Converting CUeglFrame to VPIImage

[mohamed.abdelaziz1]

Hi, I am using LibArgus to capture frames from a camera on an EGLStream. Then I use the CUDA EGL Interop API to acquire the image as CUeglFrame

output_stream_settings_.reset(i_session_->createOutputStreamSettings(Argus::STREAM_TYPE_EGL));

....

cuEGLStreamConsumerAcquireFrame(&cuda_connection_, &cuda_resource_, &cuda_stream_, CUDA_EGL_INFINITE_TIMEOUT);

cuGraphicsResourceGetMappedEglFrame(&cuda_egl_frame_, cuda_resource_, 0, 0);

Debugging the output frame returns the following

CUeglFrame:
width: 1920
height: 1200
depth: 0
pitch: 0
planeCount: 2
numChannels: 1
frameType: array
colorFormat: YUV420 semi-planar ER
cuFormat: uint8

I need to wrap this image into a VPIImage. I attempted to create a VPI_IMAGE_BUFFER_CUDA_ARRAY buffer and set its data to the array contained in the CUeglFrame, but this didn’t work since VPICreateWrapper doesn’t work with VPI_IMAGE_BUFFER_CUDA_ARRAY buffers

VPIImageBuffer vpi_image_buffer;
vpi_image_buffer.cudaarray = reinterpret_cast<cudaArray_t>(cuda_egl_frame_.frame.pArray[0]);
VPIImageData vpi_image_data;
vpi_image_data.buffer = vpi_image_buffer;
vpi_image_data.bufferType = VPI_IMAGE_BUFFER_CUDA_ARRAY;

VPI_ERROR_INVALID_ARGUMENT: Wrapping of VPI_IMAGE_BUFFER_CUDA_ARRAY type isn't supported

Is there a way to achieve this EGL ↔ CUDA ↔ VPI Interoperability? Maybe by copying the data first to an NVBUFFER which is compatible with VPI, but I am not sure how to copy the data to that NvBuffer.

[mohamed.abdelaziz1]

I attempted to allocate pitched memory and copy the data into it from the array like this

VPIImageBuffer vpi_image_buffer;
vpi_image_buffer.pitch.format = VPI_IMAGE_FORMAT_NV12_ER;
vpi_image_buffer.pitch.numPlanes = 2;
// Plane 0 (Y)
vpi_image_buffer.pitch.planes[0].width = cuda_egl_frame_.width;
vpi_image_buffer.pitch.planes[0].height = cuda_egl_frame_.height;
vpi_image_buffer.pitch.planes[0].pixelType = VPI_PIXEL_TYPE_U8;
// Plane 1 (UV)
vpi_image_buffer.pitch.planes[1].width = cuda_egl_frame_.width / 2;
vpi_image_buffer.pitch.planes[1].height = cuda_egl_frame_.height / 2;
vpi_image_buffer.pitch.planes[1].pixelType = VPI_PIXEL_TYPE_2U8;

// Allocate memory
size_t y_pitch;
auto cuda_response = cudaMallocPitch(
     &vpi_image_buffer.pitch.planes[0].data,
     &y_pitch,
     vpi_image_buffer.pitch.planes[0].width,
     vpi_image_buffer.pitch.planes[0].height
);
vpi_image_buffer.pitch.planes[0].pitchBytes = y_pitch;

size_t uv_pitch;
cuda_response = cudaMallocPitch(
      &vpi_image_buffer.pitch.planes[1].data,
      &uv_pitch,
      vpi_image_buffer.pitch.planes[1].width,
      vpi_image_buffer.pitch.planes[1].height
);
vpi_image_buffer.pitch.planes[0].pitchBytes = uv_pitch;

// Copy data
cuda_response = cudaMemcpy2DFromArray(
    vpi_image_buffer.pitch.planes[0].data,
    uv_pitch,
    reinterpret_cast<cudaArray_t>(cuda_egl_frame_.frame.pArray[0]),
    0, 0, vpi_image_buffer.pitch.planes[0].width,
    vpi_image_buffer.pitch.planes[0].height,
    cudaMemcpyDeviceToDevice
);

cuda_response = cudaMemcpy2DFromArray(
    vpi_image_buffer.pitch.planes[1].data,
    uv_pitch,
    reinterpret_cast<cudaArray_t>(cuda_egl_frame_.frame.pArray[1]),
    0, 0, vpi_image_buffer.pitch.planes[1].width,
    vpi_image_buffer.pitch.planes[1].height,
    cudaMemcpyDeviceToDevice
);

// Create Image data
VPIImageData vpi_image_data;
vpi_image_data.buffer = vpi_image_buffer;
vpi_image_data.bufferType = VPI_IMAGE_BUFFER_CUDA_PITCH_LINEAR;

// Create VPIImage
VPIImage vpi_image;
CHECK_STATUS(
  vpiImageCreate(1920, 1200, 
    VPI_IMAGE_FORMAT_NV12_ER, VPI_BACKEND_CUDA, &vpi_image));

// Wrap the image
CHECK_STATUS(vpiImageCreateWrapper(&vpi_image_data, nullptr, VPI_BACKEND_CUDA, &vpi_image));

The wrapping step results in the following error

VPI_ERROR_INVALID_ARGUMENT: Plane row stride must be greater or equal row length in bytes

Edit:

I got it to work. The UV plane requires the width to be set to double during allocation and copying like this

cuda_response = cudaMallocPitch(
      &vpi_image_buffer.pitch.planes[1].data,
      &uv_pitch,
      vpi_image_buffer.pitch.planes[1].width * 2,
      vpi_image_buffer.pitch.planes[1].height
);

cuda_response = cudaMemcpy2DFromArray(
    vpi_image_buffer.pitch.planes[1].data,
    uv_pitch,
    reinterpret_cast<cudaArray_t>(cuda_egl_frame_.frame.pArray[1]),
    0, 0, vpi_image_buffer.pitch.planes[1].width * 2,
    vpi_image_buffer.pitch.planes[1].height,
    cudaMemcpyDeviceToDevice
);