ACL GEMM API Migration Sample【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/geFunction DescriptionDemonstrates how toreplace deprecated ACLBLAS and ACLOP APIs with ACLNN APIs. Uses the GEMM (General Matrix Multiply) operator to compare equivalent implementations across three ACL API families, helping developers understand the migration workflow.APIStatusInvocationaclblasGemmExDeprecatedExecute GEMM via ACLBLAS interfaceaclopExecuteV2(GEMM)DeprecatedInvoke GEMM operator via ACLOP interfaceaclnnMatmulaclnnMulsaclnnAddRecommendedCompose atomic ACLNN operationsThe sample runs the same input data through all three APIs sequentially, automatically compares precision, and outputsmax_error.GEMM FormulaC α·A·B β·CParameterValueDescriptionM64Rows of A / Rows of CN64Columns of B / Columns of CK64Columns of A / Rows of Balpha2.0Scale factor for matmul resultbeta2.0Scale factor for original CData typeFP16All matrices and scalars use FP16ACLNN API combination: ACLNN APIs are closer to atomic operator operations, requiring GEMM to be decomposed into:aclnnMatmul→A BaclnnMuls→alpha × (A B)andbeta × CaclnnAdd→ sum the two partsQuick Startbash scripts/build.sh bash scripts/run.shDirectory Structure4_sample_acl_gemm/ ├── CMakeLists.txt # Top-level build script ├── scripts/ │ ├── build.sh # Build script │ └── run.sh # Run script └── src/ ├── CMakeLists.txt # Build configuration └── sample_acl_gemm.cpp # Main program (three-API comparison)Environment RequirementsInstalltoolkitandopspackages per Environment PreparationSet environment variables:source /usr/local/Ascend/cann/set_env.sh(adjust path to your installation)(Optional) Specify chip version:export SOC_VERSIONAscend910B3(defaults toAscend910B3if not set)Detailed StepsStep 1: Buildbash scripts/build.shStep 2: Runbash scripts/run.shExpected Output[INFO] ACL GEMM sample starts [INFO] Running ACLBLAS GEMM... [INFO] Running ACLOP GEMM... [INFO] Running ACLNN GEMM... [INFO] Comparing results... max_error: 0 [INFO] VERIFICATION PASSED [INFO] SAMPLE PASSEDFAQQ: Build fails with missing ACLNN header filesMake sure you have runsource /usr/local/Ascend/cann/set_env.shand that your installed toolkit version includes ACLNN APIs.Q: max_error is not zeroSmall errors may occur under FP16 precision. If max_error is small (e.g., 1e-3), it is a normal floating-point error. If it is large, check whetherSOC_VERSIONmatches your current chip.Q: How to migrate my own ACLOP code to ACLNN?Follow the ACLNN implementation pattern insample_acl_gemm.cpp: first callaclnnXxxGetWorkspaceSizeto obtain the workspace size, then callaclnnXxxto execute the operation.【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
ACL GEMM API Migration Sample【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/geFunction DescriptionDemonstrates how toreplace deprecated ACLBLAS and ACLOP APIs with ACLNN APIs. Uses the GEMM (General Matrix Multiply) operator to compare equivalent implementations across three ACL API families, helping developers understand the migration workflow.APIStatusInvocationaclblasGemmExDeprecatedExecute GEMM via ACLBLAS interfaceaclopExecuteV2(GEMM)DeprecatedInvoke GEMM operator via ACLOP interfaceaclnnMatmulaclnnMulsaclnnAddRecommendedCompose atomic ACLNN operationsThe sample runs the same input data through all three APIs sequentially, automatically compares precision, and outputsmax_error.GEMM FormulaC α·A·B β·CParameterValueDescriptionM64Rows of A / Rows of CN64Columns of B / Columns of CK64Columns of A / Rows of Balpha2.0Scale factor for matmul resultbeta2.0Scale factor for original CData typeFP16All matrices and scalars use FP16ACLNN API combination: ACLNN APIs are closer to atomic operator operations, requiring GEMM to be decomposed into:aclnnMatmul→A BaclnnMuls→alpha × (A B)andbeta × CaclnnAdd→ sum the two partsQuick Startbash scripts/build.sh bash scripts/run.shDirectory Structure4_sample_acl_gemm/ ├── CMakeLists.txt # Top-level build script ├── scripts/ │ ├── build.sh # Build script │ └── run.sh # Run script └── src/ ├── CMakeLists.txt # Build configuration └── sample_acl_gemm.cpp # Main program (three-API comparison)Environment RequirementsInstalltoolkitandopspackages per Environment PreparationSet environment variables:source /usr/local/Ascend/cann/set_env.sh(adjust path to your installation)(Optional) Specify chip version:export SOC_VERSIONAscend910B3(defaults toAscend910B3if not set)Detailed StepsStep 1: Buildbash scripts/build.shStep 2: Runbash scripts/run.shExpected Output[INFO] ACL GEMM sample starts [INFO] Running ACLBLAS GEMM... [INFO] Running ACLOP GEMM... [INFO] Running ACLNN GEMM... [INFO] Comparing results... max_error: 0 [INFO] VERIFICATION PASSED [INFO] SAMPLE PASSEDFAQQ: Build fails with missing ACLNN header filesMake sure you have runsource /usr/local/Ascend/cann/set_env.shand that your installed toolkit version includes ACLNN APIs.Q: max_error is not zeroSmall errors may occur under FP16 precision. If max_error is small (e.g., 1e-3), it is a normal floating-point error. If it is large, check whetherSOC_VERSIONmatches your current chip.Q: How to migrate my own ACLOP code to ACLNN?Follow the ACLNN implementation pattern insample_acl_gemm.cpp: first callaclnnXxxGetWorkspaceSizeto obtain the workspace size, then callaclnnXxxto execute the operation.【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
