chundoong-lab-ta/SamsungDS22/submissions/HW5/hj614.yoo/mat_mul.cpp

306 lines
9.8 KiB
C++

#include "mat_mul.h"
#include <stdio.h>
#include <CL/cl.h>
#define WORK_SIZE (16)
#define VECT_SIZE (4)
#define MAX_DEV (4)
#define CHECK_ERROR(err) \
if (err != CL_SUCCESS) { \
printf("[%s:%d] OpenCL error %d\n", __FILE__, __LINE__, err); \
exit(EXIT_FAILURE); \
}
static cl_int err;
static cl_platform_id platform;
static cl_device_id device[MAX_DEV];
static cl_context context;
static cl_command_queue queue[MAX_DEV];
static cl_program program;
static cl_kernel kernel[MAX_DEV];
static cl_kernel addPad[MAX_DEV], delPad[MAX_DEV];
static cl_mem a_d[MAX_DEV], b_d[MAX_DEV], c_d[MAX_DEV];
static cl_mem aaa[MAX_DEV], bbb[MAX_DEV], ccc[MAX_DEV];
static float *A, *B, *C;
static int M, N, K;
static int aM, aN, aK;
static uint nDev;
static int base[MAX_DEV];
static int slice;
void SetPaddingArg(cl_kernel *kernel, int P, int Q, int nP, int nQ, cl_mem *input, cl_mem *output) {
err = clSetKernelArg(*kernel, 0, sizeof(int), &P);
CHECK_ERROR(err);
err = clSetKernelArg(*kernel, 1, sizeof(int), &Q);
CHECK_ERROR(err);
err = clSetKernelArg(*kernel, 2, sizeof(int), &nP);
CHECK_ERROR(err);
err = clSetKernelArg(*kernel, 3, sizeof(int), &nQ);
CHECK_ERROR(err);
err = clSetKernelArg(*kernel, 4, sizeof(cl_mem), input);
CHECK_ERROR(err);
err = clSetKernelArg(*kernel, 5, sizeof(cl_mem), output);
CHECK_ERROR(err);
}
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;
size_t gws[2], lws[2];
// A padding
// Setup kernel arguments
for (uint i = 0; i < nDev; ++i) {
SetPaddingArg(&addPad[i],slice,K,slice,aK,&a_d[i],&aaa[i]);
}
// Setup global work size and local work size
gws[0] = slice, gws[1] = aK;
lws[0] = WORK_SIZE, lws[1] = WORK_SIZE;
for (int i = 0; i < 2; ++i) {
gws[i] = (gws[i] + lws[i] - 1) / lws[i] * lws[i];
}
// Run kernel
for (uint i = 0; i < nDev; ++i) {
if (base[i] > M) break;
err = clEnqueueNDRangeKernel(queue[i], addPad[i], 2, NULL, gws, lws, 0, NULL, NULL);
CHECK_ERROR(err);
}
// B padding
// Setup kernel arguments
for (uint i = 0; i < nDev; ++i) {
SetPaddingArg(&addPad[i],K,N,aK,aN,&b_d[i],&bbb[i]);
}
// Setup global work size and local work size
gws[0] = aK, gws[1] = aN;
lws[0] = WORK_SIZE, lws[1] = WORK_SIZE;
for (int i = 0; i < 2; ++i) {
gws[i] = (gws[i] + lws[i] - 1) / lws[i] * lws[i];
}
// Run kernel
for (uint i = 0; i < nDev; ++i) {
if (base[i] > M) break;
err = clEnqueueNDRangeKernel(queue[i], addPad[i], 2, NULL, gws, lws, 0, NULL, NULL);
CHECK_ERROR(err);
}
// mat_mul
// Setup kernel arguments
for (uint i = 0; i < nDev; i++) {
err = clSetKernelArg(kernel[i], 0, sizeof(cl_mem), &aaa[i]);
CHECK_ERROR(err);
err = clSetKernelArg(kernel[i], 1, sizeof(cl_mem), &bbb[i]);
CHECK_ERROR(err);
err = clSetKernelArg(kernel[i], 2, sizeof(cl_mem), &ccc[i]);
CHECK_ERROR(err);
err = clSetKernelArg(kernel[i], 3, sizeof(int), &slice);
CHECK_ERROR(err);
err = clSetKernelArg(kernel[i], 4, sizeof(int), &aN);
CHECK_ERROR(err);
err = clSetKernelArg(kernel[i], 5, sizeof(int), &aK);
CHECK_ERROR(err);
err = clSetKernelArg(kernel[i], 6, sizeof(int), &base[i]);
CHECK_ERROR(err);
}
// Setup global work size and local work size
gws[0] = slice, gws[1] = aN/VECT_SIZE;
lws[0] = WORK_SIZE, lws[1] = WORK_SIZE/VECT_SIZE;
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];
}
// Run kernel
for (uint i = 0; i < nDev; i++) {
if (base[i] > M) break;
err = clEnqueueNDRangeKernel(queue[i], kernel[i], 2, NULL, gws, lws, 0, NULL, NULL);
CHECK_ERROR(err);
}
// Remove padding and gathering
// Setup kernel arguments
for (uint i = 0; i < nDev; ++i) {
SetPaddingArg(&delPad[i],slice,aN,slice,N,&ccc[i],&c_d[i]);
}
// Setup global work size and local work size
gws[0] = slice, gws[1] = aN;
lws[0] = WORK_SIZE, lws[1] = WORK_SIZE;
for (int i = 0; i < 2; ++i) {
gws[i] = (gws[i] + lws[i] - 1) / lws[i] * lws[i];
}
// Run kernel
