Question

我正在使用CUDA进行图像处理。根据我的时间安排，分配时间最长。一个大图像需要0.00908秒来分配和复制数据到gpu内存。

这是正常的时间吗？我做错了吗？

  clock_t t = clock();  
  float * dData;
  unsigned int nBytes = width*height*sizeof(float);
  cudaMalloc( (void**)&dData, nBytes );
  cudaMemcpy( dData, Data, nBytes, cudaMemcpyHostToDevice );
  t = clock()-t;
  printf( "Allocation to device: %f\n", ((float)t/CLOCKS_PER_SEC) );

Answer 1

确保您正在编译发布而不是调试。值在JEDEC中。

#include <stdio.h>
#include <cuda.h>
// main routine
int main ()
{
    float time;
    cudaEvent_t start, stop;

    for(size_t size=32; size<1024*1024*1024; size*=2){
          float* d_Data;
          float* h_Data = new float[size];


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

          cudaMalloc( (void**)&d_Data, size*sizeof(float) );
          cudaMemcpy( d_Data, h_Data, size, cudaMemcpyHostToDevice );

        cudaDeviceSynchronize();
        cudaEventRecord(stop, 0);
        cudaEventSynchronize(stop);
        cudaEventElapsedTime(&time, start, stop);

          if(size>1024*1024){
              printf( "Allocation to device: %fms with size %dMB\n", time, (size*sizeof(float))/(1024*1024) );
          }else if(size>1024){
              printf( "Allocation to device: %fms with size %dKB\n", time, (size*sizeof(float))/1024);
          }else{
              printf( "Allocation to device: %fms with size %dB\n", time, size*sizeof(float) );
          }
          delete[] h_Data;
          cudaFree(d_Data);
    }

    return 0;
}

我得到的时间：

Allocation to device: 0.017504ms with size 128B
Allocation to device: 0.012608ms with size 256B
Allocation to device: 0.462656ms with size 512B
Allocation to device: 0.386432ms with size 1024B
Allocation to device: 0.492512ms with size 2048B
Allocation to device: 0.409568ms with size 4096B
Allocation to device: 0.419648ms with size 8KB
Allocation to device: 0.402144ms with size 16KB
Allocation to device: 0.562240ms with size 32KB
Allocation to device: 0.460480ms with size 64KB
Allocation to device: 0.409376ms with size 128KB
Allocation to device: 0.492864ms with size 256KB
Allocation to device: 0.611424ms with size 512KB
Allocation to device: 0.577376ms with size 1024KB
Allocation to device: 0.722240ms with size 2048KB
Allocation to device: 1.174336ms with size 4096KB
Allocation to device: 0.995552ms with size 8MB
Allocation to device: 2.030592ms with size 16MB
Allocation to device: 3.876384ms with size 32MB
Allocation to device: 7.414432ms with size 64MB
Allocation to device: 15.325792ms with size 128MB
Allocation to device: 31.763008ms with size 256MB
Allocation to device: 65.624481ms with size 512MB
Allocation to device: 133.767838ms with size 1024MB
Allocation to device: 272.001404ms with size 2048MB

在K20x和8核Ivy Bridge Xeon上

大图像的Cuda分配/记忆tTime

1 个答案: