Merged
feat:将cv仓的aicpu算子从merge每个文件夹下的json配置算子信息更改到def文件配置算子信息自动生成json #622
zhaowenruiAdd on Mar 28
feat:将cv仓的aicpu算子从merge每个文件夹下的json配置算子信息更改到def文件配置算子信息自动生成json #622
Merged
Merged fromzhaowenrui666/ops-cvopdefintocann/ops-cvmaster
Total: 14 files changed+389-144
| @@ -379,7 +379,7 @@ macro(add_modules_sources) | |||
| 379 | target_sources(${OPHOST_NAME}_tiling_obj PRIVATE ${OPTILING_SRCS}) | 379 | target_sources(${OPHOST_NAME}_tiling_obj PRIVATE ${OPTILING_SRCS}) |
| 380 | endif() | 380 | endif() |
| 381 | 381 | ||
| 382 | - file(GLOB AICPU_SRCS ${SOURCE_DIR}/*_aicpu*.cpp) | 382 | + file(GLOB AICPU_SRCS ${SOURCE_DIR}/*_aicpu.cpp) |
| 383 | if(AICPU_SRCS) | 383 | if(AICPU_SRCS) |
| 384 | add_aicpu_kernel_modules() | 384 | add_aicpu_kernel_modules() |
| 385 | target_sources(${OPHOST_NAME}_aicpu_obj PRIVATE ${AICPU_SRCS}) | 385 | target_sources(${OPHOST_NAME}_aicpu_obj PRIVATE ${AICPU_SRCS}) |
| @@ -192,3 +192,42 @@ function(gen_aclnn_with_opdef) | |||
| 192 | target_include_directories(opbuild_gen_aclnn_all PRIVATE ${OPAPI_INCLUDE}) | 192 | target_include_directories(opbuild_gen_aclnn_all PRIVATE ${OPAPI_INCLUDE}) |
| 193 | endif() | 193 | endif() |
| 194 | endfunction() | 194 | endfunction() |
| 195 | + | ||
| 196 | +function(opbuild_aicpu_ini) | ||
| 197 | + message(STATUS "Opbuild generating sources") | ||
 | |||
| 198 | + message("ARGN: ${ARGN}") | ||
| 199 | + cmake_parse_arguments(OPBUILD "" "OUT_DIR;PROJECT_NAME;ACCESS_PREFIX;ENABLE_SOURCE" "OPS_SRC" ${ARGN}) | ||
| 200 | + execute_process(COMMAND ${CMAKE_CXX_COMPILER} -g -fPIC -shared -std=c++11 ${OPBUILD_OPS_SRC} -D_GLIBCXX_USE_CXX11_ABI=0 | ||
缩进不对    zhaowenrui Apr 1 Comments： Apr 1 Comments： C++11标准是否还适用，看看库上intf_pub里面的版本 zhaowenrui Apr 1 Comments： Apr 1 Comments： | |||
| 201 | + -I ${ASCEND_CANN_PACKAGE_PATH}/include -I ${CMAKE_CURRENT_SOURCE_DIR}/op_kernel -L ${ASCEND_CANN_PACKAGE_PATH}/lib64 -lexe_graph -lregister -ltiling_api | ||
| 202 | + -o ${OPBUILD_OUT_DIR}/libaicpu_ops.so | ||
| 203 | + RESULT_VARIABLE EXEC_RESULT | ||
| 204 | + OUTPUT_VARIABLE EXEC_INFO | ||
| 205 | + ERROR_VARIABLE EXEC_ERROR | ||
| 206 | + ) | ||
| 207 | + if (${EXEC_RESULT}) | ||
| 208 | + message("build ops lib info: ${EXEC_INFO}") | ||
| 209 | + message("build ops lib error: ${EXEC_ERROR}") | ||
| 210 | + message(FATAL_ERROR "opbuild run failed!") | ||
| 211 | + endif() | ||
| 212 | + set(proj_env "") | ||
| 213 | + set(prefix_env "") | ||
| 214 | + if (NOT "${OPBUILD_PROJECT_NAME}x" STREQUAL "x") | ||
| 215 | + set(proj_env "OPS_PROJECT_NAME=${OPBUILD_PROJECT_NAME}") | ||
| 216 | + endif() | ||
| 217 | + if (NOT "${OPBUILD_ACCESS_PREFIX}x" STREQUAL "x") | ||
| 218 | + set(prefix_env "OPS_DIRECT_ACCESS_PREFIX=${OPBUILD_ACCESS_PREFIX}") | ||
| 219 | + endif() | ||
| 220 | + set(ENV{ENABLE_SOURCE_PACKAGE} ${OPBUILD_ENABLE_SOURCE}) | ||
| 221 | + execute_process(COMMAND ${proj_env} ${prefix_env} ${ASCEND_CANN_PACKAGE_PATH}/toolkit/tools/opbuild/op_build | ||
| 222 | + ${OPBUILD_OUT_DIR}/libaicpu_ops.so ${OPBUILD_OUT_DIR} --aicpu | ||
这个中间的空格是不是有问题，两个？ | |||
| 223 | + RESULT_VARIABLE EXEC_RESULT | ||
| 224 | + OUTPUT_VARIABLE EXEC_INFO | ||
| 225 | + ERROR_VARIABLE EXEC_ERROR | ||
| 226 | + ) | ||
| 227 | + unset(ENV{ENABLE_SOURCE_PACKAGE}) | ||
| 228 | + if (${EXEC_RESULT}) | ||
| 229 | + message("opbuild ops info: ${EXEC_INFO}") | ||
| 230 | + message("opbuild ops error: ${EXEC_ERROR}") | ||
| 231 | + endif() | ||
| 232 | + message(STATUS "Opbuild generating sources - done") | ||
| 233 | +endfunction() | ||
| @@ -380,30 +380,44 @@ function(gen_cust_proto_symbol) | |||
| 380 | ) | 380 | ) |
| 381 | endfunction() | 381 | endfunction() |
| 382 | 382 | ||
| 383 | -function(gen_aicpu_json_symbol enable_built_in) | 383 | +function(gen_aicpu_op_def_symbol) |
| 384 | - get_property(ALL_AICPU_JSON_FILES GLOBAL PROPERTY AICPU_JSON_FILES) | 384 | + message(STATUS "enter gen_aicpu_op_def_symbol") |
| 385 | - if(NOT ALL_AICPU_JSON_FILES) | 385 | + get_property(ALL_AICPU_OP_DEF_FILES GLOBAL PROPERTY AICPU_OP_DEF_FILES) |
| 386 | - message(STATUS "No aicpu json files to merge, skipping.") | 386 | + if(NOT ALL_AICPU_OP_DEF_FILES) |
| 387 | + message(STATUS "No aicpu op def files, skipping.") | ||
| 387 | return() | 388 | return() |
| 388 | endif() | 389 | endif() |
| 390 | + message(STATUS "ALL_AICPU_OP_DEF_FILES:${ALL_AICPU_OP_DEF_FILES}") | ||
| 391 | + # 调用opbuild命令生成固定文件 aicpu_kernel.ini | ||
| 392 | + opbuild_aicpu_ini(OPS_SRC ${ALL_AICPU_OP_DEF_FILES} | ||
| 393 | + OUT_DIR ${ASCEND_AUTOGEN_PATH} | ||
| 394 | + ) | ||
| 395 | +endfunction() | ||
| 389 | 396 | ||
| 397 | +function(gen_aicpu_json_symbol enable_built_in) | ||
| 398 | + gen_aicpu_op_def_symbol() | ||
| 390 | set(MERGED_JSON ${CMAKE_BINARY_DIR}/cust_aicpu_kernel.json) | 399 | set(MERGED_JSON ${CMAKE_BINARY_DIR}/cust_aicpu_kernel.json) |
| 391 | if(enable_built_in) | 400 | if(enable_built_in) |
| 392 | set(MERGED_JSON ${CMAKE_BINARY_DIR}/aicpu_cv.json) | 401 | set(MERGED_JSON ${CMAKE_BINARY_DIR}/aicpu_cv.json) |
| 393 | endif() | 402 | endif() |
| 394 | 403 | ||
| 395 | - add_custom_command( | 404 | + set(cmd "python3 ${CMAKE_SOURCE_DIR}/scripts/kernel/binary_script/parser_ini.py ${ASCEND_AUTOGEN_PATH}/aicpu_kernel.ini ${MERGED_JSON}") |
| 396 | - OUTPUT ${MERGED_JSON} | 405 | + message("cmd: ${cmd}") |
| 397 | - COMMAND bash ${CMAKE_SOURCE_DIR}/scripts/util/merge_aicpu_info_json.sh ${CMAKE_SOURCE_DIR} ${MERGED_JSON} ${ALL_AICPU_JSON_FILES} | 406 | + |
| 398 | - DEPENDS ${ALL_AICPU_JSON_FILES} | 407 | + execute_process( |
| 399 | - COMMENT "Merging Json files into ${MERGED_JSON}" | 408 | + COMMAND bash -c "${cmd}" |
| 400 | - VERBATIM | 409 | + WORKING_DIRECTORY ${ASCEND_AUTOGEN_PATH} |
| 410 | + OUTPUT_VARIABLE result | ||
| 411 | + ERROR_VARIABLE error | ||
| 412 | + RESULT_VARIABLE code | ||
| 413 | + OUTPUT_STRIP_TRAILING_WHITESPACE | ||
| 401 | ) | 414 | ) |
| 402 | - add_custom_target(merge_aicpu_json ALL DEPENDS ${MERGED_JSON}) | 415 | + message("convert ini to json finish.") |
| 416 | + | ||
| 403 | install( | 417 | install( |
| 404 | - FILES ${MERGED_JSON} | 418 | + FILES ${MERGED_JSON} |
| 405 | - DESTINATION ${AICPU_JSON_CONFIG} | 419 | + DESTINATION ${AICPU_JSON_CONFIG} |
| 406 | - OPTIONAL | 420 | + OPTIONAL |
| 407 | ) | 421 | ) |
| 408 | endfunction() | 422 | endfunction() |
| 409 | 423 | ||
| @@ -0,0 +1,46 @@ | |||
| 1 | +/** | ||
| 2 | + * Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 3 | + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 4 | + * CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 5 | + * Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 6 | + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 7 | + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 8 | + * See LICENSE in the root of the software repository for the full text of the License. | ||
| 9 | + */ | ||
| 10 | + | ||
| 11 | +/*! | ||
| 12 | + * \file aicpu_op_def.h | ||
| 13 | + * \brief | ||
| 14 | + */ | ||
| 15 | +#ifndef CV_COMMON_AICPU_OP_DEF_H | ||
| 16 | +#define CV_COMMON_AICPU_OP_DEF_H | ||
| 17 | +namespace ops { | ||
| 18 | + const std::string CV_AICPU_KERNEL_SO = "libcv_aicpu_kernels.so"; | ||
| 19 | + const std::string OP_INFO_WORKSPACESIZE = "opInfo.workspaceSize"; | ||
| 20 | + const std::string DEFAULT_WORKSPACE_SIZE = "100"; // 记录AICPU算子日志信息内存大小，单位 KB | ||
| 21 | + const std::string OP_INFO_FORMAT_AGNOSTIC = "opInfo.formatAgnostic";// 是否格式敏感，默认不敏感 | ||
| 22 | + const std::string FALSE_FORMAT_AGNOSTIC = "False"; | ||
| 23 | + const std::string TRUE_FORMAT_AGNOSTIC = "True"; | ||
| 24 | + const std::string OP_INFO_OPS_FLAG = "opInfo.opsFlag"; // 是否支持常量折叠，默认不支持 | ||
| 25 | + const std::string CLOSE_OPS_FLAG = "OPS_FLAG_CLOSE"; | ||
| 26 | + const std::string OPEN_OPS_FLAG = "OPS_FLAG_OPEN"; | ||
| 27 | + const std::string OP_INFO_SUB_TYPE_OF_INFERSHAPE = "opInfo.subTypeOfInferShape"; // 算子类型，默认是1类算子 | ||
| 28 | + const std::string DEFAULT_SUB_TYPE_OF_INFERSHAPE_1 = "1"; | ||
| 29 | + const std::string DEFAULT_SUB_TYPE_OF_INFERSHAPE_2 = "2"; | ||
| 30 | + const std::string DEFAULT_SUB_TYPE_OF_INFERSHAPE_3 = "3"; | ||
| 31 | + const std::string DEFAULT_SUB_TYPE_OF_INFERSHAPE_4 = "4"; | ||
| 32 | + | ||
| 33 | + | ||
| 34 | + template <typename TOpDef> | ||
| 35 | + inline void ApplyCvAicpuDefaultCfg(TOpDef &opDef) | ||
| 36 | + { | ||
| 37 | + opDef.AICPU().OpKernelLib("CUSTAICPUKernel"); | ||
| 38 | + opDef.AICPU().KernelSo(CV_AICPU_KERNEL_SO.c_str()); | ||
| 39 | + opDef.AICPU().UserDefined(true); | ||
| 40 | + opDef.AICPU().ExtendCfgInfo(OP_INFO_WORKSPACESIZE.c_str(), DEFAULT_WORKSPACE_SIZE.c_str()); | ||
| 41 | + opDef.AICPU().ExtendCfgInfo(OP_INFO_FORMAT_AGNOSTIC.c_str(), FALSE_FORMAT_AGNOSTIC.c_str()); | ||
| 42 | + opDef.AICPU().ExtendCfgInfo(OP_INFO_OPS_FLAG.c_str(), CLOSE_OPS_FLAG.c_str()); | ||
| 43 | + opDef.AICPU().ExtendCfgInfo(OP_INFO_SUB_TYPE_OF_INFERSHAPE.c_str(), DEFAULT_SUB_TYPE_OF_INFERSHAPE_1.c_str()); | ||
| 44 | + } | ||
| 45 | +} // namespace op | ||
| 46 | +#endif | ||
| @@ -14,12 +14,12 @@ if (BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG) | |||
| 14 | set(CMAKE_CXX_COMPILER ${ASCEND_DIR}/toolkit/toolchain/hcc/bin/aarch64-target-linux-gnu-g++) | 14 | set(CMAKE_CXX_COMPILER ${ASCEND_DIR}/toolkit/toolchain/hcc/bin/aarch64-target-linux-gnu-g++) |
| 15 | endif() | 15 | endif() |
| 16 | 16 | ||
| 17 | - # aicpu json | 17 | + # aicpu op_def |
| 18 | - file(GLOB_RECURSE JSON_FILE ${CMAKE_CURRENT_SOURCE_DIR}/*.json) | 18 | + file(GLOB_RECURSE OP_DEF_FILE ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu_def.cpp) |
| 19 | - set_property(GLOBAL APPEND PROPERTY AICPU_JSON_FILES ${JSON_FILE}) | 19 | + set_property(GLOBAL APPEND PROPERTY AICPU_OP_DEF_FILES ${OP_DEF_FILE}) |
| 20 | 20 | ||
| 21 | # aicpu cust kernel | 21 | # aicpu cust kernel |
| 22 | - file(GLOB AICPU_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu*.cpp) | 22 | + file(GLOB AICPU_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu.cpp) |
| 23 | message(STATUS "[crop_and_resize] Found aicpu sources: ${AICPU_SRC}, ascend dir: ${ASCEND_DIR}, ophsot name: ${OPHOST_NAME}") | 23 | message(STATUS "[crop_and_resize] Found aicpu sources: ${AICPU_SRC}, ascend dir: ${ASCEND_DIR}, ophsot name: ${OPHOST_NAME}") |
| 24 | 24 | ||
| 25 | set(OBJ_NAME crop_and_resize_cust_obj) | 25 | set(OBJ_NAME crop_and_resize_cust_obj) |
| @@ -1,35 +0,0 @@ | |||
| 1 | -{ | ||
| 2 | - "CropAndResize":{ | ||
| 3 | - "opInfo":{ | ||
| 4 | - "computeCost":"100", | ||
| 5 | - "engine":"DNN_VM_AICPU", | ||
| 6 | - "flagAsync":"False", | ||
| 7 | - "flagPartial":"False", | ||
| 8 | - "functionName":"RunCpuKernel", | ||
| 9 | - "kernelSo":"libcv_aicpu_kernels.so", | ||
| 10 | - "opKernelLib":"CUSTAICPUKernel", | ||
| 11 | - "userDefined":"True", | ||
| 12 | - "workspaceSize":"100" | ||
| 13 | - }, | ||
| 14 | - "input0":{ | ||
| 15 | - "name":"x", | ||
| 16 | - "type":"DT_FLOAT,DT_FLOAT16,DT_DOUBLE,DT_UINT8,DT_INT8,DT_UINT16,DT_INT16,DT_INT32,DT_INT64" | ||
| 17 | - }, | ||
| 18 | - "input1":{ | ||
| 19 | - "name":"boxes", | ||
| 20 | - "type":"DT_FLOAT" | ||
| 21 | - }, | ||
| 22 | - "input2":{ | ||
| 23 | - "name":"box_index", | ||
| 24 | - "type":"DT_INT32" | ||
| 25 | - }, | ||
| 26 | - "input3":{ | ||
| 27 | - "name":"crop_size", | ||
| 28 | - "type":"DT_INT32" | ||
| 29 | - }, | ||
| 30 | - "output0":{ | ||
| 31 | - "name":"y", | ||
| 32 | - "type":"DT_FLOAT" | ||
| 33 | - } | ||
| 34 | - } | ||
| 35 | -} | ||
| @@ -0,0 +1,46 @@ | |||
| 1 | +/** | ||
| 2 | + * Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 3 | + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 4 | + * CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 5 | + * Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 6 | + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 7 | + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 8 | + * See LICENSE in the root of the software repository for the full text of the License. | ||
| 9 | + */ | ||
| 10 | + | ||
| 11 | +/*! | ||
| 12 | + * \file crop_and_resize_aicpu_def.cpp | ||
| 13 | + * \brief crop_and_resize def | ||
| 14 | + */ | ||
| 15 | + | ||
| 16 | +#include "register/op_def_registry.h" | ||
| 17 | +#include "../../../common/inc/aicpu/aicpu_op_def.h" | ||
| 18 | + | ||
| 19 | +namespace ops { | ||
| 20 | +/* | ||
| 21 | +此配置项针对动态shape的场景，配置AI CPU的infershape方式，有以下几种取值： | ||
| 22 | +1：根据inputshape可直接推导outputshape。 | ||
