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 47 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
4 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
4 years ago
5 years ago
4 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
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 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
4 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
4 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
4 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
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
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
4 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982
  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",
  111. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str(), index);
  112. GELOGE(graph_status, "[Get][TensorMemorySize] In Bytes failed for op:%s(%s) index:%u",
  113. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str(), index);
  114. return FAILED;
  115. }
  116. TensorUtils::SetSize(desc_temp, size_temp);
  117. if (node_op_desc->UpdateOutputDesc(index, desc_temp) != SUCCESS) {
  118. REPORT_CALL_ERROR("E19999", "Update Output desc size failed for op:%s(%s) index:%u",
  119. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str(), index);
  120. GELOGE(FAILED, "[Update][OutputDesc] failed for op:%s(%s) index:%u",
  121. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str(), index);
  122. return FAILED;
  123. }
  124. GELOGD("update output desc, dim_size: %u, mem_size: %ld, format: %s, type: %s, node name:%s", dim_num, size_temp,
  125. TypeUtils::FormatToSerialString(desc_temp.GetFormat()).c_str(),
  126. TypeUtils::DataTypeToSerialString(desc_temp.GetDataType()).c_str(), node_op_desc->GetName().c_str());
  127. index++;
  128. }
  129. return SUCCESS;
  130. }
  131. bool ModelBuilder::SetInputConst(const OpDescPtr &op_desc, const NodePtr &src_node, size_t index,
  132. vector<bool> &is_input_const) {
  133. GELOGI("SetIsInputConst const: %s, source node: %s", op_desc->GetName().c_str(), src_node->GetName().c_str());
  134. for (size_t i = is_input_const.size(); i <= index; ++i) {
  135. is_input_const.push_back(false);
  136. }
  137. is_input_const[index] = true;
  138. vector<GeTensorPtr> weights = OpDescUtils::MutableWeights(src_node);
  139. if (weights.empty()) {
  140. GELOGW("SetInputIsConst weights is empty, node: %s", src_node->GetName().c_str());
  141. return false;
  142. }
  143. GeTensorPtr weight = weights[0];
  144. GE_IF_BOOL_EXEC(weight == nullptr, return true);
  145. GeTensorDesc &tensor_desc = weight->MutableTensorDesc();
  146. int64_t data_offset = 0;
  147. if (TensorUtils::GetDataOffset(tensor_desc, data_offset) != GRAPH_SUCCESS) {
  148. GELOGW("Get Offset from weight failed");
  149. return false;
  150. }
  151. auto input_tensor = op_desc->MutableInputDesc(static_cast<uint32_t>(index));
  152. if (input_tensor == nullptr) {
  153. GELOGW("Get input_tensor failed");
  154. return false;
  155. }
  156. TensorUtils::SetDataOffset(*input_tensor, data_offset);
  157. return true;
  158. }
  159. void ModelBuilder::SetInputIsConst(const ge::NodePtr &n) {
  160. auto node_op_desc = n->GetOpDesc();
  161. GE_CHECK_NOTNULL_JUST_RETURN(node_op_desc);
  162. auto is_input_const = node_op_desc->GetIsInputConst();
  163. // must set all true input_const to false
  164. for (size_t i = 0; i < is_input_const.size(); i++) {
  165. is_input_const[i] = false;
  166. }
  167. std::string const_type;
  168. auto in_data_anchors = n->GetAllInDataAnchors();
  169. for (size_t index = 0; index < in_data_anchors.size(); index++) {
  170. auto in_data_anchor = in_data_anchors.at(index);
  171. const auto &peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  172. GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
  173. const auto &src_node = peer_out_anchor->GetOwnerNode();
  174. if (!NodeUtils::GetConstOpType(src_node, const_type)) {
  175. continue;
  176. }
  177. if (const_type == CONSTANT) {
  178. if (!SetInputConst(node_op_desc, src_node, index, is_input_const)) {
  179. return;
  180. }
  181. } else {
  182. if ((index < is_input_const.size()) && is_input_const[index]) {
  183. is_input_const[index] = false;
  184. }
  185. }
  186. }
  187. GELOGD("update opdesc:%s InputConst:%s", node_op_desc->GetName().c_str(), ToString(is_input_const).c_str());
  188. node_op_desc->SetIsInputConst(is_input_const);
  189. }
  190. Status ModelBuilder::AdjustConstWeightSize(const ge::NodePtr &node, size_t &mem_offset) {
  191. GE_CHECK_NOTNULL(node);
  192. if (node->GetType() == CONSTANT) {
  193. vector<GeTensorPtr> weights = OpDescUtils::MutableWeights(node);
  194. if (weights.empty()) {
  195. REPORT_INNER_ERROR("E19999", "Check weights size of node %s(%s) is empty",
  196. node->GetName().c_str(), node->GetType().c_str());
  197. GELOGE(FAILED, "[Check][Param] weights size of node %s is empty", node->GetName().c_str());
  198. return FAILED;
  199. }
  200. GeTensorPtr weight = weights[0];
  201. if (weight == nullptr) {
  202. REPORT_INNER_ERROR("E19999", "Check weight of node %s(%s) is nullptr",
  203. node->GetName().c_str(), node->GetType().c_str());
  204. GELOGE(FAILED, "[Check][Param] weights[0] is null, node:%s.", node->GetName().c_str());
  205. return FAILED;
  206. }
  207. GeTensorDesc &tensor_desc = weight->MutableTensorDesc();
  208. size_t output_size = weight->GetData().size();
  209. TensorUtils::SetDataOffset(tensor_desc, mem_offset);
  210. GELOGD("Node: %s, weight size: %zu.", node->GetName().c_str(), output_size);
  211. mem_offset += output_size;
  212. }
  213. return SUCCESS;
  214. }
  215. Status ModelBuilder::SetInputOutputDesc() {
  216. Status ret;
  217. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  218. auto node_op_desc = n->GetOpDesc();
  219. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  220. if (!is_loop_graph_ && node_op_desc->GetType() == LOOPCOND) {
  221. is_loop_graph_ = true;
  222. }
  223. // if user set input node format ND, the expected node for data and netoutput format is ND in
  224. // final graph.
