Example of using NVX_MEMORY_TYPE_CUDA_ARRAY

Autonomous Machines Jetson & Embedded Systems Jetson TX2
#1

Can someone provide an example of mapping/binding vx_image to a cuda array? There are some words in the Visionworks document section cuda-openvx interop but they are not very helpful. Mapping function, vxMapImagePatch(…) does exist but I don’t know how to get “struct nvx_cuarray_handle_2d_t” from it?

My goal is to be able to bind vx_image to a cuda texture so I can access top bottom pixels with cache misses. The only viable solution I can see is this flag, NVX_MEMORY_TYPE_CUDA_ARRAY. I can bind a CuArray as a cuda texutre and access cuda texture in a kernel. Please help.

In general, Visionworks document is not really strong. I hope it could be improved in the future.

#2

Use a different approach. Use “cudaBindTexture2D” to bind 2D device memory to a texture memory.

#3

Hi,

You can check our low level sample:
[VisionWorks-1.6-Samples]/demos/feature_tracker_nvxcu

The CUDA pointer can be found:

nvxcu_pitch_linear_image_t image;
image.base.image_type = NVXCU_PITCH_LINEAR_IMAGE;
image.base.format = NVXCU_DF_IMAGE_U8;
image.base.width = width;
image.base.height = height;
image.planes[0].dev_ptr = dev_ptr;                        <b><- array pointer</b>
image.planes[0].pitch_in_bytes = pitch;

More, if VisionWorks is not essential, it’s recommended to use our latest MMAPI.
It includes secveral optimized sample for image -> CUDA pipeline.

Thanks.

#4

Sorry, I’m not familiar with MMAPI. Please, can you tell me more about it? I don’t have to use VisonWorks.

What is the best practice for binding a cuda 2D memory to a cudaTexture memory in openvx kernel? The goal for me is preventing cache misses while directly accessing 2D memory in a Cuda kernel. Is my method a correct approach to these problems?

I’m writing a Cuda kernel for simplifying the algorithm used in belief propagation by parallel acceleration. Do you have any resource on solving Gauss-Seidel problem on GPU?

Thank you for the fast reply.

#5

Hi,

You can install MMAPI directly from the JetPack installer.

If pure CUDA kerenl is acceptable(no openvx), check this sample for information:
${HOME}/tegra_multimedia_api/samples/v4l2cuda

This demonstrates v4l2 source -> CUDA buffers.

Thanks.