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.

label_goto_task.cc 3.4 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101
  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 "ge_runtime/task/label_goto_task.h"
  17. #include "ge_runtime/task/task_factory.h"
  18. namespace ge {
  19. namespace model_runner {
  20. LabelGotoTask::LabelGotoTask(const ModelContext &model_context, const std::shared_ptr<LabelGotoTaskInfo> &task_info)
  21. : TaskRepeater<LabelGotoTaskInfo>(model_context, task_info),
  22. task_info_(task_info),
  23. stream_(nullptr),
  24. index_value_(nullptr) {
  25. if (task_info_ == nullptr) {
  26. GELOGW("task_info_ is null!");
  27. return;
  28. }
  29. auto stream_list = model_context.stream_list();
  30. auto label_list = model_context.label_list();
  31. rt_model_handle_ = model_context.rt_model_handle();
  32. uint32_t stream_id = task_info->stream_id();
  33. label_id_ = task_info->label_id();
  34. GELOGI("Stream list size:%zu, stream id:%u.", stream_list.size(), stream_id);
  35. GELOGI("Label list size:%zu, label id:%u.", label_list.size(), label_id_);
  36. if (stream_id >= stream_list.size() || label_id_ >= label_list.size()) {
  37. GELOGW("Stream/Label id invalid.");
  38. return;
  39. }
  40. stream_ = stream_list[stream_id];
  41. label_manager_ = LabelManager::GetInstance();
  42. if (label_manager_ == nullptr) {
  43. GELOGW("Get label manager instance failed.");
  44. return;
  45. }
  46. label_info_ = label_manager_->GetLabelInfo(rt_model_handle_, {label_id_}, label_list);
  47. }
  48. LabelGotoTask::~LabelGotoTask() {
  49. if (index_value_ != nullptr) {
  50. rtError_t rt_ret = rtFree(index_value_);
  51. if (rt_ret != RT_ERROR_NONE) {
  52. GELOGE(RT_FAILED, "rtFree index_value_ failed! ret: 0x%X.", rt_ret);
  53. }
  54. index_value_ = nullptr;
  55. }
  56. }
  57. bool LabelGotoTask::Distribute() {
  58. GELOGI("LabelGotoTask Distribute start.");
  59. if (stream_ == nullptr) {
  60. GELOGE(PARAM_INVALID, "stream is null!");
  61. return false;
  62. }
  63. if (label_info_ == nullptr) {
  64. GELOGE(PARAM_INVALID, "label info is null!");
  65. return false;
  66. }
  67. if (index_value_ == nullptr) {
  68. rtError_t rt_ret = rtMalloc(&index_value_, sizeof(uint64_t), RT_MEMORY_HBM);
  69. if (rt_ret != RT_ERROR_NONE) {
  70. GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
  71. return false;
  72. }
  73. uint64_t index = 0;
  74. rt_ret = rtMemcpy(index_value_, sizeof(uint64_t), &index, sizeof(index), RT_MEMCPY_HOST_TO_DEVICE);
  75. if (rt_ret != RT_ERROR_NONE) {
  76. GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
  77. return false;
  78. }
  79. }
  80. void *label_info = label_info_->GetLabelInfo();
  81. rtError_t rt_ret = rtLabelSwitchByIndex(index_value_, 1, label_info, stream_);
  82. if (rt_ret != RT_ERROR_NONE) {
  83. GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
  84. return false;
  85. }
  86. GELOGI("DistributeTask end.");
  87. return true;
  88. }
  89. REGISTER_TASK(TaskInfoType::LABEL_GOTO, LabelGotoTask, LabelGotoTaskInfo);
  90. } // namespace model_runner
  91. } // 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