  225. if ((compute_graph_->GetParentGraph() == nullptr) && (GetLocalOmgContext().format == domi::DOMI_TENSOR_ND) && (!node_op_desc->HasAttr("_is_single_op")) &&
  226. ((node_op_desc->GetType() == DATA_TYPE) || (node_op_desc->GetType() == NETOUTPUT))) {
  227. auto inputDescsPtr = node_op_desc->GetAllInputsDescPtr();
  228. auto outputDescsPtr = node_op_desc->GetAllOutputsDescPtr();
  229. ge::Format format = ge::FORMAT_ND;
  230. for (auto &inputDescPtr : inputDescsPtr) {
  231. GE_CHECK_NOTNULL(inputDescPtr);
  232. inputDescPtr->SetFormat(format);
  233. inputDescPtr->SetOriginFormat(format);
  234. }
  235. for (auto &outputDescPtr : outputDescsPtr) {
  236. GE_CHECK_NOTNULL(outputDescPtr);
  237. outputDescPtr->SetFormat(format);
  238. outputDescPtr->SetOriginFormat(format);
  239. }
  240. }
  241. if (node_op_desc->GetType() == DATA_TYPE || node_op_desc->GetType() == AIPP_DATA_TYPE) {
  242. GELOGD("Data node: %s.", n->GetName().c_str());
  243. continue;
  244. }
  245. GE_IF_BOOL_EXEC(n->GetInAllNodes().empty() && n->GetOutAllNodes().empty(), continue;);
  246. SetInputIsConst(n);
  247. bool is_unknow = false;
  248. (void)NodeUtils::GetNodeUnknownShapeStatus(*n, is_unknow);
  249. if ((IsGeLocalOp(n->GetOpDesc())) && (!is_unknow)) {
  250. GE_CHK_STATUS_RET(CalcOutputSize(n), "[Calc][OutputSize] failed, node:%s", n->GetName().c_str());
  251. }
  252. ret = AdjustConstWeightSize(n, weight_offset_);
  253. GE_CHK_STATUS_RET(ret, "[Adjust][ConstWeightSize] failed, node:%s", n->GetName().c_str());
  254. GE_IF_BOOL_EXEC(((weight_offset_ > 0) && (weight_offset_ % MEM_ALIGN_SIZE != 0)),
  255. weight_offset_ = (weight_offset_ + MEM_ALIGN_SIZE - 1) / MEM_ALIGN_SIZE * MEM_ALIGN_SIZE);
  256. }
  257. GE_CHK_STATUS_RET(compute_graph_->TopologicalSorting(), "[Call][TopologicalSorting] failed, graph:%s",
  258. compute_graph_->GetName().c_str());
  259. return SUCCESS;
  260. }
  261. void ModelBuilder::AddNodeInputProperty() {
  262. for (const ge::NodePtr &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  263. auto node_op_desc = node->GetOpDesc();
  264. GE_IF_BOOL_EXEC(node_op_desc == nullptr, GELOGW("node_op_desc is nullptr!"); return);
  265. vector<string> src_name_list;
  266. vector<int64_t> src_index_list;
  267. for (const auto &in_data_anchor : node->GetAllInDataAnchors()) {
  268. auto peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
  269. GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
  270. GE_IF_BOOL_EXEC(node_op_desc->HasAttr(MERGE_PRENODE_FLAG), continue);
  271. ge::NodePtr src_node = peer_out_anchor->GetOwnerNode();
  272. src_name_list.emplace_back(src_node->GetName());
  273. src_index_list.emplace_back(peer_out_anchor->GetIdx());
  274. }
  275. auto in_control_anchor = node->GetInControlAnchor();
  276. if (in_control_anchor != nullptr) {
  277. string src_name_temp;
  278. for (const auto &out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
  279. ge::NodePtr src_node = out_control_anchor->GetOwnerNode();
  280. src_name_temp = src_name_temp.empty() ? src_node->GetName() : src_name_temp + ":" + src_node->GetName();
  281. }
  282. GE_IF_BOOL_EXEC(!src_name_temp.empty(), src_name_list.emplace_back(src_name_temp);)
  283. }
  284. node_op_desc->SetSrcName(src_name_list);
  285. node_op_desc->SetSrcIndex(src_index_list);
  286. }
  287. for (const ge::NodePtr &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  288. auto node_op_desc = node->GetOpDesc();
  289. GE_IF_BOOL_EXEC(node_op_desc == nullptr, GELOGW("node_op_desc is nullptr!"); return);
  290. GE_IF_BOOL_EXEC(node_op_desc->GetType() == NETOUTPUT, continue);
  291. auto out_control_anchor = node->GetOutControlAnchor();
  292. GE_IF_BOOL_EXEC(out_control_anchor == nullptr, GELOGW("out_control_anchor is nullptr"); return);
  293. vector<string> dst_name_list;
  294. vector<int64_t> dst_index_list;
  295. string dst_name_temp;
  296. for (const auto &in_control_anchor : out_control_anchor->GetPeerInControlAnchors()) {
  297. ge::NodePtr dst_node = in_control_anchor->GetOwnerNode();
  298. dst_name_temp = dst_name_temp.empty() ? dst_node->GetName() : dst_name_temp + ":" + dst_node->GetName();
  299. }
  300. GE_IF_BOOL_EXEC(!dst_name_temp.empty(), dst_name_list.emplace_back(dst_name_temp));
  301. GE_IF_BOOL_EXEC(!out_control_anchor->GetPeerInControlAnchors().empty(),
  302. dst_index_list.emplace_back(kInvalidIndexNum));
  303. for (const auto &out_data_anchor : node->GetAllOutDataAnchors()) {
  304. GE_IF_BOOL_EXEC(node_op_desc->HasAttr(MERGE_PRENODE_FLAG), break);
  305. dst_name_temp = "";
  306. int64_t dst_index = kWrongIndex; // assign an impossible value to dst_index.
