CUDA学习(九) 想好好解释一下利用shared memory去做matrixMul

 代码来自CUDA Samples\v9.1\0_Simple\matrixMul

主要注释了matrixMulCUDA（）

C=A*B

它的理念是把矩阵分成小块，每个线程利用两层循环，大循环在迭代subMatrix，小循环则是迭代每个小subMatrix的横纵坐标，

可以理解为一个线程计算出的Csub就是对应一个元素，然后你想象所有线程是并行的，所以所有的元素计算就可以理解过来了，但是单纯的一个线程是解决不了这个问题的，因为它小循环迭代的时候是需要所有元素的内容的，这也是使用__syncthreads()的作用，而你在使用global memory的时候，不需要用这个，因为每个线程各做各的，都需要去读global memory,这也是使用shared memory的作用，它可以减少对global的访问。

说的不对的地方请大家给我指出来，这各shard memory的使用是我觉得最难的

/**
 * Copyright 1993-2015 NVIDIA Corporation.  All rights reserved.
 *
 * Please refer to the NVIDIA end user license agreement (EULA) associated
 * with this source code for terms and conditions that govern your use of
 * this software. Any use, reproduction, disclosure, or distribution of
 * this software and related documentation outside the terms of the EULA
 * is strictly prohibited.
 *
 */

/**
 * Matrix multiplication: C = A * B.
 * Host code.
 *
 * This sample implements matrix multiplication as described in Chapter 3
 * of the programming guide.
 * It has been written for clarity of exposition to illustrate various CUDA
 * programming principles, not with the goal of providing the most
 * performant generic kernel for matrix multiplication.
 *
 * See also:
 * V. Volkov and J. Demmel, "Benchmarking GPUs to tune dense linear algebra,"
 * in Proc. 2008 ACM/IEEE Conf. on Supercomputing (SC '08),
 * Piscataway, NJ: IEEE Press, 2008, pp. Art. 31:1-11.
 */

// System includes
#include <stdio.h>
#include <assert.h>

// CUDA runtime
#include <cuda_runtime.h>

// Helper functions and utilities to work with CUDA
#include <helper_functions.h>
#include <helper_cuda.h>

/**
 * Matrix multiplication (CUDA Kernel) on the device: C = A * B
 * wA is A's width and wB is B's width
 */
template <int BLOCK_SIZE> __global__ void
matrixMulCUDA(float *C, float *A, float *B, int wA, int wB)
{
    // Block index
    int bx = blockIdx.x;
    int by = blockIdx.y;

    // Thread index
    int tx = threadIdx.x;
    int ty = threadIdx.y;

    // Index of the first sub-matrix of A processed by the block
    //这是A的第一个细分矩阵的索引,利用block处理（by=blockIdx.y）
    //也是从这里让aBegin根据不同block进行处理
    //不同block拥有不同的aBegin，即不同的block线程下的循环初始值的不同
    int aBegin = wA * BLOCK_SIZE * by;

    // Index of the last sub-matrix of A processed by the block
    //这是A的最后一个细分矩阵的索引，利用block处理
    //注意wA是A的width
    int aEnd   = aBegin + wA - 1;

    // Step size used to iterate through the sub-matrices of A
    //A的细分矩阵的步距
    int aStep  = BLOCK_SIZE;

    // Index of the first sub-matrix of B processed by the block
    //这是B的第一个细分矩阵的索引，也是利用block处理（bx=blockIdx.x）
    int bBegin = BLOCK_SIZE * bx;

    // Step size used to iterate through the sub-matrices of B
    //B的细分矩阵的迭代用步距
    int bStep  = BLOCK_SIZE * wB;

    // Csub is used to store the element of the block sub-matrix
    // that is computed by the thread
    float Csub = 0;

    // Loop over all the sub-matrices of A and B
    // required to compute the block sub-matrix
    // 用于计算细分矩阵（一个block）
    // aEnd只是迭代了一行的block
    for (int a = aBegin, b = bBegin;
         a <= aEnd;
         a += aStep, b += bStep)
    {

