Adapt to GPUCompiler#master.

Author: maleadt

Manifest.toml

@@ -2,7 +2,7 @@
 
 julia_version = "1.8.5"
 manifest_format = "2.0"
-project_hash = "e8fd324aba09b72a771344d3910c4e1284ae62a0"
+project_hash = "c38f3e404ea342fb740cb16012407cbac90360fd"
 
 [[deps.AbstractFFTs]]
 deps = ["ChainRulesCore", "LinearAlgebra", "Test"]
@@ -170,9 +170,11 @@ version = "0.1.5"
 
 [[deps.GPUCompiler]]
 deps = ["ExprTools", "InteractiveUtils", "LLVM", "Libdl", "Logging", "Scratch", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "8de395b1243771bbb79ac832ec96c7def7a4586f"
+git-tree-sha1 = "2bec2c336b65f1ca50c4fcb2933712503db8eb2d"
+repo-rev = "master"
+repo-url = "https://github.com/JuliaGPU/GPUCompiler.jl.git"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "0.22.0"
+version = "0.23.0"
 
 [[deps.InlineStrings]]
 deps = ["Parsers"]

Project.toml

@@ -31,10 +31,15 @@ RandomNumbers = "e6cf234a-135c-5ec9-84dd-332b85af5143"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
-SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 UnsafeAtomicsLLVM = "d80eeb9a-aca5-4d75-85e5-170c8b632249"
 
+[weakdeps]
+SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
+
+[extensions]
+SpecialFunctionsExt = "SpecialFunctions"
+
 [compat]
 AbstractFFTs = "0.4, 0.5, 1.0"
 Adapt = "3.3"
@@ -47,7 +52,7 @@ Crayons = "4"
 DataFrames = "1"
 ExprTools = "0.1"
 GPUArrays = "8.6"
-GPUCompiler = "0.22"
+GPUCompiler = "0.23"
 KernelAbstractions = "0.9.2"
 LLVM = "6"
 NVTX = "0.3.2"
@@ -60,3 +65,6 @@ Requires = "0.5, 1.0"
 SpecialFunctions = "1.3, 2"
 UnsafeAtomicsLLVM = "0.1"
 julia = "1.8"
+
+[extras]
+SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"

src/device/intrinsics/special_math.jl renamed to ext/SpecialFunctionsExt.jl

@@ -1,5 +1,12 @@
 # math functionality corresponding to SpecialFunctions.jl
 
+module SpecialFunctionsExt
+
+using CUDA
+using CUDA: @device_override
+
+isdefined(Base, :get_extension) ? (using SpecialFunctions) : (using ..SpecialFunctions)
+
 
 ## error
 
@@ -46,3 +53,6 @@
 
 @device_override SpecialFunctions.bessely(n::Int32, x::Float64) = ccall("extern __nv_yn", llvmcall, Cdouble, (Int32, Cdouble), n, x)
 @device_override SpecialFunctions.bessely(n::Int32, x::Float32) = ccall("extern __nv_ynf", llvmcall, Cfloat, (Int32, Cfloat), n, x)
+
+
+end

src/CUDAKernels.jl

@@ -157,7 +157,7 @@ macro device_override(ex)
         error()
     end
     code = quote
-        $GPUCompiler.@override($CUDA.method_table, $ex)
+        Base.Experimental.@overlay($CUDA.method_table, $ex)
     end
     if isdefined(Base.Experimental, Symbol("@overlay"))
         return esc(code)

src/device/utils.jl

@@ -13,24 +13,11 @@ else
 const method_table = nothing
 end
 
-# list of overrides (only for Julia 1.6)
-const overrides = Expr[]
-
 macro device_override(ex)
     ex = macroexpand(__module__, ex)
-    if Meta.isexpr(ex, :call)
-        @show ex = eval(ex)
-        error()
-    end
-    code = quote
-        $GPUCompiler.@override(CUDA.method_table, $ex)
-    end
-    if isdefined(Base.Experimental, Symbol("@overlay"))
-        return esc(code)
-    else
-        push!(overrides, code)
-        return
-    end
+    esc(quote
+        Base.Experimental.@overlay(CUDA.method_table, $ex)
+    end)
 end
 
 macro device_function(ex)

src/initialization.jl

@@ -147,15 +147,9 @@ function __init__()
         return
     end
 
-    # register device overrides
-    if !precompiling
-        eval(Expr(:block, overrides...))
-        empty!(overrides)
-
+    @static if !isdefined(Base, :get_extension)
         @require SpecialFunctions="276daf66-3868-5448-9aa4-cd146d93841b" begin
-            include("device/intrinsics/special_math.jl")
-            eval(Expr(:block, overrides...))
-            empty!(overrides)
+            include("../ext/SpecialFunctionsExt.jl")
         end
     end