  307. for (const auto &in_data_anchor : out_data_anchor->GetPeerInDataAnchors()) {
  308. GE_IF_BOOL_EXEC(in_data_anchor == nullptr, GELOGW("in_data_anchor is nullptr"); return);
  309. ge::NodePtr dst_node = in_data_anchor->GetOwnerNode();
  310. dst_name_temp = dst_name_temp.empty() ? dst_node->GetName() : dst_name_temp + ":" + dst_node->GetName();
  311. dst_index = in_data_anchor->GetIdx();
  312. }
  313. GE_IF_BOOL_EXEC(dst_index != kWrongIndex, dst_index_list.emplace_back(dst_index)); // not found
  314. GE_IF_BOOL_EXEC(!dst_name_temp.empty(), dst_name_list.emplace_back(dst_name_temp));
  315. }
  316. node_op_desc->SetDstName(dst_name_list);
  317. node_op_desc->SetDstIndex(dst_index_list);
  318. }
  319. }
  320. Status ModelBuilder::AdjustInputTensorFlag() {
  321. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  322. if ((n->GetType() == DATA_TYPE) || (n->GetType() == AIPP_DATA_TYPE)) {
  323. GELOGD("Data node: %s.", n->GetName().c_str());
  324. for (const auto &anchor : n->GetAllOutDataAnchors()) {
  325. for (const auto &in_anchors : anchor->GetPeerInDataAnchors()) {
  326. GE_IF_BOOL_EXEC(in_anchors == nullptr, continue);
  327. auto owner_node = in_anchors->GetOwnerNode();
  328. auto owner_node_op_desc = owner_node->GetOpDesc();
  329. GE_IF_BOOL_EXEC(owner_node_op_desc == nullptr, continue);
  330. auto input_desc = owner_node_op_desc->GetInputDesc(in_anchors->GetIdx());
  331. ge::TensorUtils::SetInputTensor(input_desc, true);
  332. if (owner_node_op_desc->UpdateInputDesc(in_anchors->GetIdx(), input_desc) != SUCCESS) {
  333. REPORT_CALL_ERROR("E19999", "Update Input desc size failed for op:%s(%s) index:%u",
  334. owner_node_op_desc->GetName().c_str(), owner_node_op_desc->GetType().c_str(),
  335. in_anchors->GetIdx());
  336. GELOGE(FAILED, "[Update][InputDesc] failed for op:%s(%s) index:%u",
  337. owner_node_op_desc->GetName().c_str(), owner_node_op_desc->GetType().c_str(), in_anchors->GetIdx());
  338. return FAILED;
  339. }
  340. }
  341. }
  342. }
  343. }
  344. return SUCCESS;
  345. }
  346. void ModelBuilder::InitL1FusionOption() {
  347. string buffer_optimize = "off_optimize";
  348. graphStatus ret = ge::GetContext().GetOption(BUFFER_OPTIMIZE, buffer_optimize);
  349. if (ret == GRAPH_SUCCESS) {
  350. bool off_superkernel = false;
  351. (void)AttrUtils::GetBool(compute_graph_, ATTR_NAME_OFF_SUPERKERNEL_ATTR, off_superkernel);
  352. is_l1_fusion_enable_ = ((buffer_optimize == "l1_optimize") && (!off_superkernel));
  353. GELOGI("Compute graph %s the value of %s is %s, superkernel flag %d.", compute_graph_->GetName().c_str(),
  354. BUFFER_OPTIMIZE.c_str(), buffer_optimize.c_str(), is_l1_fusion_enable_);
  355. } else {
  356. GELOGW("The value of %s is empty.", kEnableL1Fusion.c_str());
  357. }
  358. }
  359. Status ModelBuilder::BuildModelDef(ge::Model &model) {
  360. ClearOriginalFormat();
  361. max_mem_offset_ = mem_type_to_mem_offset_[RT_MEMORY_HBM];
  362. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_MEMORY_SIZE, max_mem_offset_),
  363. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_MEMORY_SIZE.c_str());
  364. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_MEMORY_SIZE.c_str());
  365. return FAILED);
  366. if (mem_type_to_mem_offset_.find(RT_MEMORY_P2P_DDR) != mem_type_to_mem_offset_.end()) {
  367. p2p_mem_offset_ = mem_type_to_mem_offset_[RT_MEMORY_P2P_DDR];
  368. }
  369. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_P2P_MEMORY_SIZE, p2p_mem_offset_),
  370. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_P2P_MEMORY_SIZE.c_str());
  371. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_P2P_MEMORY_SIZE.c_str());
  372. return FAILED);
  373. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_WEIGHT_SIZE, weight_offset_),
  374. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_WEIGHT_SIZE.c_str());
  375. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_WEIGHT_SIZE.c_str());
  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", ATTR_MODEL_STREAM_NUM.c_str());
  379. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_STREAM_NUM.c_str());
  380. return FAILED);
  381. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_EVENT_NUM, event_num_),
  382. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_EVENT_NUM.c_str());
  383. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_EVENT_NUM.c_str());
  384. return FAILED);
  385. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(&model, ATTR_MODEL_HUGE_STREAM_LIST, huge_streams_),
  386. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_HUGE_STREAM_LIST.c_str());
  387. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_HUGE_STREAM_LIST.c_str());
  388. return FAILED);
  389. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_LABEL_NUM, label_num_),
  390. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_LABEL_NUM.c_str());
  391. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_LABEL_NUM.c_str());
  392. return FAILED);
  393. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetInt(&model, ATTR_MODEL_ZERO_COPY_MEMORY_SIZE, zero_copy_mem_size_),
  394. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed",
