cudaMallocManaged (统一内存)和cuBLAS

Question

我试图在cudaMallocManaged()和cuBLAS库中使用统一内存。我正在执行一个简单的矩阵向量乘法作为一个简单的例子，并将结果存储在一个数组results中。然而，当打印results数组时，我会得到所有的0，而不是矩阵mat乘以矢量vec的结果。

我使用的流程是：

用cudaMallocManaged()

• Initializing分配内存用data

• Allocating的数组cuBLAS handle

• Calling cublasDgemv执行在results

中存储结果的乘法

当使用new，然后使用cublasSetMatrix()或cublasSetVector()时，效果很好。

如何在cuBLAS中使用统一内存？

以下是最低限度的工作实例：

统一内存尝试(这将返回results中的所有0)：

#include <cuda.h>
#include <cuda_runtime.h>
#include <iostream>
#include <ctime>
#include "cublas_v2.h"

#define cudaErrChk(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);
        if (abort) exit(code);
    }
}

static const char *cublasErrChk(cublasStatus_t error)
{
    switch (error)
    {
        case CUBLAS_STATUS_SUCCESS:
            return "CUBLAS_STATUS_SUCCESS";

        case CUBLAS_STATUS_NOT_INITIALIZED:
            return "CUBLAS_STATUS_NOT_INITIALIZED";

        case CUBLAS_STATUS_ALLOC_FAILED:
            return "CUBLAS_STATUS_ALLOC_FAILED";

        case CUBLAS_STATUS_INVALID_VALUE:
            return "CUBLAS_STATUS_INVALID_VALUE";

        case CUBLAS_STATUS_ARCH_MISMATCH:
            return "CUBLAS_STATUS_ARCH_MISMATCH";

        case CUBLAS_STATUS_MAPPING_ERROR:
            return "CUBLAS_STATUS_MAPPING_ERROR";

        case CUBLAS_STATUS_EXECUTION_FAILED:
            return "CUBLAS_STATUS_EXECUTION_FAILED";

        case CUBLAS_STATUS_INTERNAL_ERROR:
            return "CUBLAS_STATUS_INTERNAL_ERROR";
    }

    return "<unknown>";
}

int main() {

    size_t dims = 4;

    double *vec, *mat, *results;

    cudaErrChk( cudaMallocManaged(&vec, dims * sizeof(double)) );
    cudaErrChk( cudaMallocManaged(&mat, dims * dims * sizeof(double)) );
    cudaErrChk( cudaMallocManaged(&results, dims * sizeof(double)) );

    printf("Vector:\n");
    for (int i = 1; i < dims + 1; i++) {
        vec[i] = 0.5 * i;
        printf("%.2lf ", vec[i]);
    } 
    printf("\n\nMatrix:\n");

    for (int i = 1; i < dims * dims + 1; i++) {
        mat[i] = 1.0 * i;
        printf("%.2lf ", mat[i]);

        if (i % dims == 0)
            printf("\n");
    }
    printf("\n");

    cublasHandle_t handle;
    cublasErrChk( cublasCreate(&handle) );

    double alpha = 1.f, beta = 1.f;

    // multiply mat by vec to get results
    cublasErrChk(
        cublasDgemv(
            handle, CUBLAS_OP_N,
            dims, dims,
            &alpha,
            mat, dims,
            vec, 1,
            &beta,
            results, 1
        )
    );

    for (int i = 0; i < dims; i++)
        printf("%.2lf ", results[i]);
    printf("\n");

    cudaErrChk( cudaFree(vec) );
    cudaErrChk( cudaFree(mat) );
    cudaErrChk( cudaFree(results) );

    return 0;
}

普通malloc/setMatrix()尝试：

#include <cuda.h>
#include <cuda_runtime.h>
#include <iostream>
#include <ctime>
#include "cublas_v2.h"

#define cudaErrChk(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);
        if (abort) exit(code);
    }
}

static const char *cublasErrChk(cublasStatus_t error)
{
    switch (error)
    {
        case CUBLAS_STATUS_SUCCESS:
            return "CUBLAS_STATUS_SUCCESS";

        case CUBLAS_STATUS_NOT_INITIALIZED:
            return "CUBLAS_STATUS_NOT_INITIALIZED";

        case CUBLAS_STATUS_ALLOC_FAILED:
            return "CUBLAS_STATUS_ALLOC_FAILED";

        case CUBLAS_STATUS_INVALID_VALUE:
            return "CUBLAS_STATUS_INVALID_VALUE";

        case CUBLAS_STATUS_ARCH_MISMATCH:
            return "CUBLAS_STATUS_ARCH_MISMATCH";

        case CUBLAS_STATUS_MAPPING_ERROR:
            return "CUBLAS_STATUS_MAPPING_ERROR";

        case CUBLAS_STATUS_EXECUTION_FAILED:
            return "CUBLAS_STATUS_EXECUTION_FAILED";

        case CUBLAS_STATUS_INTERNAL_ERROR:
            return "CUBLAS_STATUS_INTERNAL_ERROR";
    }

    return "<unknown>";
}

int main() {

    size_t dims = 4;

    double *h_vec, *h_mat, *h_results;

