Cross process problem in libargus

[happy886]

tegra_multimedia_api and argus samples show interop : argus->IEGLOutputStream->eglstream->Acquire frame
or argus->IEGLOutputStream->IFrameConsumer->Acquire frame
In cross process ,just as my other topic:Cross eglstream copy CUeglFrame failed
this is the cpu performance:producer 2.0，consumer 5.3
(just gpu to cpu copy cuMemcpy3D
producer：argus->…
consumer:>cuEGLStreamConsumerAcquireFrame->cuGraphicsResourceGetMappedEglFrame-> cuMemcpy3D ->cuEGLStreamConsumerReleaseFrame
)

if i need 3 comsuer(yolo detect/face detect /ros usage ) for the 1 camera argus producer,it will cost much(2.0+3*5.3=17.9),how can i deal with this problem on my cross process ? any zero-copy method?The yolo/face/ros usage are developed by different person on 3 separate processes.
IImageNativeBuffer dma-buf can Cross process processing？any zero-copy method

[ShaneCCC]

Have you check the cudaBayerDemosaic and cudaHistogram MMAPI sample.

[happy886]

cudaHistogram cpu【1.3% ~2.0%】,so the argus and eglstream cost little;gpu data process will not cost cpu;

en,cpu cost much for the GPU to CPU memcpy copy!so just for my project usecase,i need to avoid date copy (zero copy)or just copy once!
so how can i deal with my problems on the best performance? one process for argus to producer camera frame;one yolo detect frame consumer process,the second face detect frame consumer process,and the third ros frame consumer process.

[happy886]

Mainly data copy takes up CPU,
Is there a better way for cross process usecase? Is there a more appropriate technical route for one producer process and more consumer process?

[AastaLLL]

Hi,

Not sure if we understand your question correctly.

In /usr/src/jetson_multimedia_api/argus/samples/cudaHistogram.
After mounting the buffer with cuGraphicsResourceGetMappedEglFrame, you can get a GPU-accessible buffer pointer.

This indicates that you don’t need to copy the buffer from CPU to GPU.
Instead, just create a wrapper and use it as zero-copy data.

cuResult = cuGraphicsResourceGetMappedEglFrame(&cudaEGLFrame, cudaResource, 0, 0);
...
CUDA_RESOURCE_DESC cudaResourceDesc;
memset(&cudaResourceDesc, 0, sizeof(cudaResourceDesc));
cudaResourceDesc.resType = CU_RESOURCE_TYPE_ARRAY;
cudaResourceDesc.res.array.hArray = cudaEGLFrame.frame.pArray[0];
CUsurfObject cudaSurfObj = 0;
cuResult = cuSurfObjectCreate(&cudaSurfObj, &cudaResourceDesc);

Thanks.

[happy886]

cudaHistogram sample,the wrapper of CUDA_RESOURCE_DESC for “histogram.cu”,it is simple applications;how can i use the wrapper of CUDA_RESOURCE_DESC on complex applications?
Eg:the cudaHistogram process get the wapper of CUDA_RESOURCE_DESC
the yolo detect process need the frame date,the wapper data ;
the face detect process also need the frame date,the wapper data ;
the Ros Visual navigation process also need the frame date,the wapper data ;
Is there a more appropriate technical route for one producer process and more consumer process?the cross process application
the argus camera frame data need to shared by multiple processes applications with zero copy

[happy886]

In fact,if my project is interop,multiple thread,it is easy;just use NvBuffer to get dma-buf fd,and use NvBufferTransform to copy the dma-buf;
But on cross process applications,the dma-buf can not Cross-process sharing

[happy886]

how about it?
No interop ,no multithreading，no single process，no one consumer!
Cross process! one argus camera producer,multiple consumer（about 3 consumer）

[AastaLLL]

Hi,

Do you want an example for across process NvBuffer access.
Or an example that access the NvBuffer in multi-thread but the same process?

Thanks.

[happy886]

@AastaLLL

ulti-thread but the s

I think cross process NvBuffer can can solve this!

[happy886]

@happy886

I think cross process NvBuffer can can solve this!

I think cross process NvBuffer can can solve this!
But no example,could you give me some advices?

[DaneLLL]

Hi,
The sample is under internal review. For implementation, you can check discussion in
How to share the buffer in process context？