  395. ATTR_MODEL_ZERO_COPY_MEMORY_SIZE.c_str());
  396. GELOGE(FAILED, "[Set][Attr] %s in model failed.", ATTR_MODEL_ZERO_COPY_MEMORY_SIZE.c_str());
  397. return FAILED);
  398. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&model, ATTR_MODEL_OUT_NODES_NAME, GetLocalOmgContext().net_out_nodes),
  399. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_OUT_NODES_NAME.c_str());
  400. GELOGE(FAILED, "[Set][Str] %s in model failed.", ATTR_MODEL_OUT_NODES_NAME.c_str());
  401. return FAILED);
  402. GELOGI("For model, max_mem_offset_: %zu, p2p_mem_size: %zu, zero_copy_mem_size_: %zu", max_mem_offset_,
  403. p2p_mem_offset_, zero_copy_mem_size_);
  404. string fp_ceiling_mode;
  405. if (ge::GetContext().GetOption("ge.fpCeilingMode", fp_ceiling_mode) == SUCCESS) {
  406. if (!ge::AttrUtils::SetStr(&model, ATTR_FP_CEILING_MODE, fp_ceiling_mode)) {
  407. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_FP_CEILING_MODE.c_str());
  408. GELOGE(FAILED, "[Set][Str] %s in model failed", ATTR_FP_CEILING_MODE.c_str());
  409. return FAILED;
  410. }
  411. GELOGI("Set attr ATTR_FP_CEILING_MODE to model, value is %s.", fp_ceiling_mode.c_str());
  412. }
  413. string ge_core_type;
  414. Status ret = ge::GetContext().GetOption(kCoreType, ge_core_type);
  415. if (ret != SUCCESS) {
  416. GELOGW("get the option CORE_TYPE fail, set it to default value VECTOR_ENGINE");
  417. }
  418. int64_t core_type = (ge_core_type == kVectorCore) ? 1 : 0;
  419. GELOGI("core_type: %ld", core_type);
  420. if (!ge::AttrUtils::SetInt(&model, ATTR_MODEL_CORE_TYPE, core_type)) {
  421. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_MODEL_CORE_TYPE.c_str());
  422. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_MODEL_CORE_TYPE.c_str());
  423. }
  424. InitL1FusionOption();
  425. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetBool(&model, ATTR_NAME_SWITCH_FOR_L1_FUSION, is_l1_fusion_enable_),
  426. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_NAME_SWITCH_FOR_L1_FUSION.c_str());
  427. GELOGE(FAILED, "[Set][Attr] %s in model failed.", ATTR_NAME_SWITCH_FOR_L1_FUSION.c_str());
  428. return FAILED);
  429. const DumpProperties &dump_properties = DumpManager::GetInstance().GetDumpProperties(session_id_);
  430. bool is_op_debug = dump_properties.IsOpDebugOpen();
  431. if (is_op_debug) {
  432. if (!ge::AttrUtils::SetBool(&model, ATTR_OP_DEBUG_FLAG, is_op_debug)) {
  433. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_OP_DEBUG_FLAG.c_str());
  434. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_OP_DEBUG_FLAG.c_str());
  435. return FAILED;
  436. }
  437. uint32_t op_debug_mode = dump_properties.GetOpDebugMode();
  438. GELOGI("Get op debug mode:%d", op_debug_mode);
  439. if (!ge::AttrUtils::SetInt(&model, ATTR_OP_DEBUG_MODE, op_debug_mode)) {
  440. REPORT_INNER_ERROR("E19999", "Set Attr:%s in model failed", ATTR_OP_DEBUG_MODE.c_str());
  441. GELOGE(FAILED, "[Set][Attr] %s in model failed", ATTR_OP_DEBUG_MODE.c_str());
  442. return FAILED;
  443. }
  444. }
  445. model.SetName(compute_graph_->GetName());
  446. model.SetGraph(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_));
  447. GELOGI("weight_offset_: %zu", weight_offset_);
  448. GELOGI("Set event num: %ld.", event_num_);
  449. if (Params::Instance() == nullptr) {
  450. return FAILED;
  451. }
  452. platform_type_ = Params::Instance()->GetTarget_8bit();
  453. return SUCCESS;
  454. }
  455. void ModelBuilder::ClearOriginalFormat() {
  456. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  457. auto node_op_desc = n->GetOpDesc();
  458. if (node_op_desc != nullptr) {
  459. if (node_op_desc->HasAttr(ATTR_NAME_FORMAT)) {
  460. if (node_op_desc->DelAttr(ATTR_NAME_FORMAT) != SUCCESS) {
  461. GELOGW("DelAttr ATTR_NAME_FORMAT failed.");
  462. }
  463. }
  464. GE_IF_BOOL_EXEC(
  465. node_op_desc->HasAttr(ATTR_NAME_INFERRED_FORMAT),
  466. if (node_op_desc->DelAttr(ATTR_NAME_INFERRED_FORMAT) != SUCCESS) {
  467. GELOGW("DelAttr ATTR_NAME_INFERRED_FORMAT failed.");
  468. });
  469. GE_IF_BOOL_EXEC(
  470. node_op_desc->HasAttr(ATTR_NAME_PRED_PERMUTE_DELETED),
  471. if (node_op_desc->DelAttr(ATTR_NAME_PRED_PERMUTE_DELETED) != SUCCESS) {
  472. GELOGW("DelAttr ATTR_NAME_PRED_PERMUTE_DELETED failed.");
  473. });
  474. GE_IF_BOOL_EXEC(
  475. node_op_desc->HasAttr(ATTR_NAME_IGNORE_PRED_FORMAT),
  476. if (node_op_desc->DelAttr(ATTR_NAME_IGNORE_PRED_FORMAT) != SUCCESS) {
  477. GELOGW("DelAttr ATTR_NAME_IGNORE_PRED_FORMAT failed.");
  478. });
  479. }
  480. }
  481. }
  482. Status ModelBuilder::MergeWeights() {
  483. if (weight_offset_ == 0) {
  484. return SUCCESS;
  485. }
  486. ge::Buffer buffer(weight_offset_);
  487. weight_buffer_ = buffer;
  488. auto base_addr = weight_buffer_.GetData();
  489. for (const ge::NodePtr &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  490. auto op_desc = node->GetOpDesc();
  491. GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
  492. if (node->GetType() != CONSTANT) {
  493. continue;
  494. }
  495. // Get const op weight pointer
  496. ge::GeTensorPtr weight = nullptr;
  497. // If MutableTensor failed, weight is nullptr.
