[Do not merge] Switch to GPUArrays.jl reduction implementation

[christiangnrd]

Don't remove the file yet to avoid merge conflict with #627

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/perf/runbenchmarks.jl b/perf/runbenchmarks.jl
index ba5e0d40..1d7901c5 100644
--- a/perf/runbenchmarks.jl
+++ b/perf/runbenchmarks.jl
@@ -1,6 +1,6 @@
 # benchmark suite execution and codespeed submission
 using Pkg
-Pkg.add(url="https://github.com/christiangnrd/GPUArrays.jl", rev="akreduce")
+Pkg.add(url = "https://github.com/christiangnrd/GPUArrays.jl", rev = "akreduce")
 
 using Metal
 
diff --git a/test/runtests.jl b/test/runtests.jl
index 4ee51134..fb376e4f 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -6,7 +6,7 @@ import REPL
 using Test
 
 using Pkg
-Pkg.add(url="https://github.com/christiangnrd/GPUArrays.jl", rev="akreduce")
+Pkg.add(url = "https://github.com/christiangnrd/GPUArrays.jl", rev = "akreduce")
 
 # Quit without erroring if Metal loaded without issues on unsupported platforms
 if !Sys.isapple()

[christiangnrd]

Leaving the current mapreducedim! implementation present, we can transition in two parts. First once AK supports broadcasted reductions, and then remove implementations from this repo after AK supports >1 input dims.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.63%. Comparing base (1942968) to head (c0eddd1).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #628   +/-   ##
=======================================
  Coverage   80.63%   80.63%           
=======================================
  Files          61       61           
  Lines        2722     2722           
=======================================
  Hits         2195     2195           
  Misses        527      527           

☔ 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.

