问计算得不到结果

EN

__global__
void
convolve_component_EXTEND_kern(const JSAMPLE *data, // image data
                           ssize_t data_width, // image width
                           ssize_t data_height, // image height
                           const float *kern, // convolution kernel data
                           ssize_t kern_w_f, // convolution kernel has a width of 2 * kern_w_f + 1
                           ssize_t kern_h_f, // convolution_kernel has a height of 2 * kern_h_f + 1
                           JSAMPLE *res) // array to store the result
{
ssize_t i = ::blockIdx.x * ::blockDim.x + ::threadIdx.x;
ssize_t j = ::blockIdx.y * ::blockDim.y + ::threadIdx.y;

float value = 0;

for (ssize_t m = 0; m < 2 * kern_w_f + 1; m++) {
    for (ssize_t n = 0; n < 2 * kern_h_f + 1; n++) {
            ssize_t x = i + m - kern_w_f; // column index for this contribution to convolution sum for (i, j)
            ssize_t y = j + n - kern_h_f; // row index for ...
            x = x < 0 ? 0 : (x >= data_width ? data_width - 1 : x);
            y = y < 0 ? 0 : (y >= data_height ? data_height - 1 : y);
            value += ((float) data[data_width * y + x]) * kern[(2 * kern_w_f + 1) * n + m];
    }
}

res[data_width * j + i] = (JSAMPLE) value;
}

void
convolve_component_EXTEND_cuda(const JSAMPLE *data,
                           ssize_t data_width,
                           ssize_t data_height,
                           const float *kern,
                           ssize_t kern_w_f,
                           ssize_t kern_h_f,
                           JSAMPLE *res)
{
JSAMPLE *d_data;
cudaMallocManaged(&d_data,
                  data_width * data_height * sizeof(JSAMPLE));
cudaMemcpy(d_data,
           data,
           data_width * data_height * sizeof(JSAMPLE),
           cudaMemcpyHostToDevice);

float *d_kern;
cudaMallocManaged(&d_kern,
                  (2 * kern_w_f + 1) * (2 * kern_h_f + 1) * sizeof(float));
cudaMemcpy(d_kern,
           kern,
           (2 * kern_w_f + 1) * (2 * kern_h_f + 1) * sizeof(float),
           cudaMemcpyHostToDevice);

JSAMPLE *d_res;
cudaMallocManaged(&d_res,
                  data_width * data_height * sizeof(JSAMPLE));

dim3 threadsPerBlock(16, 16);  // can be adjusted to 32, 32 (1024 threads per block is the maximum)
dim3 numBlocks(data_width / threadsPerBlock.x,
               data_height / threadsPerBlock.y);
convolve_component_EXTEND_kern<<<numBlocks, threadsPerBlock>>>(d_data,
                                                               data_width,
                                                               data_height,
                                                               d_kern,
                                                               kern_w_f,
                                                               kern_h_f,
                                                               d_res);

cudaDeviceSynchronize();

cudaMemcpy(d_res,
           res,
           data_width * data_height * sizeof(JSAMPLE),
           cudaMemcpyDeviceToHost);
cudaFree(d_data);
cudaFree(d_kern);
cudaFree(d_res);
}