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.

kernel_store.cc 4.2 kB

4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118
  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 "common/kernel_store.h"
  17. namespace ge {
  18. void KernelStore::AddKernel(const KernelBinPtr &kernel) {
  19. if (kernel != nullptr) {
  20. kernels_[kernel->GetName()] = kernel;
  21. }
  22. }
  23. bool KernelStore::Build() {
  24. buffer_.clear();
  25. size_t total_len = 0;
  26. for (const auto &item : kernels_) {
  27. auto kernel = item.second;
  28. total_len += sizeof(KernelStoreItemHead);
  29. total_len += kernel->GetName().length();
  30. total_len += kernel->GetBinDataSize();
  31. }
  32. try {
  33. buffer_.resize(total_len);
  34. } catch (std::bad_alloc &e) {
  35. GELOGE(ge::MEMALLOC_FAILED, "All build memory failed, memory size %zu", total_len);
  36. return false;
  37. }
  38. uint8_t *next_buffer = buffer_.data();
  39. size_t remain_len = total_len;
  40. errno_t mem_ret;
  41. for (const auto &item : kernels_) {
  42. auto kernel = item.second;
  43. KernelStoreItemHead kernel_head{};
  44. kernel_head.magic = kKernelItemMagic;
  45. kernel_head.name_len = static_cast<uint32_t>(kernel->GetName().length());
  46. kernel_head.bin_len = static_cast<uint32_t>(kernel->GetBinDataSize());
  47. GELOGI("get kernel bin name %s, addr %p, size %u",
  48. kernel->GetName().c_str(), kernel->GetBinData(), kernel->GetBinDataSize());
  49. mem_ret = memcpy_s(next_buffer, remain_len, &kernel_head, sizeof(kernel_head));
  50. GE_CHK_BOOL_EXEC_NOLOG(mem_ret == EOK, return false);
  51. next_buffer += sizeof(kernel_head);
  52. mem_ret = memcpy_s(next_buffer, remain_len - sizeof(kernel_head), kernel->GetName().data(), kernel_head.name_len);
  53. GE_CHK_BOOL_EXEC_NOLOG(mem_ret == EOK, return false);
  54. next_buffer += kernel_head.name_len;
  55. mem_ret = memcpy_s(next_buffer, remain_len - sizeof(kernel_head) - kernel_head.name_len, kernel->GetBinData(),
  56. kernel_head.bin_len);
  57. GE_CHK_BOOL_EXEC_NOLOG(mem_ret == EOK, return false);
  58. next_buffer += kernel_head.bin_len;
  59. remain_len = remain_len - sizeof(kernel_head) - kernel_head.name_len - kernel_head.bin_len;
  60. }
  61. kernels_.clear();
  62. return true;
  63. }
  64. const uint8_t *KernelStore::Data() const { return buffer_.data(); }
  65. size_t KernelStore::DataSize() const { return buffer_.size(); }
  66. bool KernelStore::Load(const uint8_t *data, const size_t &len) {
  67. if (data == nullptr || len == 0) {
  68. return false;
  69. }
  70. size_t buffer_len = len;
  71. while (buffer_len > sizeof(KernelStoreItemHead)) {
  72. const char *next_buffer = reinterpret_cast<const char *>(data) + (len - buffer_len);
  73. const auto *kernel_head = reinterpret_cast<const KernelStoreItemHead *>(next_buffer);
  74. if (buffer_len < kernel_head->name_len + kernel_head->bin_len + sizeof(KernelStoreItemHead)) {
  75. GELOGW("Invalid kernel block remain buffer len %zu, name len %u, bin len %u", buffer_len, kernel_head->name_len,
  76. kernel_head->bin_len);
  77. break;
  78. }
  79. next_buffer += sizeof(KernelStoreItemHead);
  80. std::string name(next_buffer, kernel_head->name_len);
  81. next_buffer += kernel_head->name_len;
  82. GELOGI("Load kernel from om:%s,%u,%u", name.c_str(), kernel_head->name_len, kernel_head->bin_len);
  83. std::vector<char> kernel_bin(next_buffer, next_buffer + kernel_head->bin_len);
  84. KernelBinPtr teb_kernel_ptr = ge::MakeShared<KernelBin>(name, std::move(kernel_bin));
  85. if (teb_kernel_ptr != nullptr) {
  86. kernels_.emplace(name, teb_kernel_ptr);
  87. }
  88. buffer_len -= sizeof(KernelStoreItemHead) + kernel_head->name_len + kernel_head->bin_len;
  89. }
  90. return true;
  91. }
  92. KernelBinPtr KernelStore::FindKernel(const std::string &name) const {
  93. auto it = kernels_.find(name);
  94. if (it != kernels_.end()) {
  95. return it->second;
  96. }
  97. return nullptr;
  98. }
  99. } // 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