    h_vec = new double[dims];
    h_mat = new double[dims * dims];
    h_results = new double[dims];

    printf("Vector:\n");
    for (int i = 1; i < dims + 1; i++) {
        h_vec[i] = 0.5 * i;
        printf("%.2lf ", h_vec[i]);
    } 
    printf("\n\nMatrix:\n");

    for (int i = 1; i < dims * dims + 1; i++) {
        h_mat[i] = 1.0 * i;
        printf("%.2lf ", h_mat[i]);

        if (i % dims == 0)
            printf("\n");
    }
    printf("\n");

    double *d_vec, *d_mat, *d_results;

    cudaErrChk( cudaMalloc(&d_vec, dims * sizeof(double)) );
    cudaErrChk( cudaMalloc(&d_mat, dims * dims * sizeof(double)) );
    cudaErrChk( cudaMalloc(&d_results, dims * sizeof(double)) );

    cublasHandle_t handle;
    cublasErrChk( cublasCreate(&handle) );

    // copy the data manually to the GPUs
    cublasErrChk( cublasSetVector(dims, sizeof(*d_vec), h_vec, 1, d_vec, 1) );
    cublasErrChk( cublasSetMatrix(dims, dims, sizeof(double), h_mat, dims, d_mat, dims) );

    double alpha = 1.f, beta = 1.f;

    // // multiply mat by vec to get results
    cublasErrChk(
        cublasDgemv(
            handle, CUBLAS_OP_N,
            dims, dims,
            &alpha,
            d_mat, dims,
            d_vec, 1,
            &beta,
            d_results, 1
        )
    );

    cublasErrChk( cublasGetVector(dims, sizeof(*h_results), d_results, 1, h_results, 1) );

    for (int i = 0; i < dims; i++)
        printf("%.2lf ", h_results[i]);
    printf("\n");

    cudaErrChk( cudaFree(d_vec) );
    cudaErrChk( cudaFree(d_mat) );
    cudaErrChk( cudaFree(d_results) );

    delete [] h_vec;
    delete [] h_mat;
    delete [] h_results;

    return 0;
}

用

nvcc -o main main.cu -lcublas

Answer 1

正如@talonmies所指出的，问题在于我使用的是异步调用，而没有及时获得结果。这是通过在cudaDeviceSynchronize()调用之后添加cublasDgemv()来修正的：

#include <cuda.h>
#include <cuda_runtime.h>
#include <iostream>
#include <ctime>
#include "cublas_v2.h"

#define cudaErrChk(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);
        if (abort) exit(code);
    }
}

static const char *cublasErrChk(cublasStatus_t error)
{
    switch (error)
    {
        case CUBLAS_STATUS_SUCCESS:
            return "CUBLAS_STATUS_SUCCESS";

        case CUBLAS_STATUS_NOT_INITIALIZED:
            return "CUBLAS_STATUS_NOT_INITIALIZED";

        case CUBLAS_STATUS_ALLOC_FAILED:
            return "CUBLAS_STATUS_ALLOC_FAILED";

        case CUBLAS_STATUS_INVALID_VALUE:
            return "CUBLAS_STATUS_INVALID_VALUE";

        case CUBLAS_STATUS_ARCH_MISMATCH:
            return "CUBLAS_STATUS_ARCH_MISMATCH";

        case CUBLAS_STATUS_MAPPING_ERROR:
            return "CUBLAS_STATUS_MAPPING_ERROR";

        case CUBLAS_STATUS_EXECUTION_FAILED:
            return "CUBLAS_STATUS_EXECUTION_FAILED";

        case CUBLAS_STATUS_INTERNAL_ERROR:
            return "CUBLAS_STATUS_INTERNAL_ERROR";
    }

    return "<unknown>";
}

int main() {

    size_t dims = 4;

    double *vec, *mat, *results;

    cudaErrChk( cudaMallocManaged(&vec, dims * sizeof(double)) );
    cudaErrChk( cudaMallocManaged(&mat, dims * dims * sizeof(double)) );
    cudaErrChk( cudaMallocManaged(&results, dims * sizeof(double)) );

    printf("Vector:\n");
    for (int i = 1; i < dims + 1; i++) {
        vec[i] = 0.5 * i;
        printf("%.2lf ", vec[i]);
    } 
    printf("\n\nMatrix:\n");

    for (int i = 1; i < dims * dims + 1; i++) {
        mat[i] = 1.0 * i;
        printf("%.2lf ", mat[i]);

        if (i % dims == 0)
            printf("\n");
    }
    printf("\n");

    cublasHandle_t handle;
    cublasErrChk( cublasCreate(&handle) );

    double alpha = 1.f, beta = 1.f;

    // multiply mat by vec to get results
    cublasErrChk(
        cublasDgemv(
            handle, CUBLAS_OP_N,
            dims, dims,
            &alpha,
            mat, dims,
            vec, 1,
            &beta,
            results, 1
        )
    );
    cudaDeviceSynchronize();

    for (int i = 0; i < dims; i++)
        printf("%.2lf ", results[i]);
    printf("\n");

    cudaErrChk( cudaFree(vec) );
    cudaErrChk( cudaFree(mat) );
    cudaErrChk( cudaFree(results) );

    return 0;
}