KernelAbstractions 0.10

[christiangnrd]

Not a draft to also run benchmarks

[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/MetalKernels.jl b/src/MetalKernels.jl
index 4e856194..7573c5e1 100644
--- a/src/MetalKernels.jl
+++ b/src/MetalKernels.jl
@@ -136,26 +136,26 @@ end
 
 KI.argconvert(::MetalBackend, arg) = mtlconvert(arg)
 
-function KI.kernel_function(::MetalBackend, f::F, tt::TT=Tuple{}; name=nothing, kwargs...) where {F,TT}
+function KI.kernel_function(::MetalBackend, f::F, tt::TT = Tuple{}; name = nothing, kwargs...) where {F, TT}
     kern = mtlfunction(f, tt; name, kwargs...)
-    KI.Kernel{MetalBackend, typeof(kern)}(MetalBackend(), kern)
+    return KI.Kernel{MetalBackend, typeof(kern)}(MetalBackend(), kern)
 end
 
-function (obj::KI.Kernel{MetalBackend})(args...; numworkgroups=1, workgroupsize=1)
+function (obj::KI.Kernel{MetalBackend})(args...; numworkgroups = 1, workgroupsize = 1)
     KI.check_launch_args(numworkgroups, workgroupsize)
 
-    obj.kern(args...; threads=workgroupsize, groups=numworkgroups)
+    return obj.kern(args...; threads = workgroupsize, groups = numworkgroups)
 end
 
 
-function KI.kernel_max_work_group_size(kikern::KI.Kernel{<:MetalBackend}; max_work_items::Int=typemax(Int))::Int
-    Int(min(kikern.kern.pipeline.maxTotalThreadsPerThreadgroup, max_work_items))
+function KI.kernel_max_work_group_size(kikern::KI.Kernel{<:MetalBackend}; max_work_items::Int = typemax(Int))::Int
+    return Int(min(kikern.kern.pipeline.maxTotalThreadsPerThreadgroup, max_work_items))
 end
 function KI.max_work_group_size(::MetalBackend)::Int
-    Int(device().maxThreadsPerThreadgroup.width)
+    return Int(device().maxThreadsPerThreadgroup.width)
 end
 function KI.multiprocessor_count(::MetalBackend)::Int
-    Metal.num_gpu_cores()
+    return Metal.num_gpu_cores()
 end
 
 
diff --git a/src/broadcast.jl b/src/broadcast.jl
index 5d107ec2..3455fad2 100644
--- a/src/broadcast.jl
+++ b/src/broadcast.jl
@@ -66,9 +66,9 @@ end
     if _broadcast_shapes[Is] > BROADCAST_SPECIALIZATION_THRESHOLD
         ## COV_EXCL_START
         function broadcast_cartesian_static(dest, bc, Is)
-             i = KI.get_global_id().x
-             stride = KI.get_global_size().x
-             while 1 <= i <= length(dest)
+            i = KI.get_global_id().x
+            stride = KI.get_global_size().x
+            while 1 <= i <= length(dest)
                 I = @inbounds Is[i]
                 @inbounds dest[I] = bc[I]
                 i += stride
@@ -91,13 +91,13 @@ end
        (isa(IndexStyle(dest), IndexLinear) && isa(IndexStyle(bc), IndexLinear))
         ## COV_EXCL_START
         function broadcast_linear(dest, bc)
-             i = KI.get_global_id().x
-             stride = KI.get_global_size().x
-             while 1 <= i <= length(dest)
-                 @inbounds dest[i] = bc[i]
-                 i += stride
-             end
-             return
+            i = KI.get_global_id().x
+            stride = KI.get_global_size().x
+            while 1 <= i <= length(dest)
+                @inbounds dest[i] = bc[i]
+                i += stride
+            end
+            return
         end
         ## COV_EXCL_STOP
 
@@ -168,9 +168,9 @@ end
     else
         ## COV_EXCL_START
         function broadcast_cartesian(dest, bc)
-             i = KI.get_global_id().x
-             stride = KI.get_global_size().x
-             while 1 <= i <= length(dest)
+            i = KI.get_global_id().x
+            stride = KI.get_global_size().x
+            while 1 <= i <= length(dest)
                 I = @inbounds CartesianIndices(dest)[i]
                 @inbounds dest[I] = bc[I]
                 i += stride
diff --git a/src/device/random.jl b/src/device/random.jl
index 12b053a2..edc999cd 100644
--- a/src/device/random.jl
+++ b/src/device/random.jl
@@ -89,8 +89,8 @@ end
         @inbounds global_random_counters()[simdgroupId]
     elseif field === :ctr2
         globalId = KI.get_global_id().x +
-                   (KI.get_global_id().y - 1i32) * KI.get_global_size().x +
-                   (KI.get_global_id().z - 1i32) * KI.get_global_size().x * KI.get_global_size().y
+            (KI.get_global_id().y - 1i32) * KI.get_global_size().x +
+            (KI.get_global_id().z - 1i32) * KI.get_global_size().x * KI.get_global_size().y
         globalId % UInt32
     end::UInt32
 end
diff --git a/src/mapreduce.jl b/src/mapreduce.jl
index 7be5ef43..a737e8d0 100644
--- a/src/mapreduce.jl
+++ b/src/mapreduce.jl
@@ -224,7 +224,8 @@ function GPUArrays.mapreducedim!(f::F, op::OP, R::WrappedMtlArray{T},
     # we might not be able to launch all those threads to reduce each slice in one go.
     # that's why each threads also loops across their inputs, processing multiple values
     # so that we can span the entire reduction dimension using a single item group.
-    kernel = KI.@kernel backend launch = false partial_mapreduce_device(f, op, init, Val(maxthreads), Val(Rreduce), Val(Rother),
+    kernel = KI.@kernel backend launch = false partial_mapreduce_device(
+        f, op, init, Val(maxthreads), Val(Rreduce), Val(Rother),
                                                           Val(UInt64(length(Rother))), Val(grain), Val(shuffle), R, A)
 
