VPI 3.2 Image Wrapper Performance Problems when using CudaMallocManaged

[marc.haubenstock]

Hello, similar to Performance issues using vpiImageCreate*Wrapper

I have performance problems when using the following code

                        VPIImage img;
                        VPIImageData data = {};
                        data.bufferType = VPI_IMAGE_BUFFER_CUDA_PITCH_LINEAR;
                        data.buffer.pitch.numPlanes = 1;
                        data.buffer.pitch.format = VPI_IMAGE_FORMAT_U8;
                        data.buffer.pitch.planes[0].width  = mvImagePyramid[level]->getWidth();
                        data.buffer.pitch.planes[0].height = mvImagePyramid[level]->getHeight();
                        data.buffer.pitch.planes[0].data   = mvImagePyramid[level]->getRaw(); // get ptr allocated by CudaMallocManaged
                        data.buffer.pitch.planes[0].pitchBytes = mvImagePyramid[level]->getStep();
                        //data.buffer.pitch.planes[0].pixelType = VPI_PIXEL_TYPE_DEFAULT;

                        vpiImageCreateWrapper(&data, nullptr, 0, &img);

I just wanted to ask if there is anything else I can do/if this is a problem/ if you are aware of this problem?

Best,
Marc

[AastaLLL]

Hi,

Would you mind sharing more info about your issues?

Do you meet the performance issues for the wrapper?
Or an VPI algorithm? If so, could you share a sample cod so we can check?

Thanks

[marc.haubenstock]

The wrapper seems to be slower than expect. Granted: I double checked and I was running this in a very hot path. But I would like to know: Does the Wrapper do any memory allocations/check any memory region?

Ill try to generate a better example

[AastaLLL]

Hi,

A sample will help us to understand your question.

As VPI supports various backends, it’s possible to have some internal memory allocation/movement.
But if the buffer won’t be destroyed during the pipeline, you can do the wrap and unwrap just in the beginning and terminated.

Thanks.

[marc.haubenstock]

Up to this point I didnt set backend yet. All the VPI code is as you see. But I think we will use a cuda kernel solution instead of VPI

[AastaLLL]

Hi

Setting the flags to 0 indicates the image can be accessed via all possible backends.

vpiImageCreateWrapper

So the wrapper will automatically trigger some mechanism (ex. prepare buffer for different backends).

Thanks.