Switch to CompilerCaching.jl.

[maleadt]

No description provided.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 4915289	Previous: 240107f	Ratio
amdgpu/synchronization/context/device	610 ns	610 ns	1
amdgpu/synchronization/stream/blocking	250 ns	250 ns	1
amdgpu/synchronization/stream/nonblocking	330 ns	350 ns	0.94
array/accumulate/Float32/1d	86291 ns	84951 ns	1.02
array/accumulate/Float32/dims=1	355675 ns	342295 ns	1.04
array/accumulate/Float32/dims=1L	135692 ns	137362 ns	0.99
array/accumulate/Float32/dims=2	133702 ns	126782 ns	1.05
array/accumulate/Float32/dims=2L	2828710 ns	2825642 ns	1.00
array/accumulate/Int64/1d	97981 ns	98651 ns	0.99
array/accumulate/Int64/dims=1	339875 ns	329425 ns	1.03
array/accumulate/Int64/dims=1L	167512 ns	168173 ns	1.00
array/accumulate/Int64/dims=2	127022 ns	125522 ns	1.01
array/accumulate/Int64/dims=2L	3004263 ns	3003954 ns	1.00
array/broadcast	101982 ns	92211 ns	1.11
array/construct	1770 ns	1760 ns	1.01
array/copy	38751 ns	37590 ns	1.03
array/copyto!/cpu_to_gpu	183723 ns	114062 ns	1.61
array/copyto!/gpu_to_cpu	161533 ns	183553 ns	0.88
array/copyto!/gpu_to_gpu	66871 ns	60371 ns	1.11
array/iteration/findall/bool	182373 ns	181042 ns	1.01
array/iteration/findall/int	201543 ns	188513 ns	1.07
array/iteration/findfirst/bool	128432 ns	118562 ns	1.08
array/iteration/findfirst/int	118671 ns	118282 ns	1.00
array/iteration/findmin/1d	169972 ns	170152 ns	1.00
array/iteration/findmin/2d	156433 ns	156083 ns	1.00
array/iteration/logical	348615 ns	348075 ns	1.00
array/iteration/scalar	294094 ns	303894 ns	0.97
array/permutedims/2d	74711 ns	73381 ns	1.02
array/permutedims/3d	72981 ns	73131 ns	1.00
array/permutedims/4d	76651 ns	76332 ns	1.00
array/random/rand/Float32	51450 ns	50961 ns	1.01
array/random/rand/Int64	57430 ns	57361 ns	1.00
array/random/rand!/Float32	89911 ns	87231 ns	1.03
array/random/rand!/Int64	89632 ns	96922 ns	0.92
array/random/randn/Float32	90981 ns	86641 ns	1.05
array/random/randn!/Float32	106741 ns	80272 ns	1.33
array/reductions/mapreduce/Float32/1d	134082 ns	133782 ns	1.00
array/reductions/mapreduce/Float32/dims=1	96071 ns	95681 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	776461 ns	778072 ns	1.00
array/reductions/mapreduce/Float32/dims=2	98132 ns	97781 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	296604 ns	294594 ns	1.01
array/reductions/mapreduce/Int64/1d	135682 ns	134242 ns	1.01
array/reductions/mapreduce/Int64/dims=1	96552 ns	95802 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	782951 ns	782372 ns	1.00
array/reductions/mapreduce/Int64/dims=2	98242 ns	96722 ns	1.02
array/reductions/mapreduce/Int64/dims=2L	297564 ns	296194 ns	1.00
array/reductions/reduce/Float32/1d	133532 ns	133413 ns	1.00
array/reductions/reduce/Float32/dims=1	96391 ns	95362 ns	1.01
array/reductions/reduce/Float32/dims=1L	770962 ns	772321 ns	1.00
array/reductions/reduce/Float32/dims=2	98442 ns	95681 ns	1.03
array/reductions/reduce/Float32/dims=2L	296314 ns	293734 ns	1.01
array/reductions/reduce/Int64/1d	133962 ns	134022 ns	1.00
array/reductions/reduce/Int64/dims=1	97791 ns	95612 ns	1.02
array/reductions/reduce/Int64/dims=1L	781442 ns	783112 ns	1.00
array/reductions/reduce/Int64/dims=2	98811 ns	96961 ns	1.02
array/reductions/reduce/Int64/dims=2L	298285 ns	296764 ns	1.01
array/reverse/1d	44371 ns	44210 ns	1.00
array/reverse/1dL	76171 ns	75981 ns	1.00
array/reverse/1dL_inplace	103922 ns	95681 ns	1.09
array/reverse/1d_inplace	71101 ns	77711 ns	0.91
array/reverse/2d	52091 ns	51820 ns	1.01
array/reverse/2dL	102171 ns	100941 ns	1.01
array/reverse/2dL_inplace	175052 ns	86952 ns	2.01
array/reverse/2d_inplace	84571 ns	132582 ns	0.64
array/sorting/1d	345175 ns	341995 ns	1.01
integration/byval/reference	39350 ns	39641 ns	0.99
integration/byval/slices=1	40230 ns	40761 ns	0.99
integration/byval/slices=2	160222 ns	148282 ns	1.08
integration/byval/slices=3	235113 ns	234153 ns	1.00
integration/volumerhs	5024212 ns	5028074 ns	1.00
kernel/indexing	65371 ns	129912 ns	0.50
kernel/indexing_checked	73611 ns	131002 ns	0.56
kernel/launch	1560 ns	1350 ns	1.16
kernel/rand	193932 ns	195323 ns	0.99
latency/import	1709436321 ns	1601453816 ns	1.07
latency/precompile	37339939945 ns	36585492651 ns	1.02
latency/ttfp	5731761660 ns	2182646656 ns	2.63

