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

[christiangnrd]

Opened to run benchmarks.

Todo:

• Add compat bound when GPUArrays version released

[github-actions]

Metal Benchmarks

Benchmark suite	Current: b296d15	Previous: 1942968	Ratio
latency/precompile	9804167000 ns	9844653958 ns	1.00
latency/ttfp	3988041166.5 ns	3972040229 ns	1.00
latency/import	1280372083 ns	1275530958.5 ns	1.00
integration/metaldevrt	822834 ns	828500 ns	0.99
integration/byval/slices=1	1544479.5 ns	1536750 ns	1.01
integration/byval/slices=3	10917542 ns	9632625 ns	1.13
integration/byval/reference	1542000 ns	1543583 ns	1.00
integration/byval/slices=2	2646833.5 ns	2621958.5 ns	1.01
kernel/indexing	544708 ns	567792 ns	0.96
kernel/indexing_checked	552500 ns	569292 ns	0.97
kernel/launch	8875 ns	9208 ns	0.96
array/construct	6000 ns	6625 ns	0.91
array/broadcast	567417 ns	583375 ns	0.97
array/random/randn/Float32	788125 ns	784333 ns	1.00
array/random/randn!/Float32	634250 ns	623250 ns	1.02
array/random/rand!/Int64	566645.5 ns	547458 ns	1.04
array/random/rand!/Float32	593833 ns	585291 ns	1.01
array/random/rand/Int64	750291.5 ns	771250 ns	0.97
array/random/rand/Float32	611167 ns	622687 ns	0.98
array/accumulate/Int64/1d	2082292 ns	1277104.5 ns	1.63
array/accumulate/Int64/dims=1	2177000 ns	1868333 ns	1.17
array/accumulate/Int64/dims=2	2116083 ns	2183625 ns	0.97
array/accumulate/Int64/dims=1L	6449187.5 ns	11737104 ns	0.55
array/accumulate/Int64/dims=2L	17938375 ns	9771416.5 ns	1.84
array/accumulate/Float32/1d	1693042 ns	1142833 ns	1.48
array/accumulate/Float32/dims=1	1982625 ns	1570458 ns	1.26
array/accumulate/Float32/dims=2	2028583 ns	1931625 ns	1.05
array/accumulate/Float32/dims=1L	4866500 ns	9864375 ns	0.49
array/accumulate/Float32/dims=2L	14886187.5 ns	7308021 ns	2.04
array/reductions/reduce/Int64/1d	1432291.5 ns	1373353.5 ns	1.04
array/reductions/reduce/Int64/dims=1	1059375 ns	1069291.5 ns	0.99
array/reductions/reduce/Int64/dims=2	1181312.5 ns	1193292 ns	0.99
array/reductions/reduce/Int64/dims=1L	2088917 ns	2113062.5 ns	0.99
array/reductions/reduce/Int64/dims=2L	3433958 ns	3456458 ns	0.99
array/reductions/reduce/Float32/1d	946750 ns	971625 ns	0.97
array/reductions/reduce/Float32/dims=1	783562.5 ns	808458 ns	0.97
array/reductions/reduce/Float32/dims=2	771312.5 ns	768979 ns	1.00
array/reductions/reduce/Float32/dims=1L	1720625 ns	1739041 ns	0.99
array/reductions/reduce/Float32/dims=2L	1760062.5 ns	1772125 ns	0.99
array/reductions/mapreduce/Int64/1d	1489271 ns	1456146 ns	1.02
array/reductions/mapreduce/Int64/dims=1	1097750 ns	1074875 ns	1.02
array/reductions/mapreduce/Int64/dims=2	1118541.5 ns	1206417 ns	0.93
array/reductions/mapreduce/Int64/dims=1L	2112562.5 ns	2119292 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	3424187.5 ns	3444375 ns	0.99
array/reductions/mapreduce/Float32/1d	995729 ns	990792 ns	1.00
array/reductions/mapreduce/Float32/dims=1	782500 ns	810062.5 ns	0.97
array/reductions/mapreduce/Float32/dims=2	763291 ns	761104 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	1735041 ns	1740812.5 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	1761854 ns	1781292 ns	0.99
array/private/copyto!/gpu_to_gpu	637416 ns	651375 ns	0.98
array/private/copyto!/cpu_to_gpu	802750 ns	805542 ns	1.00
array/private/copyto!/gpu_to_cpu	800291.5 ns	817667 ns	0.98
array/private/iteration/findall/int	1828291.5 ns	1646500 ns	1.11
array/private/iteration/findall/bool	1601167 ns	1444584 ns	1.11
array/private/iteration/findfirst/int	1853209 ns	1754958.5 ns	1.06
array/private/iteration/findfirst/bool	1771604 ns	1703625 ns	1.04
array/private/iteration/scalar	4173625 ns	4772500 ns	0.87
array/private/iteration/logical	2695062 ns	2536917 ns	1.06
array/private/iteration/findmin/1d	1856916 ns	1815666 ns	1.02
array/private/iteration/findmin/2d	1412041 ns	1431750 ns	0.99
array/private/copy	558250 ns	538167 ns	1.04
array/shared/copyto!/gpu_to_gpu	83959 ns	86375 ns	0.97
array/shared/copyto!/cpu_to_gpu	82541.5 ns	86583 ns	0.95
array/shared/copyto!/gpu_to_cpu	82792 ns	84833 ns	0.98
array/shared/iteration/findall/int	1850292 ns	1609874.5 ns	1.15
array/shared/iteration/findall/bool	1607187.5 ns	1464354 ns	1.10
array/shared/iteration/findfirst/int	1408875 ns	1377750 ns	1.02
array/shared/iteration/findfirst/bool	1347166.5 ns	1319166 ns	1.02
array/shared/iteration/scalar	205708 ns	217500 ns	0.95
array/shared/iteration/logical	2387416 ns	2288708.5 ns	1.04
array/shared/iteration/findmin/1d	1399167 ns	1421750 ns	0.98
array/shared/iteration/findmin/2d	1422937.5 ns	1430854.5 ns	0.99
array/shared/copy	249042 ns	248666 ns	1.00
array/permutedims/4d	2439875 ns	2438438 ns	1.00
array/permutedims/2d	1169750 ns	1193250 ns	0.98
array/permutedims/3d	1740250 ns	1768458 ns	0.98
metal/synchronization/stream	19042 ns	19916 ns	0.96
metal/synchronization/context	20333 ns	20375 ns	1.00