for (uint i = 0; i < nDev; ++i) {
if (base[i] > M) break;
err = clEnqueueNDRangeKernel(queue[i], delPad[i], 2, NULL, gws, lws, 0, NULL, NULL);
CHECK_ERROR(err);
}
// Check Finish
for (uint 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;
for (uint i = 0; i < nDev; i++) {
CHECK_ERROR(clGetDeviceInfo(device[i], CL_DEVICE_NAME, 0, NULL, &sz));
buf = (char*)malloc(sz);
CHECK_ERROR(clGetDeviceInfo(device[i], 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, device, "", NULL, NULL);
if (err == CL_BUILD_PROGRAM_FAILURE) {
for (uint i = 0; i < nDev; i++) {
size_t log_size;
CHECK_ERROR(clGetProgramBuildInfo(program, device[i], CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size));
char *log = (char*)malloc(log_size + 1);
CHECK_ERROR(clGetProgramBuildInfo(program, device[i], 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) {
aN = (N + WORK_SIZE - 1) / WORK_SIZE * WORK_SIZE;
aK = (K + WORK_SIZE - 1) / WORK_SIZE * WORK_SIZE;
// Get OpenCL platform
err = clGetPlatformIDs(1, &platform, NULL);
CHECK_ERROR(err);
print_platform_info(platform);
// Get OpenCL device
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &nDev);
CHECK_ERROR(err);
if (nDev > MAX_DEV || nDev < 1) nDev = 1;
aM = (M + (WORK_SIZE*nDev) - 1) / (WORK_SIZE*nDev) * (WORK_SIZE*nDev);
slice = aM / nDev;
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, nDev, device, NULL);
CHECK_ERROR(err);
print_device_info(device);
// Create OpenCL context
context = clCreateContext(NULL, nDev, device, NULL, NULL, &err);
CHECK_ERROR(err);
// Create OpenCL command queue
for (uint i = 0; i < nDev; i++) {
base[i] = i * slice;
queue[i] = clCreateCommandQueue(context, device[i], 0, &err);
CHECK_ERROR(err);
}
// Compile program from "kernel.cl"
program = create_and_build_program_with_source(context, device, "kernel.cl");
// Extract kernel from compiled program
for (uint i = 0; i < nDev; i++) {
kernel[i] = clCreateKernel(program, "sgemm", &err);
CHECK_ERROR(err);
addPad[i] = clCreateKernel(program, "addPadding", &err);
CHECK_ERROR(err);
delPad[i] = clCreateKernel(program, "delPadding", &err);
CHECK_ERROR(err);
}
// Create GPU buffers
for (uint i = 0; i < nDev; i++) {
int modM;
a_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, slice * K * sizeof(float), NULL, &err);
CHECK_ERROR(err);
aaa[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, slice * aK * 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);
bbb[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, aK * aN * sizeof(float), NULL, &err);
CHECK_ERROR(err);
c_d[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, slice * N * sizeof(float), NULL, &err);
CHECK_ERROR(err);
ccc[i] = clCreateBuffer(context, CL_MEM_READ_WRITE, slice * aN * sizeof(float), NULL, &err);
CHECK_ERROR(err);
// Write to GPU; A (cpu) -> a_d (gpu), B (cpu) -> b_d (gpu)
err = clEnqueueWriteBuffer(queue[i], b_d[i], CL_TRUE, 0, K * N * sizeof(float), B, 0, NULL, NULL);
CHECK_ERROR(err);
if (base[i] + slice < M)
modM = slice;
else if (base[i] < M)
modM = M - base[i];
else
modM = 0;
if (modM > 0) {
err = clEnqueueWriteBuffer(queue[i], a_d[i], CL_TRUE, 0, modM * K * sizeof(float), A + i * slice * K, 0, NULL, NULL);
CHECK_ERROR(err);
}
}
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
for (uint i = 0; i < nDev; i++) {
err = clFinish(queue[i]);
CHECK_ERROR(err);
}
}
void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
// Read from GPU; c_d (gpu) -> C (cpu)
for (uint i = 0; i < nDev; ++i) {
int modM;
if (base[i] + slice < M)
modM = slice;
else if (base[i] < M)
modM = M - base[i];
else
modM = 0;
if (modM > 0) {
err = clEnqueueReadBuffer(queue[i], c_d[i], CL_TRUE, 0, modM * N * sizeof(float), C + i * slice * N, 0, NULL, NULL);
CHECK_ERROR(err);
}
}
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
for (uint i = 0; i < nDev; i++) {
err = clFinish(queue[i]);
CHECK_ERROR(err);
}
}