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: 2e5239f	Previous: 28d2eb3	Ratio
latency/precompile	24914872583 ns	25451816292 ns	0.98
latency/ttfp	2147869250 ns	2366668104.5 ns	0.91
latency/import	1234056917 ns	1439498750 ns	0.86
integration/metaldevrt	874458 ns	865458 ns	1.01
integration/byval/slices=1	1570291 ns	1573625 ns	1.00
integration/byval/slices=3	8384667 ns	8948666.5 ns	0.94
integration/byval/reference	1561833 ns	1548416 ns	1.01
integration/byval/slices=2	2637792 ns	2650958 ns	1.00
kernel/indexing	640166.5 ns	647875 ns	0.99
kernel/indexing_checked	616104.5 ns	636375 ns	0.97
kernel/launch	12875 ns	11625 ns	1.11
kernel/rand	579291 ns	578375 ns	1.00
array/construct	6625 ns	6084 ns	1.09
array/broadcast	611062.5 ns	605917 ns	1.01
array/random/randn/Float32	1019604 ns	1008250 ns	1.01
array/random/randn!/Float32	757875 ns	729458 ns	1.04
array/random/rand!/Int64	560125 ns	547959 ns	1.02
array/random/rand!/Float32	589958 ns	597333 ns	0.99
array/random/rand/Int64	763500 ns	743500 ns	1.03
array/random/rand/Float32	620625 ns	636959 ns	0.97
array/accumulate/Int64/1d	1241417 ns	1254333 ns	0.99
array/accumulate/Int64/dims=1	1855917 ns	1840916.5 ns	1.01
array/accumulate/Int64/dims=2	2192625 ns	2154750 ns	1.02
array/accumulate/Int64/dims=1L	11493750 ns	11672583 ns	0.98
array/accumulate/Int64/dims=2L	9823583.5 ns	9846791.5 ns	1.00
array/accumulate/Float32/1d	1132812.5 ns	1104187.5 ns	1.03
array/accumulate/Float32/dims=1	1565292 ns	1543791.5 ns	1.01
array/accumulate/Float32/dims=2	1878958.5 ns	1875562.5 ns	1.00
array/accumulate/Float32/dims=1L	9760312.5 ns	9790417 ns	1.00
array/accumulate/Float32/dims=2L	7244917 ns	7247208 ns	1.00
array/reductions/reduce/Int64/1d	577834 ns	1573749.5 ns	0.37
array/reductions/reduce/Int64/dims=1	1044708 ns	1091271 ns	0.96
array/reductions/reduce/Int64/dims=2	1046417 ns	1133334 ns	0.92
array/reductions/reduce/Int64/dims=1L	2495396 ns	2015499.5 ns	1.24
array/reductions/reduce/Int64/dims=2L	2872562.5 ns	4237625 ns	0.68
array/reductions/reduce/Float32/1d	957791.5 ns	991562.5 ns	0.97
array/reductions/reduce/Float32/dims=1	1043042 ns	837958 ns	1.24
array/reductions/reduce/Float32/dims=2	1069833 ns	865645.5 ns	1.24
array/reductions/reduce/Float32/dims=1L	1796062.5 ns	1332167 ns	1.35
array/reductions/reduce/Float32/dims=2L	2859416 ns	1802521 ns	1.59
array/reductions/mapreduce/Int64/1d	613333 ns	1505209 ns	0.41
array/reductions/mapreduce/Int64/dims=1	1040271 ns	1092375 ns	0.95
array/reductions/mapreduce/Int64/dims=2	1079208 ns	1141166.5 ns	0.95
array/reductions/mapreduce/Int64/dims=1L	2444604.5 ns	2016167 ns	1.21
array/reductions/mapreduce/Int64/dims=2L	2875541 ns	3599334 ns	0.80
array/reductions/mapreduce/Float32/1d	1037666.5 ns	1036333.5 ns	1.00
array/reductions/mapreduce/Float32/dims=1	1049625 ns	845584 ns	1.24
array/reductions/mapreduce/Float32/dims=2	1060542 ns	849958 ns	1.25
array/reductions/mapreduce/Float32/dims=1L	1797542 ns	1305625 ns	1.38
array/reductions/mapreduce/Float32/dims=2L	2863021 ns	1811750 ns	1.58
array/private/copyto!/gpu_to_gpu	634792 ns	642208 ns	0.99
array/private/copyto!/cpu_to_gpu	803750 ns	796688 ns	1.01
array/private/copyto!/gpu_to_cpu	801250 ns	807500 ns	0.99
array/private/iteration/findall/int	1585750 ns	1567583 ns	1.01
array/private/iteration/findall/bool	1407792 ns	1400584 ns	1.01
array/private/iteration/findfirst/int	1913479.5 ns	2097125 ns	0.91
array/private/iteration/findfirst/bool	1549167 ns	2047416.5 ns	0.76
array/private/iteration/scalar	4096792 ns	3930375 ns	1.04
array/private/iteration/logical	2912313 ns	2624146.5 ns	1.11
array/private/iteration/findmin/1d	1898958 ns	2514125 ns	0.76
array/private/iteration/findmin/2d	1874854.5 ns	1793520.5 ns	1.05
array/private/copy	617417 ns	571104.5 ns	1.08
array/shared/copyto!/gpu_to_gpu	86834 ns	85708 ns	1.01
array/shared/copyto!/cpu_to_gpu	85292 ns	81542 ns	1.05
array/shared/copyto!/gpu_to_cpu	88541 ns	82333 ns	1.08
array/shared/iteration/findall/int	1583709 ns	1568292 ns	1.01
array/shared/iteration/findall/bool	1443500 ns	1421958 ns	1.02
array/shared/iteration/findfirst/int	1627354 ns	1664250 ns	0.98
array/shared/iteration/findfirst/bool	1424125 ns	1655375 ns	0.86
array/shared/iteration/scalar	204708 ns	201375 ns	1.02
array/shared/iteration/logical	2863250 ns	2429250 ns	1.18
array/shared/iteration/findmin/1d	1709291 ns	2128417 ns	0.80
array/shared/iteration/findmin/2d	1950500 ns	1798333.5 ns	1.08
array/shared/copy	250458 ns	248750 ns	1.01
array/permutedims/4d	2395167 ns	2382625 ns	1.01
array/permutedims/2d	1043250 ns	1176375 ns	0.89
array/permutedims/3d	1689062 ns	1675000 ns	1.01
metal/synchronization/stream	19792 ns	19000 ns	1.04
metal/synchronization/context	20167 ns	20125 ns	1.00

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

[maleadt]

Leaving the current mapreducedim! implementation present, we can transition in two parts. First once AK supports broadcasted reductions, and then remove implementations from this repo after AK supports >1 input dims.

I think I'd rather we do it in one pass, because the change needs to be made across back-ends.

[maleadt]

In any case, despite some regressions the overall performance seems better here than over in CUDA.jl.