemoryAssigner mem_assigner(compute_graph_);
  698. GE_CHK_STATUS_RET(mem_assigner.AssignMemory(is_loop_graph_, mem_type_to_mem_offset_, zero_copy_mem_size_),
  699. "Assign Memory Failed!");
  700. GE_TIMESTAMP_END(AssignMemory, "GraphBuilder::AssignMemory");
  701. ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kOther);
  702. GE_TIMESTAMP_START(SetInputOutputOffset);
  703. SetInputOutputOffsetPass input_output_offset;
  704. GE_CHK_STATUS_RET(input_output_offset.Run(compute_graph_), "Set input output offset failed.");
  705. GE_TIMESTAMP_END(SetInputOutputOffset, "SetInputOutputOffsetPass::Run");
  706. // Compile single op in graph build stage
  707. GE_TIMESTAMP_START(CompileSingleOp);
  708. GE_CHK_STATUS_RET(CompileSingleOp(), "ATC builder CompileSingleOp() return fail.");
  709. GE_TIMESTAMP_EVENT_END(CompileSingleOp, "GraphBuilder::CompileSingleOp");
  710. ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kStreamAlloc);
  711. // Refresh real streams and insert event nodes.
  712. GE_TIMESTAMP_START(RefreshRealStream);
  713. GE_CHK_STATUS_RET(stream_allocator.RefreshRealStream(stream_num_, event_num_), "RefreshRealStream failed.");
  714. huge_streams_ = stream_allocator.GetHugeStreams();
  715. GE_TIMESTAMP_END(RefreshRealStream, "GraphBuilder::RefreshRealStream");
  716. ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kOther);
  717. GE_TIMESTAMP_START(MergeWeights);
  718. GE_CHK_STATUS_RET(MergeWeights(), "MergeWeights Failed!");
  719. GE_TIMESTAMP_END(MergeWeights, "GraphBuilder::MergeWeights");
  720. GE_TIMESTAMP_START(BuildModelDef);
  721. GE_CHK_STATUS_RET(BuildModelDef(model), "BuildModelDef failed!");
  722. GE_TIMESTAMP_END(BuildModelDef, "GraphBuilder::BuildModelDef");
  723. SetModelVersion(model);
  724. return SUCCESS;
  725. }
  726. Status ModelBuilder::BuildModelForGetDynShapeTask(ge::Model &model_def) {
  727. GE_TIMESTAMP_START(BuildModelDef);
  728. GE_CHK_STATUS_RET(BuildModelDef(model_def), "BuildModelDef failed!");
  729. GE_TIMESTAMP_END(BuildModelDef, "GraphBuilder::BuildModelDef");
  730. SetModelVersion(model_def);
  731. return SUCCESS;
  732. }
  733. ge::Buffer ModelBuilder::GetWeightBuffer() const { return weight_buffer_; }
  734. Status ModelBuilder::CompileSingleOp() {
  735. GELOGD("Begin to compile single op.");
  736. // Create ge instance
  737. std::shared_ptr<GELib> instance = ge::GELib::GetInstance();
  738. if ((instance == nullptr) || !instance->InitFlag()) {
  739. REPORT_INNER_ERROR("E19999", "Check GELib instance not init before when %s", __FUNCTION__);
  740. GELOGE(ge::GE_CLI_GE_NOT_INITIALIZED, "CompileSingleOp failed.");
  741. return ge::GE_CLI_GE_NOT_INITIALIZED;
  742. }
  743. GE_TIMESTAMP_CALLNUM_START(BatchCompileOp);
  744. std::unordered_map<string, vector<ge::NodePtr>> node_vector_map;
  745. for (auto &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  746. auto op_desc = node->GetOpDesc();
  747. if (op_desc == nullptr) {
  748. continue;
  749. }
  750. // Graph build stage only supports the individual compilation of atomic clean operator
  751. if (op_desc->GetType() == ATOMICADDRCLEAN) {
  752. GELOGD("Begin to compile single op, op name is %s.", op_desc->GetName().c_str());
  753. string kernel_lib_name = op_desc->GetOpKernelLibName();
  754. if (kernel_lib_name.empty()) {
  755. // Reset op kernel lib
  756. (void)instance->DNNEngineManagerObj().GetDNNEngineName(node);
  757. kernel_lib_name = op_desc->GetOpKernelLibName();
  758. if (kernel_lib_name.empty()) {
  759. REPORT_INNER_ERROR("E19999", "Check kernel lib name empty of op:%s(%s) when %s",
  760. node->GetName().c_str(), node->GetType().c_str(), __FUNCTION__);
  761. GELOGE(ge::INTERNAL_ERROR, "Get node:%s(%s) kernel lib failed.", node->GetName().c_str(),
  762. node->GetType().c_str());
  763. return ge::INTERNAL_ERROR;
  764. }
  765. }
