CI fixes

[maleadt]

No description provided.

[maleadt]

Huh, looks like 11.3's compute-sanitizer doesn't get through our tests...

[codecov]

Codecov Report

Merging #950 (df08dd5) into master (2c40cb4) will increase coverage by 1.31%.
The diff coverage is 33.33%.

@@            Coverage Diff             @@
##           master     #950      +/-   ##
==========================================
+ Coverage   79.20%   80.51%   +1.31%     
==========================================
  Files         118      118              
  Lines        8002     8017      +15     
==========================================
+ Hits         6338     6455     +117     
+ Misses       1664     1562     -102     

Impacted Files	Coverage Δ
lib/cudnn/util.jl	42.85% <0.00%> (+1.94%)	⬆️
src/compiler/execution.jl	83.97% <0.00%> (-0.64%)	⬇️
lib/cudadrv/execution.jl	96.55% <50.00%> (-3.45%)	⬇️
lib/cudadrv/module/jit.jl	68.96% <0.00%> (-1.21%)	⬇️
lib/cudadrv/context.jl	73.10% <0.00%> (-0.40%)	⬇️
src/state.jl	77.90% <0.00%> (-0.07%)	⬇️
lib/curand/random.jl	93.25% <0.00%> (+0.07%)	⬆️
src/pool.jl	75.73% <0.00%> (+0.10%)	⬆️
lib/cudadrv/error.jl	83.72% <0.00%> (+0.38%)	⬆️
lib/utils/cache.jl	88.88% <0.00%> (+0.42%)	⬆️
... and 35 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2c40cb4...df08dd5. Read the comment docs.

[maleadt]

Interestingly, the 1.7 CI failures only occur on older devices, and seem to have been introduced by rc1 (they don't occur on 1.7-beta4). Not limited to changes in this PR though; also occur on CUDA.jl#master. MWE: permutedims(CuArray(rand(Float32, 1, 1, 1)), (2, 1, 3))

[maleadt]

Turns out JuliaLang/julia#42119 triggers a miscompilation on older GPUs... 😭

[maleadt]

Another case of divergent control flow due to unreachable instructions causing ptxas to emit bad code, filed with NVIDIA as #3382020. I've reverted the upstream change (yay for contextual dispatch), because we badly need to get CI working again.

For those interested, a C++-based MWE is available at https://gist.github.com/maleadt/dc59ee944952593e8d9967ad2c3543da