Skip to main content

用于调用ascendc编写的算子

Project description

1 功能描述

由于在ascendc算子开发过程中运行算子比较复杂,为了简化算子的运行,将运行算子变成可以用python直接调用的函数。所以编写了此代码。

2 安装

pip install l0n0lacl

3 运行算子实例

3.1 先切换到cann环境,比如我的环境是:

source /home/HwHiAiUser/Ascend/ascend-toolkit/set_env.sh

3.2 先安装我们编写的算子

bash custom_opp_xxx_aarch64.run

3.3 创建算子运行器

from l0n0lacl import *
ascendc_gelu = OpRunner("Gelu", op_path_prefix='customize')

3.4 调用算子

3.4.1 先看调用传参顺序

在算子工程编译后,会有代码生成,在算子工程目录: ${算子目录}/build_out/autogen/aclnn_xxx.h中可以找到aclnnXXXGetWorkspaceSize函数。以Gelu为例:

__attribute__((visibility("default")))
aclnnStatus aclnnGeluGetWorkspaceSize(
    const aclTensor *x,
    const aclTensor *out,
    uint64_t *workspaceSize,
    aclOpExecutor **executor);

可以看到参数为 x, out, workspaceSize, executor。其中 workspaceSize, executor不需要管。

3.4.2 调用算子

import torch
from l0n0lacl import *
ascendc_gelu = OpRunner("Gelu", op_path_prefix='customize')
target_dtype = torch.float
x = torch.empty(shape, dtype=target_dtype).uniform_(-1, 1)
y = torch.empty(shape, dtype=target_dtype).zero_()
out = ascendc_gelu(x.numpy(), y.numpy()).to_cpu()
print(out)

4. api参考

4.1 AclNDTensor

class AclNDTensor:
    def __init__(self, np_array: np.ndarray):
        pass
    def to_cpu(self):
        pass

numpy ndarray与ascend nd tensor间的桥梁

4.1.1 __init__

  • np_array: numpy的tensor

4.1.2 to_cpu

将运算结果从npu拷贝到cpu

4.2 OpRunner

class OpRunner:
    def __init__(self, name, op_path_prefix='customize', op_path=None, device_id=0) -> None:
        pass
    def __call__(self, *args, outCout=1, argtypes=None, stream=None) -> Union[AclNDTensor, List[AclNDTensor]]:
        pass
    def sync_stream(self)->None:
        pass

4.2.1 __init__

  • name:算子名称,
  • op_path_prefix: 算子工程中CMakePresets.json文件中vender_name的值。默认是customize,可以不传
"vendor_name": {
    "type": "STRING",
    "value": "customize"
},
  • op_path: 算子libcust_opapi.so库的绝对位置。不传。
  • device_id: 设备ID。默认0

4.2.2 __call__

  • args: 表示传给aclnnXXXGetWorkspaceSize除了workspaceSize, executor的参数
  • outCout : 表示算子的输出个数。如果输出个数为1,返回一个AclNDTensor。如果输出个数大于1,返回List[AclNDTensor]
  • argtypes: 表示aclnnXXXGetWorkspaceSize的参数ctypes参数类型,对于特别复杂的算子,如果发现调用异常,可以手动指定类型。 比如(仅用于举例,其实可以不传,自动推导就可运行。但是当发现运行异常的情况下,可以自己指定),对于:
__attribute__((visibility("default")))
aclnnStatus aclnnCumsumGetWorkspaceSize(
    const aclTensor *x,
    const aclTensor *axis,
    bool exclusiveOptional,
    bool reverseOptional,
    const aclTensor *out,
    uint64_t *workspaceSize,
    aclOpExecutor **executor);
import ctypes
from l0n0lacl import *
ascendc_cumsum = OpRunner("Cumsum")
target_dtype = np.float32
data_range = (-10, 10)
shape = [100, 3, 2304]
axis_py = 1
exclusive = True
reverse = False
x = np.random.uniform(*data_range, shape).astype(target_dtype)
axis = np.array([axis_py]).astype(np.int32)
golden: np.ndarray = tf.cumsum(x, axis_py, exclusive, reverse, argtypes=[
    ctypes.c_void_p, # x
    ctypes.c_void_p, # axis
    ctypes.c_bool,   # exclusiveOptional
    ctypes.c_bool,   # reverseOptional
    ctypes.c_void_p, # out
    ctypes.c_void_p, # workspaceSize
    ctypes.c_void_p, # executor
]).numpy()
y = np.ones_like(golden, golden.dtype) * 123
ascendc_cumsum(x, axis, exclusive, reverse,  y).to_cpu()
print(y)
  • stream 如果是多stream的情况下,可以自己指定stream: 例如:
import numpy as np
from l0n0lacl import *
ascendc_gelu = OpRunner("Gelu", op_path_prefix='customize')
target_dtype = np.float32
shape = [10, 10]
x = np.random.uniform(-10, 10, shape).astype(target_dtype)
y = np.zeros_like(x, dtype=target_dtype)
with AclStream(0) as stream:
    out = ascendc_gelu(x, y, stream=stream).to_cpu()
print(out)

4.2.3 sync_stream

用于同步stream

4.3 verify_result

参考自:https://gitee.com/ascend/samples/blob/master/operator/AddCustomSample/KernelLaunch/AddKernelInvocationNeo/scripts/verify_result.py

def verify_result(real_result:numpy.ndarray, golden:numpy.ndarray):
    pass

判断精度是否符合 float16: 千分之一 float32: 万分之一 int16,int32,int8: 0

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

l0n0lacl-1.0.0.tar.gz (12.9 kB view details)

Uploaded Source

Built Distribution

l0n0lacl-1.0.0-py3-none-any.whl (9.8 kB view details)

Uploaded Python 3

File details

Details for the file l0n0lacl-1.0.0.tar.gz.

File metadata

  • Download URL: l0n0lacl-1.0.0.tar.gz
  • Upload date:
  • Size: 12.9 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/5.1.1 CPython/3.12.4

File hashes

Hashes for l0n0lacl-1.0.0.tar.gz
Algorithm Hash digest
SHA256 3d09b3318608b4d953b641c6c6bee5d11ae725cbdf7773a90fb69d8d1a35e519
MD5 87db5293be9d8af2a01d2ba1c8a369f4
BLAKE2b-256 ca69304ba98407f10c4c24aae6cd8e0b0f5c3fc6c4cb1bfd03fd727ff56941e2

See more details on using hashes here.

File details

Details for the file l0n0lacl-1.0.0-py3-none-any.whl.

File metadata

  • Download URL: l0n0lacl-1.0.0-py3-none-any.whl
  • Upload date:
  • Size: 9.8 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/5.1.1 CPython/3.12.4

File hashes

Hashes for l0n0lacl-1.0.0-py3-none-any.whl
Algorithm Hash digest
SHA256 c96c3b3730f4418f08b3bdfcd3ee8daec648f4900a38d2e2f8220d4626403129
MD5 c6d6369d201d7acb0621c61f1ebcb192
BLAKE2b-256 b42b7cc8ef0274aa0b7c8f2b107db7f0c858722910ae2d2de39a09275fcd804c

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page