helloworld:vec-add
#include <stdio.h>
#include <cuda.h>
#include "_aux.h"
/* host, add */
void vec_add_host(float *x, float *y, float *z, int N);
/* device function */
__global__ void vec_add(float *x, float *y, float *z, int N)
{
/* 1D block */
int idx = get_tid();
if (idx < N) z[idx] = z[idx] + y[idx] + x[idx];
}
void vec_add_host(float *x, float *y, float *z, int N)
{
int i;
for (i = 0; i < N; i++) z[i] = z[i] + y[i] + x[i];
}
int main()
{
int N = 20000000;
int nbytes = N * sizeof(float);
/* 1D block */
int bs = 256;
/* 2D grid */
int s = ceil(sqrt((N + bs - 1.) / bs));
dim3 grid = dim3(s, s);
float *dx = NULL, *hx = NULL;
float *dy = NULL, *hy = NULL;
float *dz = NULL, *hz = NULL;
int itr = 30;
int i;
double th, td;
/* allocate GPU mem */
cudaMalloc((void **)&dx, nbytes);
cudaMalloc((void **)&dy, nbytes);
cudaMalloc((void **)&dz, nbytes);
if (dx == NULL || dy == NULL || dz == NULL) {
printf("couldn't allocate GPU memory\n");
return -1;
}
printf("allocated %.2f MB on GPU\n", nbytes / (1024.f * 1024.f));
/* alllocate CPU mem */
hx = (float *) malloc(nbytes);
hy = (float *) malloc(nbytes);
hz = (float *) malloc(nbytes);
if (hx == NULL || hy == NULL || hz == NULL) {
printf("couldn't allocate CPU memory\n");
return -2;
}
printf("allocated %.2f MB on CPU\n", nbytes / (1024.f * 1024.f));
/* init */
for (i = 0; i < N; i++) {
hx[i] = 1;
hy[i] = 1;
hz[i] = 1;
}
/* copy data to GPU */
cudaMemcpy(dx, hx, nbytes, cudaMemcpyHostToDevice);
cudaMemcpy(dy, hy, nbytes, cudaMemcpyHostToDevice);
cudaMemcpy(dz, hz, nbytes, cudaMemcpyHostToDevice);
/* call GPU */
cudaDeviceSynchronize();
td = get_time();
for (i = 0; i < itr; i++) vec_add<<<grid, bs>>>(dx, dy, dz, N);
cudaDeviceSynchronize();
td = get_time() - td;
/* CPU */
th = get_time();
for (i = 0; i < itr; i++) vec_add_host(hx, hy, hz, N);
th = get_time() - th;
printf("GPU time: %e, CPU time: %e, speedup: %g\n", td, th, th / td);
cudaFree(dx);
cudaFree(dy);
cudaFree(dz);
free(hx);
free(hy);
free(hz);
return 0;
}
内存管理
#include <stdio.h>
#include <cuda.h>
__global__ void sum(float *x)
{
int tid = threadIdx.x;
x[tid] += 1;
}
int main()
{
int N = 32;
int nbytes = N * sizeof(float);
float *dx = NULL, *hx = NULL;
int i;
/* allocate GPU mem */
cudaMalloc((void **)&dx, nbytes);
if (dx == NULL) {
printf("couldn't allocate GPU memory\n");
return -1;
}
/* allocate CPU host mem: memory copy is faster than malloc */
//hx = (float *)malloc(nbytes);
cudaMallocHost((void **)&hx, nbytes);
if (hx == NULL) {
printf("couldn't allocate CPU memory\n");
return -2;
}
/* init */
printf("hx original: \n");
for (i = 0; i < N; i++) {
hx[i] = i;
printf("%g\n", hx[i]);
}
/* copy data to GPU */
cudaMemcpy(dx, hx, nbytes, cudaMemcpyHostToDevice);
/* call GPU */
