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.

aicpu_kernel_task_builder.cc 3.5 kB

4 years ago
4 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192
  1. /**
  2. * Copyright 2019-2020 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 "single_op/task/aicpu_kernel_task_builder.h"
  17. #include "cce/taskdown_common.hpp"
  18. #include "graph/load/new_model_manager/model_manager.h"
  19. namespace ge {
  20. AiCpuCCTaskBuilder::AiCpuCCTaskBuilder(const OpDescPtr &op_desc, const domi::KernelDef &kernel_def)
  21. : op_desc_(op_desc), kernel_def_(kernel_def) {}
  22. Status AiCpuCCTaskBuilder::SetKernelArgs(AiCpuCCTask &task) {
  23. size_t aicpu_arg_size = kernel_def_.args_size();
  24. if (aicpu_arg_size <= 0) {
  25. GELOGE(RT_FAILED, "aicpu_arg_size is invalid, value = %zu", aicpu_arg_size);
  26. return RT_FAILED;
  27. }
  28. std::unique_ptr<uint8_t[]> aicpu_args;
  29. aicpu_args.reset(new(std::nothrow) uint8_t[aicpu_arg_size]());
  30. if (aicpu_args == nullptr) {
  31. GELOGE(RT_FAILED, "malloc failed, size = %zu", aicpu_arg_size);
  32. return RT_FAILED;
  33. }
  34. auto err = memcpy_s(aicpu_args.get(), aicpu_arg_size, kernel_def_.args().data(), aicpu_arg_size);
  35. if (err != EOK) {
  36. GELOGE(RT_FAILED, "memcpy_s args failed, size = %zu, err = %d", aicpu_arg_size, err);
  37. return RT_FAILED;
  38. }
  39. task.SetIoAddr(aicpu_args.get() + sizeof(aicpu::AicpuParamHead));
  40. task.SetKernelArgs(std::move(aicpu_args), aicpu_arg_size);
  41. return SUCCESS;
  42. }
  43. Status AiCpuCCTaskBuilder::BuildTask(AiCpuCCTask &task) {
  44. auto ret = SetKernelArgs(task);
  45. if (ret != SUCCESS) {
  46. return ret;
  47. }
  48. const std::string &so_name = kernel_def_.so_name();
  49. const std::string &kernel_name = kernel_def_.kernel_name();
  50. task.SetSoName(so_name);
  51. task.SetkernelName(kernel_name);
  52. task.op_desc_ = op_desc_;
  53. const auto &context = kernel_def_.context();
  54. auto kernel_type = static_cast<cce::ccKernelType>(context.kernel_type());
  55. if (kernel_type == cce::ccKernelType::CUST_AI_CPU) {
  56. task.is_custom_ = true;
  57. task.dump_flag_ |= RT_KERNEL_CUSTOM_AICPU;
  58. GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc_, so_name), "launch cust aicpu so failed");
  59. }
  60. task.num_inputs_ = op_desc_->GetInputsSize();
  61. task.num_outputs_ = op_desc_->GetOutputsSize();
  62. // get kernel_ext_info
  63. auto &kernel_ext_info = kernel_def_.kernel_ext_info();
  64. auto kernel_ext_info_size = kernel_def_.kernel_ext_info_size();
  65. GE_CHK_BOOL_RET_STATUS(kernel_ext_info.size() == kernel_ext_info_size, FAILED,
  66. "task def kernel_ext_info.size=%zu, but kernel_ext_info_size=%u.",
  67. kernel_ext_info.size(), kernel_ext_info_size);
  68. ret = task.SetExtInfoAndType(kernel_ext_info);
  69. if (ret != SUCCESS) {
  70. GELOGE(ret, "Init ext info failed.");
  71. return ret;
  72. }
  73. auto aicpu_param_head = reinterpret_cast<aicpu::AicpuParamHead *>(task.args_.get());
  74. if (task.ext_info_addr_dev_ != nullptr) {
  75. aicpu_param_head->extInfoLength = kernel_ext_info.size();
  76. aicpu_param_head->extInfoAddr = reinterpret_cast<uintptr_t>(task.ext_info_addr_dev_);
  77. }
  78. return SUCCESS;
  79. }
  80. } // 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