PyAsc

PyAsc (Python-Ascend) is a Python programming model for writing compute kernels that run on Huawei Ascend NPUs. Kernels are written in Python, translated through an MLIR pipeline into Ascend C, and then compiled by the Bisheng compiler into an NPU binary. Two APIs are available:

• asc2 (tile-based, NumPy-like, @asc2.jit) for high-level kernels,
• asc (1:1 Ascend C mapping, @asc.jit) for low-level control.

Installation

1. Install CANN toolkit (provides the Bisheng compiler and NPU runtime).

2. Install PyAsc:

   python3 -m pip install pyasc

   or build from source.

Quick Example

import asc
import asc2
import torch

@asc2.jit
def vadd_kernel(x_ptr: asc.GlobalAddress, y_ptr: asc.GlobalAddress, out_ptr: asc.GlobalAddress,
                size: int, tile_size: asc.ConstExpr[int], tile_per_block: asc.ConstExpr[int]):
    x_gm   = asc2.global_tensor(x_ptr,   [size])
    y_gm   = asc2.global_tensor(y_ptr,   [size])
    out_gm = asc2.global_tensor(out_ptr, [size])
    base   = asc2.block_idx() * tile_size * tile_per_block
    for i in range(tile_per_block):
        off = base + i * tile_size
        x   = asc2.copy_in(x_gm, [tile_size], offsets=[off])
        y   = asc2.copy_in(y_gm, [tile_size], offsets=[off])
        asc2.copy_out(x + y, out_gm, offsets=[off])

# Create tensors
x = torch.rand(8192, dtype=torch.float32)
y = torch.rand_like(x)
out = torch.empty_like(x)

# Launch on 16 NPU cores
asc.runtime.config.set_plarform("NPU")
tile_size, cores = 128, 16
vadd_kernel[cores](x, y, out, out.size, tile_size, asc.ceildiv(out.size // tile_size, cores))

# Verify the result
torch.testing.assert_close(out, x + y)

Documentation

Read the full documentation to learn more:

Section	Description
Installation	Environment setup (CANN), build from source
Design	Project overview, design decisions, key implementation details
Development	Coding style (C++, Python), adding new APIs, development tools

It also includes autogenerated Python API reference and MLIR dialect listing.

License

CANN Open Software License Agreement Version 2.0. See LICENSE.