c++ - 连续 block 从初始 block 读取内存

Question

所以这是我程序的一部分，我对两个类(class)进行了减和。我通过共享数组 __shared__ int nrules[max_threads * MAX_CLASSES]; 的一半索引类，所以第一类从 nrules[0] 开始，第二类从 开始nrules[blockDim.x 或 max_threads]。减少了两半。总和保存在作为参数传递的全局数组中，该数组将保留每个 block 的总和，因此由 blockIdx.x 索引。

我有一个测试用例的大小，用MAX_SIZE表示，所有测试先从1处理到MAX_SIZE，求和在全局累加每个 block 的数组。

我想调用一个 block 数等于我的测试数 (10000) 的内核，但是总和有一些问题，所以我改为按步骤进行。

我找不到解决这个问题的方法，但是每当我调用一个 block 数超过 max_threads 的内核时，它就会开始从初始 block 中求和。如果执行代码，您会看到它将打印每个 block 的值，在本例中为 64，每个 block 有 64 个线程。如果我再执行至少一个 block ，它的总和将改为 128。这是第一类总和。就好像偏移变量什么都不做，写入又发生在第一个 block 上。当 MAX_SIZE = 3 时，第一个 block 的第二类总和更改为 192。 这里的 Cuda 功能是 2.0，一张 GT 520 卡。使用 CUDA 6.5 编译。

#include <stdio.h>
#include <cuda.h>
#include <cuda_runtime.h>

#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort = true)
{
    if (code != cudaSuccess)
    {
        fprintf(stderr, "GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);

    }
}

#define MAX_CLASSES 2
#define max_threads 64
//#define MAX_FEATURES 65

__device__ __constant__ int d_MAX_SIZE;
__device__  __constant__ int offset;

__device__ void rules_points_reduction(float points[max_threads * MAX_CLASSES], int nrules[max_threads * MAX_CLASSES]){

    float psum[MAX_CLASSES];
    int nsum[MAX_CLASSES];

    for (int i = 0; i < MAX_CLASSES; i++){
        psum[i] = points[threadIdx.x + i * blockDim.x];
        nsum[i] = nrules[threadIdx.x + i * blockDim.x];
    }

    __syncthreads();

    if (blockDim.x >= 1024) {
        if (threadIdx.x < 512) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 512 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 512 + i * blockDim.x];
            }

        } __syncthreads();
    }
    if (blockDim.x >= 512) {
        if (threadIdx.x < 256) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 256 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 256 + i * blockDim.x];
            }
        } __syncthreads();
    }
    if (blockDim.x >= 256) {
        if (threadIdx.x < 128) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 128 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 128 + i * blockDim.x];
            }
        } __syncthreads();
    }
    if (blockDim.x >= 128) {
        if (threadIdx.x <  64) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 64 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 64 + i * blockDim.x];
            }
        } __syncthreads();
    }

    if (threadIdx.x < 32)
    {
        // now that we are using warp-synchronous programming (below)
        // we need to declare our shared memory volatile so that the compiler
        // doesn't reorder stores to it and induce incorrect behavior.
        //volatile int* smem = nrules;
        //volatile float* smemf = points;
        if (blockDim.x >= 64) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 32 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 32 + i * blockDim.x];
            }
        }
        if (blockDim.x >= 32) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 16 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 16 + i * blockDim.x];
            }
        }
        if (blockDim.x >= 16) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 8 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 8 + i * blockDim.x];
            }
        }
        if (blockDim.x >= 8) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 4 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 4 + i * blockDim.x];
            }
        }
        if (blockDim.x >= 4) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 2 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 2 + i * blockDim.x];
            }
        }
        if (blockDim.x >= 2) {
            for (int i = 0; i < MAX_CLASSES; i++){
                points[threadIdx.x + i * blockDim.x] = psum[i] = psum[i] + points[threadIdx.x + 1 + i * blockDim.x];
                nrules[threadIdx.x + i * blockDim.x] = nsum[i] = nsum[i] + nrules[threadIdx.x + 1 + i * blockDim.x];
            }
        }
    }

}

__device__ void d_get_THE_prediction(short k, float* finalpoints, int* gn_rules)
{   
    int max;
    short true_label, n_items;

