我是 OpenCL 的新手。
我编写了一个程序,应该对 100 万个元素的数组进行并行归约。在代码的最后一部分,我比较了 CPU sum 和 GPU sum ,它们不一样,这就是问题所在。我的本地尺码是 64。 从索引“90”开始,GPU 中的总和开始变大。
编辑:如果我对较小的数字求和(现在我对 0 - 1m 求和),假设 1 的最终和是正确的。
内核:
__kernel void gpuSumfunc( __global float *vec ,__global float* sum, int n)
{
__local float tempSum[64];
const int i;
const int globalID = get_global_id(0); //BLOCK_DIM*BLOCK_IND+THREAD_ID
const int tid = get_local_id(0); //THREAD_ID
const int BlockDIM = get_local_size(0);//BLOCK_DIM=64
if (globalID < n)
{
tempSum[tid] = vec[globalID]; //Inserting global data to local data
}
else
{
tempSum[tid] = 0;
}
barrier(CLK_LOCAL_MEM_FENCE); //Wating for all the threads to copy their data
for (i = BlockDIM / 2; i > 0; i /= 2)
{
if (tid < i)
{
tempSum[tid] += tempSum[tid + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0)
{
sum[get_group_id(0)] = tempSum[0];
}
}
主要内容:
//HOST-cpu
float *h_a;//input
float *h_b;//output
float *h_s;
//DEVICE-gpu
cl_mem d_a;//input buffer
cl_mem d_b;//Output
//Kernel File
FILE* fileKernel;
//Memory allocation - cpu input
vector = (float*)malloc(n * sizeof(float));
h_a = (float*)malloc(n * sizeof(float));
h_b = (float*)malloc(n * sizeof(float));
h_s = (float*)malloc(n * sizeof(float));
*vector = { 0 };
*h_a = { 0 };
*h_b = { 0 };
*h_s = { 0 };
//Initializing Data for gpu
for (i = 0; i < n; i++) {
h_a[i] = i;//(float)i;
}
//Initializing Data for cpu
for (i = 0; i < n; i++) {
vector[i] = i;//(float)i;
}
fileKernel = fopen("KernelCode.cl", "r");
if (!fileKernel)
{
printf("Cannot open kernel file!\n");
exit(1);
}
// Read kernel code
kernelSource = (char*)malloc(MAX_SOURCE_SIZE);
source_size = fread(kernelSource, 1, MAX_SOURCE_SIZE, fileKernel);
fclose(fileKernel);
error = clGetPlatformIDs(2, cp_Platform, NULL); //array with two devices
error = clGetDeviceIDs(cp_Platform[1], CL_DEVICE_TYPE_GPU, 1, &Device_ID, NULL); // cp_platform[1] = Nvidia GPU
context = clCreateContext(NULL, 1, &Device_ID, NULL, NULL, &error); // creating openCL context
queue = clCreateCommandQueue(context, Device_ID, 0, &error); // creating command queue, executing openCL context on device cp_Platform[1]
globalSize = ceil(n / (float)localSize)*localSize;
d_a = clCreateBuffer(context, CL_MEM_READ_ONLY, n * sizeof(float), NULL, NULL);
d_b = clCreateBuffer(context, CL_MEM_READ_ONLY, n * sizeof(float), NULL, NULL);
error = clEnqueueWriteBuffer(queue, d_a, CL_TRUE, 0, n * sizeof(float), h_a, 0, NULL, NULL); //Enqueue commands to write to a buffer object from host memory.
error |= clEnqueueWriteBuffer(queue, d_b, CL_TRUE, 0,n * sizeof(float), h_s, 0, NULL, NULL); //Enqueue commands to write to a buffer object from host memory.
program = clCreateProgramWithSource(context, 1, (const char **)& kernelSource, (const size_t *)&source_size, &error); //this function creates a program object for this specific openCL context
error = clBuildProgram(program, 0, NULL, NULL, NULL, NULL); //compiles and links a program executable from the program source
kernel = clCreateKernel(program, "gpuSumfunc", &error); //creating kernel object
error = clGetKernelWorkGroupInfo(kernel, Device_ID, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), (void*)&workGroupSize, NULL);
error = clGetKernelWorkGroupInfo(kernel, Device_ID, CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, sizeof(size_t), (void*)&pWorkGroupSize, NULL);
error = clGetDeviceInfo(Device_ID, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(NumOfCU), &NumOfCU, NULL);
error |= clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_a); //Used to set the argument value for a specific argument of a kernel.
error |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &d_b);
error |= clSetKernelArg(kernel, 2, sizeof(int), &n);
error |= clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &globalSize, &localSize, 0, NULL, NULL); // Enqueues a command to execute a kernel on a device.
clFinish(queue);
clEnqueueReadBuffer(queue, d_b, CL_TRUE, 0, n*sizeof(float) , h_b, 0, NULL, NULL); ////writing data from the device (d_b) to host(h_b)
clock_t end = clock();
for (i = 0; i < (n+localSize-1)/localSize; i++)
{
gpuSum += h_b[i];
cpuSum = cpuSumfunc(vector, 64*(i+1));
if ((gpuSum - cpuSum) > Tolerance)
{
printf("\nfailed! for index:%d",i);
printf("\nCPU sum = %f", cpuSum);
printf("\nGPU sum = %f\n", gpuSum);
}
else
{
printf("\nPassed! for index:%d",i);
printf("\nCPU sum: %.2f", cpuSum);
printf("\nGPU sum: %.2f\n", gpuSum);
}
}
// cpu
time_spent = (double)(end - begin) / CLOCKS_PER_SEC;
//printf("\nTotal program's running time is: %.2f\n", time_spent);
free(h_a);
free(h_b);
free(h_s);
free(vector);
//free(kernelSource);
clReleaseProgram(program);
clReleaseContext(context);
clReleaseKernel(kernel);
clReleaseCommandQueue(queue);
}
float cpuSumfunc(float * vec, int n)
{
float sum = 0;
int i;
for (i = 0; i < n; i++)
{
sum += vec[i];
}
return sum;
}
最佳答案
Float32
值对于您的求和运算而言不够准确,并且会出现舍入误差,这在 CPU 和 GPU 设备中会有所不同。
16956560
需要 25 位才能准确表示。
Float32
仅提供 23 位的精度。
这意味着:16956560 + 1 = 16956560 如果该操作在 Float32 中执行。
两种设备的区别在于:
- 排序:CPU 和 GPU 将以不同的顺序求和,具有不同的舍入误差。
- 准确性:大多数 CPU(x86 等)使用内部 48 位浮点运算,然后将其保存为 32 位。而 GPU 以纯 32 位进行所有数学运算。
您可以使用 Float64 (double
) 或使用整数 (int64_t = Long) 来求解。
注意:实际上,您的 GPU 总和比 CPU 总和更准确,因为它首先将小值打包在一起,然后将那些大值与最终总和相加。
关于c++ - OpenCL - GPU 总和和 CPU 总和不一样,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/48095602/