[DO NOT MERGE] Test accumulate acceleratedkernel

[christiangnrd]

I'm getting errors locally

[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/src/accumulate.jl b/src/accumulate.jl
index dee031fb..781864ce 100644
--- a/src/accumulate.jl
+++ b/src/accumulate.jl
@@ -170,17 +170,17 @@ end
 ## Base interface
 
 Base._accumulate!(op, output::WrappedMtlArray, input::WrappedMtlVector, dims::Nothing, init::Nothing) =
-    @inline AK.accumulate!(op, output, input; dims, init=AK.neutral_element(op, eltype(output)), alg=AK.ScanPrefixes())
+    @inline AK.accumulate!(op, output, input; dims, init = AK.neutral_element(op, eltype(output)), alg = AK.ScanPrefixes())
 
 Base._accumulate!(op, output::WrappedMtlArray, input::WrappedMtlArray, dims::Integer, init::Nothing) =
-    @inline AK.accumulate!(op, output, input; dims, init=AK.neutral_element(op, eltype(output)), alg=AK.ScanPrefixes())
+    @inline AK.accumulate!(op, output, input; dims, init = AK.neutral_element(op, eltype(output)), alg = AK.ScanPrefixes())
 Base._accumulate!(op, output::WrappedMtlArray, input::MtlVector, dims::Nothing, init::Some) =
-    @inline AK.accumulate!(op, output, input; dims, init=something(init), alg=AK.ScanPrefixes())
+    @inline AK.accumulate!(op, output, input; dims, init = something(init), alg = AK.ScanPrefixes())
 
 Base._accumulate!(op, output::WrappedMtlArray, input::WrappedMtlArray, dims::Integer, init::Some) =
-    @inline AK.accumulate!(op, output, input; dims, init=something(init), alg=AK.ScanPrefixes())
+    @inline AK.accumulate!(op, output, input; dims, init = something(init), alg = AK.ScanPrefixes())
 
-Base.accumulate_pairwise!(op, result::WrappedMtlVector, v::WrappedMtlVector) = @inline AK.accumulate!(op, result, v; init=AK.neutral_element(op, eltype(result)), alg=AK.ScanPrefixes())
+Base.accumulate_pairwise!(op, result::WrappedMtlVector, v::WrappedMtlVector) = @inline AK.accumulate!(op, result, v; init = AK.neutral_element(op, eltype(result)), alg = AK.ScanPrefixes())
 
 # default behavior unless dims are specified by the user
 function Base.accumulate(op, A::WrappedMtlArray;
@@ -188,7 +188,7 @@ function Base.accumulate(op, A::WrappedMtlArray;
     nt = values(kw)
     if dims === nothing && !(A isa AbstractVector)
         # This branch takes care of the cases not handled by `_accumulate!`.
-        return reshape(AK.accumulate(op, A[:]; init = (:init in keys(kw) ? nt.init : AK.neutral_element(op, eltype(A))), alg=AK.ScanPrefixes()), size(A))
+        return reshape(AK.accumulate(op, A[:]; init = (:init in keys(kw) ? nt.init : AK.neutral_element(op, eltype(A))), alg = AK.ScanPrefixes()), size(A))
     end
     if isempty(kw)
         out = similar(A, Base.promote_op(op, eltype(A), eltype(A)))

[maleadt]

Why not implement this at the GPUArrays.jl level? Or is this purely for debugging?

[christiangnrd]

@maleadt purely for debugging. I’ll add do not merge in the title

[christiangnrd]

The goal behind this was to get an idea of what the benchmark results look like, but there seem to be strange failures that I haven't yet determined if they're a Metal issue or an AcceleratedKernels issue.

[christiangnrd]

Just for benchmarks, this still shouldn't be merged

[github-actions]

