Question

这是我的代码：

using namespace std;
#include <iostream>
#include <stdio.h>
#include <stdlib.h>

const int N = 8000;

void fillArray(int *data, int count) {
    for (int i = 0; i < count; i++)
        data[i] = rand() % 100;
}

__global__ void add(int* a, int *b, int *c) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    if (tid < N) {
        c[tid] = a[tid] + b[tid];
    }
}

__global__ void subtract(int* a, int *b, int *c) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    if (tid < N) {
        c[tid] = a[tid] - b[tid];
    }
}

__global__ void multiply(int* a, int *b, int *c) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    if (tid < N) {
        c[tid] = a[tid] * b[tid];
    }
}

__global__ void divide(int* a, int *b, int *c) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    if (tid < N) {
        c[tid] = a[tid] / b[tid];
    }
}

__global__ void modu(int* a, int *b, int *c) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    if (tid < N) {
        c[tid] = a[tid] % b[tid];
    }
}

__global__ void neg(int *data, int *c) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    if (tid < N) {
        c[tid] = -data[tid];
    }
}

float duration(int *devA, int *devB, int *devC, int blocksPerGrid, int threadsPerBlock) {

    cudaEvent_t start, stop;
    float elapsedTime;

    cudaEventCreate(&start);
    cudaEventCreate(&stop);
    cudaEventRecord(start, 0);

    int hArrayC[N];

    add<<<blocksPerGrid, threadsPerBlock>>>(devA, devB,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    subtract<<<blocksPerGrid, threadsPerBlock>>>(devA, devB,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    multiply<<<blocksPerGrid, threadsPerBlock>>>(devA, devB,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    divide<<<blocksPerGrid, threadsPerBlock>>>(devA, devB,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    modu<<<blocksPerGrid, threadsPerBlock>>>(devA, devB,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    neg<<<blocksPerGrid, threadsPerBlock>>>(devA,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    neg<<<blocksPerGrid, threadsPerBlock>>>(devB,devC);
    cudaMemcpy(hArrayC,devC,N*sizeof(int),cudaMemcpyDeviceToHost);

    cudaEventRecord(stop, 0);
    cudaEventSynchronize(stop);
    cudaEventElapsedTime(&elapsedTime, start, stop);

    cudaEventDestroy(start);
    cudaEventDestroy(stop);

    return elapsedTime;
}

int main(void) {

    int *a, *b;
    a = new int[N];
    b = new int [N];

    float dur = 0;

    int *devA, *devB,*devC;

    cudaMalloc((void**) &devA, N * sizeof(int));
    cudaMalloc((void**) &devB, N * sizeof(int));
    cudaMalloc((void**) &devC, N * sizeof(int));

    fillArray(a, N);
    fillArray(b, N);

    cudaMemcpy(devA, a, N * sizeof(int), cudaMemcpyHostToDevice);
    cudaMemcpy(devB, b, N * sizeof(int), cudaMemcpyHostToDevice);


    dur = duration(devA, devB, devC,N, 1);

    cout << "Global memory version:\n";
    cout << "Process completed in " << dur;
    cout << " for a data set of " << N << " integers.";



    cudaFree(devA);
    cudaFree(devB);
    delete [] a;
    delete [] b;

    return 0;
}

我想知道持续时间函数的总毫秒数。但是，毫秒总是以不同的值返回。有时它是10毫秒有时它是0.78652有时它是30毫秒。为什么？我的代码出了什么问题？

Answer 1

这可能是由NVIDIA驱动程序的加载/卸载引起的。将其视为GPU的初始化步骤。

您可以将GPU设置为持久模式：

nvidia-smi -pm 1

或者您可以在计算GPU代码以触发加载驱动程序之前运行虚拟内核：

__global__ void dummy()
{
    // This kernel does nothing, this is just a "warm-up"
}

// Before your cudaEventRecord etc.
dummy<<<blocksPerGrid, threadsPerBlock>>>();

或者可以在计算内核之前使用cudaThreadSynchronize()。

CUDA性能 - 始终返回不同的值

1 个答案: