《CUDA By Example》【Chapter 09】原子性？

9.1 概述

• 了解不同NVIDIA GPU的计算功能集
• 了解原子操作以及为什么需要使用
• 了解如何在CUDA C核函数中执行带有原子操作的运算

9.2 计算功能集

不同架构的CPU有不同的功能和指令集（如MMX,SSE,SSE2），对于CUDA支持的GPU也一样。NVIDIA将GPU支持的各种功能统称为计算功能集（Compute Capability）。

9.2.1 NVIDIA GPU计算功能集

计算功能集包括1.0, 1.1, 1.2, 1.3以及2.0。高版本计算功能集是低版本计算功能集的超集。
本章介绍硬件在内存上执行原子操作的能力。从功能集1.2开始，既支持共享内存原子操作又支持全局内存原子操作。

9.2.2 基于最小计算功能集的编译

告诉编译器代码需要使用某一版本如（1.2/1.1）版本或者更高的计算功能集。

nvcc -arch=sm_12
nvcc -arch=sm_11

9.3原子操作简介

9.4计算直方图

9.4.1在CPU上计算直方图

hist_cpu.cu

#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;
}

9.4.2在GPU上计算直方图

使用全局内存原子操作，性能可能会下降。
核函数中计算很少，很可能是全局内存上的原子操作引起了性能的降低。当数千个线程尝试访问少量的内存位置时，将产生大量的竞争。为了确保递增操作的原子性，对相同内存位置的操作都将被硬件串行化。

#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;
}

使用共享内存原子操作和全局内存原子操作。上面代码的性能问题是由于原子操作带来的，有意思的是，解决的 办法是增加原子操作来优化性能。这里引入了共享内存来优化。比单纯使用全局内存原子操作好很多。

#include "../common/book.h"

#define SIZE    (100*1024*1024)


__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;
    }
}

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.
    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!  Off by %d\n", i, histo[i] );
    }

    HANDLE_ERROR( cudaEventDestroy( start ) );
    HANDLE_ERROR( cudaEventDestroy( stop ) );
    cudaFree( dev_histo );
    cudaFree( dev_buffer );
    free( buffer );
    return 0;
}

9.5 小结

有时候以来原子操作会带来性能问题，并且这些问题只能通过对算法的部分重构来加以解决。在直方图中，使用了一种两阶段算法，从而降低了在全局内存访问上竞争程度。通常，这种降低内存竞争程度的策略总能带来不错的效果。