     # how many threads do we want?
@@ -260,7 +261,8 @@ function GPUArrays.mapreducedim!(f::F, op::OP, R::WrappedMtlArray{T},
         # we can cover the dimensions to reduce using a single group
         kernel(f, op, init, Val(maxthreads), Val(Rreduce), Val(Rother),
                Val(UInt64(length(Rother))), Val(grain), Val(shuffle), R, A;
-               workgroupsize = threads, numworkgroups = groups)
+            workgroupsize = threads, numworkgroups = groups
+        )
     else
         # temporary empty array whose type will match the final partial array
 	    partial = similar(R, ntuple(_ -> 0, Val(ndims(R)+1)))
@@ -287,7 +289,8 @@ function GPUArrays.mapreducedim!(f::F, op::OP, R::WrappedMtlArray{T},
         partial_kernel(f, op, init, Val(threads), Val(Rreduce),
                         Val(Rother), Val(UInt64(length(Rother))),
                         Val(grain), Val(shuffle), partial, A;
-                        numworkgroups = partial_groups, workgroupsize = partial_threads)
+            numworkgroups = partial_groups, workgroupsize = partial_threads
+        )
 
         GPUArrays.mapreducedim!(identity, op, R, partial; init=init)
     end
diff --git a/test/kernelabstractions.jl b/test/kernelabstractions.jl
index cda5b249..339fcbc8 100644
--- a/test/kernelabstractions.jl
+++ b/test/kernelabstractions.jl
@@ -7,6 +7,6 @@ Testsuite.testsuite(MetalBackend, "Metal", Metal, MtlArray, Metal.MtlDeviceArray
     "Convert",           # depends on https://github.com/JuliaGPU/Metal.jl/issues/69
     "SpecialFunctions",  # gamma and erfc not currently supported on Metal.jl
     "sparse",            # not supported yet
-    "CPU synchronization",
-    "fallback test: callable types",
+            "CPU synchronization",
+            "fallback test: callable types",
 ]))
diff --git a/test/runtests.jl b/test/runtests.jl
index 32b45c8c..14fcfb93 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -1,6 +1,6 @@
 @static if VERSION < v"1.11" && get(ENV, "BUILDKITE_PIPELINE_NAME", "Metal.jl") == "Metal.jl"
     using Pkg
-    Pkg.add(url="https://github.com/JuliaGPU/KernelAbstractions.jl", rev="main")
+    Pkg.add(url = "https://github.com/JuliaGPU/KernelAbstractions.jl", rev = "main")
 end
 