        // Declaration of the shared memory array As used to
        // store the sub-matrix of A
        // 为一个block即A的sub-matrix分配一个As[32][32]的共享内存
        __shared__ float As[BLOCK_SIZE][BLOCK_SIZE];

        // Declaration of the shared memory array Bs used to
        // store the sub-matrix of B
        // 为一个block即B的sub-matrix分配一个Bs[32][32]
        __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];

        // Load the matrices from device memory
        // to shared memory; each thread loads
        // one element of each matrix
        //每个线程读入一个元素到shared memory,但是这是在一个循环体中，这是我最难理解的
        //第一次循环的时候，想象一下，每个线程都执行到这里
        //这里显示的是一个block里面所有的内容
        //但是考虑到所有线程，在不同block里面可以复制不同的submatrix
        As[ty][tx] = A[a + wA * ty + tx];
        Bs[ty][tx] = B[b + wB * ty + tx];

        // Synchronize to make sure the matrices are loaded
        //这是是等待所有的都读完
        __syncthreads();

        // Multiply the two matrices together;
        // each thread computes one element
        // of the block sub-matrix
#pragma unroll
        
        
        //不同的线程ty和tx不同，这里一个线程第一次就完成一个submatrix的计算
        //一个线程对应一个CSub
        //大循环相当于block之间的迭代
        //小循环相当于block之内的迭代
        //一个大循环第一次计算这个小循环相当于完成了一个小block的计算，即submatrix的计算
        //不同线程，大循环的位置不同，对应的CSub不同
        //这里也是很难理解的地方，因为这个循环每个线程都会执行
        //这里也是速度提高的直接体现，因为可以看到不同线程共用了As里面的内容
        //如果采用了global memory，没个线程都会到global memory里面去找，这会降低访问速度
        for (int k = 0; k < BLOCK_SIZE; ++k)
        {
            Csub += As[ty][k] * Bs[k][tx];
        }

        // Synchronize to make sure that the preceding
        // computation is done before loading two new
        // sub-matrices of A and B in the next iteration
        __syncthreads();
        //等待所有的block执行完
    }

    // Write the block sub-matrix to device memory;
    // each thread writes one element
    //计算出所有的数据，然后传递给C，用于输出
    int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
    C[c + wB * ty + tx] = Csub;
}

void constantInit(float *data, int size, float val)
{
    for (int i = 0; i < size; ++i)
    {
        data[i] = val;
    }
}

/**
 * Run a simple test of matrix multiplication using CUDA
 */
int matrixMultiply(int argc, char **argv, int block_size, dim3 &dimsA, dim3 &dimsB)
{
    // Allocate host memory for matrices A and B
    //分配host端内存
    unsigned int size_A = dimsA.x * dimsA.y;
    unsigned int mem_size_A = sizeof(float) * size_A;
    float *h_A = (float *)malloc(mem_size_A);
    unsigned int size_B = dimsB.x * dimsB.y;
    unsigned int mem_size_B = sizeof(float) * size_B;
    float *h_B = (float *)malloc(mem_size_B);

    // Initialize host memory
    //初始化host端的内存
    const float valB = 0.01f;
    constantInit(h_A, size_A, 1.0f);
    constantInit(h_B, size_B, valB);

    // Allocate device memory
    //分配device端的内存
    float *d_A, *d_B, *d_C;

    // Allocate host matrix C
    //分配矩阵C的内存，用于输出
    // dimsB.x=5*4*32 dimsA.y=5*2*32
    dim3 dimsC(dimsB.x, dimsA.y, 1);
    unsigned int mem_size_C = dimsC.x * dimsC.y * sizeof(float);
    float *h_C = (float *) malloc(mem_size_C);

    if (h_C == NULL)
    {
        fprintf(stderr, "Failed to allocate host matrix C!\n");
        exit(EXIT_FAILURE);
    }

    cudaError_t error;