    __shared__ float points[max_threads * MAX_CLASSES]; 
    __shared__ int nrules[max_threads * MAX_CLASSES];
    //__shared__ short  items[MAX_FEATURES], ele[MAX_FEATURES];
    __shared__ int max2;

    for (int i = 0; i < MAX_CLASSES; i++)
    {
        points[threadIdx.x + i * blockDim.x] = 1;
        nrules[threadIdx.x + i * blockDim.x] = 1;
    }

    if (threadIdx.x == 0) {
        if (k == 1){
            nrules[0] = 1;
            nrules[blockDim.x] = 1;
        }
        //max2 = GetBinCoeff_l_d(n_items, k);
    }
    __syncthreads();

    //max = max2;

    //d_induce_rules(items, ele, n_items, k, max, nrules, points);

    __syncthreads();

    rules_points_reduction(points, nrules);

    if (threadIdx.x == 0){

        for (int i = 0; i < MAX_CLASSES; i++){
            gn_rules[(blockIdx.x + offset) + i * blockDim.x] += nrules[i * blockDim.x];
            finalpoints[(blockIdx.x + offset) + i * blockDim.x] += points[i * blockDim.x];

        }       
        printf("block %d k%d %f %f %d %d\n", (blockIdx.x + offset), k, finalpoints[(blockIdx.x + offset)],
            finalpoints[(blockIdx.x + offset) + blockDim.x], gn_rules[(blockIdx.x + offset)], gn_rules[(blockIdx.x + offset) + blockDim.x]);

    }
}

__global__ void lazy_supervised_classification_kernel(int k, float* finalpoints, int* n_rules){

    d_get_THE_prediction( k, finalpoints, n_rules);

}


int main() {
    //freopen("output.txt", "w", stdout);

    int N_TESTS = 10000;
    int MAX_SIZE = 3;

    float *finalpoints = (float*)calloc(MAX_CLASSES * N_TESTS, sizeof(float));
    float *d_finalpoints = 0;

    int *d_nruls = 0;
    int *nruls = (int*)calloc(MAX_CLASSES * N_TESTS, sizeof(int));  

    gpuErrchk(cudaMalloc(&d_finalpoints, MAX_CLASSES * N_TESTS * sizeof(float)));
    gpuErrchk(cudaMemset(d_finalpoints, 0, MAX_CLASSES * N_TESTS * sizeof(float)));

    gpuErrchk(cudaMalloc(&d_nruls, MAX_CLASSES * N_TESTS * sizeof(int)));
    gpuErrchk(cudaMemset(d_nruls, 0, MAX_CLASSES * N_TESTS * sizeof(int)));

    gpuErrchk(cudaMemcpyToSymbol(d_MAX_SIZE, &MAX_SIZE, sizeof(int), 0, cudaMemcpyHostToDevice));

    int step = max_threads, ofset = 0;

    for (int k = 1; k < MAX_SIZE; k++){ 

                               //N_TESTS-step
        for (ofset = 0; ofset < max_threads; ofset += step){

            gpuErrchk(cudaMemcpyToSymbol(offset, &ofset, sizeof(int), 0, cudaMemcpyHostToDevice));
            lazy_supervised_classification_kernel <<<step, max_threads >>>(k, d_finalpoints, d_nruls);
            gpuErrchk(cudaDeviceSynchronize());
        }

        gpuErrchk(cudaMemcpyToSymbol(offset, &ofset, sizeof(int), 0, cudaMemcpyHostToDevice));//comment these lines
                                          //N_TESTS - step      
        lazy_supervised_classification_kernel <<<3, max_threads >> >(k, d_finalpoints, d_nruls);//
        gpuErrchk(cudaDeviceSynchronize());//

    }
    gpuErrchk(cudaFree(d_finalpoints));
    gpuErrchk(cudaFree(d_nruls));
    free(finalpoints);
    free(nruls);    

    gpuErrchk(cudaDeviceReset());   
    return(0);
}

Answer 1

我不相信这个索引是你想要的:

 gn_rules[(blockIdx.x + offset) + i * blockDim.x] += ...;
 finalpoints[(blockIdx.x + offset) + i * blockDim.x] += ...;

对于 MAX_CLASSES = 2，每个 block 需要存储 2 个 finalpoints 值和 2 个 gn_rules 值。因此，当 offset 不为零时，它需要按 MAX_CLASSES 值缩放，以便索引到该 block 正确存储的开始。

因此，如果您将上面的代码行更改为:

 gn_rules[(blockIdx.x + (offset*MAX_CLASSES)) + i * blockDim.x] += nrules[i * blockDim.x];
 finalpoints[(blockIdx.x + (offset*MAX_CLASSES)) + i * blockDim.x] += points[i * blockDim.x];

我相信你会得到你期望的输出。