  766. OpsKernelInfoStorePtr kernel_info = instance->OpsKernelManagerObj().GetOpsKernelInfoStore(kernel_lib_name);
  767. if (kernel_info != nullptr) {
  768. node_vector_map[kernel_lib_name].emplace_back(node);
  769. } else {
  770. REPORT_INNER_ERROR("E19999", "Get ops kernel info store failed for op:%s(%s), op_kernel_name:%s, when %s",
  771. node->GetName().c_str(), node->GetType().c_str(), kernel_lib_name.c_str(), __FUNCTION__);
  772. GELOGE(ge::GE_GRAPH_PARAM_NULLPTR, "Get op %s ops kernel info store failed", node->GetName().c_str());
  773. return ge::GE_GRAPH_PARAM_NULLPTR;
  774. }
  775. }
  776. }
  777. for (auto &it : node_vector_map) {
  778. auto &kernel_lib_name = it.first;
  779. auto &node_vector = it.second;
  780. OpsKernelInfoStorePtr kernel_info = instance->OpsKernelManagerObj().GetOpsKernelInfoStore(kernel_lib_name);
  781. GE_CHECK_NOTNULL(kernel_info);
  782. GE_TIMESTAMP_RESTART(BatchCompileOp);
  783. auto ret = kernel_info->CompileOp(node_vector);
  784. GELOGI("[GEPERFTRACE] The node size of compile op of %s is %zu", kernel_lib_name.c_str(), node_vector.size());
  785. GE_TIMESTAMP_ADD(BatchCompileOp);
  786. if (ret != ge::SUCCESS) {
  787. REPORT_CALL_ERROR("E19999", "Batch compile op failed, kernel lib name, node size:%u, when %s",
  788. node_vector.size(), __FUNCTION__);
  789. GELOGE(ret, "Compile op failed, kernel lib name is %s", kernel_lib_name.c_str());
  790. return ret;
  791. }
  792. }
  793. GE_TIMESTAMP_CALLNUM_END(BatchCompileOp, "GraphBuild::CompileOp");
  794. return ge::SUCCESS;
  795. }
  796. void ModelBuilder::CollectCheckAicpuAttr(const OpDescPtr &op_desc, std::set<std::string> &aicpu_op_types,
  797. std::set<std::string> &aicpu_tf_op_types) {
  798. std::string aicpu_optype;
  799. bool has_attr_check_cpu = ge::AttrUtils::GetStr(op_desc, "needCheckCpu", aicpu_optype);
  800. std::vector<std::string> tf_optypes;
  801. bool has_attr_check_tf = ge::AttrUtils::GetListStr(op_desc, "needCheckTf", tf_optypes);
  802. if (has_attr_check_cpu && !aicpu_optype.empty()) {
  803. aicpu_op_types.insert(aicpu_optype);
  804. }
  805. if (has_attr_check_tf && !tf_optypes.empty()) {
  806. aicpu_tf_op_types.insert(tf_optypes.begin(), tf_optypes.end());
  807. }
  808. return;
  809. }
  810. void ModelBuilder::SetModelCheckAicpuAttr(ge::Model &model, std::set<std::string> &aicpu_op_types,
  811. std::set<std::string> &aicpu_tf_op_types) {
  812. std::vector<std::string> aicpu_optype_list;
  813. std::vector<std::string> aicpu_tf_optype_list;
  814. if (ge::AttrUtils::GetListStr(&model, "needCheckCpu", aicpu_optype_list)) {
  815. GELOGI("Already have aicpu optype size: %zu", aicpu_optype_list.size());
  816. aicpu_op_types.insert(aicpu_optype_list.begin(), aicpu_optype_list.end());
  817. }
  818. if (ge::AttrUtils::GetListStr(&model, "needCheckTf", aicpu_tf_optype_list)) {
  819. GELOGI("Already have aicpu tf optype size: %zu", aicpu_tf_optype_list.size());
  820. aicpu_tf_op_types.insert(aicpu_tf_optype_list.begin(), aicpu_tf_optype_list.end());
  821. }
  822. // reset list with set
  823. aicpu_optype_list.assign(aicpu_op_types.begin(), aicpu_op_types.end());
  824. aicpu_tf_optype_list.assign(aicpu_tf_op_types.begin(), aicpu_tf_op_types.end());
  825. GELOGI(
  826. "Check Aicpu op types ComputeGraph: %s aicpu_op_types: %zu, aicpu_optype_list: %zu, aicpu_tf_op_types: %zu, "
  827. "aicpu_tf_optype_list:%zu.",
  828. compute_graph_->GetName().c_str(), aicpu_op_types.size(), aicpu_optype_list.size(), aicpu_tf_op_types.size(),
  829. aicpu_tf_optype_list.size());
  830. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&model, "needCheckCpu", aicpu_optype_list), return,
  831. "Set attr needCheckCpu fail.");
  832. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&model, "needCheckTf", aicpu_tf_optype_list), return,
  833. "Set attr needCheckTf fail.");
  834. return;
  835. }
  836. } // namespace ge

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示