* Copyright (c) 2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#include "stable_part_provider.h"
namespace ge {
Status ResolveStablePart(StablePartId id, std::string &output) {
switch (id) {
case StablePartId::kChkStatusMacro:
output =
"#define OM2_CHK_STATUS(expr, ...) \\\n"
"do { \\\n"
" const aclError _chk_status = (expr); \\\n"
" if (_chk_status != ACL_SUCCESS) { \\\n"
" return _chk_status; \\\n"
" } \\\n"
"} while (false)\n";
return SUCCESS;
case StablePartId::kChkNotNullMacro:
output =
"#define OM2_CHK_NOTNULL(ptr, ...) \\\n"
"do { \\\n"
" if ((ptr) == nullptr) { \\\n"
" return ACL_ERROR_FAILURE; \\\n"
" } \\\n"
"} while (false)\n";
return SUCCESS;
case StablePartId::kChkTrueMacro:
output =
"#define OM2_CHK_TRUE(expr, ...) \\\n"
"do { \\\n"
" if (!(expr)) { \\\n"
" return ACL_ERROR_FAILURE; \\\n"
" } \\\n"
"} while (false)\n";
return SUCCESS;
case StablePartId::kGetAddrMacro:
output =
"#define GET_ADDR(mem_ptr, offset) \\\n"
"(reinterpret_cast<void *>( \\\n"
" reinterpret_cast<uintptr_t>(mem_ptr) + \\\n"
" static_cast<uintptr_t>(offset)))\n";
return SUCCESS;
case StablePartId::kMakeGuardMacro:
output = "#define OM2_MAKE_GUARD(var, callback) const ::om2::ScopeGuard const_guard_##var(callback)\n";
return SUCCESS;
case StablePartId::kInterfaceMacros:
output =
"#define OM2_CHK_STATUS(expr, ...) \\\n"
"do { \\\n"
" const aclError _chk_status = (expr); \\\n"
" if (_chk_status != ACL_SUCCESS) { \\\n"
" return _chk_status; \\\n"
" } \\\n"
"} while (false)\n"
"\n"
"#define OM2_CHK_NOTNULL(ptr, ...) \\\n"
"do { \\\n"
" if ((ptr) == nullptr) { \\\n"
" return ACL_ERROR_FAILURE; \\\n"
" } \\\n"
"} while (false)\n"
"\n"
"#define OM2_CHK_TRUE(expr, ...) \\\n"
"do { \\\n"
" if (!(expr)) { \\\n"
" return ACL_ERROR_FAILURE; \\\n"
" } \\\n"
"} while (false)\n"
"\n"
"#define OM2_CHK_RT(expr, ...) \\\n"
"do { \\\n"
" const rtError_t _rt_err = (expr); \\\n"
" if (_rt_err != RT_ERROR_NONE) { \\\n"
" return ACL_ERROR_FAILURE; \\\n"
" } \\\n"
"} while (false)\n"
"\n"
"#define GET_ADDR(mem_ptr, offset) \\\n"
"(reinterpret_cast<void *>( \\\n"
" reinterpret_cast<uintptr_t>(mem_ptr) + \\\n"
" static_cast<uintptr_t>(offset)))\n"
"\n"
"#define OM2_MAKE_GUARD(var, callback) const ::om2::ScopeGuard const_guard_##var(callback)\n";
return SUCCESS;
case StablePartId::kPointerHelpers:
output =
"template<typename T>\n"
"inline T *PtrAdd(T *const ptr, const size_t max_buf_len, const size_t idx) {\n"
" if ((ptr != nullptr) && (idx < max_buf_len)) {\n"
" return reinterpret_cast<T *>(ptr + idx);\n"
" }\n"
" return nullptr;\n"
"}\n"
"template<typename TI, typename TO>\n"
"inline TO *PtrToPtr(TI *const ptr) {\n"
" return reinterpret_cast<TO *>(ptr);\n"
"}\n"
"inline uint64_t PtrToValue(const void *const ptr) {\n"
" return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(ptr));\n"
"}\n"
"inline void *ValueToPtr(const uint64_t value) {\n"
" return reinterpret_cast<void *>(static_cast<uintptr_t>(value));\n"