sum<<<1, N>>>(dx);
/* let GPU finish */
cudaDeviceSynchronize();
/* copy data from GPU */
cudaMemcpy(hx, dx, nbytes, cudaMemcpyDeviceToHost);
printf("\nhx from GPU: \n");
for (i = 0; i < N; i++) {
printf("%g\n", hx[i]);
}
cudaFree(dx);
//free(hx);
cudaFreeHost(hx);
return 0;
}
规约算法
volatile关键字能防止编译器优化循环展开,造成结果错误。
注意:下面的程序若去掉volatile关键字,则函数reduction_2和reduction_3的计算结果有误。
#include <stdio.h>
#include <cuda.h>
__global__ void reduction_1(const float* x, float* y)
{
__shared__ float sdata[256];
int tid = threadIdx.x;
/*load data to shared mem*/
sdata[tid] = x[tid];
__syncthreads();
/*reduction using shared mem*/
if (tid < 128)sdata[tid] += sdata[tid + 128]; __syncthreads();
if (tid < 64) sdata[tid] += sdata[tid + 64]; __syncthreads();
if (tid < 32) sdata[tid] += sdata[tid + 32]; __syncthreads();
if (tid < 16) sdata[tid] += sdata[tid + 16]; __syncthreads();
if (tid < 8) sdata[tid] += sdata[tid + 8];__syncthreads();
if (tid < 4) sdata[tid] += sdata[tid + 4]; __syncthreads();
if (tid < 2) sdata[tid] += sdata[tid + 2]; __syncthreads();
if (tid == 0) *y = sdata[0] + sdata[1];
}
__global__ void reduction_2(const float* x, float* y)
{
__shared__ volatile float sdata[256];
int tid = threadIdx.x;
/*load data to shared mem*/
sdata[tid] = x[tid];
__syncthreads();
/*reduction using shared mem*/
if (tid < 128) sdata[tid] += sdata[tid + 128]; __syncthreads();
if (tid < 64) sdata[tid] += sdata[tid + 64]; __syncthreads();
if (tid < 32) {
sdata[tid] += sdata[tid + 32];
sdata[tid] += sdata[tid + 16];
sdata[tid] += sdata[tid + 8];
sdata[tid] += sdata[tid + 4];
sdata[tid] += sdata[tid + 2];
sdata[tid] += sdata[tid + 1];
}
if (tid == 0) y[0] = sdata[0];
}
__device__ void warpReduce(volatile float* sdata, int tid)
{
sdata[tid] += sdata[tid + 32];
sdata[tid] += sdata[tid + 16];
sdata[tid] += sdata[tid + 8];
sdata[tid] += sdata[tid + 4];
sdata[tid] += sdata[tid + 2];
sdata[tid] += sdata[tid + 1];
}
__global__ void reduction_3(const float* x, float* y)
{
__shared__ float sdata[256];
int tid = threadIdx.x;
/* load data to shared mem*/
sdata[tid] = x[tid];
__syncthreads();
/*reduction using shared mem*/
if (tid < 128)sdata[tid] += sdata[tid + 128]; __syncthreads();
if (tid < 64) sdata[tid] += sdata[tid + 64]; __syncthreads();
if (tid < 32) warpReduce(sdata, tid);
if (tid == 0) y[0] = sdata[0];
}
int main()
{
int N = 256;
int nbytes = N * sizeof(float);
float* dx = NULL, *hx = NULL;
float* dy = NULL;
int i;
float as = 0;
cudaMalloc((void**)& dx, nbytes);
cudaMalloc((void**)& dy, sizeof(float));
if (dx == NULL || dy == NULL)
return -1;
hx = (float*)malloc(nbytes);
if (hx == NULL)
return -2;
for (i = 0; i < N; i++)
hx[i] = 1;
cudaMemcpy(dx, hx, nbytes, cudaMemcpyHostToDevice);
