Adopt MatX v0.4.0

[dagardner-nv]

Description

• Adopt updated utilities to pick up MatX 0.4.0
• Currently we are tracking a commit hash to adopt fixes needed for Morpheus, which were rolled into the 0.4.0 release.

fixes #909

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[cwharris]

The build error is caused by a compute capability issue. https://stackoverflow.com/questions/74201452/cuda-11-8-fails-to-compile-atomiccas-for-16-bit-unsigned-integers-is-cudas-doc

Looking for a workaround.

[cwharris]

@cliffburdick has fixed the compilation error, but the PR does not add support for half reduction to pascal architecture.

NVIDIA/MatX#412

[cliffburdick]

@cliffburdick has fixed the compilation error, but the PR does not add support for half reduction to pascal architecture.

NVIDIA/MatX#412

Please let us know if there's a use case for this on Pascal. Specifically this would be argmax/argmin with fp16/bf16 only I believe.

[dagardner-nv]

@cliffburdick has fixed the compilation error, but the PR does not add support for half reduction to pascal architecture.
NVIDIA/MatX#412

Please let us know if there's a use case for this on Pascal. Specifically this would be argmax/argmin with fp16/bf16 only I believe.

Thanks for the quick fix!
We use matx::rmax, I'm not sure if that uses argmax or not, but even then we only use matx::rmax for 32bit and 64bit floats.

[dagardner-nv]

Closing this PR as the goal was to move off of tracking a commit hash on to an official release tag.

[cliffburdick]

@cliffburdick has fixed the compilation error, but the PR does not add support for half reduction to pascal architecture.
NVIDIA/MatX#412

Please let us know if there's a use case for this on Pascal. Specifically this would be argmax/argmin with fp16/bf16 only I believe.

Thanks for the quick fix! We use matx::rmax, I'm not sure if that uses argmax or not, but even then we only use matx::rmax for 32bit and 64bit floats.

Those are just max/min and will use cub, so I don't think you're using anything that would be affected.