Cub 库支持的最大大小

Question

有谁知道 cub::scan 支持的最大大小是多少？我得到了输入大小超过 5 亿的核心转储。我想确保我没有做错任何事...

这是我的代码：

#define CUB_STDERR
#include <stdio.h>
#include "cub/util_allocator.cuh"
#include "cub/device/device_scan.cuh"
#include <sys/time.h>
using namespace cub;

bool                    g_verbose = false;  // Whether to display input/output to console
CachingDeviceAllocator  g_allocator(true);  // Caching allocator for device memory
typedef int mytype;

/**
 * Solve inclusive-scan problem
 */

static void solve(mytype *h_in, mytype *h_cpu, int n)
{
    mytype inclusive = 0;
    for (int i = 0; i < n; ++i) {
      inclusive += h_in[i];
      h_cpu[i] = inclusive;
    }
}
static int compare(mytype *h_cpu, mytype *h_o, int n)
{
    for (int i = 0; i < n; i++) {
      if (h_cpu[i] != h_o[i]) {
        return i + 1;
      }
    }
    return 0;
}

/**
 * Main
 */
int main(int argc, char** argv)
{
    cudaSetDevice(0);
    struct timeval start, end;
    int num_items = 1073741824;
    const int repetitions = 5;
    mytype *h_in, *h_out, *h_cpu;
    const int size = num_items * sizeof(mytype);
    // Allocate host arrays
    h_in = (mytype *)malloc(size);
    h_out = (mytype *)malloc(size);
    h_cpu = (mytype *)malloc(size);


    // Initialize problem and solution
    for (int i = 0; i < num_items; i++) {
        h_in[i] = i;
        h_out[i] = 0;
        h_cpu[i] = 0;
    }

    solve(h_in, h_cpu, num_items);

    // Allocate problem device arrays
    mytype *d_in = NULL;
    CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(mytype) * num_items));

    // Initialize device input
    CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(mytype) * num_items, cudaMemcpyHostToDevice));

    // Allocate device output array
    mytype *d_out = NULL;
    CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(mytype) * num_items));

    // Allocate temporary storage
    void            *d_temp_storage = NULL;
    size_t          temp_storage_bytes = 0;


    CubDebugExit(DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));
    CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));

    // Run
    gettimeofday(&start, NULL);
    for (long i = 0; i < repetitions; i++) 
        DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
    cudaThreadSynchronize();
    gettimeofday(&end, NULL);
    double ctime = end.tv_sec + end.tv_usec / 1000000.0 - start.tv_sec - start.tv_usec / 1000000.0;

    cudaMemcpy(h_out, d_out, sizeof(mytype) * num_items, cudaMemcpyDeviceToHost);
    int cmp = compare(h_cpu, h_out, num_items);
    printf("%d\t", num_items);
    if (!cmp)
        printf("\t%7.4fs \n", ctime);
    printf("\n");
    if (h_in) delete[] h_in;
    if (h_out) delete[] h_out;
    if (h_cpu) delete[] h_cpu;
    if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));
    if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));
    if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));

    printf("\n\n");

    return 0;
}

Answer 1

问题在这里：

const int size = num_items * sizeof(mytype);

并且可以通过将其更改为来修复：

const size_t size = num_items * sizeof(mytype);

代码中num_items的价值超过10亿。当我们将其乘以 sizeof(mytype) 时，我们将其乘以 4，因此结果超过 40 亿。该值不能存储在 int 变量中。如果您仍然尝试那样使用它，那么您后续的主机代码将会做坏事。这个问题（核心转储）实际上与 CUDA 无关。如果删除所有 CUB 元素，代码将进行核心转储。

当我修改上面的代码行并针对正确的 GPU 进行编译时（例如，在我的情况下为 -arch=sm_35，对于 Titan X GPU 为 -arch=sm_52），然后我得到了正确的答案（并且没有段 fault/core 转储）。

一般来说，追查 seg fault/core 转储类型错误的正确起点是识别此错误是由 host code 引起的，您应该尝试本地化生成此错误的确切源代码行。这可以通过在您的代码中放置许多 printf 语句来完成 trivially/tediously ，直到您确定您的代码行之后您看不到任何 printf 输出，或者通过使用主机代码调试器，例如作为 linux.

上的 gdb

另请注意，编写的这段代码在主机上需要略多于 12GB 的内存，在 GPU 上需要略多于 8GB 的内存，因此在此类设置中它只会运行正确。

供参考，这里是固定代码（基于 OP 发布的 here）：

#define CUB_STDERR
#include <stdio.h>
#include "cub/util_allocator.cuh"
#include "cub/device/device_scan.cuh"
#include <sys/time.h>
using namespace cub;

bool                    g_verbose = false;  // Whether to display input/output to console
CachingDeviceAllocator  g_allocator(true);  // Caching allocator for device memory
typedef int mytype;

/**
 * Solve inclusive-scan problem
 */

static void solve(mytype *h_in, mytype *h_cpu, int n)
{
    mytype inclusive = 0;
    for (int i = 0; i < n; ++i) {
      inclusive += h_in[i];
      h_cpu[i] = inclusive;
    }
}
static int compare(mytype *h_cpu, mytype *h_o, int n)
{
    for (int i = 0; i < n; i++) {
      if (h_cpu[i] != h_o[i]) {
        return i + 1;
      }
    }
    return 0;
}

/**
 * Main
 */
int main(int argc, char** argv)
{
    cudaSetDevice(0);
    struct timeval start, end;
    int num_items = 1073741824;
    const int repetitions = 5;
    mytype *h_in, *h_out, *h_cpu;
    const size_t size = num_items * sizeof(mytype);
    // Allocate host arrays
    h_in = (mytype *)malloc(size);
    h_out = (mytype *)malloc(size);
    h_cpu = (mytype *)malloc(size);


    // Initialize problem and solution
    for (int i = 0; i < num_items; i++) {
        h_in[i] = i;
        h_out[i] = 0;
        h_cpu[i] = 0;
    }

    solve(h_in, h_cpu, num_items);

    // Allocate problem device arrays
    mytype *d_in = NULL;
    CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(mytype) * num_items));

    // Initialize device input
    CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(mytype) * num_items, cudaMemcpyHostToDevice));

    // Allocate device output array
    mytype *d_out = NULL;
    CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(mytype) * num_items));

    // Allocate temporary storage
    void            *d_temp_storage = NULL;
    size_t          temp_storage_bytes = 0;


    CubDebugExit(DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));
    CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));

    // Run
    gettimeofday(&start, NULL);
    for (long i = 0; i < repetitions; i++) 
        DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
    cudaThreadSynchronize();
    gettimeofday(&end, NULL);
    double ctime = end.tv_sec + end.tv_usec / 1000000.0 - start.tv_sec - start.tv_usec / 1000000.0;

    cudaMemcpy(h_out, d_out, sizeof(mytype) * num_items, cudaMemcpyDeviceToHost);
    int cmp = compare(h_cpu, h_out, num_items);
    printf("%d\t", num_items);
    if (!cmp)
        printf("\t%7.4fs \n", ctime);
    printf("\n");
    if (h_in) delete[] h_in;
    if (h_out) delete[] h_out;
    if (h_cpu) delete[] h_cpu;
    if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));
    if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));
    if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));

    printf("\n\n");

    return 0;
}

Cub 库支持的最大大小

maximum supported size for cub library

cuda

nvidia

cub

prefix-sum