[Do not merge] Test KernelIntrinsics

[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 72ced3ed..e90f5826 100644
--- a/src/broadcast.jl
+++ b/src/broadcast.jl
@@ -66,8 +66,8 @@ 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
+            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]
@@ -91,8 +91,8 @@ 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
+            i = KI.get_global_id().x
+            stride = KI.get_global_size().x
              while 1 <= i <= length(dest)
                  @inbounds dest[i] = bc[i]
                  i += stride
@@ -150,8 +150,8 @@ end
     else
         ## COV_EXCL_START
         function broadcast_cartesian(dest, bc)
-             i = KI.get_global_id().x
-             stride = KI.get_global_size().x
+            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]
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 221ee680..6f9d5a2c 100644
--- a/test/kernelabstractions.jl
+++ b/test/kernelabstractions.jl
@@ -7,6 +7,6 @@ Testsuite.testsuite(()->MetalBackend(), "Metal", Metal, MtlArray, Metal.MtlDevic
     "Convert",           # depends on https://github.com/JuliaGPU/Metal.jl/issues/69
     "SpecialFunctions",  # no equivalent Metal intrinsics for gamma, erf, etc
     "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 6805205e..99ccadac 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -1,6 +1,6 @@
 @static if VERSION < v"1.11"
     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

Benchmark suite	Current: ce67b4c	Previous: 1232d45	Ratio
latency/precompile	29018226583 ns	24992031042 ns	1.16
latency/ttfp	2316989604.5 ns	2131427167 ns	1.09
latency/import	1389818917 ns	1225716666 ns	1.13
integration/metaldevrt	848875 ns	834458 ns	1.02
integration/byval/slices=1	1598042 ns	1547145.5 ns	1.03
integration/byval/slices=3	19072604 ns	8632292 ns	2.21
integration/byval/reference	1575125 ns	1525292 ns	1.03
integration/byval/slices=2	2686750 ns	2592229 ns	1.04
kernel/indexing	454500.5 ns	624250 ns	0.73
kernel/indexing_checked	507750 ns	613041 ns	0.83
kernel/launch	12666 ns	12000 ns	1.06
kernel/rand	532292 ns	560375 ns	0.95
array/construct	6375 ns	6375 ns	1
array/broadcast	529604 ns	596750 ns	0.89
array/random/randn/Float32	895500 ns	814458.5 ns	1.10
array/random/randn!/Float32	578416 ns	620208 ns	0.93
array/random/rand!/Int64	538333 ns	552375 ns	0.97
array/random/rand!/Float32	532812.5 ns	583750 ns	0.91
array/random/rand/Int64	911604.5 ns	756208.5 ns	1.21
array/random/rand/Float32	850145.5 ns	583604 ns	1.46
array/accumulate/Int64/1d	1374375 ns	1252709 ns	1.10
array/accumulate/Int64/dims=1	1963500 ns	1815125 ns	1.08
array/accumulate/Int64/dims=2	2340750 ns	2160708 ns	1.08
array/accumulate/Int64/dims=1L	12117709 ns	11656187 ns	1.04
array/accumulate/Int64/dims=2L	10014854 ns	9645833 ns	1.04
array/accumulate/Float32/1d	1114125 ns	1067958 ns	1.04
array/accumulate/Float32/dims=1	1718459 ns	1541250 ns	1.11
array/accumulate/Float32/dims=2	2159917 ns	1840146 ns	1.17
array/accumulate/Float32/dims=1L	10564334 ns	9832792 ns	1.07
array/accumulate/Float32/dims=2L	7796354 ns	7218458.5 ns	1.08
