[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: 84f519a	Previous: de3fd23	Ratio
latency/precompile	25018084416 ns	25055876416 ns	1.00
latency/ttfp	2129990500 ns	2125052000 ns	1.00
latency/import	1225508166 ns	1219352833 ns	1.01
integration/metaldevrt	956625 ns	968354.5 ns	0.99
integration/byval/slices=1	1644625 ns	1660375 ns	0.99
integration/byval/slices=3	10295687.5 ns	8945875 ns	1.15
integration/byval/reference	1633875 ns	1638208 ns	1.00
integration/byval/slices=2	2747625 ns	2721062.5 ns	1.01
kernel/indexing	692437.5 ns	703875 ns	0.98
kernel/indexing_checked	681375 ns	694208 ns	0.98
kernel/launch	13020.5 ns	12875 ns	1.01
array/construct	6292 ns	6083 ns	1.03
array/broadcast	660542 ns	670666.5 ns	0.98
array/random/randn/Float32	849938 ns	879916 ns	0.97
array/random/randn!/Float32	621917 ns	639812.5 ns	0.97
array/random/rand!/Int64	554792 ns	567000 ns	0.98
array/random/rand!/Float32	589083 ns	602916 ns	0.98
array/random/rand/Int64	752104.5 ns	754292 ns	1.00
array/random/rand/Float32	545291 ns	574541 ns	0.95
array/accumulate/Int64/1d	2378188 ns	1336875 ns	1.78
array/accumulate/Int64/dims=1	2295312.5 ns	1912291.5 ns	1.20
array/accumulate/Int64/dims=2	2555417 ns	2256916.5 ns	1.13
array/accumulate/Int64/dims=1L	6595145.5 ns	11644666.5 ns	0.57
array/accumulate/Int64/dims=2L	18580062.5 ns	9900979.5 ns	1.88
array/accumulate/Float32/1d	1685084 ns	1245625 ns	1.35
array/accumulate/Float32/dims=1	2124459 ns	1630541.5 ns	1.30
array/accumulate/Float32/dims=2	2386125 ns	1968750 ns	1.21
array/accumulate/Float32/dims=1L	5082146 ns	9898709 ns	0.51
array/accumulate/Float32/dims=2L	14983750 ns	7337354 ns	2.04
array/reductions/reduce/Int64/1d	1349687.5 ns	1381500.5 ns	0.98
array/reductions/reduce/Int64/dims=1	1177333 ns	1154562.5 ns	1.02
array/reductions/reduce/Int64/dims=2	1291041 ns	1287541 ns	1.00
array/reductions/reduce/Int64/dims=1L	2127500 ns	2078000 ns	1.02
array/reductions/reduce/Int64/dims=2L	3575749.5 ns	3569083 ns	1.00
array/reductions/reduce/Float32/1d	1015125 ns	1047333.5 ns	0.97
array/reductions/reduce/Float32/dims=1	885375 ns	899875 ns	0.98
array/reductions/reduce/Float32/dims=2	800416 ns	801708.5 ns	1.00
array/reductions/reduce/Float32/dims=1L	1386084 ns	1393042 ns	1.00
array/reductions/reduce/Float32/dims=2L	1909291 ns	1903875 ns	1.00
array/reductions/mapreduce/Int64/1d	1338875 ns	1353375 ns	0.99
array/reductions/mapreduce/Int64/dims=1	1142416 ns	1160042 ns	0.98
array/reductions/mapreduce/Int64/dims=2	1287270.5 ns	1282979 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	2103667 ns	2111146 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	3442541.5 ns	3466062 ns	0.99
array/reductions/mapreduce/Float32/1d	975541 ns	1083604 ns	0.90
array/reductions/mapreduce/Float32/dims=1	890208.5 ns	902542 ns	0.99
array/reductions/mapreduce/Float32/dims=2	788042 ns	819041.5 ns	0.96
array/reductions/mapreduce/Float32/dims=1L	1385458.5 ns	1404791.5 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	1918042 ns	1904375 ns	1.01
array/private/copyto!/gpu_to_gpu	642271 ns	661417 ns	0.97
array/private/copyto!/cpu_to_gpu	819417 ns	827708 ns	0.99
array/private/copyto!/gpu_to_cpu	818854.5 ns	823833 ns	0.99
array/private/iteration/findall/int	1746687.5 ns	1654645.5 ns	1.06
array/private/iteration/findall/bool	1575458 ns	1502750 ns	1.05
array/private/iteration/findfirst/int	1933875 ns	2023208 ns	0.96
array/private/iteration/findfirst/bool	1745458 ns	1852750 ns	0.94
array/private/iteration/scalar	4033375 ns	5040709 ns	0.80
array/private/iteration/logical	2604125 ns	2707041 ns	0.96
array/private/iteration/findmin/1d	1990291 ns	2059979 ns	0.97
array/private/iteration/findmin/2d	1640084 ns	1638750 ns	1.00
array/private/copy	580229.5 ns	566958.5 ns	1.02
array/shared/copyto!/gpu_to_gpu	80708 ns	79375 ns	1.02
array/shared/copyto!/cpu_to_gpu	79708 ns	81333 ns	0.98
array/shared/copyto!/gpu_to_cpu	80000 ns	78750 ns	1.02
array/shared/iteration/findall/int	1761916.5 ns	1657354 ns	1.06
array/shared/iteration/findall/bool	1683500 ns	1507000 ns	1.12
array/shared/iteration/findfirst/int	1539875 ns	1648125 ns	0.93
array/shared/iteration/findfirst/bool	1427729.5 ns	1429542 ns	1.00
array/shared/iteration/scalar	161459 ns	159083 ns	1.01
array/shared/iteration/logical	2442375 ns	2359208 ns	1.04
array/shared/iteration/findmin/1d	1511833 ns	1598729.5 ns	0.95
array/shared/iteration/findmin/2d	1630792 ns	1642520.5 ns	0.99
array/shared/copy	250604 ns	253958 ns	0.99
array/permutedims/4d	2465500 ns	2460792 ns	1.00
array/permutedims/2d	1249208.5 ns	1249583.5 ns	1.00
array/permutedims/3d	1743167 ns	1743375 ns	1.00
metal/synchronization/stream	14875 ns	14875 ns	1
metal/synchronization/context	15708 ns	15500 ns	1.01

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 9b6b0c3d..6d16c110 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -11,7 +11,7 @@ if parse(Bool, get(ENV, "BUILDKITE", "false"))
 end
 
 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.