[QDP] Fix invalid CUDA kernel launch when num_samples exceeds grid dimension limit

[viiccwen]

Purpose of PR

Fixes a bug in launch_l2_norm_batch (f64) where attempting to process more than limited samples would result in an invalid CUDA kernel launch. The fix adds early validation to return an error when num_samples exceeds the CUDA 1D grid dimension limit.

Related Issues or PRs

closes #967

Changes Made

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Breaking Changes

• Yes
• No

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes
• Successfully built and ran all unit tests or manual tests locally
• PR title follows "MAHOUT-XXX: Brief Description" format (if related to an issue)
• Code follows ASF guidelines

[viiccwen]

cc @rich7420, @ryankert01

[ryankert01]

using 65,535, you might be using a limit inherited from the Fermi architecture (pre-2012).

On any modern GPU (Compute Capability 3.0 or higher, which is virtually everything in use today), the 1D grid limit for the X-dimension is significantly higher. (at 2^31-1)

Which may or may not be touch before out-of-memory. (but it's good to add a check tho)

[viiccwen]

@ryankert01 thx for suggestion! What if we use cudaDeviceGetAttribute(...) to get the cudaDevAttrMaxGridDimX from user's CUDA? That prevent Hard-coded attribute CUDA_MAX_GRID_DIM_1D.

[ryankert01]

Thanks for the update! Some comments

[rich7420]

@viiccwen thanks for the patch!
I think we could add tests for num_samples==0/sample_len==0.

[rich7420]

LGTM

[rich7420]

@ryankert01 do you want to take another look?

[ryankert01]

@viiccwen lg, ty for the contribution