CompatHelper: bump compat for "CUDA" to "5"

[github-actions]

This pull request changes the compat entry for the CUDA package from 3.5, 4 to 3.5, 4, 5.

This keeps the compat entries for earlier versions.

Note: I have not tested your package with this new compat entry. It is your responsibility to make sure that your package tests pass before you merge this pull request.

[maleadt]

Benchmark results for commit dedaad6 (comparing to bca8c30):

test	master	PR	Δmin
BLAS WMMA FP16'*FP16'=FP16 (256×256×256, α)	45.6 ± 0.14 μs	36.7 ± 0.1 μs	19.7% ✅
BLAS WMMA FP16'*FP16'=FP16 (4096×4096×4096, α)	5.89 ± 0.035 ms	4.28 ± 0.026 ms	29.2% ✅
BLAS WMMA FP16'*FP16'=FP32 (256×256×256, α)	34.2 ± 0.42 μs	27.96 ± 0.089 μs	15.2% ✅
BLAS WMMA FP16*FP16'=FP16 (4096×4096×4096, α)	5.355 ± 0.0021 ms	4.07 ± 0.019 ms	26.2% ✅
BLAS WMMA FP16*FP16'=FP32 (256×256×256, α)	33.5 ± 0.49 μs	27.99 ± 0.09 μs	14.8% ✅
BLAS WMMA FP16*FP16=FP16 (256×256×256, α, β)	38.1 ± 0.56 μs 16 bytes local mem	37.39 ± 0.077 μs 8 bytes local mem	0.2%
BLAS WMMA FP16*FP16=FP16 (256×256×256, β)	34.5 ± 0.54 μs 153 regs	34.2 ± 0.11 μs 155 regs	-0.2%
BLAS WMMA FP16*FP16=FP16 (4096×4096×4096, α, β)	4.16 ± 0.024 ms 16 bytes local mem	4.11 ± 0.03 ms 8 bytes local mem	3.3%
BLAS WMMA FP16*FP16=FP16 (4096×4096×4096, β)	2.986 ± 0.0021 ms 153 regs	3.034 ± 0.0048 ms 155 regs	-1.4%
BLAS WMMA FP16*FP16=FP32 (256×256×256, α, β)	29.8 ± 0.53 μs 182 regs	29.39 ± 0.083 μs 178 regs	-0.8%
BLAS WMMA FP16*FP16=FP32 (4096×4096×4096, α, β)	4.43 ± 0.07 ms 182 regs	4.81 ± 0.04 ms 178 regs	-3.4%

[codecov]

Codecov Report

Patch and project coverage have no change.

Comparison is base (bca8c30) 32.17% compared to head (dedaad6) 32.17%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #155   +/-   ##
=======================================
  Coverage   32.17%   32.17%           
=======================================
  Files          11       11           
  Lines         889      889           
=======================================
  Hits          286      286           
  Misses        603      603           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[maleadt]

The change in registers is due to the removal of a quirk in JuliaGPU/CUDA.jl@760c2bd#diff-d1deb39259ffe01ecfe5db7c6c2331af73b81772539dd8ef49076d58fa2529d2, but AFAICT that actually improves performance. I still have to find the culprit of the other slowdowns, but I can't reproduce it locally.

[maleadt]

Painfully bisected to JuliaGPU/CUDA.jl#2025. Which doesn't make sense, as we manually synchronize the stream, and thus do not rely on non-blocking synchronization. The only point nonblocking synchronization is used, is when allocating the arrays during the set-up of each benchmark. Maybe we're just introducing lots of noise there right before each benchmark?

Getting rid of that seems to greatly improve timings, so hopefully that fixes the issue here.