CUDA links, instructions?

[arothenberg]

---

Required Info
Camera Model	D435
Firmware Version	05.12.02.100
Operating System & Version	Win (10)
Platform	PC
SDK Version	2.32
Language	C++
Segment	desktop app

Issue Description

Is there a step by step for setting up CUDA?
I read that I should cmake from a CUDA file? I've been cmaking from the root I think and everything works great. How should I do it for this case?

Also, I read that once the project is made(maked) using the Cuda config there is nothing more that needs to be done, is that right? IOW the code I write stays the same.

[MartyG-RealSense]

Hi @arothenberg If you are installing librealsense from packages then CUDA support is built into them. An Nvidia graphics GPU is required to take advantage of CUDA. Nvidia Jetson developer boards have an Nvidia GPU.

If you are building librealsense from source code with CMake then you can enable CUDA support by including the build flag DBUILD_WITH_CUDA:bool=true in the CMake instruction.

When CUDA support is enabled, SDK operations that can benefit from CUDA (pointcloud, alignment and color conversion) should be automatically accelerated.

#1866

[arothenberg]

ok thanks

[arothenberg]

I need the developer board, not just an nvidea card like a Gforce?
Also does it affect the standard texture rendering or just the point cloud rendering.

[arothenberg]

By color conversion, you mean the color mapping of the depth data?

[MartyG-RealSense]

As far as I know, any detectable Nvidia GPU, including video cards, is usable by librealsense's CUDA support.

The color conversion is about converting a color image from the native image format (YUY2) to another. The formats that have support for conversion speed acceleration include those in the table linked to earlier.

#1866 (comment)

[arothenberg]

Ah ok. thanks.

[arothenberg]

Wow, there is almost nothing on realsense and cuda. I'll make it and see what happens. Are there a list of cuda functions?

[MartyG-RealSense]

If CUDA support is enabled and an Nvidia GPU is present then you do not need a list of CUDA functions, as the accelerations are applied automatically.

---

It is different for the SDK's alternate graphics acceleration method, which is called GLSL processing blocks. With GLSL, which works with any brand of GPU instead of just Nvidia as it is 'vendor-neutral', you have to amend script instructions to take advantage of the acceleration - for example, changing rs2::align to rs2::gl::align or changing rs2::pointcloud to rs2::gl::pointcloud

The link below has a very good pros and cons analysis of GLSL processing blocks and when it can be used.

#3654