Adding extension fix and allocate

JuliaGPU · Jan 29, 2024 · b82dfb4 · b82dfb4
1 parent 9dd121d
commit b82dfb4
Show file tree

Hide file tree

Showing 7 changed files with 55 additions and 45 deletions.
diff --git a/ext/CUDAEnzymeExt.jl b/ext/CUDAEnzymeExt.jl
@@ -9,6 +9,8 @@ module CUDAEnzymeExt
     import KernelAbstractions: Kernel, StaticSize, launch_config, __groupsize, __groupindex, blocks, mkcontext, CompilerMetadata, CPU, GPU
     using CUDA
 
+    include("enzyme_utils.jl")
+
     function EnzymeRules.augmented_primal(config::Config, func::Const{<:Kernel{CUDABackend}}, type::Type{Const{Nothing}}, args::Vararg{Any, N}; ndrange=nothing, workgroupsize=nothing) where N
         println("Custom rule GPU")
         kernel = func.val
@@ -52,8 +54,8 @@ module CUDAEnzymeExt
         TapeType = EnzymeCore.tape_type(job, ReverseSplitModified(ReverseSplitWithPrimal, ModifiedBetween), FT, Const,  Const{ctxTy}, map(Core.Typeof, args2)...)
         @show TapeType
 
-
-        subtape = Array{TapeType}(undef, size(blocks(iterspace)))
+        subtape = allocate(CUDABackend(), TapeType, size(blocks(iterspace)))
+        # subtape = Array{TapeType}(undef, size(blocks(iterspace)))
 
         aug_kernel = similar(kernel, aug_fwd)
 
@@ -67,4 +69,4 @@ module CUDAEnzymeExt
         return res
     end
 
-end
+end
diff --git a/ext/EnzymeExt.jl b/ext/EnzymeExt.jl
@@ -8,25 +8,9 @@ module EnzymeExt
     end
     import KernelAbstractions: Kernel, StaticSize, launch_config, __groupsize, __groupindex, blocks, mkcontext, CompilerMetadata, CPU, GPU
 
-    EnzymeRules.inactive(::Type{StaticSize}, x...) = nothing
-
-    function fwd(ctx, f, args...)
-        EnzymeCore.autodiff_deferred(Forward, Const(f), Const, Const(ctx), args...)
-        return nothing
-    end
+    include("enzyme_utils.jl")
 
-    function aug_fwd(ctx, f::FT, ::Val{ModifiedBetween}, subtape, args...) where {ModifiedBetween, FT}
-        forward, reverse = EnzymeCore.autodiff_deferred_thunk(ReverseSplitModified(ReverseSplitWithPrimal, Val(ModifiedBetween)), Const{Core.Typeof(f)}, Const, Const{Core.Typeof(ctx)}, map(Core.Typeof, args)...)
-        subtape[__groupindex(ctx)] = forward(Const(f), Const(ctx), args...)[1]
-        return nothing
-    end
-
-    function rev(ctx, f::FT, ::Val{ModifiedBetween}, subtape, args...) where {ModifiedBetween, FT}
-        forward, reverse = EnzymeCore.autodiff_deferred_thunk(ReverseSplitModified(ReverseSplitWithPrimal, Val(ModifiedBetween)), Const{Core.Typeof(f)}, Const, Const{Core.Typeof(ctx)}, map(Core.Typeof, args)...)
-        tp = subtape[__groupindex(ctx)]
-        reverse(Const(f), Const(ctx), args..., tp)
-        return nothing
-    end
+    EnzymeRules.inactive(::Type{StaticSize}, x...) = nothing
 
     function EnzymeRules.forward(func::Const{<:Kernel}, ::Type{Const{Nothing}}, args...; ndrange=nothing, workgroupsize=nothing)
         kernel = func.val
@@ -36,24 +20,6 @@ module EnzymeExt
         fwd_kernel(f, args...; ndrange, workgroupsize)
     end
 
-    @inline function make_active_byref(f::F, ::Val{ActiveTys}) where {F, ActiveTys}
-        if !any(ActiveTys)
-            return f
-        end
-        function inact(ctx, args2::Vararg{Any, N}) where N
-            args3 = ntuple(Val(N)) do i
-                Base.@_inline_meta
-                if ActiveTys[i]
-                    args2[i][]
-                else
-                    args2[i]
-                end
-            end
-            f(ctx, args3...)
-        end
-        return inact
-    end
-
     function EnzymeRules.augmented_primal(config::Config, func::Const{<:Kernel{CPU}}, ::Type{Const{Nothing}}, args::Vararg{Any, N}; ndrange=nothing, workgroupsize=nothing) where N
         println("Custom rule CPU")
         kernel = func.val

