How to access DMA buffer from CUDA device?

[cailiudel]

Hi,

I was trying to use Jetson NANO with a USB camera. I want to somehow feed the camera image into CUDA. I was following the 12_camera_v4l2_cuda (https://docs.nvidia.com/jetson/l4t-multimedia/l4t_mm_12_camera_v4l2_cuda.html) example. It says I can apply CUDA code in the function HandleEGLImage, but it is applying the CUDA process on EGL images. I don’t want to use EGL images. Can I somehow map a device pointer to a DAM buffer?

In the camera_v4l2_cuda.cpp example file, it gets a dmabuff_fd during the camera stream.

if (ctx->capture_dmabuf) {
  // Cache sync for VIC operation
  NvBufferMemSyncForDevice(ctx->g_buff[v4l2_buf.index].dmabuff_fd, 0,
      (void**)&ctx->g_buff[v4l2_buf.index].start);
} else {
  Raw2NvBuffer(ctx->g_buff[v4l2_buf.index].start, 0,
       ctx->cam_w, ctx->cam_h, ctx->g_buff[v4l2_buf.index].dmabuff_fd);
}

I wanted to use device to device CUDA memory copy from ctx->g_buff[v4l2_buf.index].start directly to a allocated device pointer, assuming it’s ready for the GPU usage. However, it seems not the case and I always get zeros in the copied memory.

I’m wondering what the correct way is. Thank you very much!

Best,
Caili

[DaneLLL]

Hi,
Please look at cuda_postprocess() in 12_camera_v4l2_cuda, and HandleEGLImage() in

/usr/src/jetson_multimedia_api/samples/common/algorithm/cuda

You can get GPU pointer by following the function calls.