array/reductions/reduce/Int64/1d	1251833 ns	1536834 ns	0.81
array/reductions/reduce/Int64/dims=1	1114500 ns	1068916 ns	1.04
array/reductions/reduce/Int64/dims=2	1346250 ns	1114812.5 ns	1.21
array/reductions/reduce/Int64/dims=1L	2059959 ns	2015792 ns	1.02
array/reductions/reduce/Int64/dims=2L	4082500 ns	4220062.5 ns	0.97
array/reductions/reduce/Float32/1d	757125 ns	1050021 ns	0.72
array/reductions/reduce/Float32/dims=1	832104 ns	814312.5 ns	1.02
array/reductions/reduce/Float32/dims=2	861792 ns	840000 ns	1.03
array/reductions/reduce/Float32/dims=1L	1371458 ns	1300250 ns	1.05
array/reductions/reduce/Float32/dims=2L	1883979.5 ns	1785812.5 ns	1.05
array/reductions/mapreduce/Int64/1d	1241917 ns	1544645.5 ns	0.80
array/reductions/mapreduce/Int64/dims=1	1102729 ns	1075416 ns	1.03
array/reductions/mapreduce/Int64/dims=2	1347375 ns	1124750 ns	1.20
array/reductions/mapreduce/Int64/dims=1L	1949020.5 ns	2014958 ns	0.97
array/reductions/mapreduce/Int64/dims=2L	3882666 ns	3606874.5 ns	1.08
array/reductions/mapreduce/Float32/1d	795187 ns	1030000 ns	0.77
array/reductions/mapreduce/Float32/dims=1	813042 ns	804459 ns	1.01
array/reductions/mapreduce/Float32/dims=2	875063 ns	816145.5 ns	1.07
array/reductions/mapreduce/Float32/dims=1L	1374708 ns	1307375 ns	1.05
array/reductions/mapreduce/Float32/dims=2L	1882291 ns	1894042 ns	0.99
array/private/copyto!/gpu_to_gpu	535604.5 ns	641459 ns	0.83
array/private/copyto!/cpu_to_gpu	742062 ns	784042 ns	0.95
array/private/copyto!/gpu_to_cpu	728375 ns	798458 ns	0.91
array/private/iteration/findall/int	1649042 ns	1556333 ns	1.06
array/private/iteration/findall/bool	1563770.5 ns	1422437.5 ns	1.10
array/private/iteration/findfirst/int	2163500 ns	2047375 ns	1.06
array/private/iteration/findfirst/bool	2084958 ns	2034375 ns	1.02
array/private/iteration/scalar	2989917 ns	3970666 ns	0.75
array/private/iteration/logical	2755750 ns	2580125 ns	1.07
array/private/iteration/findmin/1d	2333333 ns	2224833 ns	1.05
array/private/iteration/findmin/2d	1591875 ns	1504083 ns	1.06
array/private/copy	806438 ns	579459 ns	1.39
array/shared/copyto!/gpu_to_gpu	85209 ns	85625 ns	1.00
array/shared/copyto!/cpu_to_gpu	82291.5 ns	82000 ns	1.00
array/shared/copyto!/gpu_to_cpu	83167 ns	84417 ns	0.99
array/shared/iteration/findall/int	1666417 ns	1572333 ns	1.06
array/shared/iteration/findall/bool	1568541.5 ns	1439166.5 ns	1.09
array/shared/iteration/findfirst/int	1765645.5 ns	1659750 ns	1.06
array/shared/iteration/findfirst/bool	1710375 ns	1615833 ns	1.06
array/shared/iteration/scalar	204625 ns	206709 ns	0.99
array/shared/iteration/logical	2561208 ns	2254958 ns	1.14
array/shared/iteration/findmin/1d	1959417 ns	1802791 ns	1.09
array/shared/iteration/findmin/2d	1586271 ns	1506104.5 ns	1.05
array/shared/copy	210042 ns	251209 ns	0.84
array/permutedims/4d	3132333 ns	2351833 ns	1.33
array/permutedims/2d	1176417 ns	1144750.5 ns	1.03
array/permutedims/3d	1820208 ns	1650917 ns	1.10
metal/synchronization/stream	19334 ns	19209 ns	1.01
metal/synchronization/context	19750 ns	20125 ns	0.98

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