Hardware Platform: dGPU **• DeepStream Version: 7.0 ** • TensorRT Version: 8.6.1 • NVIDIA GPU Driver Version: 550 • Issue Type: question We have been trying to use GPU preprocessing on nvdspreprocess to avoid mem copying. Currently we have 2 option: CVCUDA, VPI. May i ask, what is the preferred…

The sample of operation CVCUDA APIs are in the future plan of DeepStream SDK.

From DeepStream point, it already provides the NvBufSurface interface for you to access the CUDA buffer containing video data. It is OK for you to handle the CUDA buffer in the way you prefer. Unfortunately, we don’t provide DeepStream sample for operating with other SDKs. There is already sample…

Yes we can get buffer with opencv Cuda already but we want to use CVCUDA for better performance but there are no working samples even though we found some information on using CVCUDA with Deepstream in GTC: Advancing AI Applications with Custom GPU-Powered Plugins for NVIDIA DeepStream. But other t…

There is no such plan. If you want to use CVCUDA to handle the CUDA buffer you got from DeepStream, you can refer to CVCUDA samples. If you want to feed CUDA buffer you got from CVCUDA to DeepStream, it is not supported. There are lots of buffer alignment limitations with the NvBufSurface.

For CVCUDA, i have tried using this code (similar to Advancing AI Applications with Custom GPU-Powered Plugins for NVIDIA DeepStream S51612 | GTC Digital Spring 2023 | NVIDIA On-Demand ). We can warp the buffer but resize/cropping cause segmentation fault. NvDsFrameMeta *frame_meta = (NvDsFr…

[image] namogg1012: cv::Mat input_frame(surf_params->height, surf_params->width, CV_8UC4, surf_params->dataPtr, surf_params->pitch); int frame_height = static_cast<int>(surf_params->height); int frame_width = static_cast<int>(surf_params->width); cv::Mat rgb_frame; …

I wouldn’t waste my time with CVCUDA, it’s obviously not supported because I asked a question 5 months ago and it’s never been answered: [QUESTION] Can this be used with the Nvidia Deepstream NvBufSurface? · Issue #230 · CVCUDA/CV-CUDA · GitHub

I have, could never figure out how to properly map an NvBufSurface to the NvSurface that VPI offers(they are different).

Me too, i have seen some samples on jetson but never got my code on dGPU to map correctly.

After some consideration, using NPP is actually very straight forward and we have successfully warp image using NPP. Using NPP and some custom cuda kernal is enough to replace open-cv CPU.

CVCUDA or VPI on Deepstream dGPU?

Accelerated Computing Intelligent Video Analytics DeepStream SDK

system Closed September 5, 2025, 7:01am 16

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.

Topic		Replies	Views
VPI on dGPU and Jetson DeepStream SDK deepstream	6	194	February 24, 2025
Deepstream VPI 3.0 Wrapper Compatibility dGPU DeepStream SDK deepstream	3	179	January 14, 2025
Copy DeepStream frame to cv2.cuda_GpuMat object DeepStream SDK	9	979	April 4, 2023
Using VPI in a Custom DeepStream Plugin DeepStream SDK gstreamer , vpi	4	1983	October 12, 2021
How to use pyds and opencv-python with cuda_GpuMat mapped with NvBufSurface DeepStream SDK opencv	6	743	November 9, 2022
VPI + DeepStream is slower then expected (only wrapping) DeepStream SDK	4	813	November 15, 2022
Deepstream SDK + VPI on Jetson tx2 DeepStream SDK nvbugs	19	4109	October 12, 2021
Offsetting NVMM RGBA Image using VPI and DeepStream DeepStream SDK vpi , deepstream	10	1291	April 24, 2022
Deepstream surface with pycuda DeepStream SDK	2	587	November 26, 2022
From nvBufSurface to gpu::Mat or cuda stream DeepStream SDK cuda , deepstream	15	378	February 6, 2025

CVCUDA or VPI on Deepstream dGPU?

Related topics