"}\n"
"\n";
return SUCCESS;
case StablePartId::kInterfacePointerHelpers:
output =
"template<typename T>\n"
"inline T *PtrAdd(T *const ptr, const size_t max_buf_len, const size_t idx) {\n"
" if ((ptr != nullptr) && (idx < max_buf_len)) {\n"
" return reinterpret_cast<T *>(ptr + idx);\n"
" }\n"
" return nullptr;\n"
"}\n"
"template<typename TI, typename TO>\n"
"inline TO *PtrToPtr(TI *const ptr) {\n"
" return reinterpret_cast<TO *>(ptr);\n"
"}\n"
"inline uint64_t PtrToValue(const void *const ptr) {\n"
" return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(ptr));\n"
"}\n"
"inline void *ValueToPtr(const uint64_t value) {\n"
" return reinterpret_cast<void *>(static_cast<uintptr_t>(value));\n"
"}\n"
"inline uint64_t PtrToU64(const void *ptr) {\n"
" return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(ptr));\n"
"}\n"
"\n"
"inline void *ResolveOpAddr(uint32_t mem_src, uint64_t offset,\n"
" void *total_dev_mem_ptr, void *session_scope_mem_ptr,\n"
" void **constants) {\n"
" void *base_ptr;\n"
" if (mem_src == 0xFFFFFFFFU) {\n"
" base_ptr = session_scope_mem_ptr;\n"
" } else if (mem_src == 0U) {\n"
" base_ptr = total_dev_mem_ptr;\n"
" } else {\n"
" base_ptr = constants[mem_src - 1U];\n"
" }\n"
" return GET_ADDR(base_ptr, offset);\n"
"}\n"
"\n"
"extern \"C\" {\n"
"struct Om2Tensor {\n"
" uint64_t device_address;\n"
" uint64_t size;\n"
" int32_t data_type;\n"
" int32_t format;\n"
" const int64_t* shape_dims;\n"
" uint64_t shape_dims_num;\n"
"};\n"
"}\n"
"\n"
"template<typename... Args>\n"
"inline std::vector<uint64_t> FlattenHostArgs(Args&&... args) {\n"
" std::vector<uint64_t> buf;\n"
" auto append_arg = [&](auto&& arg) {\n"
" using T = std::decay_t<decltype(arg)>;\n"
" if constexpr (std::is_pointer_v<T>) {\n"
" buf.push_back(reinterpret_cast<uint64_t>(arg));\n"
" } else if constexpr (std::is_same_v<T, Om2Tensor>) {\n"
" buf.push_back(arg.device_address);\n"
" } else if constexpr (std::is_same_v<T, std::vector<int64_t>>) {\n"
" for (auto d : arg) buf.push_back(static_cast<uint64_t>(d));\n"
" } else if constexpr (std::is_integral_v<T>) {\n"
" buf.push_back(static_cast<uint64_t>(arg));\n"
" } else {\n"
" static_assert(sizeof(T) == 0, \"Unsupported type in FlattenHostArgs\");\n"
" }\n"
" };\n"
" (append_arg(std::forward<Args>(args)), ...);\n"
" return buf;\n"
"}\n";
return SUCCESS;
case StablePartId::kFlattenHostArgs:
output =
"extern \"C\" {\n"
"struct Om2Tensor {\n"
" uint64_t device_address;\n"
" uint64_t size;\n"
" int32_t data_type;\n"
" int32_t format;\n"
" const int64_t* shape_dims;\n"
" uint64_t shape_dims_num;\n"
"};\n"
"}\n"
"\n"
"template<typename... Args>\n"
"inline std::vector<uint64_t> FlattenHostArgs(Args&&... args) {\n"
" std::vector<uint64_t> buf;\n"
" auto append_arg = [&](auto&& arg) {\n"
" using T = std::decay_t<decltype(arg)>;\n"
" if constexpr (std::is_pointer_v<T>) {\n"
" buf.push_back(reinterpret_cast<uint64_t>(arg));\n"