reduction_1 << <1, N >> > (dx, dy);
//reduction_2 << <1, N >> > (dx, dy);
//reduction_3 << <1, N >> > (dx, dy);
cudaDeviceSynchronize();
cudaMemcpy(&as, dy, sizeof(float), cudaMemcpyDeviceToHost);
printf("answer: 256, calculated by GPU:%g\n", as);
cudaFree(dx);
cudaFree(dy);
free(hx);
return 0;
}
完整的三阶段规约算法
1.块大小, 256:数组长度降低256倍 ——大规模数组依旧很长,例如256万降低到1万
2.对部分和继续使用上一步的算法
3.使用一个块,将最后结果规约
/* asum: sum of all entries of a vector */
#include "aux.h"
typedef double FLOAT;
__device__ void warpReduce(volatile FLOAT *sdata, int tid)
{
sdata[tid] += sdata[tid + 32];
sdata[tid] += sdata[tid + 16];
sdata[tid] += sdata[tid + 8];
sdata[tid] += sdata[tid + 4];
sdata[tid] += sdata[tid + 2];
sdata[tid] += sdata[tid + 1];
}
/* sum all entries in x and asign to y
* block dim must be 256 */
__global__ void asum_stg_1(const FLOAT *x, FLOAT *y, int N)
{
__shared__ FLOAT sdata[256];
int idx = get_tid();
int tid = threadIdx.x;
int bid = get_bid();
/* load data to shared mem */
if (idx < N) {
sdata[tid] = x[idx];
}
else {
sdata[tid] = 0;
}
__syncthreads();
/* reduction using shared mem */
if (tid < 128) sdata[tid] += sdata[tid + 128];
__syncthreads();
if (tid < 64) sdata[tid] += sdata[tid + 64];
__syncthreads();
if (tid < 32) warpReduce(sdata, tid);
if (tid == 0) y[bid] = sdata[0];
}
__global__ void asum_stg_3(FLOAT *x, int N)
{
__shared__ FLOAT sdata[128];
int tid = threadIdx.x;
int i;
sdata[tid] = 0;
/* load data to shared mem */
for (i = 0; i < N; i += 128) {
if (tid + i < N) sdata[tid] += x[i + tid];
}
__syncthreads();
/* reduction using shared mem */
if (tid < 64) sdata[tid] = sdata[tid] + sdata[tid + 64];
__syncthreads();
if (tid < 32) warpReduce(sdata, tid);
__syncthreads();
if (tid == 0) x[0] = sdata[0];
}
/* dy and dz serve as cache: result stores in dz[0] */
void asum(FLOAT *dx, FLOAT *dy, FLOAT *dz, int N)
{
/* 1D block */
int bs = 256;
/* 2D grid */
int s = ceil(sqrt((N + bs - 1.) / bs));
dim3 grid = dim3(s, s);
int gs = 0;
/* stage 1 */
asum_stg_1<<<grid, bs>>>(dx, dy, N);
/* stage 2 */
{
/* 1D grid */
int N2 = (N + bs - 1) / bs;
int s2 = ceil(sqrt((N2 + bs - 1.) / bs));
dim3 grid2 = dim3(s2, s2);
asum_stg_1<<<grid2, bs>>>(dy, dz, N2);
/* record gs */
gs = (N2 + bs - 1.) / bs;
}
/* stage 3 */
asum_stg_3<<<1, 128>>>(dz, gs);
}
FLOAT asum_host(FLOAT *x, int N);
FLOAT asum_host(FLOAT *x, int N)
{
int i;
FLOAT t = 0;
for (i = 0; i < N; i++) t += x[i];
return t;
}
int main(int argc, char **argv)
{
int N = 10000070;
int nbytes = N * sizeof(FLOAT);
FLOAT *dx = NULL, *hx = NULL;
FLOAT *dy = NULL, *dz;
int i, itr = 20;
FLOAT asd = 0, ash;
double td, th;
if (argc == 2) {
int an;
an = atoi(argv[1]);
if (an > 0) N = an;
}
/* allocate GPU mem */
cudaMalloc((void **)&dx, nbytes);