    //把host端的内存分配到device端，这里还是用的global memory
    error = cudaMalloc((void **) &d_A, mem_size_A);

    if (error != cudaSuccess)
    {
        printf("cudaMalloc d_A returned error %s (code %d), line(%d)\n", cudaGetErrorString(error), error, __LINE__);
        exit(EXIT_FAILURE);
    }

    error = cudaMalloc((void **) &d_B, mem_size_B);

    if (error != cudaSuccess)
    {
        printf("cudaMalloc d_B returned error %s (code %d), line(%d)\n", cudaGetErrorString(error), error, __LINE__);
        exit(EXIT_FAILURE);
    }

    error = cudaMalloc((void **) &d_C, mem_size_C);

    if (error != cudaSuccess)
    {
        printf("cudaMalloc d_C returned error %s (code %d), line(%d)\n", cudaGetErrorString(error), error, __LINE__);
        exit(EXIT_FAILURE);
    }

    // copy host memory to device
    error = cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice);

    if (error != cudaSuccess)
    {
        printf("cudaMemcpy (d_A,h_A) returned error %s (code %d), line(%d)\n", cudaGetErrorString(error), error, __LINE__);
        exit(EXIT_FAILURE);
    }

    error = cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice);

    if (error != cudaSuccess)
    {
        printf("cudaMemcpy (d_B,h_B) returned error %s (code %d), line(%d)\n", cudaGetErrorString(error), error, __LINE__);
        exit(EXIT_FAILURE);
    }

    // Setup execution parameters
    // main函数传入的 block_size为32
    // thread为（32，32）
    dim3 threads(block_size, block_size);
    // dimsB.x=5*4*32 dimsA.y=5*2*32
    // grid为（20，10）
    dim3 grid(dimsB.x / threads.x, dimsA.y / threads.y);

    // Create and start timer
    printf("Computing result using CUDA Kernel...\n");

    // Performs warmup operation using matrixMul CUDA kernel
    if (block_size == 16)
    {
        matrixMulCUDA<16><<< grid, threads >>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
    }
    else
    {
        matrixMulCUDA<32><<< grid, threads >>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
    }

    printf("done\n");

    cudaDeviceSynchronize();

    // Allocate CUDA events that we'll use for timing
    cudaEvent_t start;
    error = cudaEventCreate(&start);

    if (error != cudaSuccess)
    {
        fprintf(stderr, "Failed to create start event (error code %s)!\n", cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }

    cudaEvent_t stop;
    error = cudaEventCreate(&stop);

    if (error != cudaSuccess)
    {
        fprintf(stderr, "Failed to create stop event (error code %s)!\n", cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }

    // Record the start event
    error = cudaEventRecord(start, NULL);

    if (error != cudaSuccess)
    {
        fprintf(stderr, "Failed to record start event (error code %s)!\n", cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }

    // Execute the kernel
    int nIter = 300;

    for (int j = 0; j < nIter; j++)
    {
        if (block_size == 16)
        {
            matrixMulCUDA<16><<< grid, threads >>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
        }
        else
        {
            matrixMulCUDA<32><<< grid, threads >>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
        }
    }

    // Record the stop event
    error = cudaEventRecord(stop, NULL);

    if (error != cudaSuccess)
    {
        fprintf(stderr, "Failed to record stop event (error code %s)!\n", cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }

    // Wait for the stop event to complete
    error = cudaEventSynchronize(stop);

    if (error != cudaSuccess)
    {
        fprintf(stderr, "Failed to synchronize on the stop event (error code %s)!\n", cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }

    float msecTotal = 0.0f;
    error = cudaEventElapsedTime(&msecTotal, start, stop);

    if (error != cudaSuccess)
    {
        fprintf(stderr, "Failed to get time elapsed between events (error code %s)!\n", cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }

    // Compute and print the performance
    float msecPerMatrixMul = msecTotal / nIter;
    double flopsPerMatrixMul = 2.0 * (double)dimsA.x * (double)dimsA.y * (double)dimsB.x;
    double gigaFlops = (flopsPerMatrixMul * 1.0e-9f) / (msecPerMatrixMul / 1000.0f);
    printf(
        "Performance= %.2f GFlop/s, Time= %.3f msec, Size= %.0f Ops, WorkgroupSize= %u threads/block\n",
        gigaFlops,
        msecPerMatrixMul,
        flopsPerMatrixMul,
        threads.x * threads.y);