  498. (void)ge::AttrUtils::MutableTensor(op_desc, ATTR_NAME_WEIGHTS, weight);
  499. if (weight == nullptr) {
  500. REPORT_INNER_ERROR("E19999", "Can't get const weight in op:%s(%s)",
  501. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  502. GELOGE(FAILED, "[Call][MutableTensor] Can't get const op weight, name:%s", node->GetName().c_str());
  503. return FAILED;
  504. }
  505. // Get const op weight offset
  506. int64_t offset = 0;
  507. if (ge::TensorUtils::GetDataOffset(weight->GetTensorDesc(), offset) != SUCCESS) {
  508. GELOGW("Can't get const op offset, name: %s", node->GetName().c_str());
  509. continue; // continue to merge if can not get offset
  510. }
  511. // Get const op weight data
  512. auto weight_data = weight->MutableData();
  513. // copy const op weight data to buffer
  514. GELOGI("Move to buffer, name: %s offset: %ld size: %zu", node->GetName().c_str(), offset, weight_data.size());
  515. ge::TensorUtils::SetWeightSize(weight->MutableTensorDesc(), static_cast<uint32_t>(weight_data.size()));
  516. if ((offset == 0) || (weight_data.size() == 0)) {
  517. GELOGI("Size or offset is 0. size: %lu offset: %ld", weight_data.size(), offset);
  518. continue;
  519. }
  520. if (weight_data.data() != nullptr) {
  521. GE_IF_BOOL_EXEC(base_addr == nullptr,
  522. REPORT_INNER_ERROR("E19999", "Check weight in op:%s(%s) is nullptr",
  523. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  524. GELOGE(FAILED, "[Check][Param] weight in op:%s(%s) is nullptr",
  525. op_desc->GetName().c_str(), op_desc->GetType().c_str());
  526. return FAILED);
  527. if (weight_offset_ - offset < weight_data.size()) {
  528. REPORT_INNER_ERROR("E19999", "left weight size not enough for op:%s(%s) left_size:%zu, weight_size:%zu",
  529. op_desc->GetName().c_str(), op_desc->GetType().c_str(),
  530. weight_offset_ - offset, weight_data.size());
  531. GELOGE(FAILED, "[Check][Param] left weight size not enough for op:%s(%s). left_size:%lu, weight_size:%lu",
  532. op_desc->GetName().c_str(), op_desc->GetType().c_str(), weight_offset_ - offset, weight_data.size());
  533. return FAILED;
  534. }
  535. uintptr_t dst_ptr = reinterpret_cast<uintptr_t>(base_addr) + offset;
  536. uintptr_t src_ptr = reinterpret_cast<uintptr_t>(weight_data.data());
  537. size_t left_size = weight_data.size();
  538. while (left_size > SECUREC_MEM_MAX_LEN) {
  539. auto err = memcpy_s(reinterpret_cast<void *>(dst_ptr), SECUREC_MEM_MAX_LEN, reinterpret_cast<void *>(src_ptr),
  540. SECUREC_MEM_MAX_LEN);
  541. if (err != EOK) {
  542. REPORT_CALL_ERROR("E19999", "mem copy failed. errret:%u, "
  543. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu,",
  544. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN);
  545. GELOGE(FAILED, "mem copy failed. errret:%u, "
  546. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu",
  547. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN);
  548. return FAILED;
  549. }
  550. left_size -= SECUREC_MEM_MAX_LEN;
  551. dst_ptr = dst_ptr + SECUREC_MEM_MAX_LEN;
  552. src_ptr = src_ptr + SECUREC_MEM_MAX_LEN;
  553. }
  554. auto err = memcpy_s(reinterpret_cast<void *>(dst_ptr), left_size, reinterpret_cast<void *>(src_ptr), left_size);
  555. if (err != EOK) {
  556. REPORT_CALL_ERROR("E19999", "mem copy failed. errret:%u, "
  557. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu,",
  558. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN);
  559. GELOGE(FAILED, "[Update][Data] mem copy failed. errret:%u, "
  560. "dst_ptr:%lx, dst_size:%lu, src_ptr:%lx, src_size:%lu",
  561. err, dst_ptr, SECUREC_MEM_MAX_LEN, src_ptr, SECUREC_MEM_MAX_LEN);
  562. return FAILED;
  563. }
  564. }
  565. weight->ClearData();
  566. }
  567. return SUCCESS;
  568. }
  569. Status ModelBuilder::SaveAtomicTBEKernel(const OpDescPtr &op_desc) {
  570. ge::NodePtr atomic_clean_node = nullptr;
  571. atomic_clean_node = op_desc->TryGetExtAttr("atomic_clean_node_ptr", atomic_clean_node);
  572. if (atomic_clean_node == nullptr) {
  573. return SUCCESS;
  574. }
  575. ge::OpDescPtr atomic_op_desc = atomic_clean_node->GetOpDesc();
  576. GE_CHECK_NOTNULL(atomic_op_desc);
  577. TBEKernelPtr tbe_kernel = atomic_op_desc->TryGetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
  578. if (tbe_kernel == nullptr) {
  579. std::string kernel_name;
  580. GeAttrValue::BYTES kernel_buffer;
  581. (void) AttrUtils::GetStr(atomic_op_desc, ATTR_NAME_TBE_KERNEL_NAME, kernel_name);
  582. (void) AttrUtils::GetBytes(atomic_op_desc, ATTR_NAME_TBE_KERNEL_BUFFER, kernel_buffer);
  583. if (!kernel_name.empty() && (kernel_buffer.GetSize() > 0)) {
  584. GE_CHECK_NOTNULL(kernel_buffer.GetData());
  585. std::vector<char> data(kernel_buffer.GetData(), kernel_buffer.GetData() + kernel_buffer.GetSize());
  586. tbe_kernel = MakeShared<OpKernelBin>(kernel_name, std::move(data));
  587. GE_CHECK_NOTNULL(tbe_kernel);
  588. GELOGI("Node [%s][%s] start recovery extra attr %s from %s", atomic_op_desc->GetName().c_str(),
  589. atomic_op_desc->GetType().c_str(), ge::OP_EXTATTR_NAME_TBE_KERNEL, ATTR_NAME_TBE_KERNEL_NAME.c_str());
  590. if (!(atomic_op_desc->SetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, tbe_kernel))) {
  591. std::string error = "Node" + FmtToStr(atomic_op_desc->GetName()) + "set extra tbeKernel attr failed";
  592. GE_ERRORLOG_AND_ERRORMSG(ge::FAILED, error.c_str());
  593. return ge::FAILED;
  594. }
  595. }
  596. }
  597. if (tbe_kernel == nullptr) {
  598. GELOGD("Atomic_clean_node doesn't have tbe_kernel.");
  599. return SUCCESS;
  600. }
  601. tbe_kernel_store_.AddTBEKernel(tbe_kernel);
  602. GELOGD("Atomic_clean_node tbe_kernel_name %s!", tbe_kernel->GetName().c_str());
  603. (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TBE_KERNEL_NAME, tbe_kernel->GetName());
  604. std::string kernel_name;
  605. (void) AttrUtils::GetStr(atomic_op_desc, atomic_op_desc->GetName() + "_kernelname", kernel_name);
  606. (void) AttrUtils::SetStr(op_desc, op_desc->GetName() + "_atomic_kernelname", kernel_name);
  607. std::string meta_data;
  608. (void) AttrUtils::GetStr(atomic_op_desc, TVM_ATTR_NAME_METADATA, meta_data);
  609. (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TVM_METADATA, meta_data);
  610. std::string json_string;
  611. (void) AttrUtils::GetStr(atomic_op_desc, TVM_ATTR_NAME_MAGIC, json_string);
  612. (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TVM_MAGIC, json_string);
  613. return SUCCESS;
  614. }
  615. Status ModelBuilder::SaveDataToModel(ge::Model &model, ge::GeModel &ge_model) {
  616. // Add weight
