cudaMemcpy study notes

The host is the computer Host

Device is the gpu device.

cudaMemcpy is used to transfer data back and forth between the host (Host) and the device (Device), the usage is as follows:

Host to Device: cudaMemcpy(d_A,h_A,nBytes,cudaMemcpyHostToDevice)

Device to host: cudaMemcpy(h_A,d_A,nBytes,cudaMemcpyDeviceToHost)

Note: This function is a synchronous execution function. It will be locked before the data transfer operation is completed and will always occupy the control of the CPU process, so there is no need to add the cudaDeviceSynchronize() function.

Example:

void VectorReal_copyBuffer(realVecHandle* addr, float hostArr[], long buffer, bool tohost, int copyMode){realVecHandle * native =addr;float * host = hostArr;if ( copyMode == 0 ) {if (!tohost) {cudaMemcpy( native->data,host, native->size, cudaMemcpyHostToDevice) ;} else {cudaMemcpy( host, native->data, native->size, cudaMemcpyDeviceToHost) ;}}}

//Allocate space on video memory

CUDA_SAFE_CALL(cudaMalloc((void**)&Dst_d,sizeof(float3) * totalPNum));

//Transfer data on video memory

CUDA_SAFE_CALL(cudaMemcpy(Dst_d, Srcdata0_h, sizeof(float3) * numdata0, cudaMemcpyHostToDevice));
CUDA_SAFE_CALL(cudaMemcpy(Dst_d + numpoint0, Srcdata1_d, sizeof(float3) * numdata1, cudaMemcpyDeviceToDevice));
CUDA_SAFE_CALL(cudaMemcpy(Dst_d + numpoint0 + numpoint1, Srcdata2_d, sizeof(float3) * numdata2, cudaMemcpyDeviceToDevice));

During the data transmission process, the destination address is before the source address (the address may have an offset calculation, when the data source has the same purpose for different data sources), corresponding to the latter parameter, devicetohost or hosttodevice.

Original link: https://blog.csdn.net/xiaoheibaqi/article/details/44001273