Add device overrides for FastMath.pow_fast with integer exponents

[maleadt]

Base.FastMath.pow_fast(::IEEEFloat, ::Int32) emits the llvm.powi intrinsic, which the NVPTX backend cannot lower (it has no runtime libcalls). Override with inline small-exponent specializations and fast libdevice fallbacks (__nv_fast_powf for Float32, __nv_powi for Float64).

Also makes both the fast and regular versions use compiler flags that allow folding them.

Fixes #3065

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: ceaa278	Previous: 7a46bf3	Ratio
array/accumulate/Float32/1d	100053 ns	100894.5 ns	0.99
array/accumulate/Float32/dims=1	75944.5 ns	76463 ns	0.99
array/accumulate/Float32/dims=1L	1583630 ns	1585055 ns	1.00
array/accumulate/Float32/dims=2	142636 ns	143523 ns	0.99
array/accumulate/Float32/dims=2L	656828 ns	656818 ns	1.00
array/accumulate/Int64/1d	118052 ns	118288.5 ns	1.00
array/accumulate/Int64/dims=1	79154 ns	79382 ns	1.00
array/accumulate/Int64/dims=1L	1694167 ns	1693416 ns	1.00
array/accumulate/Int64/dims=2	155297 ns	155618 ns	1.00
array/accumulate/Int64/dims=2L	961289 ns	961230.5 ns	1.00
array/broadcast	20163 ns	20246 ns	1.00
array/construct	1251.8 ns	1266.7 ns	0.99
array/copy	18051 ns	17864 ns	1.01
array/copyto!/cpu_to_gpu	212833.5 ns	213682 ns	1.00
array/copyto!/gpu_to_cpu	281079 ns	280934 ns	1.00
array/copyto!/gpu_to_gpu	10758 ns	10761 ns	1.00
array/iteration/findall/bool	134443.5 ns	134853 ns	1.00
array/iteration/findall/int	149111 ns	149561.5 ns	1.00
array/iteration/findfirst/bool	81062.5 ns	81079 ns	1.00
array/iteration/findfirst/int	83079 ns	83035 ns	1.00
array/iteration/findmin/1d	83387 ns	85177.5 ns	0.98
array/iteration/findmin/2d	116504 ns	116610 ns	1.00
array/iteration/logical	197089.5 ns	198335.5 ns	0.99
array/iteration/scalar	68052.5 ns	66209 ns	1.03
array/permutedims/2d	51991 ns	51867 ns	1.00
array/permutedims/3d	52575.5 ns	52676 ns	1.00
array/permutedims/4d	51231.5 ns	51384 ns	1.00
array/random/rand/Float32	12426 ns	12428 ns	1.00
array/random/rand/Int64	36410 ns	36390 ns	1.00
array/random/rand!/Float32	8395 ns	8508.333333333334 ns	0.99
array/random/rand!/Int64	33906 ns	34064.5 ns	1.00
array/random/randn/Float32	36883.5 ns	41681.5 ns	0.88
array/random/randn!/Float32	30714 ns	30603.5 ns	1.00
array/reductions/mapreduce/Float32/1d	34143 ns	34414 ns	0.99
array/reductions/mapreduce/Float32/dims=1	45465 ns	40677.5 ns	1.12
array/reductions/mapreduce/Float32/dims=1L	51053.5 ns	50902 ns	1.00
array/reductions/mapreduce/Float32/dims=2	55876 ns	56099.5 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	68830 ns	69125 ns	1.00
array/reductions/mapreduce/Int64/1d	42576 ns	41725 ns	1.02
array/reductions/mapreduce/Int64/dims=1	50585 ns	42515 ns	1.19
array/reductions/mapreduce/Int64/dims=1L	86969 ns	86920 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59351.5 ns	59341 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	84579 ns	84456 ns	1.00
array/reductions/reduce/Float32/1d	34383 ns	34580 ns	0.99
array/reductions/reduce/Float32/dims=1	45725 ns	48846.5 ns	0.94
array/reductions/reduce/Float32/dims=1L	51063 ns	51087 ns	1.00
array/reductions/reduce/Float32/dims=2	56301 ns	56436 ns	1.00
array/reductions/reduce/Float32/dims=2L	69369 ns	69633 ns	1.00
array/reductions/reduce/Int64/1d	42563 ns	41868 ns	1.02
array/reductions/reduce/Int64/dims=1	50473.5 ns	49828.5 ns	1.01
array/reductions/reduce/Int64/dims=1L	86809 ns	86965 ns	1.00
array/reductions/reduce/Int64/dims=2	59227 ns	59126 ns	1.00
array/reductions/reduce/Int64/dims=2L	84097 ns	84186 ns	1.00
array/reverse/1d	17715 ns	17569 ns	1.01
array/reverse/1dL	68284 ns	68149 ns	1.00
array/reverse/1dL_inplace	65572 ns	65619 ns	1.00
array/reverse/1d_inplace	8362.333333333334 ns	10156.5 ns	0.82
array/reverse/2d	20217 ns	20487 ns	0.99
array/reverse/2dL	72232 ns	72580.5 ns	1.00
array/reverse/2dL_inplace	65611 ns	65669 ns	1.00
array/reverse/2d_inplace	9911 ns	10363 ns	0.96
array/sorting/1d	2734237 ns	2732994 ns	1.00
array/sorting/2d	1068460 ns	1074852 ns	0.99
array/sorting/by	3303186 ns	3327030 ns	0.99
cuda/synchronization/context/auto	1205.3 ns	1115.1 ns	1.08
cuda/synchronization/context/blocking	998.0526315789474 ns	926.9285714285714 ns	1.08
cuda/synchronization/context/nonblocking	6937.8 ns	7166.9 ns	0.97
cuda/synchronization/stream/auto	1028.9 ns	972.7368421052631 ns	1.06
cuda/synchronization/stream/blocking	865.6724137931035 ns	783.65 ns	1.10
cuda/synchronization/stream/nonblocking	8062.6 ns	7202.5 ns	1.12
integration/byval/reference	143582 ns	143622 ns	1.00
integration/byval/slices=1	145650 ns	145542 ns	1.00
integration/byval/slices=2	284361 ns	284225 ns	1.00
integration/byval/slices=3	423033.5 ns	422875 ns	1.00
integration/cudadevrt	102246 ns	102233 ns	1.00
integration/volumerhs	23420457.5 ns	23473989 ns	1.00
kernel/indexing	13358 ns	13079 ns	1.02
kernel/indexing_checked	13847 ns	13831 ns	1.00
kernel/launch	2022.9 ns	2051.222222222222 ns	0.99
kernel/occupancy	667.7861635220125 ns	709.1025641025641 ns	0.94
kernel/rand	14268 ns	17700 ns	0.81
latency/import	3800820720 ns	3836580868 ns	0.99
latency/precompile	4587283028.5 ns	4602357164.5 ns	1.00
latency/ttfp	4387699614.5 ns	4418505339 ns	0.99

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.42%. Comparing base (01a0795) to head (ceaa278).
⚠️ Report is 4 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3098      +/-   ##
==========================================
- Coverage   90.43%   90.42%   -0.01%     
==========================================
  Files         141      141              
  Lines       12025    12025              
==========================================
- Hits        10875    10874       -1     
- Misses       1150     1151       +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.