cuda 記憶體拷貝速率

cuda 記憶體拷貝速率問題

`` CPU與GPU之間記憶體拷貝速率，一個9M的圖檔，CPU->GPU耗時9ms，GPU->CPU耗時11ms，項目要求處理的是大圖，對每一個像素點做處理，以400M的圖檔為例子，cpu與gpu之間拷貝均耗時200多ms。顯示卡帶寬為192GB/s，(我認為的)正常情況應該傳輸速率很快，請問一下有什麼需要修改的地方嗎（剛剛接觸GPU的萌新）。下圖為GPU型号：

void Binaryzation_GPU(CppColorRange* range_host,byte* src_host, byte* dst_host, int width, int height)
{
	int length = width * height;
	int memSize = length * sizeof(uchar4);
	CppColorRange range_devive = *range_host;
	uchar4* src_device;
	uchar4* dst_device;
	cudaMalloc((void**)&src_device, memSize);
	cudaMallocHost((void**)&dst_device, memSize);
	//copy
	cudaMemcpy((void**)src_device, src_host, memSize, cudaMemcpyHostToDevice);
	cudaMemcpy((void**)dst_device, dst_host, memSize, cudaMemcpyHostToDevice);
	dim3 threadsPerBlock(32, 32);
	dim3 blocksPerGrid((width + threadsPerBlock.x - 1) / threadsPerBlock.x, (height + threadsPerBlock.y - 1) / threadsPerBlock.y);
	cudaDeviceSynchronize();
	kernel_Binaryzation << <blocksPerGrid, threadsPerBlock >> >(range_devive, src_device, dst_device, width, height);
	cudaMemcpy((void**)dst_host, dst_device, memSize, cudaMemcpyDeviceToHost);
	cudaFree(src_device);
	cudaFree(dst_device);
}