Skip to content

TODO.md

Latest commit

 

History

History
153 lines (103 loc) · 3.93 KB

TODO.md

File metadata and controls

153 lines (103 loc) · 3.93 KB

TODO

short term todo (Nov 2025)

  1. fill_kernel() -> aclnnop_FillScalar GeneralOp 类型

  2. backends -> cupy.backends, git mv + _features.py disable _preflight, so ignore Cutensor submodule

  3. enable most of math api (done), aclnnop register (almost done) matmul (done), dot

  4. BitwiseAddScalar op register, _kernel.pyx need update (done) pytest

  5. reduction kernel, replaced by aclnnop ===

  6. concat/pad/reshape op numpy_to_acl_dtype -> numpy_dtype_to_acl_dtype

  7. triton-fusion (add data adaptor API)

  8. aclBlas integration

  9. test build on diff OS, hardware with float32 support (310P?)

  10. testing (after eanble most numpy API)

  11. inplace op : _kernel.pyx need update

intermediate (within 6 months)

  1. templated ascend kernel JIT
  2. compile customer kernel
  3. impl missing numpy op for ascend
  4. random
  5. FFT
  6. single node multiple NPU distribution test

longterm (within one year)

  1. multi-node multiple NPU
  2. double datatype (float32, int64)
  3. sparse matrix

TODO (详细方案)

CuPy在AMD ROCm平台上的数组创建问题分析与解决方案 - GitCode博客

自定义算子JIT ：编译和动态加载Kernel

不清楚CANN 和 triton-ascend 路标

ElementwiseKernel 可以做到类似 CUDA的层面, 就是有些工作量.

随机数 (不紧急)

可以numpy来生成随机数, AsNumpy has impl

rand： #include <aclnnop/aclnn_rand.h> 有算子

    HIP_random = {
        'name': 'random',
        'required': True,
        'file': [
            'cupy.random._bit_generator',
            ('cupy.random._generator_api',
             ['cupy/random/cupy_distributions.cu']),
        ],
        'include': [
            'hiprand/hiprand.h',
        ],
        'libraries': [
            # Dependency from cuRAND header files
            'amdhip64',  # was hiprtc and hip_hcc before ROCm 3.8.0
            'hiprand',
        ],
        'check_method': build.check_hip_version,
        'version_method': build.get_hip_version,
    }

FFT 信号处理：价值很大

AsdSip也支持FFT一些算子

hccl 多卡支持 (rocm 没有做迁移)

CUPY的多卡, 本身需要测试工作量.

HCCL和NCCL的API兼容性, 粗看很相似.

稠密和稀疏矩阵求解 (cupyx.scipy)

cusparse + cusolver： aicpu， torch-npu

我中心在稀疏矩阵乘算子研发中取得新进展--中国科学院计算机网络信息中心

#ifdef defined(CUPY_USE_ASCEND)
// not sure if CANN support solver, leave it later
//#include "ascend/cupy_ascend_solver.h"
#include "stub/cupy_cusolver.h"  // gracefully give error message

TODO: CUBLAS_COMPUTE , cudaDataType 下沉到backend层次

_numpy_to_backend_dtype() 不同的backend实现不同的pyx文件

TODO: HCCL NCCL, MPI, RCCL 应该可以抽象

sparse, 等ascend缺失的暂不做中性化处理, 直接放入cuda/libs

split out cuda enum from backend

这些基本GPU专用, 或者不是核心必须得代码, 拆出到cuda backend去维护 cupy.backends.backend._runtime.pyx

IF CUPY_CANN_VERSION <= 0:
    # Provide access to constants from Python.
    # TODO(kmaehashi): Deprecate aliases above so that we can just do:
    # from cupy.backends.cuda.api._runtime_enum import *
    # from cupy.backends.cuda.api._device_prop import *
    def _export_enum():
        import sys
        import cupy.backends.backend.api._runtime_enum as _runtime_enum
        this = sys.modules[__name__]
        for key in dir(_runtime_enum):
            if not key.startswith('_'):
                setattr(this, key, getattr(_runtime_enum, key))

    _export_enum()
ELSE:
    # in  the future, add ascend cann enum here

cuTensor： 昇腾CANN对应？

#ifdef CUPY_USE_HIP

// Since ROCm/HIP does not have cuTENSOR, we simply include the stubs here
// to avoid code dup.
#include "stub/cupy_cutensor.h"