" } else if constexpr (std::is_same_v<T, Om2Tensor>) {\n"
" buf.push_back(arg.device_address);\n"
" } else if constexpr (std::is_same_v<T, std::vector<int64_t>>) {\n"
" for (auto d : arg) buf.push_back(static_cast<uint64_t>(d));\n"
" } else if constexpr (std::is_integral_v<T>) {\n"
" buf.push_back(static_cast<uint64_t>(arg));\n"
" } else {\n"
" static_assert(sizeof(T) == 0, \"Unsupported type in FlattenHostArgs\");\n"
" }\n"
" };\n"
" (append_arg(std::forward<Args>(args)), ...);\n"
" return buf;\n"
"}\n";
return SUCCESS;
case StablePartId::kScopeGuard:
output =
"class ScopeGuard {\n"
" public:\n"
" ScopeGuard(const ScopeGuard &) = delete;\n"
" ScopeGuard &operator=(const ScopeGuard &) = delete;\n"
"\n"
" explicit ScopeGuard(const std::function<void()> &on_exit_scope)\n"
" : on_exit_scope_(on_exit_scope) {}\n"
"\n"
" ~ScopeGuard() {\n"
" if (on_exit_scope_) {\n"
" try {\n"
" on_exit_scope_();\n"
" } catch (std::bad_function_call &) {\n"
" } catch (...) {\n"
" }\n"
" }\n"
" }\n"
"\n"
" private:\n"
" std::function<void()> on_exit_scope_;\n"
"};\n";
return SUCCESS;
case StablePartId::kReadBinaryFileToBuffer:
output =
"BinaryBuffer ReadBinaryFileToBuffer(const std::string &file_path) {\n"
" BinaryBuffer result;\n"
" std::ifstream file(file_path, std::ios::binary | std::ios::ate);\n"
" if (!file.is_open()) {\n"
" return result;\n"
" }\n"
" std::streamsize file_size = file.tellg();\n"
" if (file_size <= 0) {\n"
" return result;\n"
" }\n"
" result.size = static_cast<size_t>(file_size);\n"
" result.data = std::make_unique<uint8_t[]>(result.size);\n"
" file.seekg(0, std::ios::beg);\n"
" file.read(reinterpret_cast<char *>(result.data.get()), result.size);\n"
" if (!file.good()) {\n"
" file.close();\n"
" result.data.reset();\n"
" result.size = 0;\n"
" }\n"
" return result;\n"
"}\n";
return SUCCESS;
case StablePartId::kGenerateJsonFile:
output =
"aclError GenerateJsonFile(const AicpuRegisterInfo ®ister_info, std::string &json_path) {\n"
" using namespace std::chrono;\n"
" int64_t cur_timestamp = duration_cast<milliseconds>(steady_clock::now().time_since_epoch()).count();\n"
" json_path = \"/tmp/temp_ops_info_\" + std::to_string(cur_timestamp) + \".json\";\n"
" std::string json_data_format = R\"(\n"
"{\n"
" \"%s\":{\n"
" \"opInfo\":{\n"
" \"opKernelLib\":\"%s\",\n"
" \"kernelSo\":\"%s\",\n"
" \"functionName\":\"%s\"\n"
" }\n"
" }\n"
"}\n"
")\";\n"
" char json_data[kMaxJsonFileLen];\n"
" auto ret = snprintf_s(json_data, kMaxJsonFileLen, kMaxJsonFileLen - 1U, json_data_format.c_str(),\n"
" register_info.op_type, register_info.op_kernel_lib, register_info.so_name,\n"
" register_info.kernel_name);\n"
" OM2_CHK_TRUE(ret >= 0);\n"
" std::ofstream ofs(json_path.c_str(), std::ios::trunc);\n"
" OM2_CHK_TRUE(ofs);\n"
" ofs << json_data;\n"
" return ACL_SUCCESS;\n"
"}\n";
return SUCCESS;
case StablePartId::kLoadAndRunExternalApis:
output =
