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.

single_op_manager.cc 6.3 kB

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
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173
  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/single_op_manager.h"
  17. #include <mutex>
  18. #include <string>
  19. #include "graph/manager/graph_mem_manager.h"
  20. namespace ge {
  21. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY SingleOpManager::~SingleOpManager() {
  22. for (auto &it : stream_resources_) {
  23. delete it.second;
  24. it.second = nullptr;
  25. }
  26. }
  27. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status SingleOpManager::GetOpFromModel(const std::string &model_name,
  28. const ModelData &model_data,
  29. void *stream,
  30. SingleOp **single_op,
  31. const uint64_t model_id) {
  32. GELOGI("GetOpFromModel in. model name = %s, model id = %lu", model_name.c_str(), model_id);
  33. if (single_op == nullptr) {
  34. GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "[Check][Param:single_op] is null.");
  35. REPORT_INNER_ERROR("E10412", "input param single_op is nullptr, check invalid");
  36. return ACL_ERROR_GE_INTERNAL_ERROR;
  37. }
  38. uintptr_t resource_id = 0;
  39. GE_CHK_STATUS_RET(GetResourceId(stream, resource_id));
  40. StreamResource *res = GetResource(resource_id, stream);
  41. if (res == nullptr) {
  42. GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "[Get][Resource] failed.");
  43. REPORT_CALL_ERROR("E19999", "GetOpFromModel fail because GetResource return nullptr.");
  44. return ACL_ERROR_GE_MEMORY_ALLOCATION;
  45. }
  46. SingleOp *op = res->GetOperator(model_id);
  47. if (op != nullptr) {
  48. GELOGD("Got operator from stream cache");
  49. *single_op = op;
  50. return SUCCESS;
  51. }
  52. return res->BuildOperator(model_data, single_op, model_id);
  53. }
  54. FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status SingleOpManager::ReleaseResource(void *stream) {
  55. auto resource_id = reinterpret_cast<uintptr_t>(stream);
  56. GELOGI("ReleaseResource in. resource id = 0x%lx", static_cast<uint64_t>(resource_id));
  57. std::lock_guard<std::mutex> lock(mutex_);
  58. auto it = stream_resources_.find(resource_id);
  59. if (it == stream_resources_.end()) {
  60. MemManager::Instance().CachingInstance(RT_MEMORY_HBM).TryFreeBlocks();
  61. return SUCCESS;
  62. }
  63. delete it->second;
  64. it->second = nullptr;
  65. (void)stream_resources_.erase(it);
  66. MemManager::Instance().CachingInstance(RT_MEMORY_HBM).TryFreeBlocks();
  67. return SUCCESS;
  68. }
  69. StreamResource *SingleOpManager::GetResource(uintptr_t resource_id, rtStream_t stream) {
  70. std::lock_guard<std::mutex> lock(mutex_);
  71. auto it = stream_resources_.find(resource_id);
  72. StreamResource *res = nullptr;
  73. if (it == stream_resources_.end()) {
  74. res = new(std::nothrow) StreamResource(resource_id);
  75. if (res != nullptr) {
  76. if (res->Init() != SUCCESS) {
  77. GELOGE(FAILED, "[Malloc][Memory]Failed to malloc device buffer.");
  78. delete res;
  79. return nullptr;
  80. }
  81. res->SetStream(stream);
  82. stream_resources_.emplace(resource_id, res);
  83. }
  84. } else {
  85. res = it->second;
  86. }
  87. return res;
  88. }
  89. StreamResource *SingleOpManager::TryGetResource(uintptr_t resource_id) {
  90. std::lock_guard<std::mutex> lock(mutex_);
  91. auto it = stream_resources_.find(resource_id);
  92. if (it == stream_resources_.end()) {
  93. return nullptr;
  94. }
  95. return it->second;
  96. }
  97. Status SingleOpManager::GetDynamicOpFromModel(const string &model_name,
  98. const ModelData &model_data,
  99. void *stream,
  100. DynamicSingleOp **single_op,
  101. const uint64_t model_id) {
  102. GELOGI("GetOpFromModel in. model name = %s, model id = %lu", model_name.c_str(), model_id);
  103. if (!tiling_func_registered_) {
  104. RegisterTilingFunc();
  105. }
  106. GE_CHECK_NOTNULL(single_op);
  107. uintptr_t resource_id = 0;
  108. GE_CHK_STATUS_RET(GetResourceId(stream, resource_id));
  109. StreamResource *res = GetResource(resource_id, stream);
  110. if (res == nullptr) {
  111. GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "[Get][Resource] failed.");
  112. REPORT_CALL_ERROR("E19999", "GetDynamicOpFromModel fail because GetResource return nullptr.");
  113. return ACL_ERROR_GE_MEMORY_ALLOCATION;
  114. }
  115. DynamicSingleOp *op = res->GetDynamicOperator(model_id);
  116. if (op != nullptr) {
  117. GELOGD("Got operator from stream cache");
  118. *single_op = op;
  119. return SUCCESS;
  120. }
  121. return res->BuildDynamicOperator(model_data, single_op, model_id);
  122. }
  123. void SingleOpManager::RegisterTilingFunc() {
  124. std::lock_guard<std::mutex> lk(mutex_);
  125. if (tiling_func_registered_) {
  126. return;
  127. }
  128. op_tiling_manager_.LoadSo();
  129. tiling_func_registered_ = true;
  130. }
  131. Status SingleOpManager::GetResourceId(rtStream_t stream, uintptr_t &resource_id) {
  132. // runtime uses NULL to denote a default stream for each device
  133. if (stream == nullptr) {
  134. // get current context
  135. rtContext_t rt_cur_ctx = nullptr;
  136. auto rt_err = rtCtxGetCurrent(&rt_cur_ctx);
  137. if (rt_err != RT_ERROR_NONE) {
  138. GELOGE(rt_err, "[Get][CurrentContext] failed, runtime result is %d", static_cast<int>(rt_err));
  139. REPORT_CALL_ERROR("E19999",
  140. "GetResourceId failed because rtCtxGetCurrent result is %d", static_cast<int>(rt_err));
  141. return RT_ERROR_TO_GE_STATUS(rt_err);
  142. }
  143. // use current context as resource key instead
  144. GELOGI("use context as resource key instead when default stream");
  145. resource_id = reinterpret_cast<uintptr_t>(rt_cur_ctx);
  146. } else {
  147. GELOGI("use stream as resource key instead when create stream");
  148. resource_id = reinterpret_cast<uintptr_t>(stream);
  149. }
  150. return SUCCESS;
  151. }
  152. } // 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