    // Copy result from device to host
    error = cudaMemcpy(h_C, d_C, mem_size_C, cudaMemcpyDeviceToHost);

    if (error != cudaSuccess)
    {
        printf("cudaMemcpy (h_C,d_C) returned error %s (code %d), line(%d)\n", cudaGetErrorString(error), error, __LINE__);
        exit(EXIT_FAILURE);
    }

    printf("Checking computed result for correctness: ");
    bool correct = true;

    // test relative error by the formula
    //     |<x, y>_cpu - <x,y>_gpu|/<|x|, |y|>  < eps
    double eps = 1.e-6 ; // machine zero

    for (int i = 0; i < (int)(dimsC.x * dimsC.y); i++)
    {
        double abs_err = fabs(h_C[i] - (dimsA.x * valB));
        double dot_length = dimsA.x;
        double abs_val = fabs(h_C[i]);
        double rel_err = abs_err/abs_val/dot_length ;

        if (rel_err > eps)
        {
            printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > %E\n", i, h_C[i], dimsA.x*valB, eps);
            correct = false;
        }
    }

    printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");

    // Clean up memory
    free(h_A);
    free(h_B);
    free(h_C);
    cudaFree(d_A);
    cudaFree(d_B);
    cudaFree(d_C);

    printf("\nNOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.\n");

    if (correct)
    {
        return EXIT_SUCCESS;
    }
    else
    {
        return EXIT_FAILURE;
    }
}


/**
 * Program main
 */
int main(int argc, char **argv)
{
    printf("[Matrix Multiply Using CUDA] - Starting...\n");

    if (checkCmdLineFlag(argc, (const char **)argv, "help") ||
        checkCmdLineFlag(argc, (const char **)argv, "?"))
    {
        printf("Usage -device=n (n >= 0 for deviceID)\n");
        printf("      -wA=WidthA -hA=HeightA (Width x Height of Matrix A)\n");
        printf("      -wB=WidthB -hB=HeightB (Width x Height of Matrix B)\n");
        printf("  Note: Outer matrix dimensions of A & B matrices must be equal.\n");

        exit(EXIT_SUCCESS);
    }

    // This will pick the best possible CUDA capable device, otherwise override the device ID based on input provided at the command line
    int dev = findCudaDevice(argc, (const char **)argv);

    int block_size = 32;

    dim3 dimsA(5*2*block_size, 5*2*block_size, 1);
    dim3 dimsB(5*4*block_size, 5*2*block_size, 1);

    // width of Matrix A
    if (checkCmdLineFlag(argc, (const char **)argv, "wA"))
    {
        dimsA.x = getCmdLineArgumentInt(argc, (const char **)argv, "wA");
    }

    // height of Matrix A
    if (checkCmdLineFlag(argc, (const char **)argv, "hA"))
    {
        dimsA.y = getCmdLineArgumentInt(argc, (const char **)argv, "hA");
    }

    // width of Matrix B
    if (checkCmdLineFlag(argc, (const char **)argv, "wB"))
    {
        dimsB.x = getCmdLineArgumentInt(argc, (const char **)argv, "wB");
    }

    // height of Matrix B
    if (checkCmdLineFlag(argc, (const char **)argv, "hB"))
    {
        dimsB.y = getCmdLineArgumentInt(argc, (const char **)argv, "hB");
    }

    if (dimsA.x != dimsB.y)
    {
        printf("Error: outer matrix dimensions must be equal. (%d != %d)\n",
               dimsA.x, dimsB.y);
        exit(EXIT_FAILURE);
    }

    printf("MatrixA(%d,%d), MatrixB(%d,%d)\n", dimsA.x, dimsA.y, dimsB.x, dimsB.y);

    int matrix_result = matrixMultiply(argc, argv, block_size, dimsA, dimsB);

    exit(matrix_result);
}