"aclError Om2ModelCreate(om2::Om2ModelHandle *model_handle, aclmdlRI *rt_model_handle, "
"const char **bin_files, const void **bin_data, size_t *bin_size, int bin_num, void **constants, "
"void *work_ptr, uint64_t *session_id, uint32_t model_id, void *instance_handle) {\n"
" if ((model_handle == nullptr) || (rt_model_handle == nullptr) || (*model_handle != nullptr)) {\n"
" return ACL_ERROR_FAILURE;\n"
" }\n"
" auto *obj = new om2::Om2Model(bin_files, bin_data, bin_size, bin_num, constants, work_ptr, session_id,\n"
" model_id, instance_handle);\n"
" if (obj == nullptr) {\n"
" return ACL_ERROR_FAILURE;\n"
" }\n"
" auto ret = obj->InitResources();\n"
" if (ret != ACL_SUCCESS) {\n"
" delete obj;\n"
" return ret;\n"
" }\n"
" ret = obj->RegisterKernels();\n"
" if (ret != ACL_SUCCESS) {\n"
" delete obj;\n"
" return ret;\n"
" }\n"
" *model_handle = reinterpret_cast<om2::Om2ModelHandle>(obj);\n"
" *rt_model_handle = obj->GetRtModelHandle();\n"
" return ACL_SUCCESS;\n"
"}\n"
"\n"
"aclError Om2ModelLoad(om2::Om2ModelHandle *model_handle) {\n"
" if ((model_handle == nullptr) || (*model_handle == nullptr)) {\n"
" return ACL_ERROR_FAILURE;\n"
" }\n"
" return static_cast<om2::Om2Model*>(*model_handle)->Load();\n"
"}\n"
"\n"
"aclError Om2ModelRunAsync(om2::Om2ModelHandle *model_handle, aclrtStream stream, int input_count, "
"void **input_data, int output_count, void **output_data) {\n"
" return static_cast<om2::Om2Model*>(*model_handle)->RunAsync(stream, input_count, input_data, output_count, "
"output_data);\n"
"}\n"
"\n"
"aclError Om2ModelRun(om2::Om2ModelHandle *model_handle, int input_count, void **input_data, int output_count, "
"void **output_data, int32_t stream_sync_timeout) {\n"
" return static_cast<om2::Om2Model*>(*model_handle)->Run(input_count, input_data, output_count, output_data, stream_sync_timeout);\n"
"}\n"
"\n"
"aclError Om2ModelDestroy(om2::Om2ModelHandle *model_handle) {\n"
" delete static_cast<om2::Om2Model*>(*model_handle);\n"
" return ACL_SUCCESS;\n"
"}\n";
return SUCCESS;
case StablePartId::kInterfaceDumpApis:
output =
"struct Om2TaskIoEntry {\n"
" const struct Om2Tensor *tensor;\n"
" uint64_t offset;\n"
"};\n"
"\n"
"enum Om2L0ArgKind {\n"
" OM2_L0_ARG_INPUT = 0,\n"
" OM2_L0_ARG_OUTPUT = 1,\n"
" OM2_L0_ARG_WORKSPACE = 2,\n"
" OM2_L0_ARG_TILING = 3,\n"
" OM2_L0_ARG_SHAPE_INFO = 4,\n"
" OM2_L0_ARG_LEVEL1_DESC = 5,\n"
" OM2_L0_ARG_PLACEHOLDER = 6,\n"
" OM2_L0_ARG_CUSTOM_VALUE = 7,\n"
" OM2_L0_ARG_FFTS_ADDR = 8,\n"
" OM2_L0_ARG_EVENT_ADDR = 9,\n"
" OM2_L0_ARG_OVERFLOW_ADDR = 10,\n"
" OM2_L0_ARG_EMPTY_ADDR = 11\n"
"};\n"
"\n"
"struct Om2L0ArgSlotInfo {\n"
" uint32_t kind;\n"
" uint32_t flags;\n"
" uint64_t args_offset;\n"
" uint64_t value;\n"
" uint32_t related_index;\n"
" uint32_t event_id;\n"
" uint64_t level1_target_offset;\n"
"};\n"
"\n"
"struct Om2L0TaskRawInfo {\n"
" uint32_t version;\n"
" uint32_t need_assert_or_printf;\n"
" uint64_t arg_num;\n"
" const struct Om2L0ArgSlotInfo* args;\n"
