AddExamplePypto

English | 简体中文

This is the PyPTO version of examples/add_example. The function is also element-wise addition. The difference is that the kernel is written in the PyPTO DSL (op_kernel/add_example_pypto.py) instead of AscendC C++. This operator is used to demonstrate how PyPTO operators are integrated in ops-cv.

Compilation

source <ascend-toolkit>/set_env.sh
bash build.sh --pkg --ops=add_example_pypto --soc=ascend950

Integration Method

In the op_host/CMakeLists.txt file of the operator, add two lines before add_modules_sources (or add_all_modules_sources):

require_pypto_pro(add_example_pypto)
enable_pypto_kernel(add_example_pypto)

add_modules_sources(OPTYPE add_example_pypto ACLNNTYPE aclnn)

• require_pypto_pro(<op_name>): If pypto_pro does not exist in the environment, a warning is reported and this operator is skipped. The entire configure does not fail.
• enable_pypto_kernel(<op_file>): <op_file> must be consistent with op_kernel/<op_file>.py and ExtendCfgInfo("opFile.value", ...) in the operator definition.

Build Process

enable_pypto_kernel calls scripts/util/pypto_codegen.py in the configure phase to perform codegen for the .py file of the kernel:

Artifact	Purpose
<Tiling>_tiling.h	Tiling struct, which is directly used by the host-side tiling
<TilingKey>_tilingkey.h	ASCENDC_TPL tiling key declaration
<op_file>_pypto_infer.cpp	Self-contained source code for .i infer

The preceding two header files are force-included in the *_tiling*.cpp files of this operator in the -include mode (in the order: tilingkey first, then tiling). Therefore, the host-side tiling can directly use the tiling struct defined in the .py file without an explicit include:

// AddExamplePyptoTiling comes from op_kernel/add_example_pypto.py and is visible through force-include
AddExamplePyptoTiling* tilingData = context->GetTilingData<AddExamplePyptoTiling>();

Therefore, the host and kernel share the same tiling layout and tiling key bit allocation, and no manually written copy is required.

In the build phase, the wrapper generated for this operator (<build>/tbe/dynamic/<op_file>.py) calls pypto_compile_op instead of compile_op. It performs codegen based on the tiling key and reuses the AscendC backend to compile the kernel binary.

Chip and Arch

The codegen arch of PyPTO is automatically derived by the platform (a2/a3/a5). This operator registers only ascend950 (a5), which is consistent with the arch automatically derived by the current toolkit. When compiling the kernel binary, cmake exports PYPTO_JIT_ARCH based on the soc being compiled (ascend910b/ascend910_93 → a3, ascend950 → a5), preventing pypto_compile_op from falling back to the default a5 and generating instructions not supported by the target chip.