gpt4 book ai didi

cuda - CudaMalloc 如何工作?

转载 作者:行者123 更新时间:2023-12-01 05:44:36 26 4
gpt4 key购买 nike

我正在尝试修改 CUDA SDK 中的 imageDenosing 类,我需要多次重复过滤器以捕捉时间。但是我的代码不能正常工作。

//开始

__global__ void F1D(TColor *image,int imageW,int imageH, TColor *buffer)
{

const int ix = blockDim.x * blockIdx.x + threadIdx.x;
const int iy = blockDim.y * blockIdx.y + threadIdx.y;

if(iy != 0 && iy < imageH-1 && ix < imageW)
{

float4 fresult = get_color(image[imageW * iy + ix]);
float4 fresult4 = get_color(image[imageW * (iy+1) + ix]);
float4 fresult5 = get_color(image[imageW * (iy-1) + ix]);

float4 fresult7;
fresult7.x = fresult.x*0.5+fresult4.x*.25+fresult5.x*.25;
fresult7.y = fresult.y*0.5+fresult4.y*.25+fresult5.y*.25;
fresult7.z = fresult.z*0.5+fresult4.z*.25+fresult5.z*.25;

buffer[imageW * iy + ix] =
make_color(fresult7.x,fresult7.y,fresult7.z,0);

}

image[imageW * iy + ix] = buffer[imageW * iy + ix];
//should be use cudaMemcpy, But it fails
}

//外部

extern "C" void
cuda_F1D(TColor *dst, int imageW, int imageH)
{
dim3 threads(BLOCKDIM_X, BLOCKDIM_Y);
dim3 grid(iDivUp(imageW, BLOCKDIM_X), iDivUp(imageH, BLOCKDIM_Y));
Copy<<<grid, threads>>>(dst, imageW, imageH);

size_t size = imageW*imageH*sizeof(TColor);
TColor *host =(TColor*) malloc(size);
TColor *dst2;
//TColor *dst3;
//TColor *d = new TColor(imageW*imageH*sizeof(TColor));
dim3 threads2(imageW,1);
dim3 grid2(iDivUp(imageW, imageW), iDivUp(imageH, 1));

*for(int i = 0;i<1;i++)
{
cudaMalloc( (void **)&dst2, size);
cudaMemcpy(dst2, dst, imageW*imageH*sizeof(TColor),cudaMemcpyHostToDevice);
//cudaMalloc( (void **)&dst3, imageW*imageH*sizeof(TColor));
//cudaMemcpy(dst3, dst, imageW*imageH*sizeof(TColor),cudaMemcpyHostToDevice);
F1D<<<grid2, threads2>>>(dst, imageW, imageH,dst2);
//cudaMemcpy(dst, dst3, imageW*imageH*sizeof(TColor),cudaMemcpyDeviceToHost);
cudaFree(dst2);
}*

}

此代码有效,但无法同步图像数组。并导致许多同步问题

最佳答案

您的内核正在异步运行 - 您需要等待它完成,例如

cudaMalloc((void **)&dst2, size);
cudaMemcpy(dst2, dst, imageW * imageH * sizeof(TColor), cudaMemcpyHostToDevice);
F1D<<<grid2, threads2>>>(dst, imageW, imageH, dst2);
cudaThreadSynchronize(); // *** wait for kernel to complete ***
cudaFree(dst2);

关于cuda - CudaMalloc 如何工作?,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/2938165/

26 4 0
Copyright 2021 - 2024 cfsdn All Rights Reserved 蜀ICP备2022000587号
广告合作:1813099741@qq.com 6ren.com