332 lines
8.9 KiB
C++
332 lines
8.9 KiB
C++
|
#include "mat_mul.h"
|
||
|
|
|
||
|
|
#include <stdio.h>
|
||
|
|
#include <CL/cl.h>
|
||
|
|
|
||
|
|
#define CHECK_ERROR(err) \
|
||
|
|
if (err != CL_SUCCESS) { \
|
||
|
|
printf("[%s:%d] OpenCL error %d\n", __FILE__, __LINE__, err); \
|
||
|
|
exit(EXIT_FAILURE); \
|
||
|
|
}
|
||
|
|
|
||
|
|
#define MAX_DEV 4
|
||
|
|
|
||
|
|
static cl_int err;
|
||
|
|
static cl_platform_id platform;
|
||
|
|
static cl_device_id devices[MAX_DEV];
|
||
|
|
static cl_device_id device;
|
||
|
|
static cl_context context;
|
||
|
|
static cl_command_queue queue[MAX_DEV];
|
||
|
|
static cl_program program;
|
||
|
|
static cl_kernel kernel[MAX_DEV];
|
||
|
|
static cl_mem a_d[MAX_DEV], b_d[MAX_DEV], c_d[MAX_DEV];
|
||
|
|
#define CURRENT_DEVICE 0
|
||
|
|
int ndev;
|
||
|
|
static float *A, *B, *C;
|
||
|
|
static int M, N, K;
|
||
|
|
|
||
|
|
|
||
|
|
#define TS 32
|
||
|
|
#define WPT 8 // The amount of work-per-thread, i.e. the thread-coarsening factor
|
||
|
|
// #define TS 4
|
||
|
|
// #define WPT 2
|
||
|
|
#define RTS (TS/WPT) // The reduced tile-size in one dimension
|
||
|
|
|
||
|
|
#define KERNEL 3
|
||
|
|
#define PADDINGX 16
|
||
|
|
#define PADDINGY 16
|
||
|
|
#define TRANSPOSEX 16
|
||
|
|
#define TRANSPOSEY 16
|
||
|
|
|
||
|
|
|
||
|
|
void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {
|
||
|
|
A = _A, B = _B, C = _C;
|
||
|
|
M = _M, N = _N, K = _K;
|
||
|
|
int i,modi_M,end_M;
|
||
|
|
// if(M%ndev!=0)
|
||
|
|
// {
|
||
|
|
// modi_M=((M/ndev)+1);
|
||
|
|
// printf("%s %d modi_M %d \n",__func__,__LINE__,modi_M);
|
||
|
|
// }
|
||
|
|
|
||
|
|
modi_M=M/ndev;
|
||
|
|
|
||
|
|
// Setup kernel arguments
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
err = clSetKernelArg(kernel[i], 0, sizeof(cl_mem), &a_d[i]);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
err = clSetKernelArg(kernel[i], 1, sizeof(cl_mem), &b_d[i]);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
err = clSetKernelArg(kernel[i], 2, sizeof(cl_mem), &c_d[i]);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
if(i==(ndev-1))
|
||
|
|
{
|
||
|
|
end_M=M-modi_M*(ndev-1);
|
||
|
|
err = clSetKernelArg(kernel[i], 3, sizeof(int), &end_M);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
else
|
||
|
|
{
|
||
|
|
err = clSetKernelArg(kernel[i], 3, sizeof(int), &modi_M);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
err = clSetKernelArg(kernel[i], 4, sizeof(int), &N);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
err = clSetKernelArg(kernel[i], 5, sizeof(int), &K);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
// size_t gws[2] = {(size_t)(M+WPT - 1)/WPT/ndev, (size_t)N}, lws[2] = {TS/WPT, TS};
|
||
|
|
size_t gws[2] = {(size_t)((M/ndev)+WPT - 1)/WPT, (size_t)N}, lws[2] = {TS/WPT, TS};
|
||
|
|
// size_t gws[2] = {(size_t)(M/ndev), (size_t)N}, lws[2] = {TS, TS};
|
||
|
|
|
||
|
|
for (int i = 0; i < 2; ++i) {
|
||
|
|
// By OpenCL spec, global work size should be MULTIPLE of local work size
|
||
|
|
// Formula below achieve it
|
||
|
|
// e.g., gws = 25, lws = 16, then (25 + 16 - 1) / 16 * 16 = 40 / 16 * 16 = 2 * 16 = 32
|
||
|
|
gws[i] = (gws[i] + lws[i] - 1) / lws[i] * lws[i];
|
||
|
|
}
|
||
|
|
//printf("gws %ld %ld lws %ld %ld \n",gws[0],gws[1],lws[0],lws[1]);
|
||
|
|
// Run kernel
|
||
|
|
|
||
|
|
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
|
||
|
|
if(i==(ndev-1))
|
||
|
|
{
|
||
|
|
gws[0] = (size_t)(end_M+WPT - 1)/WPT;
|
||
|
|
for (int j = 0; j < 2; ++j) {
|
||
|
|
// By OpenCL spec, global work size should be MULTIPLE of local work size
|
||
|
|
// Formula below achieve it
|
||
|
|
// e.g., gws = 25, lws = 16, then (25 + 16 - 1) / 16 * 16 = 40 / 16 * 16 = 2 * 16 = 32
|
||
|
|
gws[j] = (gws[j] + lws[j] - 1) / lws[j] * lws[j];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
err = clEnqueueNDRangeKernel(queue[i], kernel[i], 2, NULL, gws, lws, 0, NULL, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
|
||
|
|
|
||
|
|
}
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
err = clFinish(queue[i]);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
}
|
||
|
|
|
||
|
|
static void print_platform_info(cl_platform_id platform) {
|
||
|
|
size_t sz;
|
||
|
|
char *buf;
|
||
|
|
CHECK_ERROR(clGetPlatformInfo(platform, CL_PLATFORM_NAME, 0, NULL, &sz));
|
||
|
|
buf = (char*)malloc(sz);
|
||
|
|
CHECK_ERROR(clGetPlatformInfo(platform, CL_PLATFORM_NAME, sz, buf, NULL));
|
||
|
|
printf("Detected OpenCL platform: %s\n", buf);
|
||
|
|
free(buf);
|
||
|
|
}
|
||
|
|
|
||
|
|
static void print_device_info(cl_device_id device) {
|
||
|
|
size_t sz;
|
||
|
|
char *buf;
|
||
|
|
CHECK_ERROR(clGetDeviceInfo(device, CL_DEVICE_NAME, 0, NULL, &sz));
|
||
|
|
buf = (char*)malloc(sz);
|
||
|
|
CHECK_ERROR(clGetDeviceInfo(device, CL_DEVICE_NAME, sz, buf, NULL));
|
||
|
|
printf("Detected OpenCL device: %s\n", buf);
|
||
|
|
free(buf);
|
||
|
|
}
|
||
|
|
|
||
|
|
static cl_program create_and_build_program_with_source(cl_context context, cl_device_id device, const char *file_name) {
|
||
|
|
FILE *file = fopen(file_name, "rb");
|
||
|
|
if (file == NULL) {
|
||
|
|
printf("Failed to open %s\n", file_name);
|
||
|
|
exit(EXIT_FAILURE);
|
||
|
|
}
|
||
|
|
fseek(file, 0, SEEK_END);
|
||
|
|
size_t source_size = ftell(file);
|
||
|
|
rewind(file);
|
||
|
|
char *source_code = (char*)malloc(source_size + 1);
|
||
|
|
size_t ntotal = 0;
|
||
|
|
while (ntotal < source_size) {
|
||
|
|
int nread = fread(source_code, sizeof(char), source_size, file);
|
||
|
|
ntotal += nread;
|
||
|
|
}
|
||
|
|
source_code[source_size] = '\0';
|
||
|
|
fclose(file);
|
||
|
|
cl_program program = clCreateProgramWithSource(context, 1, (const char **)&source_code, &source_size, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
free(source_code);
|
||
|
|
err = clBuildProgram(program, ndev, devices, "", NULL, NULL);
|
||
|
|
if (err == CL_BUILD_PROGRAM_FAILURE) {
|
||
|
|
size_t log_size;
|
||
|
|
CHECK_ERROR(clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size));
|
||
|
|
char *log = (char*)malloc(log_size + 1);
|
||
|
|
CHECK_ERROR(clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, log_size, log, NULL));
|
||
|
|
log[log_size] = 0;
|
||
|
|
printf("Compile error:\n%s\n", log);
|
||
|
|
free(log);
|
||
|
|
}
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
return program;
|
||
|
|
}
|
||
|
|
|
||
|
|
void mat_mul_init(float *A, float *B, float *C, int M, int N, int K) {
|
||
|
|
// Get OpenCL platform
|
||
|
|
int i;
|
||
|
|
err = clGetPlatformIDs(1, &platform, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
print_platform_info(platform);
|
||
|
|
|
||
|
|
// Get OpenCL device
|
||
|
|
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, (unsigned int*) &ndev);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
|
||
|
|
//printf("%s %d ndev %d\n",__func__,__LINE__,ndev);
|
||
|
|
|
||
|
|
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, ndev, devices, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
device = devices[CURRENT_DEVICE];
|
||
|
|
print_device_info(device);
|
||
|
|
|
||
|
|
// Create OpenCL context
|
||
|
|
context = clCreateContext(0, ndev, devices, NULL, NULL, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
|
||
|
|
//if(M%4 ==0)
|
||
|
|
// ndev=4;
|
||
|
|
//else if(M%3 ==0)
|
||
|
|
// ndev=3;
|
||
|
|
// else if(M%2 ==0)
|
||
|
|
// ndev=2;
|
||
|
|
//else
|
||
|
|
// ndev=1;
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
// Create OpenCL command queue
