想象一下使用关联属性的二元运算(让它命名为“+”)。当你可以计算并行a1 + a2 + a3 + a4 + ...
,第一计算在OpenCL中减少的最佳做法是什么?
b1 = a1 + a2
b2 = a3 + a4
然后
c1 = b1 + b2
c2 = b3 + b4
然后做同样的事情上一步的结果,依此类推,直到还剩下一个元素。
我在学习OpenCL并尝试实现这种方法来总结数组中的所有元素。我是这个技术的全新手,所以这个程序可能看起来很奇怪。
这是内核:
__kernel void reduce (__global float *input, __global float *output)
{
size_t gl = get_global_id (0);
size_t s = get_local_size (0);
int i;
float accum = 0;
for (i=0; i<s; i++) {
accum += input[s*gl+i];
}
output[gl] = accum;
}
这是主程序:
#include <stdio.h>
#include <stdlib.h>
#include <fcntl.h>
#include <unistd.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <CL/cl.h>
#define N (64*64*64*64)
#include <sys/time.h>
#include <stdlib.h>
double gettime()
{
struct timeval tv;
gettimeofday (&tv, NULL);
return (double)tv.tv_sec + (0.000001 * (double)tv.tv_usec);
}
int main()
{
int i, fd, res = 0;
void* kernel_source = MAP_FAILED;
cl_context context;
cl_context_properties properties[3];
cl_kernel kernel;
cl_command_queue command_queue;
cl_program program;
cl_int err;
cl_uint num_of_platforms=0;
cl_platform_id platform_id;
cl_device_id device_id;
cl_uint num_of_devices=0;
cl_mem input, output;
size_t global, local;
cl_float *array = malloc (sizeof (cl_float)*N);
cl_float *array2 = malloc (sizeof (cl_float)*N);
for (i=0; i<N; i++) array[i] = i;
fd = open ("kernel.cl", O_RDONLY);
if (fd == -1) {
perror ("Cannot open kernel");
res = 1;
goto cleanup;
}
struct stat s;
res = fstat (fd, &s);
if (res == -1) {
perror ("Cannot stat() kernel");
res = 1;
goto cleanup;
}
kernel_source = mmap (NULL, s.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
if (kernel_source == MAP_FAILED) {
perror ("Cannot map() kernel");
res = 1;
goto cleanup;
}
if (clGetPlatformIDs (1, &platform_id, &num_of_platforms) != CL_SUCCESS) {
printf("Unable to get platform_id\n");
res = 1;
goto cleanup;
}
if (clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_GPU, 1, &device_id,
&num_of_devices) != CL_SUCCESS)
{
printf("Unable to get device_id\n");
res = 1;
goto cleanup;
}
properties[0]= CL_CONTEXT_PLATFORM;
properties[1]= (cl_context_properties) platform_id;
properties[2]= 0;
context = clCreateContext(properties,1,&device_id,NULL,NULL,&err);
command_queue = clCreateCommandQueue(context, device_id, 0, &err);
program = clCreateProgramWithSource(context, 1, (const char**)&kernel_source, NULL, &err);
if (clBuildProgram(program, 0, NULL, NULL, NULL, NULL) != CL_SUCCESS) {
char buffer[4096];
size_t len;
printf("Error building program\n");
clGetProgramBuildInfo (program, device_id, CL_PROGRAM_BUILD_LOG, sizeof (buffer), buffer, &len);
printf ("%s\n", buffer);
res = 1;
goto cleanup;
}
kernel = clCreateKernel(program, "reduce", &err);
if (err != CL_SUCCESS) {
printf("Unable to create kernel\n");
res = 1;
goto cleanup;
}
// create buffers for the input and ouput
input = clCreateBuffer(context, CL_MEM_READ_ONLY,
sizeof(cl_float) * N, NULL, NULL);
output = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
sizeof(cl_float) * N, NULL, NULL);
// load data into the input buffer
clEnqueueWriteBuffer(command_queue, input, CL_TRUE, 0,
sizeof(cl_float) * N, array, 0, NULL, NULL);
size_t size = N;
cl_mem tmp;
double time = gettime();
while (size > 1)
{
// set the argument list for the kernel command
clSetKernelArg(kernel, 0, sizeof(cl_mem), &input);
clSetKernelArg(kernel, 1, sizeof(cl_mem), &output);
global = size;
local = 64;
// enqueue the kernel command for execution
clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, &global,
&local, 0, NULL, NULL);
clFinish(command_queue);
size = size/64;
tmp = output;
output = input;
input = tmp;
}
cl_float answer[1];
clEnqueueReadBuffer(command_queue, tmp, CL_TRUE, 0,
sizeof(cl_float), array, 0, NULL, NULL);
time = gettime() - time;
printf ("%f %f\n", array[0], time);
cleanup:
free (array);
free (array2);
clReleaseMemObject(input);
clReleaseMemObject(output);
clReleaseProgram(program);
clReleaseKernel(kernel);
clReleaseCommandQueue(command_queue);
clReleaseContext(context);
if (kernel_source != MAP_FAILED) munmap (kernel_source, s.st_size);
if (fd != -1) close (fd);
_Exit (res); // Kludge
return res;
}
所以我重新运行的内核,直到有只有一个缓冲件。这是计算OpenCL中元素总和的正确方法吗?我用gettime
测量的时间比CPU上一个简单循环的执行时间(编译铛4.0.0和-O2 -ffast-math
标志)慢大约10倍。我使用的硬件:Amd Ryzen 5 1600X和Amd Radeon HD 6950.
感谢这种我有用的建议,删除clFinish。至于那篇AMD文章,我可以用它来改进内核,这样它就能更好地分配工作组中的工作并利用本地内存。但我仍然觉得这篇文章令人困惑。例如:为什么我需要重新排序操作(使用操作的交换属性)?据我了解,工作元素加载更紧凑(因此它们之间没有间隙)会更好。那是对的吗?文章谈论什么是SIMD波前? –
查看来自各种GPU制造商(nVidia,AMD,Intel等)的OpenCL优化指南 - 他们很好地介绍了GPU的工作原理,包括术语。 – pmdj
顺便说一句,我发现[this](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.225.1324&rep=rep1&type=pdf)链接。很有用。 –