Fix ballot on wave64 GPUs and with GPUCompiler 2.9

[luraess]

The Documentation job has failed on main since build 4275, and so on every PR since. The four doctests that use ballot fail with "unsupported call to an unknown LLVM intrinsic". The only difference between the last green and the first red docs environment is GPUCompiler 2.8.2 → 2.9.0, which reports llvmcalls of unknown intrinsics at compile time (JuliaGPU/GPUCompiler.jl#958). This is not only a docs problem: ballot, activemask, ballot_sync, any_sync and all_sync all go through ballot, which had two bugs:

• The wave64 branch called llvm.amdgcn.ballot.w64. That is not a valid name for the overloaded intrinsic, whose suffix is the result type (.i64). Julia compiles the llvmcall to jl_error("llvmcall only supports intrinsic calls"), and GPUCompiler 2.9.0 now rejects that on every target, wave32 included. The branches now use .i64 and .i32.
• LLVM doesn't fold llvm.amdgcn.wavefrontsize before instruction selection, so the backend also selects the branch that never runs. On wave64, the i32 ballot can't be selected (Cannot select: i32 = AMDGPUISD::SETCC), so ballot has never compiled on MI250/MI300, with any GPUCompiler. finish_module! now folds llvm.amdgcn.wavefrontsize to the job's wavefront size, so that branch goes away before optimization. shfl, random and the hostcall and printing code also get it as a constant now.

Only the doctests covered ballot, and they run on the wave32 Buildkite GPUs, where the broken branch was dead code. The new "Wavefront Ballot" testset in device/wavefront runs ballot at the native wavefront size and in wave64 mode, so it covers both branches on wave32 GPUs too.

ballot over every other lane, on an MI250 (gfx90a, Julia 1.12). gfx1100 was only compiled, not run:

	main, GPUCompiler 2.8.2	main, GPUCompiler 2.9.0	This PR, GPUCompiler 2.9.0
gfx90a (wave64), run	Cannot select	unknown intrinsic	0x5555555555555555
gfx1100 wave32, compiled	ok	unknown intrinsic	ok
gfx1100 wave64, compiled	Cannot select	unknown intrinsic	ok

The full suite passes on the MI250 (2 GCDs, Julia 1.12, GPUCompiler 2.9.0): 17149 passed, 21 broken.