  617. ge_model.SetWeight(weight_buffer_);
  618. // Add TBE Kernels and custom aicpu op bin
  619. std::set<std::string> tbe_name_set;
  620. std::set<std::string> aicpu_name_set;
  621. std::set<std::string> aicpu_op_types;
  622. std::set<std::string> aicpu_tf_op_types;
  623. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  624. auto node_op_desc = n->GetOpDesc();
  625. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  626. // check aicpu op type
  627. CollectCheckAicpuAttr(node_op_desc, aicpu_op_types, aicpu_tf_op_types);
  628. TBEKernelPtr tbe_kernel = node_op_desc->TryGetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
  629. if (tbe_kernel == nullptr) {
  630. std::string kernel_name;
  631. GeAttrValue::BYTES kernel_buffer;
  632. (void) AttrUtils::GetStr(node_op_desc, ATTR_NAME_TBE_KERNEL_NAME, kernel_name);
  633. (void) AttrUtils::GetBytes(node_op_desc, ATTR_NAME_TBE_KERNEL_BUFFER, kernel_buffer);
  634. if (!kernel_name.empty() && (kernel_buffer.GetSize() > 0)) {
  635. GE_CHECK_NOTNULL(kernel_buffer.GetData());
  636. std::vector<char> data(kernel_buffer.GetData(), kernel_buffer.GetData() + kernel_buffer.GetSize());
  637. tbe_kernel = std::make_shared<OpKernelBin>(kernel_name, std::move(data));
  638. GE_CHECK_NOTNULL(tbe_kernel);
  639. GELOGI("Node [%s][%s] start recovery extra attr %s from %s", node_op_desc->GetName().c_str(),
  640. node_op_desc->GetType().c_str(), ge::OP_EXTATTR_NAME_TBE_KERNEL, ATTR_NAME_TBE_KERNEL_NAME.c_str());
  641. if (!(node_op_desc->SetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, tbe_kernel))) {
  642. std::string error = "Node" + FmtToStr(node_op_desc->GetName()) + "set extra tbeKernel attr failed";
  643. GE_ERRORLOG_AND_ERRORMSG(ge::FAILED, error.c_str());
  644. return ge::FAILED;
  645. }
  646. }
  647. }
  648. GE_IF_BOOL_EXEC(tbe_kernel == nullptr, continue);
  649. if (tbe_name_set.count(tbe_kernel->GetName()) > 0) {
  650. REPORT_INNER_ERROR("E19999", "tbe_kernel name %s can't be the same, judge for op:%s(%s)",
  651. tbe_kernel->GetName().c_str(), n->GetName().c_str(), n->GetType().c_str());
  652. GELOGE(FAILED, "[Check][Param] tbe_kernel name %s can't be the same, judge for op:%s(%s)",
  653. tbe_kernel->GetName().c_str(), n->GetName().c_str(), n->GetType().c_str());
  654. return FAILED;
  655. }
  656. tbe_name_set.insert(tbe_kernel->GetName());
  657. tbe_kernel_store_.AddTBEKernel(tbe_kernel);
  658. GE_CHK_STATUS_RET(SaveAtomicTBEKernel(node_op_desc), "[Save][TBEKernel] save atomic tbekernel failed!");
  659. }
  660. SetModelCheckAicpuAttr(model, aicpu_op_types, aicpu_tf_op_types);
  661. for (const ge::NodePtr &n : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  662. auto node_op_desc = n->GetOpDesc();
  663. GE_IF_BOOL_EXEC(node_op_desc == nullptr, continue);
  664. CustAICPUKernelPtr cust_aicpu_kernel =
  665. node_op_desc->TryGetExtAttr(ge::OP_EXTATTR_CUSTAICPU_KERNEL, CustAICPUKernelPtr());
  666. GE_IF_BOOL_EXEC(cust_aicpu_kernel == nullptr, continue);
  667. if (aicpu_name_set.count(cust_aicpu_kernel->GetName()) > 0) {
  668. REPORT_INNER_ERROR("E19999", "aicpu_kernel name %s can't be the same, judge for op:%s(%s)",
  669. cust_aicpu_kernel->GetName().c_str(), n->GetName().c_str(), n->GetType().c_str());
  670. GELOGE(FAILED, "[Check][Param] aicpu_kernel name %s can't be the same, judge for op:%s(%s)",
  671. cust_aicpu_kernel->GetName().c_str(), n->GetName().c_str(), n->GetType().c_str());
  672. return FAILED;
  673. }
  674. aicpu_name_set.insert(cust_aicpu_kernel->GetName());
  675. cust_aicpu_kernel_store_.AddCustAICPUKernel(cust_aicpu_kernel);
  676. GELOGI("Add cust aicpu kernel bin %s", cust_aicpu_kernel->GetName().c_str());
  677. }
  678. if (!tbe_kernel_store_.Build()) {
  679. GELOGE(FAILED, "[Call][Build] TBE Kernels store build failed!");
  680. return FAILED;
  681. }
  682. if (!cust_aicpu_kernel_store_.Build()) {
  683. GELOGE(FAILED, "[Call][Build] custom AICPU kernels store build failed!");
  684. return FAILED;
  685. }
  686. ge_model.SetTBEKernelStore(tbe_kernel_store_);
  687. ge_model.SetCustAICPUKernelStore(cust_aicpu_kernel_store_);
  688. // Add task
  689. GeAttrValue::BYTES task_def_bytes;
  690. if (!AttrUtils::GetZeroCopyBytes(model, MODEL_ATTR_TASKS, task_def_bytes)) {
  691. REPORT_CALL_ERROR("E19999", "Get attr:%s in model failed", MODEL_ATTR_TASKS.c_str());
  692. GELOGE(INTERNAL_ERROR, "[Get][Attr] %s in model failed", MODEL_ATTR_TASKS.c_str());
  693. return INTERNAL_ERROR;
  694. }
  695. int byte_size = static_cast<int>(task_def_bytes.GetSize());
  696. std::shared_ptr<domi::ModelTaskDef> task = ge::MakeShared<domi::ModelTaskDef>();
  697. GE_CHECK_NOTNULL(task);
  698. GE_CHK_BOOL_EXEC(ReadProtoFromArray(task_def_bytes.GetData(), byte_size, task.get()), return INTERNAL_ERROR,
  699. "[Read][Proto] From Array failed.");
  700. ge_model.SetModelTaskDef(task);
  701. // Add graph
  702. ge_model.SetName(model.GetName());
  703. ge_model.SetGraph(model.GetGraph());
  704. ge_model.SetVersion(model.GetVersion());
  705. ge_model.SetPlatformVersion(model.GetPlatformVersion());
  706. ge_model.SetPlatformType(platform_type_);
  707. ge_model.SetAttr(model.MutableAttrMap());
  708. return SUCCESS;
  709. }
  710. void ModelBuilder::SetModelVersion(ge::Model &model) {
  711. // set framework_version TO model
  712. string framework_version;
  713. uint32_t counter = 0;
  714. Status frame_rt = PlatformVersionManager::GetPlatformVersion(framework_version);
  715. GE_IF_BOOL_EXEC((frame_rt == SUCCESS),
  716. string model_framework_version = framework_version + "." + std::to_string(counter);
  717. model.SetPlatformVersion(model_framework_version););
  718. // set IR Version TO model
  719. model.SetVersion(static_cast<uint32_t>(OM_PROTO_VERSION));
  720. }
  721. Status ModelBuilder::PreBuildModel() {
  722. if ((compute_graph_ == nullptr) || !(compute_graph_->IsValid())) {
  723. REPORT_INNER_ERROR("E19999", "Check compute_graph no valid");
  724. GELOGE(FAILED, "[Check][Param] Graph_ is not valid.");
  725. return FAILED;
  726. }
  727. GE_CHK_STATUS_RET(SetInputOutputDesc(),
  728. "[Set][InputOutputDesc] Failed! graph:%s", compute_graph_->GetName().c_str());
  729. AddNodeInputProperty();
  730. return SUCCESS;
  731. }
  732. Status ModelBuilder::BuildModelForGetTask(ge::Model &model) {
  733. GE_CHK_STATUS_RET(AdjustInputTensorFlag(), "[Adjust][InputTensorFlag] failed! graph:%s",
  734. compute_graph_->GetName().c_str());
  735. ErrorManager::GetInstance().SetStage(error_message::kModelCompile, error_message::kStreamAlloc);
  736. // Assign logical streams.