| 23 | +2：部分计算的infershape，根据tensorData的值做简单的计算。 | ||
| 24 | +3：全部计算的infershape，但最大shape可预估，按照算子最大的shape申请内存。 | ||
| 25 | +4：无法由inputshape推导出算子的输出shape，也无法推导出输出range，此种情况需要在算子实现时进行内存的申请，计算完成后，返回output和输出shape。 | ||
| 26 | +默认值为1。 | ||
| 27 | +*/ | ||
| 28 | +class CropAndResize : public OpDef { | ||
李伟：构造函数中使this指针会不会抛异常 zhaowenrui Apr 1 Comments： Apr 1 Comments： | |||
| 29 | +public: | ||
| 30 | + explicit CropAndResize(const char* name) : OpDef(name) | ||
| 31 | + { | ||
| 32 | + ApplyCvAicpuDefaultCfg(*this); | ||
| 33 | + this->Input("x").DataType( | ||
| 34 | + {ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_DOUBLE, ge::DT_UINT8, ge::DT_INT8, ge::DT_UINT16, ge::DT_INT16, ge::DT_INT32, ge::DT_INT64}); | ||
| 35 | + this->Input("boxes").DataType({ge::DT_FLOAT}); | ||
| 36 | + this->Input("box_index").DataType({ge::DT_INT32}); | ||
| 37 | + this->Input("crop_size").DataType({ge::DT_INT32}); | ||
| 38 | + this->Output("y").DataType({ge::DT_FLOAT}); | ||
| 39 | + this->Attr("extrapolation_value").AttrType(OPTIONAL).Float(0); | ||
| 40 | + this->Attr("method").AttrType(OPTIONAL).String("bilinear"); | ||
D zhaowenrui Apr 1 Comments： Apr 1 Comments： | |||
| 41 | + this->AICPU().ExtendCfgInfo(OP_INFO_OPS_FLAG.c_str(), OPEN_OPS_FLAG.c_str()); | ||
| 42 | + } | ||
| 43 | +}; | ||
| 44 | + | ||
| 45 | +OP_ADD(CropAndResize); | ||
| 46 | +} // namespace ops | ||
| @@ -15,12 +15,12 @@ if (BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG) | |||
| 15 | set(CMAKE_CXX_COMPILER ${ASCEND_DIR}/toolkit/toolchain/hcc/bin/aarch64-target-linux-gnu-g++) | 15 | set(CMAKE_CXX_COMPILER ${ASCEND_DIR}/toolkit/toolchain/hcc/bin/aarch64-target-linux-gnu-g++) |
| 16 | endif() | 16 | endif() |
| 17 | 17 | ||
| 18 | - # aicpu json | 18 | + # aicpu op_def |
| 19 | - file(GLOB_RECURSE JSON_FILE ${CMAKE_CURRENT_SOURCE_DIR}/*.json) | 19 | + file(GLOB_RECURSE OP_DEF_FILE ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu_def.cpp) |
| 20 | - set_property(GLOBAL APPEND PROPERTY AICPU_JSON_FILES ${JSON_FILE}) | 20 | + set_property(GLOBAL APPEND PROPERTY AICPU_OP_DEF_FILES ${OP_DEF_FILE}) |
| 21 | 21 | ||
| 22 | # aicpu cust kernel | 22 | # aicpu cust kernel |
| 23 | - file(GLOB AICPU_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu*.cpp) | 23 | + file(GLOB AICPU_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu.cpp) |
| 24 | message(STATUS "[image_warp_offsets] Found aicpu sources: ${AICPU_SRC}, ascend dir: ${ASCEND_DIR}, ophsot name: ${OPHOST_NAME}") | 24 | message(STATUS "[image_warp_offsets] Found aicpu sources: ${AICPU_SRC}, ascend dir: ${ASCEND_DIR}, ophsot name: ${OPHOST_NAME}") |
| 25 | 25 | ||
| 26 | set(OBJ_NAME image_warp_offsets_cust_obj) | 26 | set(OBJ_NAME image_warp_offsets_cust_obj) |
| @@ -1,15 +0,0 @@ | |||
| 1 | -{ | ||
| 2 | - "IMGWarpOffsets":{ | ||
| 3 | - "opInfo":{ | ||
| 4 | - "computeCost":"100", | ||
| 5 | - "engine":"DNN_VM_AICPU", | ||
| 6 | - "flagAsync":"False", | ||
| 7 | - "flagPartial":"False", | ||
| 8 | - "functionName":"RunCpuKernel", | ||
| 9 | - "kernelSo":"libcv_aicpu_kernels.so", | ||
| 10 | - "opKernelLib":"CUSTAICPUKernel", | ||
| 11 | - "userDefined":"True", | ||
| 12 | - "workspaceSize":"100" | ||
| 13 | - } | ||
| 14 | - } | ||
| 15 | -} | ||
| @@ -0,0 +1,34 @@ | |||
| 1 | +/** | ||
| 2 | + * Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 3 | + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 4 | + * CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 5 | + * Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 6 | + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 7 | + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 8 | + * See LICENSE in the root of the software repository for the full text of the License. | ||
| 9 | + */ | ||
| 10 | + | ||
| 11 | +/*! | ||
| 12 | + * \file image_warp_offsets_aicpu_def.cpp | ||
| 13 | + * \brief image_warp_offsets def | ||
| 14 | + */ | ||
| 15 | + | ||
| 16 | +#include "register/op_def_registry.h" | ||
| 17 | +#include "../../../common/inc/aicpu/aicpu_op_def.h" | ||
| 18 | + | ||
| 19 | +namespace ops { | ||
| 20 | +class IMGWarpOffsets : public OpDef { | ||