cudaMalloc((void **)&dy, sizeof(FLOAT) * ((N + 255) / 256));
cudaMalloc((void **)&dz, sizeof(FLOAT) * ((N + 255) / 256));
if (dx == NULL || dy == NULL || dz == NULL) {
printf("couldn't allocate GPU memory\n");
return -1;
}
printf("allocated %e MB on GPU\n", nbytes / (1024.f * 1024.f));
/* alllocate CPU mem */
hx = (FLOAT *) malloc(nbytes);
if (hx == NULL) {
printf("couldn't allocate CPU memory\n");
return -2;
}
printf("allocated %e MB on CPU\n", nbytes / (1024.f * 1024.f));
/* init */
for (i = 0; i < N; i++) {
hx[i] = 1;
}
/* copy data to GPU */
cudaMemcpy(dx, hx, nbytes, cudaMemcpyHostToDevice);
/* let dust fall */
cudaDeviceSynchronize();
td = get_time();
/* call GPU */
for (i = 0; i < itr; i++) asum(dx, dy, dz, N);
/* let GPU finish */
cudaDeviceSynchronize();
td = get_time() - td;
th = get_time();
for (i = 0; i < itr; i++) ash = asum_host(hx, N);
th = get_time() - th;
/* copy data from GPU */
cudaMemcpy(&asd, dz, sizeof(FLOAT), cudaMemcpyDeviceToHost);
printf("asum, answer: %d, calculated by GPU:%f, calculated by CPU:%f\n", N, asd, ash);
printf("GPU time: %e, CPU time: %e, speedup: %g\n", td, th, th / td);
cudaFree(dx);
cudaFree(dy);
cudaFree(dz);
free(hx);
return 0;
}
#include "_aux.h"
#include <assert.h>
/* host, add */
float dot_host(float *x, float *y, int N)
{
int i;
float t = 0;
assert(x != NULL);
assert(y != NULL);
for (i = 0; i < N; i++) t += x[i] * y[i];
return t;
}
__device__ void warpReduce(volatile float *sdata, int tid)
{
sdata[tid] += sdata[tid + 32];
sdata[tid] += sdata[tid + 16];
sdata[tid] += sdata[tid + 8];
sdata[tid] += sdata[tid + 4];
sdata[tid] += sdata[tid + 2];
sdata[tid] += sdata[tid + 1];
}
/* partial dot product */
__global__ void dot_stg_1(const float *x, float *y, float *z, int N)
{
__shared__ float sdata[256];
int idx = get_tid();
int tid = threadIdx.x;
int bid = get_bid();
/* load data to shared mem */
if (idx < N) {
sdata[tid] = x[idx] * y[idx];
}
else {
sdata[tid] = 0;
}
__syncthreads();
/* reduction using shared mem */
if (tid < 128) sdata[tid] += sdata[tid + 128];
__syncthreads();
if (tid < 64) sdata[tid] += sdata[tid + 64];
__syncthreads();
if (tid < 32) warpReduce(sdata, tid);
if (tid == 0) z[bid] = sdata[0];
}
/* sum all entries in x and asign to y
* block dim must be 256 */
__global__ void dot_stg_2(const float *x, float *y, int N)
{
__shared__ float sdata[256];
int idx = get_tid();
int tid = threadIdx.x;
int bid = get_bid();
/* load data to shared mem */
if (idx < N) {
sdata[tid] = x[idx];
}
else {
sdata[tid] = 0;
}
__syncthreads();
/* reduction using shared mem */
if (tid < 128) sdata[tid] += sdata[tid + 128];
__syncthreads();
if (tid < 64) sdata[tid] += sdata[tid + 64];
__syncthreads();
if (tid < 32) warpReduce(sdata, tid);
if (tid == 0) y[bid] = sdata[0];
