书接上回,上一次实验完成了几个实验用例在本地和云端的编译,但云端始终无法调用gpu_selector_v(可能是我自己的问题),后续尝试在本地可以使用gpu_selector_v。
这次的矩阵乘法实验选需要比较cpu时间与gpu时间,故选择在本地完成实验。
实验课练习1:
代码:
(来源:oneAPI_course/gemm_basic.cpp at main · pengzhao-intel/oneAPI_course · GitHub)(经过略微修改,将gpu_selector_{} 改为gpu_selector_v)
#include <chrono>
#include <iostream>
#include <CL/sycl.hpp>
#define random_float() (rand() / double(RAND_MAX))
using namespace std;
using namespace sycl;
// return execution time
double gpu_kernel(float *A, float *B, float *C, int M, int N, int K, int block_size, sycl::queue &q) {
// define the workgroup size and mapping
auto grid_rows = (M + block_size - 1) / block_size * block_size;
auto grid_cols = (N + block_size - 1) / block_size * block_size;
auto local_ndrange = range<2>(block_size, block_size);
auto global_ndrange = range<2>(grid_rows, grid_cols);
double duration = 0.0f;
auto e = q.submit([&](sycl::handler &h) {
h.parallel_for<class k_name_t>(
sycl::nd_range<2>(global_ndrange, local_ndrange), [=](sycl::nd_item<2> index) {
int row = index.get_global_id(0);
int col = index.get_global_id(1);
float sum = 0.0f;
for (int i = 0; i < K; i++) {
sum += A[row * K + i] * B[i * N + col];
}
C[row * N + col] = sum;
});
});
e.wait();
duration += (e.get_profiling_info<info::event_profiling::command_end>() -
e.get_profiling_info<info::event_profiling::command_start>()) /1000.0f/1000.0f;
return(duration);
}
// return execution time
double cpu_kernel(float *cA, float *cB, float *cC, int M, int N, int K) {
double duration = 0.0;
std::chrono::high_resolution_clock::time_point s, e;
// Single Thread Computation in CPU
s = std::chrono::high_resolution_clock::now();
for(int i = 0; i < M; i++) {
for(int j = 0; j < N; j++) {
float sum = 0.0f;
for(int k = 0; k < K; k++) {
sum += cA[i * K + k] * cB[k * N + j];
}
cC[i * N + j] = sum;
}
}
e = std::chrono::high_resolution_clock::now();
duration = std::chrono::duration<float, std::milli>(e - s).count();
return(duration);
}
int verify(float *cpu_res, float *gpu_res, int length){
int err = 0;
for(int i = 0; i < length; i++) {
if( fabs(cpu_res[i] - gpu_res[i]) > 1e-3) {
err++;
printf("\n%lf, %lf", cpu_res[i], gpu_res[i]);
}
}
return(err);
}
int gemm(const int M,
const int N,
const int K,
const int block_size,
const int iterations,
sycl::queue &q) {
cout << "Problem size: c(" << M << "," << N << ") ="
<< " a(" << M << "," << K << ") *"
<< " b(" << K << "," << N << ")\n";
auto A = malloc_shared<float>(M * K, q);
auto B = malloc_shared<float>(K * N, q);
auto C = malloc_shared<float>(M * N, q);
auto C_host = malloc_host<float>(M * N, q);
// init the A/B/C
for(int i=0; i < M * K; i++) {
A[i] = random_float();
}
for(int i=0; i < K * N; i++) {
B[i] = random_float();
}
for(int i=0; i < M * N; i++) {
C[i] = 0.0f;
C_host[i] = 0.0f;
}
double flopsPerMatrixMul
= 2.0 * static_cast<double>(M) * static_cast<double>(N) * static_cast<double>(K);
double duration_gpu = 0.0f;
double duration_cpu = 0.0f;
// GPU compuation and timer
int warmup = 10;
for (int run = 0; run < iterations + warmup; run++) {