diff --git a/ext/enzyme_utils.jl b/ext/enzyme_utils.jl
@@ -0,0 +1,35 @@
+function fwd(ctx, f, args...)
+    EnzymeCore.autodiff_deferred(Forward, Const(f), Const, Const(ctx), args...)
+    return nothing
+end
+
+function aug_fwd(ctx, f::FT, ::Val{ModifiedBetween}, subtape, args...) where {ModifiedBetween, FT}
+    forward, reverse = EnzymeCore.autodiff_deferred_thunk(ReverseSplitModified(ReverseSplitWithPrimal, Val(ModifiedBetween)), Const{Core.Typeof(f)}, Const, Const{Core.Typeof(ctx)}, map(Core.Typeof, args)...)
+    subtape[__groupindex(ctx)] = forward(Const(f), Const(ctx), args...)[1]
+    return nothing
+end
+
+function rev(ctx, f::FT, ::Val{ModifiedBetween}, subtape, args...) where {ModifiedBetween, FT}
+    forward, reverse = EnzymeCore.autodiff_deferred_thunk(ReverseSplitModified(ReverseSplitWithPrimal, Val(ModifiedBetween)), Const{Core.Typeof(f)}, Const, Const{Core.Typeof(ctx)}, map(Core.Typeof, args)...)
+    tp = subtape[__groupindex(ctx)]
+    reverse(Const(f), Const(ctx), args..., tp)
+    return nothing
+end
+
+@inline function make_active_byref(f::F, ::Val{ActiveTys}) where {F, ActiveTys}
+    if !any(ActiveTys)
+        return f
+    end
+    function inact(ctx, args2::Vararg{Any, N}) where N
+        args3 = ntuple(Val(N)) do i
+            Base.@_inline_meta
+            if ActiveTys[i]
+                args2[i][]
+            else
+                args2[i]
+            end
+        end
+        f(ctx, args3...)
+    end
+    return inact
+end
diff --git a/test/Project.toml b/test/Project.toml
@@ -1,8 +1,10 @@
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
 Enzyme_jll = "7cc45869-7501-5eee-bdea-0790c847d4ef"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
+# KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"

diff --git a/test/extensions/enzyme.jl b/test/extensions/enzyme.jl
@@ -10,7 +10,6 @@ end
 function square_caller(A, backend)
     kernel = square!(backend)
     kernel(A, ndrange=size(A))
-    KernelAbstractions.synchronize(backend)
 end
 
 
@@ -22,7 +21,6 @@ end
 function mul_caller(A, B, backend)
     kernel = mul!(backend)
     kernel(A, B, ndrange=size(A))
-    KernelAbstractions.synchronize(backend)
 end
 
 function enzyme_testsuite(backend, ArrayT, supports_reverse=true)
@@ -36,13 +34,15 @@ function enzyme_testsuite(backend, ArrayT, supports_reverse=true)
             dA .= 1
 
             Enzyme.autodiff(Reverse, square_caller, Duplicated(A, dA), Const(backend()))
+            KernelAbstractions.synchronize(backend())
             @test all(dA .≈ (2:2:128))
 
 
             A .= (1:1:64)
             dA .= 1
 
             _, dB, _ = Enzyme.autodiff(Reverse, mul_caller, Duplicated(A, dA), Active(1.2), Const(backend()))[1]
+            KernelAbstractions.synchronize(backend())
 
             @test all(dA .≈ 1.2)
             @test dB ≈ sum(1:1:64)
@@ -52,6 +52,7 @@ function enzyme_testsuite(backend, ArrayT, supports_reverse=true)
         dA .= 1
 
         Enzyme.autodiff(Forward, square_caller, Duplicated(A, dA), Const(backend()))
+        KernelAbstractions.synchronize(backend())
         @test all(dA .≈ 2:2:128)
 
     end

diff --git a/test/reverse_gpu.jl b/test/reverse_gpu.jl
@@ -11,7 +11,6 @@ end
 function square_caller(A, backend)
     kernel = square!(backend)
     kernel(A, ndrange=size(A))
-    KernelAbstractions.synchronize(backend)
 end
 
 
@@ -23,7 +22,6 @@ end
 function mul_caller(A, B, backend)
     kernel = mul!(backend)
     kernel(A, B, ndrange=size(A))
-    KernelAbstractions.synchronize(backend)
 end
 
 function enzyme_testsuite(backend, ArrayT, supports_reverse=true)
@@ -37,13 +35,15 @@ function enzyme_testsuite(backend, ArrayT, supports_reverse=true)
             dA .= 1
 
             Enzyme.autodiff(Reverse, square_caller, Duplicated(A, dA), Const(backend()))
+            KernelAbstractions.synchronize(backend())
             @test all(dA .≈ (2:2:128))
 
 
             A .= (1:1:64)
             dA .= 1
 
             _, dB, _ = Enzyme.autodiff(Reverse, mul_caller, Duplicated(A, dA), Active(1.2), Const(backend()))[1]
+            KernelAbstractions.synchronize(backend())
 
             @test all(dA .≈ 1.2)
             @test dB ≈ sum(1:1:64)
@@ -53,11 +53,12 @@ function enzyme_testsuite(backend, ArrayT, supports_reverse=true)
         dA .= 1
 
         Enzyme.autodiff(Forward, square_caller, Duplicated(A, dA), Const(backend()))
+        KernelAbstractions.synchronize(backend())
         @test all(dA .≈ 2:2:128)
 
     end
 end
 
 # enzyme_testsuite(CPU, Array, true)
-enzyme_testsuite(CUDABackend, CuArray, false)
-# enzyme_testsuite(CUDABackend, CuArray, true)
+# enzyme_testsuite(CUDABackend, CuArray, false)
+enzyme_testsuite(CUDABackend, CuArray, true)
diff --git a/test/runtests.jl b/test/runtests.jl
@@ -1,3 +1,4 @@
+using CUDA
 using KernelAbstractions
 using Test
 
@@ -74,5 +75,7 @@ include("extensions/enzyme.jl")
 @static if VERSION >= v"1.7.0"
     @testset "Enzyme" begin
         enzyme_testsuite(CPU, Array)
+        enzyme_testsuite(CUDABackend, CuArray, false)
+        # enzyme_testsuite(CUDABackend, CuArray, true)
     end
 end