NVIDIA multimedia APIs support converting UYVY to RGB ?

Autonomous Machines Jetson & Embedded Systems Jetson TX1
1

Dear,

After spending time with multimedia APIs. Now I can capture and record videos file as below.
Camera (UYVY) → Video converter (YUV420M) → Encoder (H264) → FFmpeg (mp4).

I also want to convert UYVY to RGB data for other feature. So is it possible to convert form UYVY data to RGB data with HW accelerator?

Thanks and Best Regards,

2

Hi Vu,
Please refer to
[url]https://devtalk.nvidia.com/default/topic/989438/jetson-tx1/neon-optimised-vyuy2bgr/post/5067113/#5067113[/url]

Video converter can do UYVY → RGBA. Please refer to the code and apply your case.

3

Hi DaneLLL,

Thank for your reply. I can convert form YUVY to ABGR now.
However I can not find how to convert from ABGR to BGR with CUDA in backend samples.
Can you help me figure out.

Thanks and Best Regards,

4

Hi Vu, please refer to mapEGLImage2Float()
It is done in convertIntToFloatKernel() in tegra_multimedia_api\samples\common\algorithm\cuda\NvAnalysis.cu

5

Hi DaneLLL,

In this callback I can get buffer and size of ABGR image through buffer->planes[0].data and buffer->planes[0].bytesused. How can I put this to CUDA converter as above. I don’t want to use any EGLdisplay.

static bool conv0_rgb_capture_dqbuf_thread_callback(struct v4l2_buffer *v4l2_buf,
                                   NvBuffer * buffer, NvBuffer * shared_buffer,
                                   void *arg) {
    context_t *ctx = (context_t *) arg;

    if (ctx->conv_rgb->capture_plane.qBuffer(*v4l2_buf, buffer) < 0)
    {
        abort(ctx);
        ERROR_RETURN("Failed to queue buffer on VIC capture plane");
    }

    return true;
}

Thanks and Best Regards,

6

Hi Vu,
It requires the two APIs to get egl_image:
NvEGLImageFromFd()
NvDestroyEGLImage()

EGLDisplay is required.

7

Hi DaneLLL,

EGLDisplay is required but we no need render display?

I tried to convert with opencv and measure the result, it take about 4-5ms for converting 2560x1440 frame.
How much do you think H/W can finish conversion?

mapEGLImage2Float(void* pEGLImage, int width, int height, void* cuda_buf)
pEGLImage is ARGB input and cuda_buf is RGB output?

Thanks and Best Regards,

8

Hi Vu,
You are able to do profiling by adding gettimeofday() in backend sample. Please refer to tegra_multimedia_api\README

9

Hi DaneLLL,

mapEGLImage2Float(void* pEGLImage, int width, int height, void* cuda_buf)
pEGLImage is ARGB input and cuda_buf is RGB output?

Is it correct ?

10

Hi Vu,

convertIntToFloatKernel(CUdeviceptr pDevPtr, int width, int height,
                void* cuda_buf, int pitch)
{
    float *pdata = (float *)cuda_buf;
    char *psrcdata = (char *)pDevPtr;
    int row = blockIdx.y * blockDim.y + threadIdx.y;
    int col = blockIdx.x * blockDim.x + threadIdx.x;

    if (col < width && row < height)
    {
        for (int k = 0; k < 3; k++)
        {
            pdata[width * height * k + row * width + col] =
                (float)*(psrcdata + row * pitch + col * 4 + (3 - 1 - k));
        }
    }
}

The input is RGBA int and the output is BGR float