Reduce usage of libdevice, relying more on LLVM

[maleadt]

Building on JuliaGPU/GPUCompiler.jl#805, JuliaGPU/GPUCompiler.jl#804, JuliaGPU/GPUCompiler.jl#800, this PR aims to avoid some of the uses of libdevice's intrinsics, instead emitting vanilla LLVM IR and having GPUCompiler.jl post-process it into what we need in PTX. This has many advantages, including (potentially) better optimization, compatibility with LLVM tools like Enzyme, etc.

cc @vchuravy

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 986ca42	Previous: c0534dd	Ratio
array/accumulate/Float32/1d	99469 ns	99381 ns	1.00
array/accumulate/Float32/dims=1	74699 ns	74944 ns	1.00
array/accumulate/Float32/dims=1L	1575060 ns	1574920 ns	1.00
array/accumulate/Float32/dims=2	140156 ns	140278 ns	1.00
array/accumulate/Float32/dims=2L	652352 ns	652596 ns	1.00
array/accumulate/Int64/1d	116679 ns	116586 ns	1.00
array/accumulate/Int64/dims=1	78628 ns	78207 ns	1.01
array/accumulate/Int64/dims=1L	1683947 ns	1682646 ns	1.00
array/accumulate/Int64/dims=2	150782 ns	150838 ns	1.00
array/accumulate/Int64/dims=2L	959162 ns	959136 ns	1.00
array/broadcast	20011 ns	19887 ns	1.01
array/construct	1185.9 ns	1190.3 ns	1.00
array/copy	16565 ns	17076 ns	0.97
array/copyto!/cpu_to_gpu	212210 ns	211721 ns	1.00
array/copyto!/gpu_to_cpu	279922 ns	279606 ns	1.00
array/copyto!/gpu_to_gpu	10401 ns	10658 ns	0.98
array/iteration/findall/bool	130793 ns	131160 ns	1.00
array/iteration/findall/int	145151 ns	145504 ns	1.00
array/iteration/findfirst/bool	78686 ns	78933 ns	1.00
array/iteration/findfirst/int	80187 ns	80313 ns	1.00
array/iteration/findmin/1d	63791 ns	67478 ns	0.95
array/iteration/findmin/2d	101586 ns	111587 ns	0.91
array/iteration/logical	187789 ns	189045 ns	0.99
array/iteration/scalar	64529 ns	65291 ns	0.99
array/permutedims/2d	49215 ns	49578 ns	0.99
array/permutedims/3d	49748 ns	50060 ns	0.99
array/permutedims/4d	49496 ns	49640 ns	1.00
array/random/rand/Float32	11940 ns	11793 ns	1.01
array/random/rand/Int64	23373 ns	23510 ns	0.99
array/random/rand!/Float32	8159.333333333333 ns	8143.666666666667 ns	1.00
array/random/rand!/Int64	20269 ns	20558 ns	0.99
array/random/randn/Float32	34507 ns	35246 ns	0.98
array/random/randn!/Float32	24040 ns	24445 ns	0.98
array/reductions/mapreduce/Float32/1d	32769 ns	32850 ns	1.00
array/reductions/mapreduce/Float32/dims=1	37846 ns	38019 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	50060 ns	50260 ns	1.00
array/reductions/mapreduce/Float32/dims=2	55201 ns	55487 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	66587 ns	66867 ns	1.00
array/reductions/mapreduce/Int64/1d	40014 ns	39021 ns	1.03
array/reductions/mapreduce/Int64/dims=1	40604 ns	41069 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	86187 ns	86620 ns	1.00
array/reductions/mapreduce/Int64/dims=2	57917 ns	57870 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	82557 ns	82290 ns	1.00
array/reductions/reduce/Float32/1d	32651 ns	32691 ns	1.00
array/reductions/reduce/Float32/dims=1	37928 ns	38016 ns	1.00
array/reductions/reduce/Float32/dims=1L	50240 ns	50354 ns	1.00
array/reductions/reduce/Float32/dims=2	55386 ns	55481 ns	1.00
array/reductions/reduce/Float32/dims=2L	66932 ns	67234 ns	1.00
array/reductions/reduce/Int64/1d	39259 ns	38804 ns	1.01
array/reductions/reduce/Int64/dims=1	40346 ns	40976 ns	0.98
array/reductions/reduce/Int64/dims=1L	86048 ns	86550 ns	0.99
array/reductions/reduce/Int64/dims=2	57824 ns	57935 ns	1.00
array/reductions/reduce/Int64/dims=2L	82149 ns	82488 ns	1.00
array/reverse/1d	16626 ns	16936 ns	0.98
array/reverse/1dL	67580 ns	67830 ns	1.00
array/reverse/1dL_inplace	65280 ns	65354 ns	1.00
array/reverse/1d_inplace	8294.333333333334 ns	10092 ns	0.82
array/reverse/2d	19779 ns	19828 ns	1.00
array/reverse/2dL	71532 ns	71681 ns	1.00
array/reverse/2dL_inplace	65121 ns	65299 ns	1.00
array/reverse/2d_inplace	9697 ns	9999 ns	0.97
array/sorting/1d	2714988 ns	2721914 ns	1.00
array/sorting/2d	1064147 ns	1066145 ns	1.00
array/sorting/by	3280513 ns	3288145 ns	1.00
cuda/synchronization/context/auto	1097.2 ns	1130.5 ns	0.97
cuda/synchronization/context/blocking	894.2045454545455 ns	895.8947368421053 ns	1.00
cuda/synchronization/context/nonblocking	5957 ns	6056.2 ns	0.98
cuda/synchronization/stream/auto	968.7894736842105 ns	976.7333333333333 ns	0.99
cuda/synchronization/stream/blocking	791.1923076923077 ns	793.7916666666666 ns	1.00
cuda/synchronization/stream/nonblocking	5850.5 ns	5977.285714285715 ns	0.98
integration/byval/reference	143210 ns	143365 ns	1.00
integration/byval/slices=1	145163 ns	145406 ns	1.00
integration/byval/slices=2	283899 ns	283824 ns	1.00
integration/byval/slices=3	422144 ns	422343 ns	1.00
integration/cudadevrt	101658 ns	101785 ns	1.00
integration/volumerhs	11240669 ns	11093917 ns	1.01
kernel/indexing	12483 ns	12618 ns	0.99
kernel/indexing_checked	13360 ns	13432 ns	0.99
kernel/launch	2116.777777777778 ns	2074 ns	1.02
kernel/occupancy	698.3197278911565 ns	695.625850340136 ns	1.00
kernel/rand	15798 ns	13974 ns	1.13
latency/import	3847309915 ns	3869060589 ns	0.99
latency/precompile	4623850023 ns	4623502823 ns	1.00
latency/ttfp	4492330513 ns	4510157748 ns	1.00

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

[gbaraldi]

We should really do this upstream as well. Our mulladd could technically leak in the way we emit it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 16.41%. Comparing base (c0534dd) to head (986ca42).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3149      +/-   ##
==========================================
+ Coverage   16.40%   16.41%   +0.01%     
==========================================
  Files         124      124              
  Lines        9827     9827              
==========================================
+ Hits         1612     1613       +1     
+ Misses       8215     8214       -1     

☔ View full report in Codecov by Sentry.
📢 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.