| 21 | +public: | ||
| 22 | + explicit IMGWarpOffsets(const char* name) : OpDef(name) | ||
| 23 | + { | ||
| 24 | + ApplyCvAicpuDefaultCfg(*this); | ||
| 25 | + this->Input("images").DataType({ge::DT_UINT8, ge::DT_FLOAT16, ge::DT_FLOAT}); | ||
| 26 | + this->Input("offsets").DataType({ge::DT_FLOAT, ge::DT_INT32}); | ||
| 27 | + this->Output("warp_images").DataType({ge::DT_UINT8, ge::DT_FLOAT16, ge::DT_FLOAT}); | ||
| 28 | + this->AICPU().ExtendCfgInfo(OP_INFO_OPS_FLAG.c_str(), OPEN_OPS_FLAG.c_str()); | ||
| 29 | + | ||
| 30 | + } | ||
| 31 | +}; | ||
| 32 | + | ||
| 33 | +OP_ADD(IMGWarpOffsets); | ||
| 34 | +} // namespace ops | ||
| @@ -14,12 +14,12 @@ if (BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG) | |||
| 14 | set(CMAKE_CXX_COMPILER ${ASCEND_DIR}/toolkit/toolchain/hcc/bin/aarch64-target-linux-gnu-g++) | 14 | set(CMAKE_CXX_COMPILER ${ASCEND_DIR}/toolkit/toolchain/hcc/bin/aarch64-target-linux-gnu-g++) |
| 15 | endif() | 15 | endif() |
| 16 | 16 | ||
| 17 | - # aicpu json | 17 | + # aicpu op_def |
| 18 | - file(GLOB_RECURSE JSON_FILE ${CMAKE_CURRENT_SOURCE_DIR}/*.json) | 18 | + file(GLOB_RECURSE OP_DEF_FILE ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu_def.cpp) |
| 19 | - set_property(GLOBAL APPEND PROPERTY AICPU_JSON_FILES ${JSON_FILE}) | 19 | + set_property(GLOBAL APPEND PROPERTY AICPU_OP_DEF_FILES ${OP_DEF_FILE}) |
| 20 | 20 | ||
| 21 | # aicpu cust kernel | 21 | # aicpu cust kernel |
| 22 | - file(GLOB AICPU_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu*.cpp) | 22 | + file(GLOB AICPU_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*_aicpu.cpp) |
| 23 | message(STATUS "[non_max_suppression_v3] Found aicpu sources: ${AICPU_SRC}, ascend dir: ${ASCEND_DIR}, ophsot name: ${OPHOST_NAME}") | 23 | message(STATUS "[non_max_suppression_v3] Found aicpu sources: ${AICPU_SRC}, ascend dir: ${ASCEND_DIR}, ophsot name: ${OPHOST_NAME}") |
| 24 | 24 | ||
| 25 | set(OBJ_NAME non_max_suppression_v3_cust_obj) | 25 | set(OBJ_NAME non_max_suppression_v3_cust_obj) |
| @@ -1,15 +0,0 @@ | |||
| 1 | -{ | ||
| 2 | - "NonMaxSuppressionV3":{ | ||
| 3 | - "opInfo":{ | ||
| 4 | - "computeCost":"100", | ||
| 5 | - "engine":"DNN_VM_AICPU", | ||
| 6 | - "flagAsync":"False", | ||
| 7 | - "flagPartial":"False", | ||
| 8 | - "functionName":"RunCpuKernel", | ||
| 9 | - "kernelSo":"libcv_aicpu_kernels.so", | ||
| 10 | - "opKernelLib":"CUSTAICPUKernel", | ||
| 11 | - "userDefined":"True", | ||
| 12 | - "workspaceSize":"100" | ||
| 13 | - } | ||
| 14 | - } | ||
| 15 | -} | ||
| @@ -0,0 +1,36 @@ | |||
| 1 | +/** | ||
| 2 | + * Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 3 | + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 4 | + * CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 5 | + * Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 6 | + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 7 | + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 8 | + * See LICENSE in the root of the software repository for the full text of the License. | ||
| 9 | + */ | ||
| 10 | + | ||
| 11 | +/*! | ||
| 12 | + * \file non_max_suppression_v3_aicpu_def.cpp | ||
| 13 | + * \brief non_max_suppression_v3 def | ||
| 14 | + */ | ||
| 15 | + | ||
| 16 | +#include "register/op_def_registry.h" | ||
| 17 | +#include "../../../common/inc/aicpu/aicpu_op_def.h" | ||
| 18 | + | ||
| 19 | +namespace ops { | ||
| 20 | +class NonMaxSuppressionV3 : public OpDef { | ||
| 21 | +public: | ||
| 22 | + explicit NonMaxSuppressionV3(const char* name) : OpDef(name) | ||
| 23 | + { | ||
| 24 | + ApplyCvAicpuDefaultCfg(*this); | ||
| 25 | + this->Input("boxes").DataType({ge::DT_FLOAT, ge::DT_FLOAT16}); | ||
| 26 | + this->Input("scores").DataType({ge::DT_FLOAT, ge::DT_FLOAT16}); | ||
| 27 | + this->Input("max_output_size").DataType({ge::DT_INT32}); | ||
| 28 | + this->Input("iou_threshold").DataType({ge::DT_FLOAT, ge::DT_FLOAT16}); | ||
| 29 | + this->Input("score_threshold").DataType({ge::DT_FLOAT, ge::DT_FLOAT16}); | ||
| 30 | + this->Output("selected_indices").DataType({ge::DT_INT32}); | ||
| 31 | + this->AICPU().ExtendCfgInfo(OP_INFO_SUB_TYPE_OF_INFERSHAPE.c_str(), DEFAULT_SUB_TYPE_OF_INFERSHAPE_4.c_str()); | ||
