筆記(2)中提到當cuda需要對同一個內存地址中的值進行讀寫訪問時,需要原子操作(Atomics)(類似cpu多線程的互斥鎖mutex或原子變量),本章通過一個列子介紹原子操作。
原子操作(Atomics)
有些操作不能被拆分, 多線程要針對同一內存地址進行操作,這時需要對該內存使用原子操作,來保證線程對資源有暫時的“獨佔性”, 避免計算錯誤。原子操作是把雙刃劍,因爲原子操作使某些指令操作串行化,使用不當反而負優化。下面通過一個統計直方圖示例來說明原子操作加操作的正確使用。
統計直方圖cpu代碼:
#include "../common/book.h"
#define SIZE (100*1024*1024)
int main( void ) {
unsigned char *buffer =
(unsigned char*)big_random_block( SIZE );
// capture the start time
clock_t start, stop;
start = clock();
unsigned int histo[256];
for (int i=0; i<256; i++)
histo[i] = 0;
for (int i=0; i<SIZE; i++)
histo[buffer[i]]++;
stop = clock();
float elapsedTime = (float)(stop - start) /
(float)CLOCKS_PER_SEC * 1000.0f;
printf( "Time to generate: %3.1f ms\n", elapsedTime );
long histoCount = 0;
for (int i=0; i<256; i++) {
histoCount += histo[i];
}
printf( "Histogram Sum: %ld\n", histoCount );
free( buffer );
return 0;
}
未優化的gpu核代碼如下:
__global__ void histo_kernel( unsigned char *buffer,
long size,
unsigned int *histo ) {
// calculate the starting index and the offset to the next
// block that each thread will be processing
int i = threadIdx.x + blockIdx.x * blockDim.x;
int stride = blockDim.x * gridDim.x;
while (i < size) {
atomicAdd( &histo[buffer[i]], 1 );
i += stride;
}
}
上述代碼使用原子操作保證程序的正確性atomicAdd( &histo[buffer[i]], 1 ),由於每個線程都要執行atomicAdd( &histo[buffer[i]], 1 ),使線程串行等待,cuda並行編程沒有起到優化作用。爲了減少程序等待時間,我們使用共享內存來分成兩步,第一步,使用原子操作緩存每個block中所有線程的統計結果,此時競爭的線程只是每個block中的線程,所以競爭等待時間縮短;第二步,把所有block統計的結果加到一起,注意執行第二步之前要確保第一步執行完成,此時需要使用線程同步__syncthreads()。
__global__ void histo_kernel( unsigned char *buffer,
long size,
unsigned int *histo ) {
// clear out the accumulation buffer called temp
// since we are launched with 256 threads, it is easy
// to clear that memory with one write per thread
__shared__ unsigned int temp[256];
temp[threadIdx.x] = 0;
__syncthreads();
// calculate the starting index and the offset to the next
// block that each thread will be processing
int i = threadIdx.x + blockIdx.x * blockDim.x;
int stride = blockDim.x * gridDim.x;
while (i < size) {
atomicAdd( &temp[buffer[i]], 1 );
i += stride;
}
// sync the data from the above writes to shared memory
// then add the shared memory values to the values from
// the other thread blocks using global memory
// atomic adds
// same as before, since we have 256 threads, updating the
// global histogram is just one write per thread!
__syncthreads();
atomicAdd( &(histo[threadIdx.x]), temp[threadIdx.x] );
}
全部代碼如下:
/*
* Copyright 1993-2010 NVIDIA Corporation. All rights reserved.
*
* NVIDIA Corporation and its licensors retain all intellectual property and
* proprietary rights in and to this software and related documentation.
* Any use, reproduction, disclosure, or distribution of this software
* and related documentation without an express license agreement from
* NVIDIA Corporation is strictly prohibited.
*
* Please refer to the applicable NVIDIA end user license agreement (EULA)
* associated with this source code for terms and conditions that govern
* your use of this NVIDIA software.
*
*/
#include "../common/book.h"
#define SIZE (100*1024*1024)
__global__ void histo_kernel( unsigned char *buffer,
long size,
unsigned int *histo ) {
// clear out the accumulation buffer called temp
// since we are launched with 256 threads, it is easy
// to clear that memory with one write per thread
__shared__ unsigned int temp[256];
temp[threadIdx.x] = 0;
__syncthreads();
// calculate the starting index and the offset to the next
// block that each thread will be processing
int i = threadIdx.x + blockIdx.x * blockDim.x;
int stride = blockDim.x * gridDim.x;
while (i < size) {
atomicAdd( &temp[buffer[i]], 1 );
i += stride;
}
// sync the data from the above writes to shared memory
// then add the shared memory values to the values from
// the other thread blocks using global memory
// atomic adds
// same as before, since we have 256 threads, updating the
// global histogram is just one write per thread!
__syncthreads();
atomicAdd( &(histo[threadIdx.x]), temp[threadIdx.x] );
}
int main( void ) {
unsigned char *buffer =
(unsigned char*)big_random_block( SIZE );
// capture the start time
// starting the timer here so that we include the cost of
// all of the operations on the GPU. if the data were
// already on the GPU and we just timed the kernel
// the timing would drop from 74 ms to 15 ms. Very fast.
cudaEvent_t start, stop;
HANDLE_ERROR( cudaEventCreate( &start ) );
HANDLE_ERROR( cudaEventCreate( &stop ) );
HANDLE_ERROR( cudaEventRecord( start, 0 ) );
// allocate memory on the GPU for the file's data
unsigned char *dev_buffer;
unsigned int *dev_histo;
HANDLE_ERROR( cudaMalloc( (void**)&dev_buffer, SIZE ) );
HANDLE_ERROR( cudaMemcpy( dev_buffer, buffer, SIZE,
cudaMemcpyHostToDevice ) );
HANDLE_ERROR( cudaMalloc( (void**)&dev_histo,
256 * sizeof( int ) ) );
HANDLE_ERROR( cudaMemset( dev_histo, 0,
256 * sizeof( int ) ) );
// kernel launch - 2x the number of mps gave best timing
cudaDeviceProp prop;
HANDLE_ERROR( cudaGetDeviceProperties( &prop, 0 ) );
int blocks = prop.multiProcessorCount;
histo_kernel<<<blocks*2,256>>>( dev_buffer,
SIZE, dev_histo );
unsigned int histo[256];
HANDLE_ERROR( cudaMemcpy( histo, dev_histo,
256 * sizeof( int ),
cudaMemcpyDeviceToHost ) );
// get stop time, and display the timing results
HANDLE_ERROR( cudaEventRecord( stop, 0 ) );
HANDLE_ERROR( cudaEventSynchronize( stop ) );
float elapsedTime;
HANDLE_ERROR( cudaEventElapsedTime( &elapsedTime,
start, stop ) );
printf( "Time to generate: %3.1f ms\n", elapsedTime );
long histoCount = 0;
for (int i=0; i<256; i++) {
histoCount += histo[i];
}
printf( "Histogram Sum: %ld\n", histoCount );
// verify that we have the same counts via CPU
for (int i=0; i<SIZE; i++)
histo[buffer[i]]--;
for (int i=0; i<256; i++) {
if (histo[i] != 0)
printf( "Failure at %d!\n", i );
}
HANDLE_ERROR( cudaEventDestroy( start ) );
HANDLE_ERROR( cudaEventDestroy( stop ) );
cudaFree( dev_histo );
cudaFree( dev_buffer );
free( buffer );
return 0;
}