Bump downgrader, and add code_air.

[maleadt]

Adapts to / builds on JuliaGPU/GPUCompiler.jl#829.

Demo:

julia> function kernel(arr)
           @inbounds arr[1] = sqrt(arr[1])
           return
       end

code_llvm now shows the LLVM IR we use with our in-process LLVM back-end:

julia> Metal.code_llvm(kernel, Tuple{MtlDeviceArray{Float32,1,Metal.AS.Generic}}; debuginfo=:none)
define void @julia_kernel_6055(ptr nocapture noundef nonnull readonly align 8 dereferenceable(16) %"arr::MtlDeviceArray") local_unnamed_addr {
top:
  %"arr::MtlDeviceArray.unbox" = load ptr, ptr %"arr::MtlDeviceArray", align 8
  %0 = load float, ptr %"arr::MtlDeviceArray.unbox", align 4
  %1 = fcmp uge float %0, 0.000000e+00
  br i1 %1, label %L22, label %L20

L20:                                              ; preds = %top
  call fastcc void @julia_throw_complex_domainerror_6130() #6
  unreachable

L22:                                              ; preds = %top
  %2 = call float @llvm.sqrt.f32(float %0)
  store float %2, ptr %"arr::MtlDeviceArray.unbox", align 4
  ret void
}

... while code_air shows what comes out of the downgrader (i.e., with AIR intrinsics, typed pointers, etc):

julia> Metal.code_air(kernel, Tuple{MtlDeviceArray{Float32,1,Metal.AS.Generic}})
; ...

define void @julia_kernel_6195({}* nocapture nonnull readonly align 8 dereferenceable(16) %"arr::MtlDeviceArray") local_unnamed_addr {
top:
  %0 = bitcast {}* %"arr::MtlDeviceArray" to {}**
  %"arr::MtlDeviceArray.unbox" = load {}*, {}** %0, align 8, !tbaa !2, !alias.scope !6, !noalias !9
  %1 = bitcast {}* %"arr::MtlDeviceArray.unbox" to float*
  %2 = load float, float* %1, align 4
  %3 = fcmp uge float %2, 0.000000e+00
  br i1 %3, label %L22, label %L20

L20:                                              ; preds = %top
  call fastcc void @julia_throw_complex_domainerror_6270()
  unreachable

L22:                                              ; preds = %top
  %4 = call float @air.sqrt.f32(float %2)
  %5 = bitcast {}* %"arr::MtlDeviceArray.unbox" to float*
  store float %4, float* %5, align 4
  ret void
}

; ...

Also bumps the downgrader.

Closes #800
Closes #799

[codecov]

Codecov Report

❌ Patch coverage is 92.30769% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 81.38%. Comparing base (e2e8c0a) to head (9377316).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/compiler/compilation.jl	91.66%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #804      +/-   ##
==========================================
- Coverage   82.02%   81.38%   -0.64%     
==========================================
  Files          66       66              
  Lines        3165     3148      -17     
==========================================
- Hits         2596     2562      -34     
- Misses        569      586      +17     