"};\n"
"\n"
"struct Om2TaskInfo {\n"
" const char* op_name;\n"
" const char* op_type;\n"
" uint32_t task_id;\n"
" uint32_t stream_id;\n"
" uint32_t context_id;\n"
" uint32_t thread_id;\n"
" uint32_t block_dim;\n"
" uint64_t op_desc_id;\n"
" uintptr_t args_base;\n"
" uint64_t args_size;\n"
" uint64_t input_num;\n"
" const struct Om2TaskIoEntry* inputs;\n"
" uint32_t output_num;\n"
" const struct Om2TaskIoEntry* outputs;\n"
" uint32_t workspace_num;\n"
" const uint64_t* workspace_addrs;\n"
" const uint64_t* workspace_sizes;\n"
" uint32_t task_type;\n"
" void* stream;\n"
" uint32_t is_raw_address;\n"
" const struct Om2L0TaskRawInfo* l0_exception_dump_info;\n"
"};\n"
"\n"
"extern \"C\" {\n"
"__attribute__((weak)) int32_t ReportDfxTaskPreprocess(uint32_t model_id,\n"
" void* instance_handle,\n"
" const struct Om2TaskInfo* task_info,\n"
" const void* extended_attrs,\n"
" size_t extended_attrs_size);\n"
"\n"
"__attribute__((weak)) int32_t ReportDfxTaskPostprocess(uint32_t model_id,\n"
" void* instance_handle,\n"
" const struct Om2TaskInfo* task_info,\n"
" const void* extended_attrs,\n"
" size_t extended_attrs_size);\n"
"\n"
"__attribute__((weak)) int32_t IsDataDumpEnabled(uint32_t model_id,\n"
" void* instance_handle,\n"
" const char* op_name,\n"
" uint8_t* is_data_dump);\n"
"}\n";
return SUCCESS;
case StablePartId::kLoadAndRunDumpHelpers:
output =
"Om2Tensor BuildOm2Tensor(void *device_address, uint64_t size, int32_t data_type,\n"
" int32_t format, const int64_t *shape_dims, uint64_t shape_dims_num) {\n"
" Om2Tensor tensor{};\n"
" tensor.device_address = PtrToU64(device_address);\n"
" tensor.size = size;\n"
" tensor.data_type = data_type;\n"
" tensor.format = format;\n"
" tensor.shape_dims = shape_dims;\n"
" tensor.shape_dims_num = shape_dims_num;\n"
" return tensor;\n"
"}\n"
"\n"
"aclError AssembleOm2TaskInfo(Om2TaskInfo *task_info, const char *op_name, const char *op_type,\n"
" uint32_t task_id, uint32_t stream_id, uint32_t block_dim,\n"
" uint64_t op_desc_id, uintptr_t args_base, uint64_t args_size,\n"
" const Om2TaskIoEntry *inputs, uint64_t input_num,\n"
" const Om2TaskIoEntry *outputs, uint32_t output_num,\n"
" const uint64_t *workspace_addrs, const uint64_t *workspace_sizes,\n"
" uint32_t workspace_num, uint32_t task_type, void *stream,\n"
" uint32_t is_raw_address = 0U) {\n"
" task_info->op_name = op_name;\n"
" task_info->op_type = op_type;\n"
" task_info->task_id = task_id;\n"
" task_info->stream_id = stream_id;\n"
" task_info->context_id = 0U;\n"
" task_info->thread_id = 0U;\n"
" task_info->block_dim = block_dim;\n"
" task_info->op_desc_id = op_desc_id;\n"
" task_info->args_base = args_base;\n"
" task_info->args_size = args_size;\n"
" task_info->input_num = input_num;\n"
" task_info->inputs = inputs;\n"
" task_info->output_num = output_num;\n"
" task_info->outputs = outputs;\n"
" task_info->workspace_num = workspace_num;\n"