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

[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/test/runtests.jl b/test/runtests.jl
index 332f4a45..45c343ad 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="accumulatetests")
+Pkg.add(url = "https://github.com/christiangnrd/GPUArrays.jl", rev = "accumulatetests")
 
 # Quit without erroring if Metal loaded without issues on unsupported platforms
 if !Sys.isapple()

[christiangnrd]

As expected, some small regressions for most accumulate benchmarks, with a massive regression when accumulating along rows of a 3x1000000 matrix.

The performance improvement with column-wise accumulation with 3x1000000 matrices comes from Metal missing an easy optimization (see #626) Edit: I got confused this optimization is only present for reductions.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.35%. Comparing base (1942968) to head (b296d15).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #625      +/-   ##
==========================================
- Coverage   80.63%   80.35%   -0.29%     
==========================================
  Files          61       60       -1     
  Lines        2722     2678      -44     
==========================================
- Hits         2195     2152      -43     
+ Misses        527      526       -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.

[maleadt]

As expected, some small regressions for most accumulate benchmarks, with a massive regression when accumulating along rows of a 3x1000000 matrix.

I don't see a massive slowdown?

[christiangnrd]

@maleadt The accumulate dims=2L benchmarks show a 2x slowdown. Did I get my rows/columns mixed up in my comment?

[maleadt]

Oh OK, I didn't consider 2x a "massive slowdown" :-) Still something to look at of course, but much less dramatic than the 7x regressions we e.g. saw against CUDA.jl's reduction.