| 32 | + } | ||
| 33 | +}; | ||
| 34 | + | ||
| 35 | +OP_ADD(NonMaxSuppressionV3); | ||
| 36 | +} // namespace ops | ||
| @@ -1,26 +1,36 @@ | |||
| 1 | #!/usr/bin/env python3 | 1 | #!/usr/bin/env python3 |
| 2 | # -*- coding: utf-8 -*- | 2 | # -*- coding: utf-8 -*- |
| 3 | -# --------------------------------------------------------------------------------------------------------- | 3 | +# ---------------------------------------------------------------------------- |
| 4 | -# Copyright (c) 2025 Huawei Technologies Co., Ltd. | 4 | +# Copyright (c) 2025-2026 Huawei Technologies Co., Ltd. |
| 5 | -# This program is free software, you can redistribute it and/or modify it under the terms and conditions of | 5 | +# This program is free software, you can redistribute it and/or modify it under the terms and conditions of |
| 6 | # CANN Open Software License Agreement Version 2.0 (the "License"). | 6 | # CANN Open Software License Agreement Version 2.0 (the "License"). |
| 7 | # Please refer to the License for details. You may not use this file except in compliance with the License. | 7 | # Please refer to the License for details. You may not use this file except in compliance with the License. |
| 8 | -# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | 8 | +# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, |
| 9 | -# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | 9 | +# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. |
| 10 | # See LICENSE in the root of the software repository for the full text of the License. | 10 | # See LICENSE in the root of the software repository for the full text of the License. |
| 11 | - | ||
| 12 | -# The code snippet comes from Huawei's open-source Ascend project. | ||
| 13 | -# Copyright 2020 Huawei Technologies Co., Ltd | ||
| 14 | -# Licensed under the Apache License, Version 2.0 (the "License"); | ||
| 15 | -# You may obtain a copy of the License at | ||
| 16 | -# http://www.apache.org/licenses/LICENSE-2.0 | ||
| 17 | -# | ||
| 18 | # ---------------------------------------------------------------------------- | 11 | # ---------------------------------------------------------------------------- |
| 19 | 12 | ||
| 20 | import json | 13 | import json |
| 21 | import os | 14 | import os |
| 22 | import stat | 15 | import stat |
| 23 | import sys | 16 | import sys |
| 17 | +from operator import truediv | ||
| 18 | + | ||
| 19 | +required_op_input_info_keys = ["paramType", "name"] | ||
| 20 | +required_op_output_info_keys = ["paramType", "name"] | ||
| 21 | +param_type_valid_value = ["dynamic", "optional", "required"] | ||
| 22 | +required_attr_key = ["type", "value", "paramType"] | ||
| 23 | +aicpu_infershape_subtype_valid_value = ["1", "2", "3", "4"] | ||
| 24 | +aicpu_ops_flag_valid_value = ["OPS_FLAG_OPEN", "OPS_FLAG_CLOSE"] | ||
| 25 | +# workspaceSize is valid when it is within the inclusive range [100, 500] | ||
| 26 | +AICPU_WORKSPACE_SIZE_MIN = 100 | ||
| 27 | +AICPU_WORKSPACE_SIZE_MAX = 500 | ||
| 28 | +aicpu_kernel_so_valid_value = [ | ||
| 29 | + "libmath_aicpu_kernels.so", | ||
| 30 | + "libcv_aicpu_kernels.so", | ||
| 31 | + "libnn_aicpu_kernels.so", | ||
| 32 | + "libtransformer_aicpu_kernels.so", | ||
| 33 | +] | ||
| 24 | 34 | ||
| 25 | 35 | ||
| 26 | def parse_ini_files(ini_files): | 36 | def parse_ini_files(ini_files): |
| @@ -63,61 +73,146 @@ def parse_ini_to_obj(ini_file, tbe_ops_info): | |||
| 63 | op = {} | 73 | op = {} |
| 64 | tbe_ops_info[op_name] = op | 74 | tbe_ops_info[op_name] = op |
| 65 | else: | 75 | else: |
| 66 | - key1 = line[:line.index("=")].strip() | 76 | + key1 = line[: line.index("=")].strip() |
| 67 | - key2 = line[line.index("=")+1:].strip() | 77 | + key2 = line[line.index("=") + 1:].strip() |
| 68 | key1_0, key1_1 = key1.split(".") | 78 | key1_0, key1_1 = key1.split(".") |
| 69 | if not key1_0 in op: | 79 | if not key1_0 in op: |
| 70 | op[key1_0] = {} | 80 | op[key1_0] = {} |
| 71 | if key1_1 in op[key1_0]: | 81 | if key1_1 in op[key1_0]: |
| 72 | - raise RuntimeError("Op:" + op_name + " " + key1_0 + " " + key1_1 + " is repeated!") | 82 | + raise RuntimeError( |