" task_info->workspace_addrs = workspace_addrs;\n"
" task_info->workspace_sizes = workspace_sizes;\n"
" task_info->task_type = task_type;\n"
" task_info->stream = stream;\n"
" task_info->is_raw_address = is_raw_address;\n"
" task_info->l0_exception_dump_info = nullptr;\n"
" return ACL_SUCCESS;\n"
"}\n"
"\n"
"aclError ReportOm2TaskPreprocess(const char *op_name, const char *op_type, uint64_t op_desc_id,\n"
" uintptr_t args_base, uint64_t args_size,\n"
" const std::vector<Om2TaskIoEntry> &inputs,\n"
" const std::vector<Om2TaskIoEntry> &outputs,\n"
" const std::vector<uint64_t> &workspace_addrs,\n"
" const std::vector<uint64_t> &workspace_sizes,\n"
" uint32_t task_type, uint32_t block_dim, void *stream,\n"
" const Om2L0TaskRawInfo *l0_info,\n"
" uint32_t model_id, void *instance_handle,\n"
" uint32_t is_raw_address = 0U) {\n"
" uint32_t stream_id = 0U;\n"
" OM2_CHK_STATUS(aclrtStreamGetId(stream, reinterpret_cast<int32_t *>(&stream_id)));\n"
"\n"
" OM2_CHK_TRUE(workspace_addrs.size() == workspace_sizes.size());\n"
"\n"
" Om2TaskInfo task_info{};\n"
" OM2_CHK_STATUS(AssembleOm2TaskInfo(&task_info, op_name, op_type, 0U, stream_id, block_dim,\n"
" op_desc_id, args_base, args_size,\n"
" inputs.data(), static_cast<uint64_t>(inputs.size()),\n"
" outputs.data(), static_cast<uint32_t>(outputs.size()),\n"
" workspace_addrs.empty() ? nullptr : workspace_addrs.data(),\n"
" workspace_sizes.empty() ? nullptr : workspace_sizes.data(),\n"
" static_cast<uint32_t>(workspace_addrs.size()),\n"
" task_type, stream, is_raw_address));\n"
" task_info.l0_exception_dump_info = l0_info;\n"
" if (ReportDfxTaskPreprocess != nullptr && instance_handle != nullptr) {\n"
" OM2_CHK_STATUS(ReportDfxTaskPreprocess(model_id, instance_handle, &task_info, nullptr, 0U));\n"
" }\n"
" return ACL_SUCCESS;\n"
"}\n"
"\n"
"aclError ReportLaunchedOm2Task(const char *op_name, const char *op_type, uint64_t op_desc_id,\n"
" uintptr_t args_base, uint64_t args_size,\n"
" const Om2TaskIoEntry *inputs, uint64_t input_num,\n"
" const Om2TaskIoEntry *outputs, uint32_t output_num,\n"
" const uint64_t *workspace_addrs, const uint64_t *workspace_sizes,\n"
" uint32_t workspace_num,\n"
" uint32_t task_type, uint32_t block_dim, void *stream,\n"
" uint32_t model_id, void *instance_handle,\n"
" uint32_t is_raw_address = 0U) {\n"
" uint32_t task_id = 0U;\n"
" OM2_CHK_RT(aclrtGetThreadLastTaskId(&task_id));\n"
"\n"
" uint32_t stream_id = 0U;\n"
" OM2_CHK_STATUS(aclrtStreamGetId(stream, reinterpret_cast<int32_t *>(&stream_id)));\n"
"\n"
" Om2TaskInfo task_info{};\n"
" OM2_CHK_STATUS(AssembleOm2TaskInfo(&task_info, op_name, op_type, task_id, stream_id, block_dim,\n"
" op_desc_id, args_base, args_size,\n"
" inputs, input_num,\n"
" outputs, output_num,\n"
" workspace_addrs, workspace_sizes, workspace_num,\n"
" task_type, stream, is_raw_address));\n"
" task_info.l0_exception_dump_info = nullptr;\n"