Metal Benchmarks

Benchmark suite	Current: e38640b	Previous: caf2996	Ratio
private array/construct	24791.666666666668 ns	25048.666666666668 ns	0.99
private array/broadcast	462292 ns	457958 ns	1.01
private array/random/randn/Float32	829000 ns	800854.5 ns	1.04
private array/random/randn!/Float32	630875 ns	641125 ns	0.98
private array/random/rand!/Int64	581166 ns	568834 ns	1.02
private array/random/rand!/Float32	596459 ns	597333 ns	1.00
private array/random/rand/Int64	772687.5 ns	774854.5 ns	1.00
private array/random/rand/Float32	620812.5 ns	631333 ns	0.98
private array/copyto!/gpu_to_gpu	664458 ns	660958 ns	1.01
private array/copyto!/cpu_to_gpu	623875 ns	832125 ns	0.75
private array/copyto!/gpu_to_cpu	823875 ns	590416 ns	1.40
private array/accumulate/1d	1577458 ns	1352208 ns	1.17
private array/accumulate/2d	1725958.5 ns	1397271 ns	1.24
private array/iteration/findall/int	1681750 ns	1811500 ns	0.93
private array/iteration/findall/bool	1473708 ns	1579583 ns	0.93
private array/iteration/findfirst/int	1701417 ns	1738479.5 ns	0.98
private array/iteration/findfirst/bool	1658042 ns	1671583.5 ns	0.99
private array/iteration/scalar	3600209 ns	3925542 ns	0.92
private array/iteration/logical	2860896 ns	3003542 ns	0.95
private array/iteration/findmin/1d	1749042 ns	1771542 ns	0.99
private array/iteration/findmin/2d	1354167 ns	1358104 ns	1.00
private array/reductions/reduce/1d	1057375 ns	1034333 ns	1.02
private array/reductions/reduce/2d	658958 ns	661771 ns	1.00
private array/reductions/mapreduce/1d	1042791 ns	1042688 ns	1.00
private array/reductions/mapreduce/2d	661604.5 ns	656417 ns	1.01
private array/permutedims/4d	2529354 ns	2535708 ns	1.00
private array/permutedims/2d	1024166.5 ns	1024729.5 ns	1.00
private array/permutedims/3d	1579562.5 ns	1604959 ns	0.98
private array/copy	561083 ns	573375 ns	0.98
latency/precompile	9974760250 ns	9802163250 ns	1.02
latency/ttfp	4176340083 ns	4108894000 ns	1.02
latency/import	1402819458.5 ns	1286845416.5 ns	1.09
integration/metaldevrt	726625 ns	708167 ns	1.03
integration/byval/slices=1	1579854 ns	1642916 ns	0.96
integration/byval/slices=3	8492729 ns	9533624.5 ns	0.89
integration/byval/reference	1613187 ns	1619479 ns	1.00
integration/byval/slices=2	2662416.5 ns	2720083.5 ns	0.98
kernel/indexing	459500 ns	461167 ns	1.00
kernel/indexing_checked	466291.5 ns	459625 ns	1.01
kernel/launch	8250 ns	7958 ns	1.04
metal/synchronization/stream	14750 ns	14750 ns	1
metal/synchronization/context	15208 ns	15250 ns	1.00
shared array/construct	24267.333333333332 ns	23722.25 ns	1.02
shared array/broadcast	462084 ns	462042 ns	1.00
shared array/random/randn/Float32	855750 ns	809042 ns	1.06
shared array/random/randn!/Float32	647166 ns	641042 ns	1.01
shared array/random/rand!/Int64	570479 ns	572917 ns	1.00
shared array/random/rand!/Float32	600416 ns	603417 ns	1.00
shared array/random/rand/Int64	766541.5 ns	792625 ns	0.97
shared array/random/rand/Float32	645875 ns	625166 ns	1.03
shared array/copyto!/gpu_to_gpu	80875 ns	89375 ns	0.90
shared array/copyto!/cpu_to_gpu	83542 ns	82834 ns	1.01
shared array/copyto!/gpu_to_cpu	84125 ns	78708 ns	1.07
shared array/accumulate/1d	1560959 ns	1358229 ns	1.15
shared array/accumulate/2d	1870458.5 ns	1393083.5 ns	1.34
shared array/iteration/findall/int	1719167 ns	1846500.5 ns	0.93
shared array/iteration/findall/bool	1471958 ns	1593458 ns	0.92
shared array/iteration/findfirst/int	1399250 ns	1406500 ns	0.99
shared array/iteration/findfirst/bool	1378875 ns	1369834 ns	1.01
shared array/iteration/scalar	159541 ns	160875 ns	0.99
shared array/iteration/logical	2647333 ns	2872291.5 ns	0.92
shared array/iteration/findmin/1d	1465709 ns	1514166 ns	0.97
shared array/iteration/findmin/2d	1365541 ns	1377125 ns	0.99
shared array/reductions/reduce/1d	737000 ns	725229 ns	1.02
shared array/reductions/reduce/2d	664896 ns	664500 ns	1.00
shared array/reductions/mapreduce/1d	737396 ns	739395.5 ns	1.00
shared array/reductions/mapreduce/2d	675416 ns	666562.5 ns	1.01
shared array/permutedims/4d	2558375 ns	2514583 ns	1.02
shared array/permutedims/2d	1024916.5 ns	1027667 ns	1.00
shared array/permutedims/3d	1575834 ns	1579792 ns	1.00
shared array/copy	249625 ns	250375 ns	1.00

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