| 83 | + "Op:" + op_name + " " + key1_0 + " " + key1_1 + " is repeated!" | ||
| 84 | + ) | ||
| 73 | op[key1_0][key1_1] = key2 | 85 | op[key1_0][key1_1] = key2 |
| 74 | 86 | ||
| 75 | 87 | ||
| 88 | +def is_cpu_engine(op_engine): | ||
| 89 | + """判断是否为 AICPU 引擎""" | ||
| 90 | + return op_engine == "DNN_VM_AICPU" | ||
| 91 | + | ||
| 92 | + | ||
| 93 | +def check_required_keys(op_info, required_keys, op_engine): | ||
| 94 | + """检查必需字段是否存在，AICPU 算子跳过 paramType 检查""" | ||
| 95 | + missing_keys = [] | ||
| 96 | + for required_key in required_keys: | ||
| 97 | + # AICPU 算子不需要 paramType 字段 | ||
| 98 | + if is_cpu_engine(op_engine) and required_key == "paramType": | ||
| 99 | + continue | ||
| 100 | + if required_key not in op_info: | ||
| 101 | + missing_keys.append(required_key) | ||
| 102 | + return missing_keys | ||
| 103 | + | ||
| 104 | + | ||
| 105 | +def check_param_type(op_info, op_engine, error_message): | ||
| 106 | + """检查 paramType 是否合法，AICPU 算子跳过检查""" | ||
| 107 | + if is_cpu_engine(op_engine): | ||
| 108 | + return True | ||
| 109 | + if op_info.get("paramType") not in param_type_valid_value: | ||
| 110 | + print(error_message) | ||
| 111 | + return False | ||
| 112 | + return True | ||
| 113 | + | ||
| 114 | + | ||
| 115 | +def check_infershape_subtype(op_info): | ||
| 116 | + """检查 AICPU 的 subTypeOfInferShape 字段是否合法""" | ||
| 117 | + if op_info.get("subTypeOfInferShape") not in aicpu_infershape_subtype_valid_value: | ||
| 118 | + return False | ||
| 119 | + return True | ||
| 120 | + | ||
| 121 | + | ||
| 122 | +def check_ops_flag(op_info): | ||
| 123 | + """检查 AICPU 的 opsFlag 字段是否合法""" | ||
| 124 | + if op_info.get("opsFlag") not in aicpu_ops_flag_valid_value: | ||
| 125 | + return False | ||
| 126 | + return True | ||
| 127 | + | ||
| 128 | + | ||
| 129 | +def check_workspace_size(op_info): | ||
| 130 | + """检查 AICPU 的 workspaceSize 是否在 [100, 500] 范围内""" | ||
| 131 | + workspace_size = op_info.get("workspaceSize") | ||
| 132 | + try: | ||
| 133 | + workspace_size = int(workspace_size) | ||
| 134 | + except (TypeError, ValueError): | ||
| 135 | + return False | ||
| 136 | + return AICPU_WORKSPACE_SIZE_MIN <= workspace_size <= AICPU_WORKSPACE_SIZE_MAX | ||
| 137 | + | ||
| 138 | + | ||
| 139 | +def check_opkernel_so(op_info): | ||
| 140 | + """检查 kernelSo 是否在白名单中""" | ||
| 141 | + if op_info.get("kernelSo") not in aicpu_kernel_so_valid_value: | ||
| 142 | + return False | ||
| 143 | + return True | ||
| 144 | + | ||
| 145 | + | ||
| 146 | +def validate_io_info(op_key, op_info_key, op_info, op_engine, required_keys): | ||
| 147 | + """校验输入或输出信息""" | ||
| 148 | + is_valid = True | ||
| 149 | + missing_keys = check_required_keys(op_info, required_keys, op_engine) | ||
| 150 | + if missing_keys: | ||
| 151 | + print(f"op: {op_key} {op_info_key} missing: {','.join(missing_keys)}") | ||
| 152 | + is_valid = False | ||
| 153 | + | ||
| 154 | + error_message = f"op: {op_key} {op_info_key} paramType not valid, valid key:{param_type_valid_value}" | ||
| 155 | + if not check_param_type(op_info, op_engine, error_message): | ||
| 156 | + is_valid = False | ||
| 157 | + | ||
| 158 | + return is_valid | ||
| 159 | + | ||
| 160 | + | ||
| 161 | +def validate_optional_info(op_key, op_info_key, op_info): | ||
| 162 | + """校验 AICPU 的可选字段""" | ||
| 163 | + is_valid = True | ||
| 164 | + if "subTypeOfInferShape" in op_info and not check_infershape_subtype(op_info): | ||
必选参数不全，应该是有8个参数的。没有的补齐一下。 | |||
| 165 | + print(f"op: {op_key} {op_info_key} infershape_subtype not valid" | ||
| 166 | + f" valid key: {aicpu_infershape_subtype_valid_value}") | ||
| 167 | + is_valid = False | ||
| 168 | + if "opsFlag" in op_info and not check_ops_flag(op_info): | ||
| 169 | + print(f"op: {op_key} {op_info_key} opsFlag not valid" | ||
| 170 | + f" valid key: {aicpu_ops_flag_valid_value}") | ||
| 171 | + is_valid = False | ||
| 172 | + if "workspaceSize" in op_info and not check_workspace_size(op_info): | ||
| 173 | + print(f"op: {op_key} {op_info_key} workspaceSize not valid" | ||
| 174 | + f" valid range: [{AICPU_WORKSPACE_SIZE_MIN}, {AICPU_WORKSPACE_SIZE_MAX}]") | ||
