使用Pycuda：LogicError：cuMemAlloc失败：非法内存访问

Question

我正在使用 Pycuda 创建一个 python 程序，以与 GPU 并行实现 Sobel 算法。 我的代码是：

import pycuda.autoinit
import pycuda.driver as cuda
from pycuda.compiler import SourceModule
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt

def loadImage(image_path):
    # Load the image
    image = Image.open(image_path)  # Convert to grayscale
    image_array = np.asarray(image)
    return image_array

def applySobel(image):
    # Convert image to grayscale if needed
    if len(image.shape) == 3:
        image = np.mean(image, axis=2)

    # Create a GPU device array from the image
    image_gpu = cuda.to_device(image.astype(np.float32))

    # Allocate GPU memory for the output
    output_gpu = cuda.mem_alloc_like(image_gpu)

    # Define the Sobel filter kernels
    sobel_x_kernel = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=np.float32)
    sobel_y_kernel = np.array([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=np.float32)

    # Compile the CUDA kernel
    mod = SourceModule("""
        __global__ void sobelFilter(const float* input, float* output, int width, int height) {
            int row = blockIdx.y * blockDim.y + threadIdx.y;
            int col = blockIdx.x * blockDim.x + threadIdx.x;

            if (row < height && col < width) {
                float sum_x = 0.0;
                float sum_y = 0.0;

                for (int i = -1; i <= 1; ++i) {
                    for (int j = -1; j <= 1; ++j) {
                        int neighbor_row = row + i;
                        int neighbor_col = col + j;

                        if (neighbor_row >= 0 && neighbor_row < height && neighbor_col >= 0 && neighbor_col < width) {
                            float pixel = input[neighbor_row * width + neighbor_col];
                            sum_x += pixel * sobel_x_kernel[i + 1][j + 1];
                            sum_y += pixel * sobel_y_kernel[i + 1][j + 1];
                        }
                    }
                }

                output[row * width + col] = sqrt(sum_x * sum_x + sum_y * sum_y);
            }
        }
    """)

    # Get the compiled kernel function
    sobel_filter_func = mod.get_function("sobelFilter")

    # Set the block and grid dimensions
    block_dim = (16, 16)
    grid_dim = ((image.shape[1] - 1) // block_dim[0] + 1, (image.shape[0] - 1) // block_dim[1] + 1)

    # Call the CUDA kernel
    sobel_filter_func(image_gpu, output_gpu, np.int32(image.shape[1]), np.int32(image.shape[0]), block=(block_dim[0], block_dim[1], 1), grid=(grid_dim[0], grid_dim[1]))

    # Copy the result back to the CPU
    output = np.empty_like(image)
    cuda.memcpy_dtoh(output, output_gpu)

    return output

    # Copy the processed image back to CPU
    processed_image = np.empty_like(image)
    cuda.memcpy_dtoh(processed_image, image_gpu)

    return processed_image

def showImage(image):
    # Display the image
    plt.imshow(image, cmap="gray")
    plt.axis("off")
    plt.show()

#Test the code
image_path = "original_image.PNG"

#Load the image
image = loadImage(image_path)

#Apply Sobel filtering
processed_image = applySobel(image)

#Show the original and processed images
showImage(image)
showImage(processed_image)

尝试运行时，给出：

---------------------------------------------------------------------------
LogicError                                Traceback (most recent call last)
<ipython-input-17-404cd1292d54> in <cell line: 92>()
     90 
     91 #Apply Sobel filtering
---> 92 processed_image = applySobel(image)
     93 
     94 #Show the original and processed images

1 frames
/usr/local/lib/python3.10/dist-packages/pycuda/driver.py in to_device(bf_obj)
   1077     else:
   1078         bf = buffer(bf_obj)
-> 1079     result = mem_alloc(len(bf))
   1080     memcpy_htod(result, bf)
   1081     return result

LogicError: cuMemAlloc failed: an illegal memory access was encountered

我在其他算法中使用了相同的内存管理并且进展顺利，但这个不是。 请问有人可以帮我解决问题吗？

我希望有人提供解决我的问题的代码，或者至少提供一个可供审查的示例。

Answer 1

问题是传递给mem_alloc_like函数的对象必须支持nbytes属性。因此，您需要使用图像对象而不是image_gpu。将相关行更正为：

output_gpu = cuda.mem_alloc_like(image)

我还注意到，在 CUDA 内核中，您引用了字段 sobel_x_kernel 和 sobel_y_kernel，它们没有传递给内核。这是必须要做的。该过程与传输图像时类似。