Add CUDAEnzyme extension

JuliaGPU · Jan 29, 2024 · 9dd121d · 9dd121d
1 parent 26a72d9
commit 9dd121d
Show file tree

Hide file tree

Showing 5 changed files with 73 additions and 62 deletions.
diff --git a/Project.toml b/Project.toml
@@ -6,7 +6,6 @@ version = "0.9.16"
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 Atomix = "a9b6321e-bd34-4604-b9c9-b65b8de01458"
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
@@ -20,9 +19,11 @@ UnsafeAtomicsLLVM = "d80eeb9a-aca5-4d75-85e5-170c8b632249"
 
 [weakdeps]
 EnzymeCore = "f151be2c-9106-41f4-ab19-57ee4f262869"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 
 [extensions]
 EnzymeExt = "EnzymeCore"
+CUDAEnzymeExt = ["CUDA", "EnzymeCore"]
 
 [compat]
 Adapt = "0.4, 1.0, 2.0, 3.0, 4"
@@ -42,3 +43,4 @@ julia = "1.6"
 
 [extras]
 EnzymeCore = "f151be2c-9106-41f4-ab19-57ee4f262869"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
diff --git a/ext/CUDAEnzymeExt.jl b/ext/CUDAEnzymeExt.jl
@@ -0,0 +1,70 @@
+module CUDAEnzymeExt
+    if isdefined(Base, :get_extension)
+        using EnzymeCore
+        using EnzymeCore.EnzymeRules
+    else
+        using ..EnzymeCore
+        using ..EnzymeCore.EnzymeRules
+    end
+    import KernelAbstractions: Kernel, StaticSize, launch_config, __groupsize, __groupindex, blocks, mkcontext, CompilerMetadata, CPU, GPU
+    using CUDA
+
+    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
+        f = kernel.f
+        mi = CUDA.methodinstance(typeof(()->return), Tuple{})
+        job = CUDA.CompilerJob(mi, CUDA.compiler_config(device()))
+
+        ndrange, workgroupsize, iterspace, dynamic = launch_config(kernel, ndrange, workgroupsize)
+        block = first(blocks(iterspace))
+        ctx = mkcontext(kernel, ndrange, iterspace)
+        ctxTy = Core.Typeof(ctx) # CompilerMetadata{ndrange(kernel), Core.Typeof(dynamic)}
+
+        # TODO autodiff_deferred on the func.val
+        ModifiedBetween = Val((overwritten(config)[1], false, overwritten(config)[2:end]...))
+
+        tup = Val(ntuple(Val(N)) do i
+            Base.@_inline_meta
+            args[i] isa Active
+        end)
+        f = make_active_byref(f, tup)
+        FT = Const{Core.Typeof(f)}
+
+        arg_refs = ntuple(Val(N)) do i
+            Base.@_inline_meta
+            if args[i] isa Active
+                Ref(EnzymeCore.make_zero(args[i].val))
+            else
+                nothing
+            end
+        end
+        args2 = ntuple(Val(N)) do i
+            Base.@_inline_meta
+            if args[i] isa Active
+                Duplicated(Ref(args[i].val), arg_refs[i])
+            else
+                args[i]
+            end
+        end
+
+        # TODO in KA backends like CUDAKernels, etc have a version with a parent job type
+        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)))
+
+        aug_kernel = similar(kernel, aug_fwd)
+
+        aug_kernel(f, ModifiedBetween, subtape, args2...; ndrange, workgroupsize)
+
+        # TODO the fact that ctxTy is type unstable means this is all type unstable.
+        # Since custom rules require a fixed return type, explicitly cast to Any, rather
+        # than returning a AugmentedReturn{Nothing, Nothing, T} where T.
+
+        res =  AugmentedReturn{Nothing, Nothing, Tuple{Array, typeof(arg_refs)}}(nothing, nothing, (subtape, arg_refs))
+        return res
+    end
+
+end
diff --git a/ext/EnzymeExt.jl b/ext/EnzymeExt.jl
@@ -7,7 +7,6 @@ module EnzymeExt
         using ..EnzymeCore.EnzymeRules
     end
     import KernelAbstractions: Kernel, StaticSize, launch_config, __groupsize, __groupindex, blocks, mkcontext, CompilerMetadata, CPU, GPU
-    using CUDA
 
     EnzymeRules.inactive(::Type{StaticSize}, x...) = nothing
 
@@ -112,64 +111,6 @@ module EnzymeExt
         return res
     end
 
-    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
-        f = kernel.f
-        mi = CUDA.methodinstance(typeof(()->return), Tuple{})
-        job = CUDA.CompilerJob(mi, CUDA.compiler_config(device()))
-
-        ndrange, workgroupsize, iterspace, dynamic = launch_config(kernel, ndrange, workgroupsize)
-        block = first(blocks(iterspace))
-        ctx = mkcontext(kernel, ndrange, iterspace)
-        ctxTy = Core.Typeof(ctx) # CompilerMetadata{ndrange(kernel), Core.Typeof(dynamic)}
-
-        # TODO autodiff_deferred on the func.val
-        ModifiedBetween = Val((overwritten(config)[1], false, overwritten(config)[2:end]...))
-
-        tup = Val(ntuple(Val(N)) do i
-            Base.@_inline_meta
-            args[i] isa Active
-        end)
-        f = make_active_byref(f, tup)
-        FT = Const{Core.Typeof(f)}
-
-        arg_refs = ntuple(Val(N)) do i
-            Base.@_inline_meta
-            if args[i] isa Active
-                Ref(EnzymeCore.make_zero(args[i].val))
-            else
-                nothing
-            end
-        end
-        args2 = ntuple(Val(N)) do i
-            Base.@_inline_meta
-            if args[i] isa Active
-                Duplicated(Ref(args[i].val), arg_refs[i])
-            else
-                args[i]
-            end
-        end
-
-        # TODO in KA backends like CUDAKernels, etc have a version with a parent job type
-        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)))
-
-        aug_kernel = similar(kernel, aug_fwd)
-
-        aug_kernel(f, ModifiedBetween, subtape, args2...; ndrange, workgroupsize)
-
-        # TODO the fact that ctxTy is type unstable means this is all type unstable.
-        # Since custom rules require a fixed return type, explicitly cast to Any, rather
-        # than returning a AugmentedReturn{Nothing, Nothing, T} where T.
-
-        res =  AugmentedReturn{Nothing, Nothing, Tuple{Array, typeof(arg_refs)}}(nothing, nothing, (subtape, arg_refs))
-        return res
-    end
-
     function EnzymeRules.reverse(config::Config, func::Const{<:Kernel}, ::Type{<:EnzymeCore.Annotation}, tape, args::Vararg{Any, N}; ndrange=nothing, workgroupsize=nothing) where N
         subtape, arg_refs = tape
 

diff --git a/test/Project.toml b/test/Project.toml
@@ -3,7 +3,6 @@ Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 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/reverse_gpu.jl b/test/reverse_gpu.jl
@@ -1,5 +1,4 @@
 using Test
-using Enzyme_jll
 using Enzyme
 using KernelAbstractions
 using CUDA