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.

parallel_concat_start_op_pass.cc 3.6 kB

4 years ago
4 years ago
4 years ago
4 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586
  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/passes/parallel_concat_start_op_pass.h"
  17. #include <vector>
  18. #include "common/ge/ge_util.h"
  19. #include "framework/common/debug/ge_log.h"
  20. #include "graph/debug/ge_attr_define.h"
  21. #include "graph/node.h"
  22. #include "graph/utils/attr_utils.h"
  23. #include "graph/utils/graph_utils.h"
  24. #include "graph/utils/node_utils.h"
  25. namespace ge {
  26. namespace {
  27. const size_t kParallelConcatStartOutputSize = 1;
  28. const uint32_t kParallelConcatStartOutputDataIndex = 0;
  29. const char *const kAttrDtype = "dtype";
  30. const char *const kAttrShape = "shape";
  31. } // namespace
  32. Status ParallelConcatStartOpPass::Run(NodePtr &node) {
  33. GE_CHECK_NOTNULL(node);
  34. if (node->GetType() != PARALLELCONCATSTART) {
  35. return SUCCESS;
  36. }
  37. OpDescPtr node_op_desc = node->GetOpDesc();
  38. GE_CHECK_NOTNULL(node_op_desc);
  39. string node_name = node->GetName();
  40. GELOGI("Start to replace operator _ParallelConcatStart with Constant, node name: %s.", node_name.c_str());
  41. if (node_op_desc->GetOutputsSize() != kParallelConcatStartOutputSize) {
  42. REPORT_INNER_ERROR("E19999", "Output tensor num:%zu of node:%s(%s) != %zu, check invalid",
  43. node_op_desc->GetOutputsSize(), node_op_desc->GetName().c_str(),
  44. node_op_desc->GetType().c_str(), kParallelConcatStartOutputSize);
  45. GELOGE(PARAM_INVALID, "[Check][Param] Node[%s] output size is unexpected, the value is %zu, expected valude:%zu.",
  46. node_name.c_str(), node_op_desc->GetOutputsSize(), kParallelConcatStartOutputSize);
  47. return PARAM_INVALID;
  48. }
  49. auto output_tensor_desc = node_op_desc->GetOutputDesc(kParallelConcatStartOutputDataIndex);
  50. GeTensorPtr output_ptr = MakeShared<GeTensor>(output_tensor_desc);
  51. if (output_ptr == nullptr) {
  52. REPORT_CALL_ERROR("E19999", "New GeTensor failed");
  53. GELOGE(MEMALLOC_FAILED, "[New][GeTensor] failed");
  54. return FAILED;
  55. }
  56. ge::DataType attr_dtype;
  57. if (!ge::AttrUtils::GetDataType(node_op_desc, kAttrDtype, attr_dtype)) {
  58. REPORT_CALL_ERROR("E19999", "Get Attr:%s from op:%s(%s) failed", kAttrDtype,
  59. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str());
  60. GELOGE(PARAM_INVALID, "[Get][Attr] %s from op:%s(%s) failed", kAttrDtype,
  61. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str());
  62. return PARAM_INVALID;
  63. }
  64. output_ptr->MutableTensorDesc().SetDataType(attr_dtype);
  65. vector<int64_t> attr_shape_list;
  66. if (!ge::AttrUtils::GetListInt(node_op_desc, kAttrShape, attr_shape_list)) {
  67. REPORT_CALL_ERROR("E19999", "Get Attr:%s from op:%s(%s) failed", kAttrShape,
  68. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str());
  69. GELOGE(PARAM_INVALID, "[Get][Attr] %s from op:%s(%s) failed", kAttrShape,
  70. node_op_desc->GetName().c_str(), node_op_desc->GetType().c_str());
  71. return PARAM_INVALID;
  72. }
  73. output_ptr->MutableTensorDesc().SetShape(GeShape(attr_shape_list));
  74. vector<GeTensorPtr> outputs;
  75. outputs.emplace_back(output_ptr);
  76. return Folding(node, outputs);
  77. }
  78. } // 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