  737. StreamAllocator stream_allocator(compute_graph_, subgraphs_);
  738. GE_TIMESTAMP_START(AssignLogicalStreams);
  739. GE_CHK_STATUS_RET(stream_allocator.AssignLogicalStreams(stream_max_parallel_num_, hcom_parallel_),
  740. "[Assign][LogicalStreams] failed. graph:%s", compute_graph_->GetName().c_str());
  741. GE_TIMESTAMP_END(AssignLogicalStreams, "GraphBuilder::AssignLogicalStreams");
  742. ErrorManager::GetInstance().SetStage(error_message::kModelCompile, error_message::kMemoryAlloc);
  743. // Assign functional op labels.
  744. auto root_graph = GraphUtils::FindRootGraph(compute_graph_);
  745. (void)AttrUtils::GetInt(*root_graph, ATTR_MODEL_LABEL_NUM, label_num_);
  746. GE_TIMESTAMP_START(AssignMemory);
  747. MemoryAssigner mem_assigner(compute_graph_);
  748. GE_CHK_STATUS_RET(mem_assigner.AssignMemory(is_loop_graph_, mem_type_to_mem_offset_, zero_copy_mem_size_),
  749. "[Assign][Memory] Failed! graph:%s", compute_graph_->GetName().c_str());
  750. GE_TIMESTAMP_END(AssignMemory, "GraphBuilder::AssignMemory");
  751. ErrorManager::GetInstance().SetStage(error_message::kModelCompile, error_message::kOther);
  752. GE_TIMESTAMP_START(SetInputOutputOffset);
  753. SetInputOutputOffsetPass input_output_offset;
  754. GE_CHK_STATUS_RET(input_output_offset.Run(compute_graph_),
  755. "[Set][InputOutputOffset] failed. graph:%s", compute_graph_->GetName().c_str());
  756. GE_TIMESTAMP_END(SetInputOutputOffset, "SetInputOutputOffsetPass::Run");
  757. // Compile single op in graph build stage
  758. GE_TIMESTAMP_START(CompileSingleOp);
  759. GE_CHK_STATUS_RET(CompileSingleOp(), "[Compile][SingleOp] fail. graph:%s", compute_graph_->GetName().c_str());
  760. GE_TIMESTAMP_EVENT_END(CompileSingleOp, "GraphBuilder::CompileSingleOp");
  761. ErrorManager::GetInstance().SetStage(error_message::kModelCompile, error_message::kStreamAlloc);
  762. // Refresh real streams and insert event nodes.
  763. GE_TIMESTAMP_START(RefreshRealStream);
  764. GE_CHK_STATUS_RET(stream_allocator.RefreshRealStream(stream_num_, event_num_),
  765. "[Refresh][RealStream] failed. graph:%s", compute_graph_->GetName().c_str());
  766. huge_streams_ = stream_allocator.GetHugeStreams();
  767. GE_TIMESTAMP_END(RefreshRealStream, "GraphBuilder::RefreshRealStream");
  768. ErrorManager::GetInstance().SetStage(error_message::kModelCompile, error_message::kOther);
  769. GE_TIMESTAMP_START(MergeWeights);
  770. GE_CHK_STATUS_RET(MergeWeights(), "[Merge][Weights] Failed! graph:%s", compute_graph_->GetName().c_str());
  771. GE_TIMESTAMP_END(MergeWeights, "GraphBuilder::MergeWeights");
  772. GE_TIMESTAMP_START(BuildModelDef);
  773. GE_CHK_STATUS_RET(BuildModelDef(model), "[Build][ModelDef] failed! graph:%s", compute_graph_->GetName().c_str());
  774. GE_TIMESTAMP_END(BuildModelDef, "GraphBuilder::BuildModelDef");
  775. SetModelVersion(model);
  776. return SUCCESS;
  777. }
  778. Status ModelBuilder::BuildModelForGetDynShapeTask(ge::Model &model_def) {
  779. GE_TIMESTAMP_START(BuildModelDef);
  780. GE_CHK_STATUS_RET(BuildModelDef(model_def), "[Build][ModelDef] failed!");
  781. GE_TIMESTAMP_END(BuildModelDef, "GraphBuilder::BuildModelDef");
  782. SetModelVersion(model_def);
  783. return SUCCESS;
  784. }
  785. ge::Buffer ModelBuilder::GetWeightBuffer() const { return weight_buffer_; }
  786. Status ModelBuilder::CompileSingleOp() {
  787. GELOGD("Begin to compile single op.");
  788. // Create ge instance
  789. std::shared_ptr<GELib> instance = ge::GELib::GetInstance();
  790. if ((instance == nullptr) || !instance->InitFlag()) {
  791. REPORT_INNER_ERROR("E19999", "Check GELib instance not init before");
  792. GELOGE(ge::GE_CLI_GE_NOT_INITIALIZED, "[Check][Param] CompileSingleOp failed.");
  793. return ge::GE_CLI_GE_NOT_INITIALIZED;
  794. }
  795. GE_TIMESTAMP_CALLNUM_START(BatchCompileOp);
  796. std::unordered_map<string, vector<ge::NodePtr>> node_vector_map;
  797. for (auto &node : compute_graph_->GetNodes(compute_graph_->GetGraphUnknownFlag())) {
  798. auto op_desc = node->GetOpDesc();
  799. if (op_desc == nullptr) {
  800. continue;
  801. }
  802. // Graph build stage only supports the individual compilation of atomic clean operator
  803. if (op_desc->GetType() == ATOMICADDRCLEAN) {
  804. GELOGD("Begin to compile single op, op name is %s.", op_desc->GetName().c_str());
  805. string kernel_lib_name = op_desc->GetOpKernelLibName();
  806. if (kernel_lib_name.empty()) {
