Use KernelAbstractions.jl for gather/scatter kernels (#487)

* Use KA for gather

* Finish scatter

* Update testsuite for gather/scatter

* Cleanup

* Update tests

* Retain NNlibCUDA scatter kernels

* Fixup

* Add at-inbounds

* Add compat

* Use KA unsafe free

Project.toml

@@ -4,7 +4,9 @@ version = "0.8.19"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+Atomix = "a9b6321e-bd34-4604-b9c9-b65b8de01458"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
+GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
@@ -21,8 +23,10 @@ NNlibAMDGPUExt = "AMDGPU"
 [compat]
 AMDGPU = "0.4.8"
 Adapt = "2, 3.2"
+Atomix = "0.1"
 ChainRulesCore = "1.13"
-KernelAbstractions = "0.9"
+GPUArraysCore = "0.1"
+KernelAbstractions = "0.9.2"
 Requires = "0.5, 1.0"
 julia = "1.6"
 

ext/NNlibCUDA/src/NNlibCUDA.jl

@@ -13,7 +13,6 @@ include("batchedmul.jl")
 include("ctc.jl")
 include("fold.jl")
 include("scatter.jl")
-include("gather.jl")
 include("utils.jl")
 include("cudnn/cudnn.jl")
 include("cudnn/conv.jl")

ext/NNlibCUDA/src/scatter.jl

@@ -30,7 +30,7 @@ function scatter_kernel!(op::OP, dst, src, idx, max_idx, max_dims_idx, dims_size
     return nothing
 end
 
-function scatter_kernel!(op::OP, dst, src, idx::CUDA.CuDeviceArray{<:CartesianIndex}, 
+function scatter_kernel!(op::OP, dst, src, idx::CUDA.CuDeviceArray{<:CartesianIndex},
             max_idx, max_dims_idx, dims_size) where OP
     index = threadIdx().x + (blockIdx().x - 1) * blockDim().x
 
@@ -73,7 +73,7 @@ end
 
 ## Gradients
 
-function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx, 
+function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx,
                 rev_idx, max_idx, T::Type{TT})  where {OP,TT}
     index = threadIdx().x + (blockIdx().x - 1) * blockDim().x
 
@@ -93,7 +93,7 @@ function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx,
     return nothing
 end
 
-function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:CartesianIndex}, 
+function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:CartesianIndex},
                 rev_idx, max_idx, T::Type{TT}) where {OP,TT}
     index = threadIdx().x + (blockIdx().x - 1) * blockDim().x
 
@@ -113,7 +113,7 @@ function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:Ca
     return nothing
 end
 
-function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx, 
+function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx,
             rev_idx, pre_cart_idx, max_dims_idx, max_idx, T::Type{TT}) where {OP,TT}
     index = threadIdx().x + (blockIdx().x - 1) * blockDim().x
 
@@ -160,13 +160,13 @@ function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:Ca
 end
 
 function NNlib.∇scatter_src(op::Union{typeof(*),typeof(/)}, Δ, dst,
-                            src::AnyCuArray{Tsrc,Nsrc}, 
+                            src::AnyCuArray{Tsrc,Nsrc},
                             idx::AnyCuArray{Tidx,Nidx}) where {Tsrc,Tidx,Nsrc,Nidx}
     dims = Nsrc - Nidx
     Δsrc = NNlib.modify_src(op, NNlib.gather(Δ, idx), src)
     rev_idx = NNlib.reverse_indices(idx)
     rev_idx = CuArray(map(CUDA.cudaconvert, rev_idx))
-    
+
     if dims == 0
         max_idx = length(idx)
         args = op, Δsrc, src, idx, rev_idx, max_idx, Tsrc

ext/NNlibCUDA/test/gather.jl

@@ -1,7 +1,7 @@
-@testset "gather" begin 
+@testset "gather" begin
     T = Float32
     CT = CuArray{Float32}
-    
+
     ## 1d src, 2d index of ints -> 2d output
     src = CT([3, 4, 5, 6, 7])
     index = cu([1 2 3 4;
@@ -10,14 +10,14 @@
     output = CT([3 4 5 6;
                 6 4 3 5;
                 5 7 7 5])
-    
+
     y = NNlib.gather(src, index)
     @test y isa CuArray{Float32,2}
     @test size(y) == size(index)
     gputest(src -> NNlib.gather(src, index), src, checkgrad=true)
     @test NNlib.gather!(CUDA.zeros(T, size(index)...), src, index) == output
     @test_throws ArgumentError NNlib.gather!(zeros(T, 3, 5), src, index)
-    
+
     ## 1d src, 2d index of tuples -> 2d output
     src = CT([3, 4, 5, 6, 7])
     index = cu([(1,) (2,) (3,) (4,);
@@ -26,14 +26,14 @@
     output = CT([3 4 5 6;
                 6 4 3 5;
                 5 7 7 5])
-    
+
     y = NNlib.gather(src, index)
     @test y isa CuArray{Float32,2}
     @test size(y) == size(index)
     gputest(src -> NNlib.gather(src, index), src, checkgrad=true)
     @test NNlib.gather!(CUDA.zeros(T, size(index)...), src, index) == output
     @test_throws ArgumentError NNlib.gather!(zeros(T, 3, 5), src, index)
-    
+
     ## 1d src, 2d index of CartesianIndex -> 2d output
     src = CT([3, 4, 5, 6, 7])
     index = cu(CartesianIndex.([(1,) (2,) (3,) (4,);
@@ -42,7 +42,7 @@
     output = CT([3 4 5 6;
                 6 4 3 5;
                 5 7 7 5])
-    
+
     y = NNlib.gather(src, index)
     @test y isa CuArray{Float32,2}
     @test size(y) == size(index)
@@ -66,7 +66,7 @@
 
 
     ## 2d src, 2d index of ints -> 3d output
-    src = CT([3 5 7 
+    src = CT([3 5 7
              4 6 8])
     index = cu([1 2 3;
                 2 2 1;
@@ -79,14 +79,14 @@
 
     output[:,:,2] = [5 5 3
                     6 6 4]
-        
+
     output[:,:,3] = [7 3 7
                     8 4 8]
-              
+
     y = NNlib.gather(src, index)
     M = NNlib.typelength(eltype(index))
     Nsrc = ndims(src)
     @test y isa CuArray{Float32,3}
-    @test size(y) == (size(src)[1:Nsrc-M]..., size(index)...) 
+    @test size(y) == (size(src)[1:Nsrc-M]..., size(index)...)
     gputest(src -> NNlib.gather(src, index), src, checkgrad=true)
 end

src/NNlib.jl

@@ -1,10 +1,12 @@
 module NNlib
 
+import Atomix
 import ChainRulesCore: rrule
 
 using Base.Broadcast: broadcasted
 using Base.Threads
 using ChainRulesCore
+using GPUArraysCore
 using KernelAbstractions
 using KernelAbstractions: @atomic
 using LinearAlgebra