我正在尝试将用Cuda编写的代码转换为openCL并遇到麻烦。我的最终目标是在带有Mali T628 GPU的Odroid XU3板上实现代码。
为了简化转换并节省尝试调试openCL内核的时间，我完成了以下步骤：
我知道不同的架构可能有不同的优化，但这不是我现在主要关注的问题。我想在我的Nvidia GPU上运行openCL代码没有明显的问题，但在尝试运行Odroid板上的代码时一直遇到奇怪的错误。我知道不同的架构对异常等有不同的处理方式，但我不知道如何解决这些问题。
由于openCL代码适用于我的Nvidia，我认为我设法在线程/块之间进行了正确的转换 - ＆gt; workItems / workGroups等 我已经修复了几个与cl_device_max_work_group_size问题相关的问题，因此不能成为cuase。
运行代码时出现“CL_OUT_OF_RESOURCES”错误。我已将错误的原因缩小到代码中的2行，但不确定是否已解决这些问题。
错误是由以下行引起的：
我的内核代码是：
#define ALIGN_IMAGE_WIDTH 64
#define NUM_PIXEL_PER_THREAD 4
#define MIN_DISPARITY 0
#define MAX_DISPARITY 55
#define WINDOW_SIZE 19
#define WINDOW_RADIUS (WINDOW_SIZE / 2)
#define TILE_SHARED_MEM_WIDTH 96
#define TILE_SHARED_MEM_HEIGHT 32
#define TILE_BOUNDARY_WIDTH 64
#define TILE_BOUNDARY_HEIGHT (2 * WINDOW_RADIUS)
#define BLOCK_WIDTH (TILE_SHARED_MEM_WIDTH - TILE_BOUNDARY_WIDTH)
#define BLOCK_HEIGHT (TILE_SHARED_MEM_HEIGHT - TILE_BOUNDARY_HEIGHT)
#define THREAD_NUM_WIDTH 8
#define THREADS_NUM_HEIGHT TILE_SHARED_MEM_HEIGHT
//TODO fix input arguments
__kernel void hello_kernel( __global unsigned char* d_leftImage,
__global unsigned char* d_rightImage,
__global float* d_disparityLeft) {
int blockX = get_group_id(0);
int blockY = get_group_id(1);
int threadX = get_local_id(0);
int threadY = get_local_id(1);
__local unsigned char leftImage [TILE_SHARED_MEM_WIDTH * TILE_SHARED_MEM_HEIGHT];
__local unsigned char rightImage [TILE_SHARED_MEM_WIDTH * TILE_SHARED_MEM_HEIGHT];
__local unsigned int partialDiffSum [BLOCK_WIDTH * TILE_SHARED_MEM_HEIGHT];
int alignedImageWidth = 640;
int partialDiffSumTemp;
float bestDisparity[4] = {0,0,0,0};
int lowestDist[4];
lowestDist[0] = 214748364;
lowestDist[1] = 214748364;
lowestDist[2] = 214748364;
lowestDist[3] = 214748364;
// Read image blocks into shared memory. read is done at 32bit integers on a uchar array. each thread reads 3 integers(12byte) 96/12=8threads
int sharedMemIdx = threadY * TILE_SHARED_MEM_WIDTH + 4 * threadX;
int globalMemIdx = (blockY * BLOCK_HEIGHT + threadY) * alignedImageWidth + blockX * BLOCK_WIDTH + 4 * threadX;
for (int i = 0; i < 4; i++) {
leftImage [sharedMemIdx + i ] = d_leftImage [globalMemIdx + i];
leftImage [sharedMemIdx + 4 * THREAD_NUM_WIDTH + i ] = d_leftImage [globalMemIdx + 4 * THREAD_NUM_WIDTH + i];
leftImage [sharedMemIdx + 8 * THREAD_NUM_WIDTH + i ] = d_leftImage [globalMemIdx + 8 * THREAD_NUM_WIDTH + i];
rightImage[sharedMemIdx + i ] = d_rightImage[globalMemIdx + i];
