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.

broadcast_args_kernel.cc 2.9 kB

5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283
  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 "host_kernels/broadcast_args_kernel.h"
  17. #include <vector>
  18. #include "common/op/ge_op_utils.h"
  19. #include "common/types.h"
  20. #include "common/util.h"
  21. #include "framework/common/ge_inner_error_codes.h"
  22. #include "graph/common/bcast.h"
  23. #include "graph/passes/pass_utils.h"
  24. #include "inc/kernel_factory.h"
  25. namespace ge {
  26. namespace {
  27. const size_t kBCastArgsInputsSize = 2;
  28. const size_t kBCastArgsOutputsSize = 1;
  29. } // namespace
  30. Status BroadcastArgsKernel::Compute(const OpDescPtr op_desc_ptr, const std::vector<ConstGeTensorPtr> &input,
  31. std::vector<GeTensorPtr> &v_output) {
  32. GELOGD("BroadcastArgsKernel in");
  33. if (op_desc_ptr == nullptr) {
  34. GELOGE(PARAM_INVALID, "Parameter's invalid, Input opDescPtr is nullptr.");
  35. return PARAM_INVALID;
  36. }
  37. // check input size
  38. bool size_check =
  39. (op_desc_ptr->GetAllInputsDesc().size() != kBCastArgsInputsSize || input.size() != kBCastArgsInputsSize ||
  40. op_desc_ptr->GetAllOutputsDesc().size() != kBCastArgsOutputsSize);
  41. if (size_check) {
  42. GELOGW("input/output size error. InDesc size:%zu,"
  43. "OutDesc size:%zu, in size:%zu ",
  44. op_desc_ptr->GetAllInputsDesc().size(), op_desc_ptr->GetAllOutputsDesc().size(), input.size());
  45. return NOT_CHANGED;
  46. }
  47. vector<int64_t> x1_dims;
  48. vector<int64_t> x2_dims;
  49. const auto &op_in_desc = op_desc_ptr->MutableInputDesc(0);
  50. GE_CHECK_NOTNULL(op_in_desc);
  51. DataType data_type = op_in_desc->GetDataType();
  52. bool result = (OpUtils::GetShapeDataFromConstTensor(input[0], data_type, x1_dims) == SUCCESS) &&
  53. (OpUtils::GetShapeDataFromConstTensor(input[1], data_type, x2_dims) == SUCCESS);
  54. if (!result) {
  55. GELOGE(PARAM_INVALID, "GetShapeDataFromConstTensor fail.");
  56. return PARAM_INVALID;
  57. }
  58. BCast bcast;
  59. Status ret = bcast.GenerateBcastInfo(x1_dims, x2_dims);
  60. if (ret != SUCCESS) {
  61. GELOGE(ret, "GenerateBcastInfo fail.");
  62. return ret;
  63. }
  64. vector<int64_t> bcast_dims = bcast.GetOutputShape();
  65. ret = PassUtils::ConstructTensorDescWithData(op_desc_ptr->GetOutputDesc(0), bcast_dims, v_output);
  66. if (ret != SUCCESS) {
  67. GELOGE(ret, "BroadcastArgs kernel construct tensor desc fail");
  68. return ret;
  69. }
  70. return SUCCESS;
  71. }
  72. REGISTER_KERNEL(BROADCASTARGS, BroadcastArgsKernel);
  73. } // 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