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: 9377316	Previous: e2e8c0a	Ratio
array/accumulate/Float32/1d	1472562.5 ns	1474458 ns	1.00
array/accumulate/Float32/dims=1	1109916 ns	1109979.5 ns	1.00
array/accumulate/Float32/dims=1L	10258000 ns	10281083.5 ns	1.00
array/accumulate/Float32/dims=2	1413000 ns	1425645.5 ns	0.99
array/accumulate/Float32/dims=2L	7295938 ns	7227500 ns	1.01
array/accumulate/Int64/1d	1464333 ns	1455354 ns	1.01
array/accumulate/Int64/dims=1	1265687.5 ns	1239791.5 ns	1.02
array/accumulate/Int64/dims=1L	12438833.5 ns	12261208 ns	1.01
array/accumulate/Int64/dims=2	1565604 ns	1585813 ns	0.99
array/accumulate/Int64/dims=2L	9801208 ns	9732166 ns	1.01
array/broadcast	370959 ns	375250 ns	0.99
array/construct	5708 ns	6042 ns	0.94
array/permutedims/2d	633666.5 ns	635250 ns	1.00
array/permutedims/3d	1132833 ns	1135834 ns	1.00
array/permutedims/4d	1987500 ns	1988042 ns	1.00
array/private/copy	412583 ns	440834 ns	0.94
array/private/copyto!/cpu_to_gpu	365833 ns	366792 ns	1.00
array/private/copyto!/gpu_to_cpu	359437.5 ns	366875 ns	0.98
array/private/copyto!/gpu_to_gpu	338000 ns	341625 ns	0.99
array/private/iteration/findall/bool	1558750 ns	1562396 ns	1.00
array/private/iteration/findall/int	1694875 ns	1699000 ns	1.00
array/private/iteration/findfirst/bool	1459708 ns	1475416 ns	0.99
array/private/iteration/findfirst/int	1495187.5 ns	1508833 ns	0.99
array/private/iteration/findmin/1d	1572999.5 ns	1583083 ns	0.99
array/private/iteration/findmin/2d	1309250 ns	1312375 ns	1.00
array/private/iteration/logical	2170458 ns	2180291 ns	1.00
array/private/iteration/scalar	2693145.5 ns	2756709 ns	0.98
array/random/rand/Float32	619875 ns	611417 ns	1.01
array/random/rand/Int64	682833.5 ns	673334 ns	1.01
array/random/rand!/Float32	572833 ns	593917 ns	0.96
array/random/rand!/Int64	506875 ns	507250 ns	1.00
array/random/randn/Float32	573250 ns	599833 ns	0.96
array/random/randn!/Float32	522750 ns	548292 ns	0.95
array/reductions/mapreduce/Float32/1d	567375 ns	769083 ns	0.74
array/reductions/mapreduce/Float32/dims=1	499937.5 ns	515750 ns	0.97
array/reductions/mapreduce/Float32/dims=1L	824125 ns	804646 ns	1.02
array/reductions/mapreduce/Float32/dims=2	520541 ns	518125 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	1355333 ns	1362458 ns	0.99
array/reductions/mapreduce/Int64/1d	934541.5 ns	949209 ns	0.98
array/reductions/mapreduce/Int64/dims=1	783917 ns	805625 ns	0.97
array/reductions/mapreduce/Int64/dims=1L	1482041.5 ns	1655000 ns	0.90
array/reductions/mapreduce/Int64/dims=2	964125 ns	985645.5 ns	0.98
array/reductions/mapreduce/Int64/dims=2L	2256604.5 ns	2262604.5 ns	1.00
array/reductions/reduce/Float32/1d	750125 ns	758584 ns	0.99
array/reductions/reduce/Float32/dims=1	499708.5 ns	515917 ns	0.97
array/reductions/reduce/Float32/dims=1L	791624.5 ns	780604 ns	1.01
array/reductions/reduce/Float32/dims=2	497667 ns	513708 ns	0.97
array/reductions/reduce/Float32/dims=2L	1345125 ns	1357750 ns	0.99
array/reductions/reduce/Int64/1d	933708 ns	944708 ns	0.99
array/reductions/reduce/Int64/dims=1	797166 ns	796084 ns	1.00
array/reductions/reduce/Int64/dims=1L	1589666.5 ns	1493124.5 ns	1.06
array/reductions/reduce/Int64/dims=2	1036208 ns	973542 ns	1.06
array/reductions/reduce/Int64/dims=2L	2260958.5 ns	2253375 ns	1.00
array/shared/copy	202166.5 ns	230458 ns	0.88
array/shared/copyto!/cpu_to_gpu	40250 ns	42375 ns	0.95
array/shared/copyto!/gpu_to_cpu	42042 ns	41709 ns	1.01
array/shared/copyto!/gpu_to_gpu	50250 ns	42500 ns	1.18
array/shared/iteration/findall/bool	1563292 ns	1565208 ns	1.00
array/shared/iteration/findall/int	1687291 ns	1701000 ns	0.99
array/shared/iteration/findfirst/bool	1194041 ns	1197646 ns	1.00
array/shared/iteration/findfirst/int	1214084 ns	1229167 ns	0.99
array/shared/iteration/findmin/1d	1331166 ns	1323500 ns	1.01
array/shared/iteration/findmin/2d	1314812.5 ns	1313000 ns	1.00
array/shared/iteration/logical	2020750 ns	2026750 ns	1.00
array/shared/iteration/scalar	5937.5 ns	9708 ns	0.61
integration/byval/reference	1158375 ns	1161042 ns	1.00
integration/byval/slices=1	1159791 ns	1165291 ns	1.00
integration/byval/slices=2	2087708 ns	2092583 ns	1.00
integration/byval/slices=3	7912979.5 ns	7852292 ns	1.01
integration/metaldevrt	543333 ns	548584 ns	0.99
kernel/indexing	362166 ns	367375 ns	0.99
kernel/indexing_checked	485416.5 ns	497291.5 ns	0.98
kernel/launch	13583 ns	14833 ns	0.92
kernel/rand	499042 ns	503000 ns	0.99
latency/import	1404456958 ns	1396322937.5 ns	1.01
latency/precompile	30336193396 ns	31339498708 ns	0.97
latency/ttfp	1717121209 ns	1665409542 ns	1.03
metal/synchronization/context	835.1063829787234 ns	1175 ns	0.71
metal/synchronization/stream	444.8636363636364 ns	798.0408163265306 ns	0.56

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