传递给内核的 cupy 变量被忽略

Question

我已经修改了一个 cupy 示例来测试一个简单的函数，但是有些变量似乎没有取正确的值。这是代码：

import cupy as cp
import numpy as np
import sys

from cupy import prof
from timeit import default_timer as timer


_cupy_preprocessing_src = r"""
extern "C"
{
    __global__ void _cupy_preprocessing(
            const float * __restrict__ toNormalize,
            float * __restrict__ normalized,
            const int w,
            const int h,
            const float B,
            const float G,
            const float R)
    {
        const int tx { static_cast<int>(blockIdx.x * blockDim.x + threadIdx.x) };
        const int stride { static_cast<int>(blockDim.x * gridDim.x) };

        for(int tid = tx; tid < (w * h); tid += stride)
        {
            normalized[tid] = toNormalize[tid + w * h * 2] * 255.0 - B;
            normalized[tid + w * h] = toNormalize[tid + w * h] * 255.0 - G;
            normalized[tid + w * h * 2] = toNormalize[tid] * 255.0 - R; 
        }
    }
}
"""


def _preprocessing(toNorm, norm, w, h, B, G, R):
    device_id = cp.cuda.Device()
    numSM = device_id.attributes["MultiProcessorCount"]
    threadsperblock = (128, )
    blockspergrid = (numSM * 20, )

    module = cp.RawModule(code=_cupy_preprocessing_src, options=("-std=c++11"))
    kernel = module.get_function("_cupy_preprocessing")

    kernel_args = (toNorm, norm, w, h, B, G, R)

    kernel(blockspergrid, threadsperblock, kernel_args)

    cp.cuda.runtime.deviceSynchronize()


def gpu_preprocessing(toNorm, w, h, B, G, R):
    norm = cp.empty(toNorm.shape, dtype=toNorm.dtype)

    _preprocessing(toNorm, norm, w, h, B, G, R)

    return norm


def cpu_preprocessing(toNorm, w, h, B, G, R):
    norm = np.empty(toNorm.shape, dtype=toNorm.dtype)
    for i in range(w * h):
        norm[i] = toNorm[i + w * h * 2] * 255.0 - B;
        norm[i + w * h] = toNorm[i + w * h] * 255.0 - G;
        norm[i + w * h * 2] = toNorm[i] * 255.0 - R;

    return norm


if __name__ == "__main__":
    w = 512
    h = 512
    B = 1.0
    G = 1.0
    R = 1.0
    x = np.zeros((w * h * 3, ), dtype=np.float32)
    x[:w * h] = np.ones((w * h, ), dtype=np.float32)
    x[w * h:w * h * 2] = np.ones((w * h, ), dtype=np.float32) + 1.0
    x[w * h * 2:] = np.ones((w * h, ), dtype=np.float32) + 2.0
    d_x = cp.array(x)

    start = timer()
    cpu_ppre = cpu_preprocessing(x, w, h, B, G, R)
    end = timer()
    print("CPU time: {:f}".format(end - start))

    start = timer()
    gpu_ppre = gpu_preprocessing(d_x, w, h, B, G, R)
    end = timer()
    print("GPU time: {:f}".format(end - start))

    gpu_ppre = cp.asnumpy(gpu_ppre)

    print(cpu_ppre)
    print(gpu_ppre)

如果B = G = R = 0.0、cpu_preprocessing和gpu_preprocessing、return是同一个数组，而如果B、G和R 不同于零，cpu_preprocessing return 是期望值，而 gpu_preprocessing 似乎忽略了 B、G 和 R。我错过了什么吗？

Answer 1

尝试换行

kernel_args = (toNorm, norm, w, h, B, G, R)

到

kernel_args = (toNorm, norm, w, h, cp.float32(B), cp.float32(G), cp.float32(R))

看看结果是不是固定的。您正在传递 Python 浮点数，但我认为 CuPy 无法推断要转换为的正确位宽。

此外，您漏掉了一个逗号（应该是元组）：options=("-std=c++11", ).

传递给内核的 cupy 变量被忽略

cupy variables passed to kernel ignored

python

cupy