float duration = gpu_kernel(A, B, C, M, N, K, block_size, q);
if(run >= warmup) duration_gpu += duration;
}
duration_gpu = duration_gpu / iterations;
// CPU compuation and timer
warmup = 2;
for(int run = 0; run < iterations/2 + warmup; run++) {
float duration = cpu_kernel(A, B, C_host, M, N, K);
if(run >= warmup) duration_cpu += duration;
}
duration_cpu = duration_cpu / iterations/2;
// Compare the resutls of CPU and GPU
int errCode = 0;
errCode = verify(C_host, C, M*N);
if(errCode > 0) printf("\nThere are %d errors\n", errCode);
printf("\nPerformance Flops = %lf, \n"
"GPU Computation Time = %lf (ms); \n"
"CPU Computaiton Time = %lf (ms); \n",
flopsPerMatrixMul, duration_gpu, duration_cpu);
free(A, q);
free(B, q);
free(C, q);
free(C_host, q);
return(errCode);
}
int main() {
auto propList = cl::sycl::property_list {cl::sycl::property::queue::enable_profiling()};
queue my_gpu_queue( cl::sycl::gpu_selector_v , propList);
int errCode = gemm(2000, 2000, 2000, 4, 10, my_gpu_queue);
return(errCode);
}
可以自己尝试修改M,N,K的值(int errCode = gemm(2000, 2000, 2000, 4, 10, my_gpu_queue);中的前三个值),这几个值是修改矩阵的参数,尝试不同的值并比较结果。
下面是M=N=K=2000的示例结果:
实验课练习2:
代码:
(来源:https://github.com/pengzhao-intel/oneAPI_course/blob/main/code/gemm_tile.cpp#L12)(修改同上)
#include <chrono>
#include <iostream>
#include <CL/sycl.hpp>
#define random_float() (rand() / double(RAND_MAX))
using namespace std;
using namespace sycl;
#define tileY 2
#define tileX 2
// return execution time
double gpu_kernel(float *A, float *B, float *C,
int M, int N, int K,
int BLOCK, sycl::queue &q) {
// define the workgroup size and mapping
auto grid_rows = M / tileY;
auto grid_cols = N / tileX;
auto local_ndrange = range<2>(BLOCK, BLOCK);
auto global_ndrange = range<2>(grid_rows, grid_cols);
double duration = 0.0f;
auto e = q.submit([&](sycl::handler &h) {
h.parallel_for<class k_name_t>(
sycl::nd_range<2>(global_ndrange, local_ndrange), [=](sycl::nd_item<2> index) {
int row = tileY * index.get_global_id(0);
int col = tileX * index.get_global_id(1);
float sum[tileY][tileX] = {0.0f};
float subA[tileY] = {0.0f};
float subB[tileX] = {0.0f};
// core computation
for (int k = 0; k < N; k++) {
// read data to register
for(int m = 0; m < tileY; m++) {
subA[m] = A[(row + m) * N + k];
}
for(int p = 0; p < tileX; p++) {
subB[p] = B[k * N + p + col];
}
for (int m = 0; m < tileY; m++) {
for (int p = 0; p < tileX; p++) {
sum[m][p] += subA[m] * subB[p];
}
}
} //end of K
// write results back
for (int m = 0; m < tileY; m++) {
for (int p = 0; p < tileX; p++) {
C[(row + m) * N + col + p] = sum[m][p];
}
}
});
});
e.wait();
duration += (e.get_profiling_info<info::event_profiling::command_end>() -
e.get_profiling_info<info::event_profiling::command_start>()) /1000.0f/1000.0f;
return(duration);
}
// return execution time
double cpu_kernel(float *cA, float *cB, float *cC, int M, int N, int K) {
double duration = 0.0;
std::chrono::high_resolution_clock::time_point s, e;
// Single Thread Computation in CPU
