我正处于OpenCl的学习阶段。
问题:创建两个数组:使用值“-1”初始化的50个元素的输入和输出。
我将两个数组作为输入和输出缓冲区传递给GPU并解决(递增数组的每个元素)5次迭代。因此, 所有元素的值应该增加到4.因为,我将此作业分为10个核心
size_t global_work_size[1] = {10};
我的内核根据get_global_id(0)值解析所有数组,这意味着解决每个核心中的每5个元素。
所以,我的最终输出应该是:
output[0] = 4.0000
output[1] = 4.0000
...
...
output[48] = 4.0000
output[49] = 4.0000
但是,我的输出即将到来:
output[0] = 9.0000
output[1] = 9.0000
output[2] = 9.0000
output[3] = 9.0000
output[4] = 9.0000
output[5] = 4.0000
output[6] = 4.0000
output[7] = 4.0000
output[8] = 4.0000
output[9] = 4.0000
...
...
output[48] = 4.0000
output[49] = 4.0000
正如我们所看到的,输出的差异是前10个元素。这可能是因为我的第一个内核的内核运行了两次。 但是,我无法理解确切的原因。请建议我,我哪里出错了?可能逻辑错了,或者我在OpenCl中错过了一件大事?
program.c 档案:
#include <stdio.h>
#include <stdlib.h>
#ifdef __APPLE__
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif
#define MEM_SIZE (50)
#define MAX_SOURCE_SIZE (0x100000)
int main() {
float input[MEM_SIZE], output[MEM_SIZE];
int go, i;
for (i = 0; i < MEM_SIZE; i++) {
input[i] = -1.0;
output[i] = -1.0;
}
FILE *fp;
cl_device_id device_id = NULL;
cl_context context = NULL;
cl_command_queue command_queue = NULL;
cl_program program = NULL;
cl_kernel kernel = NULL;
cl_platform_id platform_id = NULL;
cl_uint ret_num_devices;
cl_uint ret_num_platforms;
cl_int ret;
size_t source_size;
cl_event event;
char *source_str;
fp = fopen("algebra.cl", "r");
if (!fp) {
fprintf(stderr, "Failed to load kernel.\n");
exit(1);
}
source_str = (char*)malloc(MAX_SOURCE_SIZE);
source_size = fread(source_str, 1, MAX_SOURCE_SIZE, fp);
fclose(fp);
/*Initialization*/
/* Get Platform and Device Info */
ret = clGetPlatformIDs(1, &platform_id, &ret_num_platforms);
ret = clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_DEFAULT, 1, &device_id, &ret_num_devices);
/* Create OpenCL context */
context = clCreateContext(NULL, 1, &device_id, NULL, NULL, &ret);
/* Create Command Queue */
command_queue = clCreateCommandQueue(context, device_id, 0, &ret);
/*Initialization complete*/
cl_mem inputBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR, MEM_SIZE * sizeof(float),(void *) input, NULL);
cl_mem outputBuffer = clCreateBuffer(context, CL_MEM_WRITE_ONLY , MEM_SIZE * sizeof(float), NULL, NULL);
ret = clEnqueueWriteBuffer(command_queue,
inputBuffer,
CL_FALSE,
0,
MEM_SIZE * sizeof(float),
input,
0,
NULL,
&event);
ret = clWaitForEvents(1, &event);
clReleaseEvent(event);
ret = clEnqueueWriteBuffer(command_queue,
outputBuffer,
CL_FALSE,
0,
MEM_SIZE * sizeof(float),
output,
0,
NULL,
&event);
ret = clWaitForEvents(1, &event);
clReleaseEvent(event);
/* Create Kernel Program from the source */
program = clCreateProgramWithSource(context, 1, (const char **)&source_str,(const size_t *)&source_size, &ret);
/* Build Kernel Program */
ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);
/* Create OpenCL Kernel */
kernel = clCreateKernel(program, "calc", &ret);
/* Set OpenCL Kernel Parameters */
ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&inputBuffer);
ret = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&outputBuffer);
/* Execute OpenCL Kernel*/
ret = clEnqueueTask(command_queue, kernel, 0, NULL,NULL);
double io;
size_t global_work_size[1] = {10};
inputBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR, MEM_SIZE * sizeof(float),(void *) output, NULL);
ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&inputBuffer);
ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, global_work_size, NULL, 0, NULL, NULL);
ret = clEnqueueReadBuffer(command_queue, outputBuffer, CL_TRUE, 0, MEM_SIZE * sizeof(float), output, 0, NULL, NULL);
for (go = 0; go < MEM_SIZE; go++) {
printf("output[%d] = %f\n", go, output[go]);
}
printf("\n\n");
/* Finalization */
ret = clFlush(command_queue);
ret = clFinish(command_queue);
ret = clReleaseKernel(kernel);
ret = clReleaseProgram(program);
ret = clReleaseMemObject(inputBuffer);
ret = clReleaseMemObject(outputBuffer);
ret = clReleaseCommandQueue(command_queue);
ret = clReleaseContext(context);
return 1;
}
algebra.cl (内核文件):
__kernel void euler(__global float* in, __global float* out)
{
int idx = get_global_id(0) * 5;
int end_idx = idx + 5;
double j = 0;
int i;
while (j < 5.0) {
j++;
for (i = idx; i < end_idx; i++) {
out[i] = out[i] + 1;
}
}
}
gcc编译:
gcc program.c -o test -l OpenCL -I /usr/local/opencl/AMD-APP-SDK-v2.9-RC-lnx32/include/ -L /usr/local/opencl/AMD-APP-SDK-v2.9-RC-lnx32/lib/x86
答案 0 :(得分:1)
你对自己在做什么有完全错误的理解,所以,你期待一种永远不会发生的行为。
clEnqueueTask(command_queue, kernel, 0, NULL,NULL);
这将仅使用1个工作项运行内核,所以,是的。第一个工作项目正在运行两次。这是你的主要问题。size_t global_work_size[1] = {10};
确实意味着10个工作项。但这不是使用OpenCL的方式。您永远不应该放置10M的物品来处理和运行10个工作项。您应该处理10M的项目以及尽可能多的工作项目(即:10M,1 work-item -> 1 simple action
),否则GPU将99%空闲。此外,这将更容易理解和编程,因为您的内核代码中几乎没有循环。out[i] = out[i] +1;
。因为如果许多工作项改变相同的值,则可能导致竞争条件或错误的结果。请记住,所有工作项可能同时独立运行。