}
__global__ void dot_stg_3(float *x, int N)
{
__shared__ float sdata[128];
int tid = threadIdx.x;
int i;
sdata[tid] = 0;
/* load data to shared mem */
for (i = 0; i < N; i += 128) {
if (tid + i < N) sdata[tid] += x[i + tid];
}
__syncthreads();
/* reduction using shared mem */
if (tid < 64) sdata[tid] = sdata[tid] + sdata[tid + 64];
__syncthreads();
if (tid < 32) warpReduce(sdata, tid);
if (tid == 0) x[0] = sdata[0];
}
/* dz and d serve as cache: result stores in d[0] */
void dot_device(float *dx, float *dy, float *dz, float *d, int N)
{
/* 1D block */
int bs = 256;
/* 2D grid */
int s = ceil(sqrt((N + bs - 1.) / bs));
dim3 grid = dim3(s, s);
int gs = 0;
/* stage 1 */
dot_stg_1<<<grid, bs>>>(dx, dy, dz, N);
/* stage 2 */
{
/* 1D grid */
int N2 = (N + bs - 1) / bs;
int s2 = ceil(sqrt((N2 + bs - 1.) / bs));
dim3 grid2 = dim3(s2, s2);
dot_stg_2<<<grid2, bs>>>(dz, d, N2);
/* record gs */
gs = (N2 + bs - 1.) / bs;
}
/* stage 3 */
dot_stg_3<<<1, 128>>>(d, gs);
}
int main(int argc, char **argv)
{
int N = 10000070;
int nbytes = N * sizeof(float);
float *hx = NULL, *hy = NULL;
float *dx = NULL, *dy = NULL, *dz = NULL, *d = NULL;
int i, itr = 20;
float asd = 0, ash;
double td, th;
if (argc == 2) {
int an;
an = atoi(argv[1]);
if (an > 0) N = an;
}
/* allocate GPU mem */
cudaMalloc((void **)&dx, nbytes);
cudaMalloc((void **)&dy, nbytes);
cudaMalloc((void **)&dz, sizeof(float) * ((N + 255) / 256));
cudaMalloc((void **)&d, sizeof(float) * ((N + 255) / 256));
if (dx == NULL || dy == NULL || dz == NULL || d == NULL) {
printf("couldn't allocate GPU memory\n");
return -1;
}
printf("allocated %e MB on GPU\n", nbytes / (1024.f * 1024.f));
/* alllocate CPU mem */
hx = (float *) malloc(nbytes);
hy = (float *) malloc(nbytes);
if (hx == NULL || hy == NULL) {
printf("couldn't allocate CPU memory\n");
return -2;
}
printf("allocated %e MB on CPU\n", nbytes / (1024.f * 1024.f));
/* init */
for (i = 0; i < N; i++) {
hx[i] = 1;
hy[i] = 2;
}
/* copy data to GPU */
cudaMemcpy(dx, hx, nbytes, cudaMemcpyHostToDevice);
cudaMemcpy(dy, hy, nbytes, cudaMemcpyHostToDevice);
/* let dust fall */
cudaDeviceSynchronize();
td = get_time();
/* call GPU */
for (i = 0; i < itr; i++) dot_device(dx, dy, dz, d, N);
/* let GPU finish */
cudaDeviceSynchronize();
td = get_time() - td;
th = get_time();
for (i = 0; i < itr; i++) ash = dot_host(hx, hy, N);
th = get_time() - th;
/* copy data from GPU */
cudaMemcpy(&asd, d, sizeof(float), cudaMemcpyDeviceToHost);
printf("dot, answer: %d, calculated by GPU:%f, calculated by CPU:%f\n", 2 * N, asd, ash);
printf("GPU time: %e, CPU time: %e, speedup: %g\n", td, th, th / td);
cudaFree(dx);
cudaFree(dy);
cudaFree(dz);
cudaFree(d);
free(hx);
free(hy);
return 0;
}