[OhadMeir]

Tracked on [RSDSO-21414]

[Copilot]

Pull request overview

This PR updates the CUDA pointcloud implementation to reduce per-frame GPU allocation overhead by introducing a per-instance helper that caches device buffers and intrinsics across frames (RSDSO-21414).

Changes:

• Add rscuda::pointcloud_cuda_helper to own persistent device buffers and cache uploaded intrinsics.
• Switch the pointcloud CUDA kernel launch to a 2D grid/block mapping (one thread per pixel) to avoid 1D striding/division.
• Wire the helper into librealsense::pointcloud_cuda so buffers are reused across frames.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
src/proc/cuda/cuda-pointcloud.h	Adds forward declaration and persistent helper member to keep CUDA headers out of the public processor header.
src/proc/cuda/cuda-pointcloud.cpp	Instantiates the helper and routes depth deprojection through it.
src/cuda/cuda-pointcloud.cuh	Defines the helper class and new constants used by the CUDA pointcloud path.
src/cuda/cuda-pointcloud.cu	Implements persistent-buffer deprojection and replaces 1D striding kernel with a 2D kernel launch.

[Nir-Az]

[rs-agentic-bot] Automated code review for CUDA pointcloud optimizations. See inline comments for details.

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated 4 comments.