Cleanup

Author: pxl-th

docs/Project.toml

@@ -1,6 +1,5 @@
 [deps]
 Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4"
-KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 
 [compat]
 Documenter = "1"

src/KernelAbstractions.jl

@@ -793,13 +793,13 @@ argconvert(k::Kernel{T}, arg) where {T} =
 supports_enzyme(::Backend) = false
 function __fake_compiler_job end
 
+include("groupreduction.jl")
+
 ###
 # Extras
 # - LoopInfo
 ###
 
-include("reduce.jl")
-
 include("extras/extras.jl")
 
 include("reflection.jl")

src/reduce.jl renamed to src/groupreduction.jl

@@ -4,7 +4,6 @@ export @groupreduce, @warp_groupreduce
     @groupreduce op val neutral [groupsize]
 
 Perform group reduction of `val` using `op`.
-If backend supports warp reduction, it will use it instead of thread reduction.
 
 # Arguments
 
@@ -27,13 +26,6 @@ macro groupreduce(op, val, groupsize)
     :(__thread_groupreduce($(esc(:__ctx__)), $(esc(op)), $(esc(val)), Val($(esc(groupsize)))))
 end
 
-macro warp_groupreduce(op, val, neutral)
-    :(__warp_groupreduce($(esc(:__ctx__)), $(esc(op)), $(esc(val)), $(esc(neutral)), Val(prod($groupsize($(esc(:__ctx__)))))))
-end
-macro warp_groupreduce(op, val, neutral, groupsize)
-    :(__warp_groupreduce($(esc(:__ctx__)), $(esc(op)), $(esc(val)), $(esc(neutral)), Val($(esc(groupsize)))))
-end
-
 function __thread_groupreduce(__ctx__, op, val::T, ::Val{groupsize}) where {T, groupsize}
     storage = @localmem T groupsize
 
@@ -61,10 +53,37 @@ end
 
 # Warp groupreduce.
 
-# NOTE: Backends should implement these two device functions (with `@device_override`).
+"""
+    @warp_groupreduce op val neutral [groupsize]
+
+Perform group reduction of `val` using `op`.
+Each warp within a workgroup performs its own reduction using [`shfl_down`](@ref) intrinsic,
+followed by final reduction over results of individual warp reductions.
+
+!!! note
+
+    Use [`supports_warp_reduction`](@ref) to query if given backend supports warp reduction.
+"""
+macro warp_groupreduce(op, val, neutral)
+    :(__warp_groupreduce($(esc(:__ctx__)), $(esc(op)), $(esc(val)), $(esc(neutral)), Val(prod($groupsize($(esc(:__ctx__)))))))
+end
+macro warp_groupreduce(op, val, neutral, groupsize)
+    :(__warp_groupreduce($(esc(:__ctx__)), $(esc(op)), $(esc(val)), $(esc(neutral)), Val($(esc(groupsize)))))
+end
+
+"""
+    shfl_down(val::T, offset::Integer)::T where T
+
+Read `val` from a lane with higher id given by `offset`.
+"""
 function shfl_down end
 supports_warp_reduction() = false
-# Host-variant.
+
+"""
+    supports_warp_reduction(::Backend)
+
+Query if given backend supports [`shfl_down`](@ref) intrinsic and thus warp reduction.
+"""
 supports_warp_reduction(::Backend) = false
 
 # Assume warp is 32 lanes.