GPU和CPU耗时统计方法

GPU端耗时统计

    cudaEvent_t start, stop;
    checkCudaErrors(cudaEventCreate(&start));
    checkCudaErrors(cudaEventCreate(&stop));
    checkCudaErrors(cudaDeviceSynchronize());

    float gpu_time = 0.0f;
    cudaEventRecord(start, 0);//cuda context中的操作完毕事件被记录
    //分配设备端内存
    float *d_idata;
    checkCudaErrors(cudaMalloc((void **) &d_idata, mem_size));
    
    //将主机端数据拷贝到设备端内存
    checkCudaErrors(cudaMemcpy(d_idata, h_idata, mem_size,  cudaMemcpyHostToDevice));

    //设备端为结果分配内存
    float *d_odata;
    checkCudaErrors(cudaMalloc((void **) &d_odata, mem_size));

    //设置执行参数
    dim3  grid(1, 1, 1);
    dim3  threads(num_threads, 1, 1);

    //执行内核，参数含义：grid是网格的纬度，threads是块的纬度，mem_size最多能动态分配的共享内存大小
    testKernel<<< grid, threads, mem_size >>>(d_idata, d_odata);

    //检查内核执行状态
    getLastCudaError("Kernel execution failed");

    //在主机端为结果分配内存
    float *h_odata = (float *) malloc(mem_size);
    //从设备端拷贝结果到主机端
    checkCudaErrors(cudaMemcpy(h_odata, d_odata, sizeof(float) * num_threads,
                               cudaMemcpyDeviceToHost));

    cudaEventRecord(stop, 0);
    unsigned long int counter = 0;
    while (cudaEventQuery(stop) == cudaErrorNotReady)
    {
        counter++;
    }
    checkCudaErrors(cudaEventElapsedTime(&gpu_time, start, stop));
    printf("GPU执行耗时: %.2f (ms)
", gpu_time);
    printf("CPU executed %lu iterations while waiting for GPU to finish
", counter);

CPU端耗时统计

    StopWatchInterface *timer = 0;
    sdkCreateTimer(&timer);
    sdkResetTimer(&timer);

    sdkStartTimer(&timer);
    //计算参考方案
    float *reference = (float *) malloc(mem_size);
    computeGold(reference, h_idata, num_threads);
    sdkStopTimer(&timer);
    printf("串行耗时：%f (ms)
", sdkGetTimerValue(&timer));