Use GPUCompiler 1.13.3

[vchuravy]

We want to use JuliaGPU/GPUCompiler.jl#812

[github-actions]

Benchmark Results

	main	f95f72d...	main / f95f72d...
saxpy/default/Float32/1024	0.0445 ± 0.028 ms	0.0437 ± 0.028 ms	1.02 ± 0.92
saxpy/default/Float32/1048576	0.186 ± 0.018 ms	0.184 ± 0.015 ms	1.01 ± 0.13
saxpy/default/Float32/16384	0.0498 ± 0.021 ms	0.05 ± 0.021 ms	0.996 ± 0.59
saxpy/default/Float32/2048	0.0449 ± 0.026 ms	0.0454 ± 0.026 ms	0.989 ± 0.82
saxpy/default/Float32/256	0.0375 ± 0.028 ms	0.0466 ± 0.028 ms	0.805 ± 0.77
saxpy/default/Float32/262144	0.0843 ± 0.023 ms	0.0844 ± 0.023 ms	0.999 ± 0.38
saxpy/default/Float32/32768	0.048 ± 0.022 ms	0.048 ± 0.022 ms	1 ± 0.65
saxpy/default/Float32/4096	0.0477 ± 0.024 ms	0.0468 ± 0.025 ms	1.02 ± 0.76
saxpy/default/Float32/512	0.0445 ± 0.028 ms	0.044 ± 0.028 ms	1.01 ± 0.91
saxpy/default/Float32/64	0.0381 ± 0.028 ms	0.0459 ± 0.028 ms	0.83 ± 0.79
saxpy/default/Float32/65536	0.0482 ± 0.023 ms	0.0504 ± 0.023 ms	0.956 ± 0.63
saxpy/default/Float64/1024	0.0427 ± 0.026 ms	0.0445 ± 0.026 ms	0.96 ± 0.82
saxpy/default/Float64/1048576	0.266 ± 0.017 ms	0.257 ± 0.018 ms	1.04 ± 0.098
saxpy/default/Float64/16384	0.0478 ± 0.021 ms	0.0483 ± 0.021 ms	0.988 ± 0.62
saxpy/default/Float64/2048	0.0489 ± 0.024 ms	0.0485 ± 0.026 ms	1.01 ± 0.74
saxpy/default/Float64/256	0.0383 ± 0.028 ms	0.0447 ± 0.028 ms	0.857 ± 0.82
saxpy/default/Float64/262144	0.103 ± 0.019 ms	0.109 ± 0.019 ms	0.943 ± 0.24
saxpy/default/Float64/32768	0.0518 ± 0.022 ms	0.0523 ± 0.022 ms	0.99 ± 0.59
saxpy/default/Float64/4096	0.0369 ± 0.025 ms	0.0442 ± 0.027 ms	0.836 ± 0.77
saxpy/default/Float64/512	0.0437 ± 0.028 ms	0.0466 ± 0.028 ms	0.938 ± 0.82
saxpy/default/Float64/64	0.0374 ± 0.028 ms	0.0448 ± 0.028 ms	0.834 ± 0.81
saxpy/default/Float64/65536	0.0544 ± 0.022 ms	0.058 ± 0.022 ms	0.938 ± 0.52
saxpy/static workgroup=(1024,)/Float32/1024	0.0432 ± 0.028 ms	0.046 ± 0.028 ms	0.939 ± 0.84
saxpy/static workgroup=(1024,)/Float32/1048576	0.184 ± 0.018 ms	0.185 ± 0.015 ms	0.997 ± 0.12
saxpy/static workgroup=(1024,)/Float32/16384	0.0496 ± 0.021 ms	0.0506 ± 0.021 ms	0.98 ± 0.58
saxpy/static workgroup=(1024,)/Float32/2048	0.0449 ± 0.026 ms	0.0445 ± 0.027 ms	1.01 ± 0.85
saxpy/static workgroup=(1024,)/Float32/256	0.0435 ± 0.028 ms	0.0449 ± 0.028 ms	0.97 ± 0.88
saxpy/static workgroup=(1024,)/Float32/262144	0.083 ± 0.024 ms	0.083 ± 0.022 ms	0.999 ± 0.39
saxpy/static workgroup=(1024,)/Float32/32768	0.0464 ± 0.022 ms	0.0473 ± 0.021 ms	0.98 ± 0.64
saxpy/static workgroup=(1024,)/Float32/4096	0.0474 ± 0.023 ms	0.0488 ± 0.024 ms	0.97 ± 0.67
saxpy/static workgroup=(1024,)/Float32/512	0.044 ± 0.028 ms	0.0453 ± 0.028 ms	0.971 ± 0.86
saxpy/static workgroup=(1024,)/Float32/64	0.042 ± 0.029 ms	0.0453 ± 0.028 ms	0.927 ± 0.86
saxpy/static workgroup=(1024,)/Float32/65536	0.0472 ± 0.023 ms	0.0491 ± 0.022 ms	0.96 ± 0.64
saxpy/static workgroup=(1024,)/Float64/1024	0.0426 ± 0.027 ms	0.045 ± 0.026 ms	0.947 ± 0.81
saxpy/static workgroup=(1024,)/Float64/1048576	0.267 ± 0.017 ms	0.255 ± 0.016 ms	1.05 ± 0.094
saxpy/static workgroup=(1024,)/Float64/16384	0.0467 ± 0.022 ms	0.0473 ± 0.021 ms	0.987 ± 0.64
saxpy/static workgroup=(1024,)/Float64/2048	0.0493 ± 0.024 ms	0.0473 ± 0.026 ms	1.04 ± 0.76
saxpy/static workgroup=(1024,)/Float64/256	0.0506 ± 0.027 ms	0.0433 ± 0.028 ms	1.17 ± 0.98
saxpy/static workgroup=(1024,)/Float64/262144	0.105 ± 0.02 ms	0.108 ± 0.018 ms	0.977 ± 0.25
saxpy/static workgroup=(1024,)/Float64/32768	0.0493 ± 0.023 ms	0.0519 ± 0.021 ms	0.949 ± 0.59
saxpy/static workgroup=(1024,)/Float64/4096	0.0432 ± 0.024 ms	0.0458 ± 0.026 ms	0.944 ± 0.75
saxpy/static workgroup=(1024,)/Float64/512	0.0459 ± 0.028 ms	0.0467 ± 0.027 ms	0.984 ± 0.82
saxpy/static workgroup=(1024,)/Float64/64	0.0448 ± 0.029 ms	0.0443 ± 0.028 ms	1.01 ± 0.92
saxpy/static workgroup=(1024,)/Float64/65536	0.0523 ± 0.023 ms	0.0566 ± 0.022 ms	0.924 ± 0.54
time_to_load	0.831 ± 0.0038 s	0.825 ± 0.001 s	1.01 ± 0.0047

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[vchuravy]

@maleadt didn't have time to dig in, but that failure looks new.

[maleadt]

Note that this may be a legitimate bug in KA.jl or so, because without the trap to ret PoCL was aggressive enough to optimize out the print statement too (making it fully invisible in many cases). For example, in OpenCL.jl I had to fix JuliaGPU/OpenCL.jl#436.

[maleadt]

Oh my god this is a PoCL miscompilation because of a lane potentially exiting before it hits a barrier. I do have a workaround though... I'm also having Claude investigate a fix in PoCL itself.

[maleadt]

So the workaround is benign; @vchuravy please review f95f72d. It essentially avoids some pesky bounds checks when running with --check-bounds=true; normal code is unaffected.

I'll fix this in PoCL next.

[vchuravy]

Neat! Thank you.