Skip to content

Fixes for CUDA 13.2#111

Merged
maleadt merged 8 commits intomainfrom
tb/cuda_13.2
Mar 12, 2026
Merged

Fixes for CUDA 13.2#111
maleadt merged 8 commits intomainfrom
tb/cuda_13.2

Conversation

@maleadt
Copy link
Member

@maleadt maleadt commented Mar 11, 2026

Closes #109

@AntonOresten
Copy link
Contributor

AntonOresten commented Mar 11, 2026

https://buildkite.com/julialang/cutile-dot-jl/builds/5/steps/canvas?jid=019cdd8c-4e08-4437-af5a-e3deb9e0d9fb
image

The CTAs hint failing makes sense if it doesn't have support pre-Hopper. cutile-python supports specifying hints based on arch, but in #27 I was lazy and just made it Int-or-nothing. Maybe num_ctas and occupancy should be able to be passed as Dicts, then in ct.launch one could just select based on sm_arch? cutile-python handles it at the kernel decorator level:

EDIT: ok a little further up the log includes: │ error: error at offset 14: expected 1 for sm_80.num_cta_in_cga

image

Not sure about the Ampere atomic error. I recall it having less support for atomics but it's weird that it fails silently?

maleadt and others added 7 commits March 12, 2026 05:21
The atomic_add, atomic_xchg, and atomic_cas functions required exact
type matching between the value argument and array element type (via
TileOrScalar{T}). This caused a confusing codegen error when e.g.
passing bid() (Int32) to an Int64 array. Add conversion methods that
auto-promote mismatched types.

Also fixes the atomic_xchg test to use bid directly instead of bid+1,
which produced values outside the asserted range on some hardware.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@maleadt maleadt merged commit 44c4823 into main Mar 12, 2026
9 checks passed
@maleadt maleadt deleted the tb/cuda_13.2 branch March 12, 2026 11:08
@AntonOresten
Copy link
Contributor

What's with the 1.11 error? I saw this "Failure artifact" once locally on 1.11 as well, but it just automatically tried again iirc, and now I can't seem to replicate (even after clearing .julia/artifacts)
image

@maleadt
Copy link
Member Author

maleadt commented Mar 12, 2026

A download error maybe? These artifacts are very large, and are cached on CI after a first successful download.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Pre-Blackwell support

2 participants