" if (ReportDfxTaskPostprocess != nullptr && instance_handle != nullptr) {\n"
" OM2_CHK_STATUS(ReportDfxTaskPostprocess(model_id, instance_handle, &task_info, nullptr, 0U));\n"
" }\n"
" return ACL_SUCCESS;\n"
"}\n"
"\n"
"uint8_t GetIsDataDump(const char *op_name, uint32_t model_id, void *instance_handle) {\n"
" if (IsDataDumpEnabled != nullptr && instance_handle != nullptr) {\n"
" uint8_t is_data_dump = 0U;\n"
" auto ret = IsDataDumpEnabled(model_id, instance_handle, op_name, &is_data_dump);\n"
" return ret == 0 ? is_data_dump : 0U;\n"
" }\n"
" return 0U;\n"
"}\n";
return SUCCESS;
case StablePartId::kCreateLabelListForLabelSwitch:
output =
"aclError Om2Model::CreateLabelListForLabelSwitch(uint32_t op_index, std::vector<uint32_t> label_list_indexs)\n"
"{\n"
" label_switch_label_list_[op_index] = nullptr;\n"
" uint32_t label_used_size = label_list_indexs.size();\n"
" label_used_.resize(label_used_size, nullptr);\n"
" for (uint32_t i = 0; i < label_used_size; i++) {\n"
" label_used_[i] = label_list_[label_list_indexs[i]];\n"
" }\n"
" OM2_CHK_STATUS(aclrtCreateLabelList(label_used_.data(), label_used_.size(), &(label_switch_label_list_[op_index])));\n"
" return ACL_SUCCESS;\n"
"}\n";
return SUCCESS;
case StablePartId::kCreateLabelListForLabelGotoEx:
output =
"aclError Om2Model::CreateLabelListForLabelGotoEx(uint32_t op_index, uint32_t label_index)\n"
"{\n"
" label_goto_ex_label_list_[op_index] = nullptr;\n"
" const auto it = label_goto_args_.find(label_index);\n"
" if (it != label_goto_args_.cend()) {\n"
" label_goto_ex_label_list_[op_index] = it->second.first;\n"
" } else {\n"
" label_used_.resize(1, nullptr);\n"
" label_used_[0] = label_list_[static_cast<size_t>(label_index)];\n"
" OM2_CHK_STATUS(aclrtCreateLabelList(label_used_.data(), label_used_.size(), &(label_goto_ex_label_list_[op_index])));\n"
" label_goto_args_[label_index] = {label_goto_ex_label_list_[op_index], static_cast<uint32_t>(sizeof(rtLabelDevInfo))};\n"
" }\n"
" return ACL_SUCCESS;\n"
"}\n";
return SUCCESS;
case StablePartId::kOpDefStructs:
output =
"// 算子分发类型枚举\n"
"enum OpDispatchType : uint32_t {\n"
" DISPATCH_AICORE = 0, // AI Core / AICPU 算子执行\n"
"};\n"
"\n"
"// 算子参数类型枚举\n"
"enum OpArgType : int32_t {\n"
" OP_ARG_INPUT = 0, // 输入张量\n"
" OP_ARG_OUTPUT = 1, // 输出张量\n"
" OP_ARG_WORKSPACE = 2, // 工作空间\n"
" OP_ARG_CONST_TENSOR = 3, // 常量张量\n"
" OP_ARG_LEVEL1_DESC = 4, // Level1 描述符\n"
" OP_ARG_SHAPE_INFO = 5, // Shape 信息\n"
" OP_ARG_CUSTOM_VALUE = 6, // 自定义值\n"
" OP_ARG_PLACEHOLDER = 7, // 占位符\n"
" OP_ARG_OPTIONAL_EMPTY = 8, // 可选空参数\n"
" OP_ARG_FFTS_ADDR = 9, // FFTS 地址\n"
" OP_ARG_EVENT_ADDR = 10, // 事件地址\n"
" OP_ARG_OVERFLOW_ADDR = 11, // 溢出地址\n"
" OP_ARG_TILING = 12, // Tiling 数据\n"
" OP_ARG_RAW_ADDR = 13, // 原始地址\n"
"};\n"
"\n"
"// 算子参数信息结构体\n"
"struct OpArgInfo {\n"
" int32_t type; // 参数类型（OpArgType，switch 条件）\n"
" struct { // 地址解析（INPUT/OUTPUT/WORKSPACE/CONST_TENSOR）\n"
" uint32_t mem_src; // 内存来源（0=设备内存，0xFFFFFFFF=session，≥1=常量数组索引）\n"