rightImage[sharedMemIdx + 4 * THREAD_NUM_WIDTH + i ] = d_rightImage[globalMemIdx + 4 * THREAD_NUM_WIDTH + i];
rightImage[sharedMemIdx + 8 * THREAD_NUM_WIDTH + i ] = d_rightImage[globalMemIdx + 8 * THREAD_NUM_WIDTH + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
int imageIdx = sharedMemIdx + TILE_BOUNDARY_WIDTH - WINDOW_RADIUS;
int partialSumIdx = threadY * BLOCK_WIDTH + 4 * threadX;
for(int dispLevel = MIN_DISPARITY; dispLevel <= MAX_DISPARITY; dispLevel++) {
// horizontal partial sum
partialDiffSumTemp = 0;
#pragma unroll
for(int i = imageIdx - WINDOW_RADIUS; i <= imageIdx + WINDOW_RADIUS; i++) {
//partialDiffSumTemp += calcDiff(leftImage [i], rightImage[i - dispLevel]);
partialDiffSumTemp += abs(leftImage[i] - rightImage[i - dispLevel]);
}
partialDiffSum[partialSumIdx] = partialDiffSumTemp;
barrier(CLK_LOCAL_MEM_FENCE);
for (int pixelNum = 1, i = imageIdx - WINDOW_RADIUS; pixelNum < NUM_PIXEL_PER_THREAD; pixelNum++, i++) {
partialDiffSum[partialSumIdx + pixelNum] = partialDiffSum[partialSumIdx + pixelNum - 1] +
abs(leftImage[i + WINDOW_SIZE] - rightImage[i - dispLevel + WINDOW_SIZE]) -
abs(leftImage[i] - rightImage[i - dispLevel]);
}
barrier(CLK_LOCAL_MEM_FENCE);
// vertical sum
if(threadY >= WINDOW_RADIUS && threadY < TILE_SHARED_MEM_HEIGHT - WINDOW_RADIUS) {
for (int pixelNum = 0; pixelNum < NUM_PIXEL_PER_THREAD; pixelNum++) {
int rowIdx = partialSumIdx - WINDOW_RADIUS * BLOCK_WIDTH;
partialDiffSumTemp = 0;
for(int i = -WINDOW_RADIUS; i <= WINDOW_RADIUS; i++,rowIdx += BLOCK_WIDTH) {
partialDiffSumTemp += partialDiffSum[rowIdx + pixelNum];
}
if (partialDiffSumTemp < lowestDist[pixelNum]) {
lowestDist[pixelNum] = partialDiffSumTemp;
bestDisparity[pixelNum] = dispLevel - 1;
}
}
}
}
if (threadY >= WINDOW_RADIUS && threadY < TILE_SHARED_MEM_HEIGHT - WINDOW_RADIUS && blockY < 32) {
d_disparityLeft[globalMemIdx + TILE_BOUNDARY_WIDTH - WINDOW_RADIUS + 0] = bestDisparity[0];
d_disparityLeft[globalMemIdx + TILE_BOUNDARY_WIDTH - WINDOW_RADIUS + 1] = bestDisparity[1];
d_disparityLeft[globalMemIdx + TILE_BOUNDARY_WIDTH - WINDOW_RADIUS + 2] = bestDisparity[2];
d_disparityLeft[globalMemIdx + TILE_BOUNDARY_WIDTH - WINDOW_RADIUS + 3] = bestDisparity[3];
}
}
感谢所有帮助
尤瓦
答案 0 :(得分：0)
根据我的经验，NVidia GPU并不总是在绑定访问时崩溃，很多时候内核仍会返回预期的结果。
使用printf检查索引。如果您安装了Nvidia OpenCL 1.2驱动程序printf应该可用作核心功能。据我检查，Mali-T628使用OpenCL 1.1，然后检查printf是否可用作供应商扩展。您也可以在printf可用的AMD / Intel CPU上运行内核（OpenCL 1.2 / 2.0）。
检查索引的替代方法可以是将__global int* debug数组传递到存储索引的位置，然后在主机上进行检查。确保将其分配得足够大，以便记录超出范围的索引。