🤖 Generated with Claude Code

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 0840adf	Previous: 8a66547	Ratio
amdgpu/synchronization/context/device	552.5 ns	550 ns	1.00
amdgpu/synchronization/stream/blocking	230 ns	222.5 ns	1.03
amdgpu/synchronization/stream/nonblocking	305 ns	307.5 ns	0.99
applications/bitonic_sort	1132481.25 ns	1160103.75 ns	0.98
applications/convolution	105429 ns	72376 ns	1.46
applications/floyd_warshall	9014979.25 ns	9100856.25 ns	0.99
applications/histogram	820511.75 ns	798998.75 ns	1.03
applications/prefix_sum	187982.75 ns	190535 ns	0.99
array/accumulate/Float32/1d	82093.75 ns	79346 ns	1.03
array/accumulate/Float32/dims=1	273718.75 ns	271536.25 ns	1.01
array/accumulate/Float32/dims=1L	107561.5 ns	72536 ns	1.48
array/accumulate/Float32/dims=2	102481.5 ns	74041 ns	1.38
array/accumulate/Float32/dims=2L	3031194 ns	2834326.75 ns	1.07
array/accumulate/Int64/1d	91161.25 ns	79098.5 ns	1.15
array/accumulate/Int64/dims=1	249486.25 ns	248393.25 ns	1.00
array/accumulate/Int64/dims=1L	115196.5 ns	89128.75 ns	1.29
array/accumulate/Int64/dims=2	98814 ns	76041 ns	1.30
array/accumulate/Int64/dims=2L	3389189 ns	3139866 ns	1.08
array/broadcast	49500.5 ns	50443.25 ns	0.98
array/construct	2097.5 ns	2247.5 ns	0.93
array/copy	39135.5 ns	39278 ns	1.00
array/copyto!/cpu_to_gpu	89113.75 ns	85676.25 ns	1.04
array/copyto!/gpu_to_cpu	89573.75 ns	85591.25 ns	1.05
array/copyto!/gpu_to_gpu	34480.25 ns	58288.25 ns	0.59
array/iteration/findall/bool	129201.75 ns	125954.25 ns	1.03
array/iteration/findall/int	146174.5 ns	146617 ns	1.00
array/iteration/findfirst/bool	157204.75 ns	155292.25 ns	1.01
array/iteration/findfirst/int	191912.75 ns	189260 ns	1.01
array/iteration/findmin/1d	103654 ns	103258.75 ns	1.00
array/iteration/findmin/2d	90551.25 ns	90081.25 ns	1.01
array/iteration/logical	227973.25 ns	226565.5 ns	1.01
array/iteration/scalar	307657 ns	295931.5 ns	1.04
array/permutedims/2d	47783.25 ns	70736 ns	0.68
array/permutedims/3d	57961 ns	69253.5 ns	0.84
array/permutedims/4d	74281 ns	50278.25 ns	1.48
array/random/rand/Float32	44853.25 ns	40065.75 ns	1.12
array/random/rand/Int64	54655.75 ns	48863 ns	1.12
array/random/rand!/Float32	39685.75 ns	65541 ns	0.61
array/random/rand!/Int64	49058 ns	69623.25 ns	0.70
array/random/randn/Float32	70631.25 ns	71161 ns	0.99
array/random/randn!/Float32	57285.75 ns	58721 ns	0.98
array/reductions/mapreduce/Float32/1d	82101 ns	81621 ns	1.01
array/reductions/mapreduce/Float32/dims=1	82861.25 ns	56660.75 ns	1.46
array/reductions/mapreduce/Float32/dims=1L	838612 ns	844514.25 ns	0.99
array/reductions/mapreduce/Float32/dims=2	81718.75 ns	74073.5 ns	1.10
array/reductions/mapreduce/Float32/dims=2L	136354.25 ns	136524.5 ns	1.00
array/reductions/mapreduce/Int64/1d	82266 ns	81663.5 ns	1.01
array/reductions/mapreduce/Int64/dims=1	77741.25 ns	73933.5 ns	1.05
array/reductions/mapreduce/Int64/dims=1L	842112.25 ns	839174.25 ns	1.00
array/reductions/mapreduce/Int64/dims=2	80413.5 ns	75486.25 ns	1.07
array/reductions/mapreduce/Int64/dims=2L	137434.5 ns	137611.75 ns	1.00
array/reductions/reduce/Float32/1d	81826.25 ns	81541.25 ns	1.00
array/reductions/reduce/Float32/dims=1	83023.75 ns	70296 ns	1.18
array/reductions/reduce/Float32/dims=1L	839754.5 ns	836984.25 ns	1.00
array/reductions/reduce/Float32/dims=2	81046.25 ns	75123.5 ns	1.08
array/reductions/reduce/Float32/dims=2L	136959.5 ns	136774.25 ns	1.00
array/reductions/reduce/Int64/1d	78481.25 ns	81368.75 ns	0.96
array/reductions/reduce/Int64/dims=1	85493.75 ns	74831 ns	1.14
array/reductions/reduce/Int64/dims=1L	844699.5 ns	844554 ns	1.00
array/reductions/reduce/Int64/dims=2	81663.75 ns	75411.25 ns	1.08
array/reductions/reduce/Int64/dims=2L	137884.25 ns	137047 ns	1.01
array/reverse/1d	42298.25 ns	43458 ns	0.97
array/reverse/1dL	54366 ns	70606 ns	0.77
array/reverse/1dL_inplace	55555.75 ns	79671 ns	0.70
array/reverse/1d_inplace	36373 ns	60390.75 ns	0.60
array/reverse/2d	47145.75 ns	47625.75 ns	0.99
array/reverse/2dL	90248.75 ns	57993.5 ns	1.56
array/reverse/2dL_inplace	66211 ns	79096 ns	0.84
array/reverse/2d_inplace	36985.5 ns	61191 ns	0.60
array/sorting/1d	327267.25 ns	326449.5 ns	1.00
gemm/tiled	1958523.25 ns	1961659.75 ns	1.00
gemm/tiled_unbounded	1974238.25 ns	1960334.75 ns	1.01
integration/byval/reference	39901 ns	39821 ns	1.00
integration/byval/slices=1	39931 ns	40830 ns	0.98
integration/byval/slices=2	147002 ns	130842 ns	1.12
integration/byval/slices=3	239803 ns	241884 ns	0.99
integration/volumerhs	4893940 ns	4900148 ns	1.00
kernel/indexing	29640.5 ns	41880.5 ns	0.71
kernel/indexing_checked	29840.5 ns	37075.5 ns	0.80
kernel/launch	1170 ns	1130 ns	1.04
kernel/rand	98979 ns	70343.5 ns	1.41
latency/import	1449711665 ns	1456646355 ns	1.00
latency/precompile	22905651598 ns	22642954217 ns	1.01
latency/ttfp	2201433275 ns	2251455325 ns	0.98
stencil/diffusion3d	1619138.25 ns	1611230 ns	1.00
stencil/diffusion3d_checked	1658591.5 ns	1655470.5 ns	1.00

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