  807. // Reset op kernel lib
  808. (void)instance->DNNEngineManagerObj().GetDNNEngineName(node);
  809. kernel_lib_name = op_desc->GetOpKernelLibName();
  810. if (kernel_lib_name.empty()) {
  811. REPORT_INNER_ERROR("E19999", "Check kernel lib name empty of op:%s(%s)",
  812. node->GetName().c_str(), node->GetType().c_str());
  813. GELOGE(ge::INTERNAL_ERROR, "[Get][Name] of node:%s(%s) kernel lib failed.", node->GetName().c_str(),
  814. node->GetType().c_str());
  815. return ge::INTERNAL_ERROR;
  816. }
  817. }
  818. OpsKernelInfoStorePtr kernel_info = instance->OpsKernelManagerObj().GetOpsKernelInfoStore(kernel_lib_name);
  819. if (kernel_info != nullptr) {
  820. node_vector_map[kernel_lib_name].emplace_back(node);
  821. } else {
  822. REPORT_INNER_ERROR("E19999", "Get ops kernel info store failed for op:%s(%s), op_kernel_name:%s,",
  823. node->GetName().c_str(), node->GetType().c_str(), kernel_lib_name.c_str());
  824. GELOGE(ge::GE_GRAPH_PARAM_NULLPTR, "[Get][OpsKernelInfoStore] for op %s failed", node->GetName().c_str());
  825. return ge::GE_GRAPH_PARAM_NULLPTR;
  826. }
  827. }
  828. }
  829. for (auto &it : node_vector_map) {
  830. auto &kernel_lib_name = it.first;
  831. auto &node_vector = it.second;
  832. OpsKernelInfoStorePtr kernel_info = instance->OpsKernelManagerObj().GetOpsKernelInfoStore(kernel_lib_name);
  833. GE_CHECK_NOTNULL(kernel_info);
  834. GE_TIMESTAMP_RESTART(BatchCompileOp);
  835. auto ret = kernel_info->CompileOp(node_vector);
  836. GELOGI("[GEPERFTRACE] The node size of compile op of %s is %zu", kernel_lib_name.c_str(), node_vector.size());
  837. GE_TIMESTAMP_ADD(BatchCompileOp);
  838. if (ret != ge::SUCCESS) {
  839. REPORT_CALL_ERROR("E19999", "Batch compile op failed, kernel lib name, node size:%zu,",
  840. node_vector.size());
  841. GELOGE(ret, "[Compile][Op] failed, kernel lib name is %s", kernel_lib_name.c_str());
  842. return ret;
  843. }
  844. }
  845. GE_TIMESTAMP_CALLNUM_END(BatchCompileOp, "GraphBuild::CompileOp");
  846. return ge::SUCCESS;
  847. }
  848. void ModelBuilder::CollectCheckAicpuAttr(const OpDescPtr &op_desc, std::set<std::string> &aicpu_op_types,
  849. std::set<std::string> &aicpu_tf_op_types) {
  850. std::string aicpu_optype;
  851. bool has_attr_check_cpu = ge::AttrUtils::GetStr(op_desc, "needCheckCpu", aicpu_optype);
  852. std::vector<std::string> tf_optypes;
  853. bool has_attr_check_tf = ge::AttrUtils::GetListStr(op_desc, "needCheckTf", tf_optypes);
  854. if (has_attr_check_cpu && !aicpu_optype.empty()) {
  855. aicpu_op_types.insert(aicpu_optype);
  856. }
  857. if (has_attr_check_tf && !tf_optypes.empty()) {
  858. aicpu_tf_op_types.insert(tf_optypes.begin(), tf_optypes.end());
  859. }
  860. return;
  861. }
  862. void ModelBuilder::SetModelCheckAicpuAttr(ge::Model &model, std::set<std::string> &aicpu_op_types,
  863. std::set<std::string> &aicpu_tf_op_types) {
  864. std::vector<std::string> aicpu_optype_list;
  865. std::vector<std::string> aicpu_tf_optype_list;
  866. if (ge::AttrUtils::GetListStr(&model, "needCheckCpu", aicpu_optype_list)) {
  867. GELOGI("Already have aicpu optype size: %zu", aicpu_optype_list.size());
  868. aicpu_op_types.insert(aicpu_optype_list.begin(), aicpu_optype_list.end());
  869. }
  870. if (ge::AttrUtils::GetListStr(&model, "needCheckTf", aicpu_tf_optype_list)) {
  871. GELOGI("Already have aicpu tf optype size: %zu", aicpu_tf_optype_list.size());
  872. aicpu_tf_op_types.insert(aicpu_tf_optype_list.begin(), aicpu_tf_optype_list.end());
  873. }
  874. // reset list with set
  875. aicpu_optype_list.assign(aicpu_op_types.begin(), aicpu_op_types.end());
  876. aicpu_tf_optype_list.assign(aicpu_tf_op_types.begin(), aicpu_tf_op_types.end());
  877. GELOGI(
  878. "Check Aicpu op types ComputeGraph: %s aicpu_op_types: %zu, aicpu_optype_list: %zu, aicpu_tf_op_types: %zu, "
  879. "aicpu_tf_optype_list:%zu.",
  880. compute_graph_->GetName().c_str(), aicpu_op_types.size(), aicpu_optype_list.size(), aicpu_tf_op_types.size(),
  881. aicpu_tf_optype_list.size());
  882. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&model, "needCheckCpu", aicpu_optype_list), return,
  883. "[Set][Attr] needCheckCpu fail.");
  884. GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&model, "needCheckTf", aicpu_tf_optype_list), return,
  885. "[Set][Attr] needCheckTf fail.");
  886. return;
  887. }
  888. } // 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