" uint64_t offset; // 内存偏移量\n"
" } addr;\n"
" union {\n"
" struct { // Tensor 元数据（INPUT/OUTPUT）\n"
" uint64_t size; // 数据大小（字节），WORKSPACE 也用此字段\n"
" int32_t data_type; // 数据类型\n"
" int32_t format; // 数据格式\n"
" int64_t shape[8]; // Shape 维度数组（最多8维）\n"
" uint32_t num_shape_dims; // 实际 shape 维度数\n"
" uint64_t args_offset; // args table 内字节偏移\n"
" } tensor;\n"
" uint64_t custom_value; // 自定义值（LEVEL1_DESC/SHAPE_INFO/CUSTOM_VALUE/EVENT_ADDR）\n"
" struct { // Tiling 数据（TILING）\n"
" const uint8_t *raw_data; // 原始数据指针\n"
" uint32_t raw_data_len; // 原始数据长度\n"
" } tiling;\n"
" } data;\n"
"};\n"
"\n"
"// 算子定义结构体\n"
"struct OpDef {\n"
" OpDispatchType dispatch_type; // 分发类型，决定走哪个 case 分支\n"
" const char *op_name; // 算子名称，用于日志和 Report 上报\n"
" const OpArgInfo *argsInfo; // 参数信息数组，地址解析时遍历\n"
" // 各 task 专属数据下沉到 union（后续适配其他 task 时再增加对应成员）\n"
" union {\n"
" struct { // AICORE 专属 (kernel_type=0)\n"
" const char *op_type; // 算子类型名，用于 Report 上报\n"
" uint32_t num_io_addrs; // IO 地址数量，控制地址解析循环次数\n"
" uint32_t args_idx; // 参数表索引，用于 GetArgsInfo 查找\n"
" struct { // Launch 配置，构建 LaunchKernelConfig → AssembleLaunchConfig\n"
" uint8_t schedule_mode; // 调度模式\n"
" uint32_t engine_type; // 引擎类型\n"
" uint32_t block_dim_offset;// BlockDim 偏移量\n"
" bool is_block_task_prefetch; // 是否预取 Block Task\n"
" uint16_t time_out; // 超时时间\n"
" uint32_t local_memory_size; // 本地内存大小\n"
" } launch;\n"
" struct { // Kernel 执行参数，传给 KernelTaskDistribute\n"
" uint32_t block_dim; // Block 维度\n"
" uint32_t func_idx; // 函数句柄索引，用于查找 func_handles\n"
" } kernel;\n"
" struct { // L0 信息，构建 Om2L0TaskRawInfo\n"
" uint32_t need_assert_or_printf; // 是否需要 assert/printf\n"
" uint32_t num_l0_slots; // L0 slot 数量\n"
" const Om2L0ArgSlotInfo *l0_slots; // L0 slot 信息数组\n"
" } l0;\n"
" } aicore;\n"
" struct { // AICPU 专属（共享 DISPATCH_AICORE, kernel_type=1）\n"
" uint32_t kernel_type;\n"
" uint32_t args_idx;\n"
" uint32_t num_io_addrs;\n"
" uint32_t func_idx;\n"
" uint32_t block_dim;\n"
" uint8_t schedule_mode;\n"
" uint32_t engine_type;\n"
" uint32_t block_dim_offset;\n"
" bool is_block_task_prefetch;\n"
" uint16_t time_out;\n"
" uint32_t local_memory_size;\n"
" uint32_t ext_info_len;\n"
" int32_t session_info_offset;\n"
" uint32_t aicpu_task_index;\n"
" uint32_t task_type;\n"
" } aicpu;\n"
" } task_data;\n"
"};\n"
"\n"
"// 算子分发上下文结构体\n"
"struct DispatchOpContext {\n"
" void *total_dev_mem_ptr; // 设备总内存指针\n"
" void *session_scope_mem_ptr; // Session 作用域内存指针\n"
" void **constants; // 常量数组\n"
" Om2ArgsTable &args_table; // 参数表\n"
" aclrtFuncHandle *func_handles; // 函数句柄数组\n"
" aclrtStream stream; // 执行流\n"
" uint32_t model_id; // 模型 ID\n"
" void *instance_handle; // 实例句柄\n"
" std::map<uint32_t, void *> &event_id_mem_map; // 事件 ID 到内存的映射\n"
" std::vector<void *> &dev_dynamic_mem_ptrs; // 设备动态内存指针列表\n"
" void *overflow_addr; // 溢出地址\n"
"};\n"
"\n";
return SUCCESS;
default:
output.clear();
return FAILED;
}
}
}