|
||
|
|
|
||
|
|
//queue = clCreateCommandQueue(context, device, 0, &err);
|
||
|
|
//CHECK_ERROR(err);
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
queue[i] = clCreateCommandQueue(context, devices[i], 0, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
// Compile program from "kernel.cl"
|
||
|
|
|
||
|
|
// Extract kernel from compiled program
|
||
|
|
program = create_and_build_program_with_source(context, device, "kernel.cl");
|
||
|
|
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
kernel[i] = clCreateKernel(program, "sgemm", &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
// Create GPU buffers
|
||
|
|
//printf("%s %d jjlee check ndev %d \n",__func__,__LINE__,ndev);
|
||
|
|
int modi_M=M;
|
||
|
|
if(M%ndev!=0)
|
||
|
|
{
|
||
|
|
modi_M=((M/ndev)+1)*ndev;
|
||
|
|
// printf("%s %d modi_M %d \n",__func__,__LINE__,modi_M);
|
||
|
|
}
|
||
|
|
|
||
|
|
modi_M=M/ndev;
|
||
|
|
int end_M;
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
if(i==(ndev-1))
|
||
|
|
{
|
||
|
|
end_M=M-modi_M*(ndev-1);
|
||
|
|
a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, end_M * K * sizeof(float), NULL, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
c_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, end_M * N * sizeof(float), NULL, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
else
|
||
|
|
{
|
||
|
|
a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, modi_M * K * sizeof(float), NULL, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
c_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, modi_M * N * sizeof(float), NULL, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
b_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, K * N * sizeof(float), NULL, &err);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
|
||
|
|
}
|
||
|
|
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
// Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
|
||
|
|
|
||
|
|
if(i==(ndev-1))
|
||
|
|
{
|
||
|
|
end_M=M-modi_M*(ndev-1);
|
||
|
|
err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_TRUE, 0, end_M * K * sizeof(float),(void *) ((size_t)A+(( modi_M * K * sizeof(*A))*i)), 0, NULL, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
else{
|
||
|
|
err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_TRUE, 0, modi_M * K * sizeof(float),(void *) ((size_t)A+(( modi_M * K * sizeof(*A))*i)), 0, NULL, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
|
||
|
|
}
|
||
|
|
for (i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
err = clFinish(queue[i]);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
|
||
|
|
|
||
|
|
}
|
||
|
|
|
||
|
|
void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
|
||
|
|
// Read from GPU; c_d (gpu) -> C (cpu)
|
||
|
|
int modi_M=M/ndev;
|
||
|
|
int end_M;
|
||
|
|
for (int i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
// printf("%s %d jjlee check %lf \n",__func__,__LINE__,c_d[i]);
|
||
|
|
|
||
|
|
|
||
|
|
if(i==(ndev-1))
|
||
|
|
{
|
||
|
|
end_M=M-modi_M*(ndev-1);
|
||
|
|
err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, end_M * N * sizeof(float),(void*) ((size_t) C+((modi_M * N * sizeof(float))*i)), 0, NULL, NULL);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
else
|
||
|
|
{
|
||
|
|
err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, modi_M * N * sizeof(float),(void*) ((size_t) C+((modi_M * N * sizeof(float))*i)), 0, NULL, NULL);
|
||
|
|
//printf("%s %d jjlee check i %d \n",__func__,__LINE__,i);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
}
|
||
|
|
for (int i = 0; i < ndev; i++)
|
||
|
|
{
|
||
|
|
err = clFinish(queue[i]);
|
||
|
|
CHECK_ERROR(err);
|
||
|
|
}
|
||
|
|
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
|
||
|
|
|
||
|
|
}
|