This comment was automatically generated by workflow using github-action-benchmark.

[maleadt]

Done on my end. I'll leave it to AMDGPU.jl maintainers to merge this.

[luraess]

@wsmoses we have this now failing for Enzyme tests as this PR requires now GPUCompiler v2 and Enzyme is still at 1.6.2 (see https://buildkite.com/julialang/amdgpu-dot-jl/builds/3886#019f5b1f-0fd3-4bc7-8c10-878504fd0dc9/L669). Could we bump compat and patch Enzyme?

[wsmoses]

@vchuravy has a pr starting to look into it.

That said, @maleadt as the changes here don't look huge, is it possible to conditionally gate this with something like

if gpucompilerver >= 2
a()
else
b()
end

that way downstream packages that haven't adapted yet [like enzyme/reactant] can still run ci with?

And the same for cuda/metal?

[maleadt]

I'd rather not. I've added enough cruft already in GPUCompiler just to keep supporting 1.10 for Enzyme, and I don't want that leak into the back-ends too.

[wsmoses]

Fair enough, in which case would you be able to help us adapt enzyme/reactant to gpucompiler 2?

[maleadt]

Valentin created PRs already. The changes are not tricky; point an LLM at my changes and let it update those PRs.

[luraess]

@wsmoses let me know if you can things rolling - as ideally I'd like to have Enzyme CI passing before merging this.

[wsmoses]

there's a draft pr open for enzyme by @vchuravy and no pr at all for reactant.

without some assistance from gpucompiler folks, there's likely going to be non-trivial latency to adapt to the api churn as we're still working through some higher priority blockers atm.

[christiangnrd]

@wsmoses I've opened EnzymeAD/Enzyme.jl#3354 targeting the draft PR which keeps support for GPUCompiler v1. This doesn't seem like too much to maintain at least until most things have transitioned to GPUCompiler 2.

I gave Reactant.jl a quick look and I don't think any of the GPUCompiler methods used by Reactant.jl need to adjust to a need signature/name. Obviously behaviour may have changed and require fixing but it's possible it'll just be a compat bump

[luraess]

What's the status of EnzymeAD/Enzyme.jl#3354 ? In the meantime, I am thinking of allowing for soft fail (as CUDA does) and merge this PR.

[luraess]

Thanks!