| 175 | + is_valid = False | ||
| 176 | + if "kernelSo" in op_info and not check_opkernel_so(op_info): | ||
| 177 | + print(f"op: {op_key} {op_info_key} kernelSo not valid" | ||
| 178 | + f" valid key: {aicpu_kernel_so_valid_value}") | ||
| 179 | + is_valid = False | ||
| 180 | + return is_valid | ||
| 181 | + | ||
| 182 | + | ||
| 76 | def check_op_info(tbe_ops): | 183 | def check_op_info(tbe_ops): |
| 77 | - """ | 184 | + """校验算子信息""" |
公共流程修改，上库前编译下check.sh看看aicore算子产物，不要影响了已有流程。 | |||
| 78 | - Check info info | ||
| 79 | - """ | ||
| 80 | print("\n\n==============check valid for ops info start==============") | 185 | print("\n\n==============check valid for ops info start==============") |
| 81 | - not_valid_op = [] | 186 | + |
| 82 | - required_op_input_info_keys = ["paramType", "name"] | 187 | + io_required_keys_map = { |
| 83 | - required_op_output_info_keys = ["paramType", "name"] | 188 | + "input": required_op_input_info_keys, |
| 84 | - shape_type_valid_value = ["fix", "range", "list"] | 189 | + "output": required_op_output_info_keys, |
| 85 | - need_compile_valid_value = ["true", "false"] | 190 | + } |
| 86 | - param_type_valid_value = ["dynamic", "optional", "required"] | ||
| 87 | - required_attr_key = ["type", "value", "paramType"] | ||
| 88 | is_valid = True | 191 | is_valid = True |
| 89 | for op_key in tbe_ops: | 192 | for op_key in tbe_ops: |
| 90 | op = tbe_ops[op_key] | 193 | op = tbe_ops[op_key] |
| 194 | + op_engine = op.get("opInfo", {}).get("engine", "") if isinstance(op, dict) else "" | ||
| 91 | 195 | ||
| 92 | for op_info_key in op: | 196 | for op_info_key in op: |
| 197 | + op_info_flag = False | ||
| 198 | + io_prefix = "" | ||
| 93 | if op_info_key.startswith("input"): | 199 | if op_info_key.startswith("input"): |
| 94 | - op_input_info = op[op_info_key] | 200 | + io_prefix = "input" |
| 95 | - missing_keys = [] | 201 | + elif op_info_key.startswith("output"): |
| 96 | - for required_op_input_info_key in required_op_input_info_keys: | 202 | + io_prefix = "output" |
| 97 | - if not required_op_input_info_key in op_input_info: | 203 | + elif op_info_key.startswith("opInfo"): |
| 98 | - missing_keys.append(required_op_input_info_key) | 204 | + op_info_flag = True |
| 99 | - if len(missing_keys) > 0: | 205 | + else: |
| 100 | - print("op: " + op_key + " " + op_info_key + " missing: " + ",".join(missing_keys)) | 206 | + continue |
| 101 | - is_valid = False | 207 | + |
| 102 | - else: | 208 | + op_info = op[op_info_key] |
| 103 | - if not op_input_info["paramType"] in param_type_valid_value: | 209 | + if io_prefix != "" and not validate_io_info( |
| 104 | - print("op: " + op_key + " " + op_info_key + \ | 210 | + op_key, op_info_key, op_info, op_engine, io_required_keys_map.get(io_prefix) |
| 105 | - " paramType not valid, valid key:[dynamic, optional, required]") | 211 | + ): |
| 106 | - is_valid = False | 212 | + is_valid = False |
| 107 | - if op_info_key.startswith("output"): | 213 | + if is_cpu_engine(op_engine) and op_info_flag and not validate_optional_info(op_key, op_info_key, op_info): |
| 108 | - op_input_info = op[op_info_key] | 214 | + is_valid = False |
| 109 | - missing_keys = [] | 215 | + |
| 110 | - for required_op_input_info_key in required_op_output_info_keys: | ||
| 111 | - if not required_op_input_info_key in op_input_info: | ||
| 112 | - missing_keys.append(required_op_input_info_key) | ||
| 113 | - if len(missing_keys) > 0: | ||
| 114 | - print("op: " + op_key + " " + op_info_key + " missing: " + ",".join(missing_keys)) | ||
| 115 | - is_valid = False | ||
| 116 | - else: | ||
| 117 | - if not op_input_info["paramType"] in param_type_valid_value: | ||
| 118 | - print("op: " + op_key + " " + op_info_key + \ | ||
| 119 | - " paramType not valid, valid key:[fix, range, list]") | ||
| 120 | - is_valid = False | ||
| 121 | print("==============check valid for ops info end================\n\n") | 216 | print("==============check valid for ops info end================\n\n") |
| 122 | return is_valid | 217 | return is_valid |
| 123 | 218 | ||
message太多了，保留点必要的，非必要的删除了。