.cu question

[JinD123]

trafficstars

device float do_separable_conv_y(float pixels[CY][CCX], int H, int W, bool sq = false) {//此时4242变成了4232 auto block = cg::this_thread_block(); int local_y = block.thread_index().y + 5;//这里+5是为了要找到每个核处理的像素中心位置，线程还是那个线程，只是处理的像素以哪里为中心 int local_x = block.thread_index().x + 5;//这里x为什么会有+5？？ float val = 0.0f; // if (threadIdx.x == 31 && blockIdx.x == 0 && blockIdx.y == 0 && blockIdx.z == 0) { // printf("local_y:%d,local_x:%d\n",local_y,local_x); // }

val += G_00 * pixels[local_y - 5][local_x]; val += G_01 * pixels[local_y - 4][local_x]; val += G_02 * pixels[local_y - 3][local_x]; val += G_03 * pixels[local_y - 2][local_x]; val += G_04 * pixels[local_y - 1][local_x]; val += G_05 * pixels[local_y ][local_x]; val += G_06 * pixels[local_y + 1][local_x]; val += G_07 * pixels[local_y + 2][local_x]; val += G_08 * pixels[local_y + 3][local_x]; val += G_09 * pixels[local_y + 4][local_x]; val += G_10 * pixels[local_y + 5][local_x];

return val; }

In the do_separable_conv_y function, the input shared memory array pixels has dimensions [CY][CCX], which is $42 \times 32$. Given that this function performs a vertical (column-wise) convolution, why is the column index calculated as int local_x = block.thread_index().x + 5;?"Since the valid column indices are $0$ to $31$, doesn't adding 5 cause the thread to try accessing columns $32$ through $36$, leading to an array out-of-bounds error? Shouldn't the calculation simply be int local_x = block.thread_index().x;? (Although the code actually runs without error in practice.)"