ascendc-direct-invoke

name	ascendc-direct-invoke
description	AscendC direct-invoke 工程契约：WA 使用 kernel.py + ascendc_op/，ModelNew 调用 torch.ops.npu.*，adapter 负责复制工程、CMake 构建和 npu-arch patch。适用于 dsl=ascendc 的 autoresearch 任务。

AscendC Direct-Invoke

This skill is adapted for AKG autoresearch from the CANNBot direct-invoke template. In WA, the kernel handoff is not the old three-string AscendC protocol and not a registry-invoke package. It is a Python wrapper plus a normal CMake AscendC project:

task_dir/
  kernel.py
  ascendc_op/
    CMakeLists.txt
    op_kernel/
    op_host/
    op_extension/
    scripts/

Hard Contract

• kernel.py defines exactly one public entry class named ModelNew.
• ModelNew.forward() calls the compiled extension through torch.ops.npu.<op>(...) or the namespace registered by the project.
• ascendc_op/ contains the CMake project; do not embed C++ source in Python strings.
• Do not call CMake or compile from kernel.py. AKG's AscendC adapter copies ascendc_op/, patches --npu-arch=..., configures CMake, builds, then imports ModelNew.
• Do not use .npu() in wrapper code. Inputs are already moved by the verifier; use ordinary PyTorch tensors on the selected device.
• Load the .so lazily inside ModelNew._load() or forward(), not at module import time.

Wrapper Pattern

from __future__ import annotations

from pathlib import Path

import torch
import torch_npu


class ModelNew(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self._loaded = False

    def _load(self):
        if self._loaded:
            return
        build_dir = Path(__file__).with_name("ascendc_op") / "build"
        so_files = sorted(build_dir.rglob("*.so"))
        if not so_files:
            raise RuntimeError(f"no AscendC extension found under {build_dir}")
        torch.ops.load_library(str(so_files[0]))
        self._loaded = True

    def forward(self, x, y):
        self._load()
        return torch.ops.npu.my_op(x, y)

Keep _load() deterministic. If multiple shared libraries exist, prefer a stable name match such as libmy_op_ops.so before falling back to sorted order.

Project Pattern

The direct-invoke project usually has two targets:

• an optional executable for standalone kernel launch and local debugging
• a shared library registered through TORCH_LIBRARY_FRAGMENT / TORCH_LIBRARY_IMPL

Common files:

ascendc_op/
  CMakeLists.txt
  op_kernel/<op>_kernel.asc
  op_kernel/<op>_tiling.h
  op_host/<op>.asc
  op_extension/<op>_torch.cpp
  op_extension/register.cpp
  op_extension/ops.h

The PyTorch extension side follows this shape:

TORCH_LIBRARY_FRAGMENT(npu, m) {
    m.def("my_op(Tensor x, Tensor y) -> Tensor");
}

TORCH_LIBRARY_IMPL(npu, PrivateUse1, m) {
    m.impl("my_op", TORCH_FN(ascend_kernel::my_op_torch));
}

The host launch code should:

• allocate output tensors with PyTorch on the same device as inputs
• compute tiling from tensor shape and dtype
• copy tiling data to device memory or a device tensor
• get the current NPU stream via c10_npu::getCurrentNPUStream()
• call the generated kernel launcher with ABI matching the kernel entry

CMake Requirements

AKG passes these variables to CMake:

NPU_ARCH
ASCENDC_NPU_ARCH
ASCEND_HOME_PATH
Python_EXECUTABLE
Python3_EXECUTABLE

If CMakeLists.txt contains a literal --npu-arch=dav-2201 or --npu-arch=dav-3510, the adapter patches it for the current arch. Prefer using ${NPU_ARCH} directly when writing new projects.

The project must build a .so somewhere under ascendc_op/build/. The adapter searches recursively.

Vector Template Route

For elementwise and simple reduction operators, use the CANNBot add_custom pattern as the mental model:

• op_kernel/<op>_tiling.h: tiling constants and POD struct shared by host and kernel
• op_kernel/<op>_kernel.asc: KernelXxx class with Init, CopyIn, Compute, CopyOut, Process
• op_extension/<op>_torch.cpp: PyTorch bridge, tiling computation, stream selection, kernel launch
• register.cpp: torch.ops.npu.<op> registration and Meta implementation

When adapting it, globally rename the operator and then modify only marked semantic points: input/output count, tiling fields, compute expression, output shape, dtype checks, and CMake target names.

Do Not Generate

• host_tiling_src = """..."""
• kernel_src = """..."""
• python_bind_src = """..."""
• run.sh-only workflows inside WA
• import-time compilation
• hard-coded local device IDs

WA invokes the formal KernelVerifier chain. The seed must be a reusable project, not a one-off script.