How to copy GPU buffer(from nppiMalloc_8u_C4), to dmabuf_fd(from NvBufferCreateEx)

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Xavier
1

hi,

Here is my way to copy, but it is to slow. Is there a way to device_to_device?

init() and process() can work fine with one camera(60fps), function process() just cost 2-3ms.
But when I use 4 cameras(each has 60fps), function process() cost 19 ms.

init()
{
…
NvBufferGetParams(m_argb_dmabuf_fd,&m_argb_parm);//pitch 9984
NvBufferMemMap(m_argb_dmabuf_fd, 0, NvBufferMem_Read_Write, &m_argb_dmabuf_buffer);

vesc_bayer_dev_ = nppiMalloc_8u_C1(CAMERA_IMAGE_WIDTH, CAMERA_IMAGE_HEIGHT, &resv_bayer_step_);//pitch 2560
resv_rgba_dev_ = nppiMalloc_8u_C4(CAMERA_IMAGE_WIDTH, CAMERA_IMAGE_HEIGHT, &resv_rgba_step_);  //pitch 10240

}

process()
{
…

NvBufferMemSyncForCpu(m_argb_dmabuf_fd, 0, &m_argb_dmabuf_buffer);

cudaMemcpy2D(m_argb_dmabuf_buffer, m_argb_parm.pitch[0], resv_rgba_dev_, 24484, 24484, 2048, cudaMemcpyDeviceToHost);

NvBufferMemSyncForDevice(m_argb_dmabuf_fd, 0, &m_argb_dmabuf_buffer);

}

I am trying blow code, but it doesn’t work.
cudaMemcpy2D(m_argb_parm.nv_buffer, m_argb_parm.pitch[0], resv_rgba_dev_, 24484, 24484, 2048, cudaMemcpyDeviceToDevice);

2

Hi,

m_argb_dmabuf_fd looks like a CPU pinned buffer to me.
Could you give cudaMemcpyDeviceToHost a try.

If it is still not working, would you mind to provide a simple reproducible source for us checking?
Thanks.

3

Hi afhel,

Have you managed to get it working? Any result can be shared?

4

Hi,kayccc

No,It is not worked.
The width of bayergb from camera is 2448, but in nppc width is 2560.
And The width of rgba in nppc is 10240, but in m_argb_dmabuf_buffer is 9984.

So I have to copy them one line by one line,using cudaMemcpy2D.

cudaMemcpy2D spend a lot of time.

5

Hi,AastaLLL

I will provide a simple a few days later.

6

Hi,

Thanks for the update.
A simple reproducible source will definitely help us figure out this issue.

Thanks.