s = std::chrono::high_resolution_clock::now();
for(int i = 0; i < M; i++) {
for(int j = 0; j < N; j++) {
float sum = 0.0f;
for(int k = 0; k < K; k++) {
sum += cA[i * K + k] * cB[k * N + j];
}
cC[i * N + j] = sum;
}
}
e = std::chrono::high_resolution_clock::now();
duration = std::chrono::duration<float, std::milli>(e - s).count();
return(duration);
}
int verify(float *cpu_res, float *gpu_res, int length){
int err = 0;
for(int i = 0; i < length; i++) {
if( fabs(cpu_res[i] - gpu_res[i]) > 1e-3) {
err++;
printf("\n%lf, %lf", cpu_res[i], gpu_res[i]);
}
}
return(err);
}
int gemm(const int M,
const int N,
const int K,
const int block_size,
const int iterations,
sycl::queue &q) {
cout << "Problem size: c(" << M << "," << N << ") ="
<< " a(" << M << "," << K << ") *"
<< " b(" << K << "," << N << ")\n";
auto A = malloc_shared<float>(M * K, q);
auto B = malloc_shared<float>(K * N, q);
auto C = malloc_shared<float>(M * N, q);
auto C_host = malloc_host<float>(M * N, q);
// init the A/B/C
for(int i=0; i < M * K; i++) {
A[i] = random_float();
}
for(int i=0; i < K * N; i++) {
B[i] = random_float();
}
for(int i=0; i < M * N; i++) {
C[i] = 0.0f;
C_host[i] = 0.0f;
}
double flopsPerMatrixMul
= 2.0 * static_cast<double>(M) * static_cast<double>(N) * static_cast<double>(K);
double duration_gpu = 0.0f;
double duration_cpu = 0.0f;
// GPU compuation and timer
int warmup = 10;
for (int run = 0; run < iterations + warmup; run++) {
float duration = gpu_kernel(A, B, C, M, N, K, block_size, q);
if(run >= warmup) duration_gpu += duration;
}
duration_gpu = duration_gpu / iterations;
// CPU compuation and timer
warmup = 2;
for(int run = 0; run < iterations/2 + warmup; run++) {
float duration = cpu_kernel(A, B, C_host, M, N, K);
if(run >= warmup) duration_cpu += duration;
}
duration_cpu = duration_cpu / iterations/2;
// Compare the resutls of CPU and GPU
int errCode = 0;
errCode = verify(C_host, C, M*N);
if(errCode > 0) printf("\nThere are %d errors\n", errCode);
printf("\nGEMM size M = %d, N = %d, K = %d", M, N, K);
printf("\nWork-Group size = %d * %d, tile_X = %d, tile_Y = %d", block_size, block_size, tileX, tileY);
printf("\nPerformance Flops = %lf, \n"
"GPU Computation Time = %lf (ms); \n"
"CPU Computaiton Time = %lf (ms); \n",
flopsPerMatrixMul, duration_gpu, duration_cpu);
free(A, q);
free(B, q);
free(C, q);
free(C_host, q);
return(errCode);
}
int main() {
auto propList = cl::sycl::property_list {cl::sycl::property::queue::enable_profiling()};
queue my_gpu_queue( cl::sycl::gpu_selector_v , propList);
int errCode = gemm(512, 512, 512, /* GEMM size, M, N, K */
4, /* workgroup size */
10, /* repeat time */
my_gpu_queue);
return(errCode);
}
测试不同tile_X和tile_Y大小下( tile_X和tile_Y可以不一致),矩阵计算的性能
(2,2)
(2,4)
(4,2)
(4,4)
(2,8)
(8,2)
(8,8)
分析:
1、首先不论x,y的值怎么变化对CPU时间并没有多大的影响(暂且看来,还没有尝试极大的数字)。
2、通过比对(2,2),(2,4),(2,8)这几组发现,y数值变大对于GPU的时间有影响但并不大。
3、通过比对(2,2),(4,2),(8,2)这几组发现,x数值变大对于GPU的时间影响很大。
4、通过比对(2,2),(4,4),(8,8)这几组发现,x,y数值同时变大对GPU时间影响要比单独增大x或y来的大,且有成倍减小趋势,猜测是由于2,4,8都为2的指数倍数。
5、当x,y不是2的整数幂时,没有结果(可自行尝试)。