 using Metal

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: 2d2a094	Previous: 7c5b7d4	Ratio
array/accumulate/Float32/1d	404167 ns	405125 ns	1.00
array/accumulate/Float32/dims=1	378125 ns	380709 ns	0.99
array/accumulate/Float32/dims=1L	8801292 ns	8813708 ns	1.00
array/accumulate/Float32/dims=2	439792 ns	449833 ns	0.98
array/accumulate/Float32/dims=2L	2633125 ns	2672417 ns	0.99
array/accumulate/Int64/1d	914792 ns	857042 ns	1.07
array/accumulate/Int64/dims=1	849584 ns	934334 ns	0.91
array/accumulate/Int64/dims=1L	9677125 ns	9530500 ns	1.02
array/accumulate/Int64/dims=2	1390584 ns	1253708 ns	1.11
array/accumulate/Int64/dims=2L	6523083 ns	6539791 ns	1.00
array/broadcast	210792 ns	186583 ns	1.13
array/construct	2166 ns	2208 ns	0.98
array/permutedims/2d	672167 ns	545333 ns	1.23
array/permutedims/3d	1237916 ns	1066167 ns	1.16
array/permutedims/4d	1427709 ns	1252917 ns	1.14
array/private/copy	239500 ns	235208 ns	1.02
array/private/copyto!/cpu_to_gpu	195500 ns	199791 ns	0.98
array/private/copyto!/gpu_to_cpu	196250 ns	199709 ns	0.98
array/private/copyto!/gpu_to_gpu	197584 ns	196458 ns	1.01
array/private/iteration/findall/bool	1124333 ns	1075000 ns	1.05
array/private/iteration/findall/int	1281500 ns	1238042 ns	1.04
array/private/iteration/findfirst/bool	1182208 ns	1101750 ns	1.07
array/private/iteration/findfirst/int	1165625 ns	1119041 ns	1.04
array/private/iteration/findmin/1d	1295500 ns	1239875 ns	1.04
array/private/iteration/findmin/2d	1091417 ns	1086292 ns	1.00
array/private/iteration/logical	1808459 ns	1752250 ns	1.03
array/private/iteration/scalar	1175500 ns	1229458 ns	0.96
array/random/rand/Float32	429500 ns	430042 ns	1.00
array/random/rand/Int64	519875 ns	526708 ns	0.99
array/random/rand!/Float32	320458 ns	370666 ns	0.86
array/random/rand!/Int64	316584 ns	401208 ns	0.79
array/random/randn/Float32	388458 ns	388208 ns	1.00
array/random/randn!/Float32	334958 ns	322500 ns	1.04
array/reductions/mapreduce/Float32/1d	400875 ns	410500 ns	0.98
array/reductions/mapreduce/Float32/dims=1	331584 ns	297000 ns	1.12
array/reductions/mapreduce/Float32/dims=1L	625709 ns	635084 ns	0.99
array/reductions/mapreduce/Float32/dims=2	370750 ns	332959 ns	1.11
array/reductions/mapreduce/Float32/dims=2L	1033833 ns	1023375 ns	1.01
array/reductions/mapreduce/Int64/1d	577292 ns	586875 ns	0.98
array/reductions/mapreduce/Int64/dims=1	597625 ns	620583 ns	0.96
array/reductions/mapreduce/Int64/dims=1L	1040250 ns	1044000 ns	1.00
array/reductions/mapreduce/Int64/dims=2	725625 ns	541875 ns	1.34
array/reductions/mapreduce/Int64/dims=2L	2257500 ns	2194542 ns	1.03
array/reductions/reduce/Float32/1d	406250 ns	405708 ns	1.00
array/reductions/reduce/Float32/dims=1	333542 ns	283625 ns	1.18
array/reductions/reduce/Float32/dims=1L	630416 ns	638291 ns	0.99
array/reductions/reduce/Float32/dims=2	226709 ns	226208 ns	1.00
array/reductions/reduce/Float32/dims=2L	457708 ns	470917 ns	0.97
array/reductions/reduce/Int64/1d	575834 ns	543209 ns	1.06
array/reductions/reduce/Int64/dims=1	601542 ns	633083 ns	0.95
array/reductions/reduce/Int64/dims=1L	1038167 ns	1033417 ns	1.00
array/reductions/reduce/Int64/dims=2	244792 ns	233958 ns	1.05
array/reductions/reduce/Int64/dims=2L	652083 ns	662833 ns	0.98
array/shared/copy	135292 ns	135666 ns	1.00
array/shared/copyto!/cpu_to_gpu	37958 ns	37250 ns	1.02
array/shared/copyto!/gpu_to_cpu	38333 ns	37500 ns	1.02
array/shared/copyto!/gpu_to_gpu	38625 ns	38375 ns	1.01
array/shared/iteration/findall/bool	1150416 ns	1091042 ns	1.05
array/shared/iteration/findall/int	1293833 ns	1249125 ns	1.04
array/shared/iteration/findfirst/bool	855208 ns	902250 ns	0.95
array/shared/iteration/findfirst/int	872750 ns	977458 ns	0.89
array/shared/iteration/findmin/1d	1162584 ns	866167 ns	1.34
array/shared/iteration/findmin/2d	1091625 ns	791583 ns	1.38
array/shared/iteration/logical	1679250 ns	1582166 ns	1.06
array/shared/iteration/scalar	3593.75 ns	3890.625 ns	0.92
array/sorting/1d	1785375 ns	1624625 ns	1.10
array/sorting/2d	8309958 ns	8289250 ns	1.00
integration/byval/reference	1112209 ns	1110334 ns	1.00
integration/byval/slices=1	1113333 ns	1116542 ns	1.00
integration/byval/slices=2	2032250 ns	2020917 ns	1.01
integration/byval/slices=3	6877709 ns	9519250 ns	0.72
integration/metaldevrt	374875 ns	312708 ns	1.20
kernel/indexing	182917 ns	191000 ns	0.96
kernel/indexing_checked	351667 ns	366959 ns	0.96
kernel/launch	1895.9 ns	1850 ns	1.02
kernel/rand	375208 ns	377666 ns	0.99
latency/import	2125140000 ns	2163364250 ns	0.98
latency/precompile	54332746125 ns	42157958500 ns	1.29
latency/ttfp	2562425125 ns	2484808458 ns	1.03
metal/synchronization/context	644.2095808383234 ns	665.6139240506329 ns	0.97
metal/synchronization/stream	335.6090909090909 ns	338.27853881278537 ns	0.99

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