Use contextual dispatch for device functions.

Author: maleadt

Manifest.toml

@@ -32,6 +32,12 @@ git-tree-sha1 = "215a9aa4a1f23fbd05b92769fdd62559488d70e9"
 uuid = "fa961155-64e5-5f13-b03f-caf6b980ea82"
 version = "0.4.1"
 
+[[ChainRulesCore]]
+deps = ["Compat", "LinearAlgebra", "SparseArrays"]
+git-tree-sha1 = "de4f08843c332d355852721adb1592bce7924da3"
+uuid = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
+version = "0.9.29"
+
 [[Compat]]
 deps = ["Base64", "Dates", "DelimitedFiles", "Distributed", "InteractiveUtils", "LibGit2", "Libdl", "LinearAlgebra", "Markdown", "Mmap", "Pkg", "Printf", "REPL", "Random", "SHA", "Serialization", "SharedArrays", "Sockets", "SparseArrays", "Statistics", "Test", "UUIDs", "Unicode"]
 git-tree-sha1 = "919c7f3151e79ff196add81d7f4e45d91bbf420b"
@@ -77,14 +83,21 @@ version = "6.2.0"
 
 [[GPUCompiler]]
 deps = ["DataStructures", "ExprTools", "InteractiveUtils", "LLVM", "Libdl", "Logging", "Scratch", "Serialization", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "ef2839b063e158672583b9c09d2cf4876a8d3d55"
+git-tree-sha1 = "b6c3b8e2df6ffe0da0b10e2045ce35a3cf618b8a"
+repo-rev = "1ecbe42"
+repo-url = "https://github.com/JuliaGPU/GPUCompiler.jl.git"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "0.10.0"
+version = "0.10.1"
 
 [[InteractiveUtils]]
 deps = ["Markdown"]
 uuid = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 
+[[JLLWrappers]]
+git-tree-sha1 = "a431f5f2ca3f4feef3bd7a5e94b8b8d4f2f647a0"
+uuid = "692b3bcd-3c85-4b1f-b108-f13ce0eb3210"
+version = "1.2.0"
+
 [[LLVM]]
 deps = ["CEnum", "Libdl", "Printf", "Unicode"]
 git-tree-sha1 = "b616937c31337576360cb9fb872ec7633af7b194"
@@ -150,6 +163,12 @@ uuid = "14a3606d-f60d-562e-9121-12d972cd8159"
 [[NetworkOptions]]
 uuid = "ca575930-c2e3-43a9-ace4-1e988b2c1908"
 
+[[OpenSpecFun_jll]]
+deps = ["Artifacts", "CompilerSupportLibraries_jll", "JLLWrappers", "Libdl", "Pkg"]
+git-tree-sha1 = "9db77584158d0ab52307f8c04f8e7c08ca76b5b3"
+uuid = "efe28fd5-8261-553b-a9e1-b2916fc3738e"
+version = "0.5.3+4"
+
 [[OrderedCollections]]
 git-tree-sha1 = "4fa2ba51070ec13fcc7517db714445b4ab986bdf"
 uuid = "bac558e1-5e72-5ebc-8fee-abe8a469f55d"
@@ -205,6 +224,12 @@ uuid = "6462fe0b-24de-5631-8697-dd941f90decc"
 deps = ["LinearAlgebra", "Random"]
 uuid = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 
+[[SpecialFunctions]]
+deps = ["ChainRulesCore", "OpenSpecFun_jll"]
+git-tree-sha1 = "5919936c0e92cff40e57d0ddf0ceb667d42e5902"
+uuid = "276daf66-3868-5448-9aa4-cd146d93841b"
+version = "1.3.0"
+
 [[Statistics]]
 deps = ["LinearAlgebra", "SparseArrays"]
 uuid = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"

Project.toml

@@ -24,6 +24,7 @@ Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 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"
 TimerOutputs = "a759f4b9-e2f1-59dc-863e-4aeb61b1ea8f"
 

src/CUDA.jl

@@ -18,6 +18,19 @@ using BFloat16s
 
 using Memoize
 
+using ExprTools
+
+
+##
+
+const ci_cache = GPUCompiler.CodeCache()
+
+@static if VERSION >= v"1.7-"
+Base.Experimental.@MethodTable(method_table)
+else
+const method_table = nothing
+end
+
 
 ## source code includes
 

src/accumulate.jl

@@ -134,8 +134,6 @@ function scan!(f::Function, output::AnyCuArray{T}, input::AnyCuArray;
     dims > ndims(input) && return copyto!(output, input)
     isempty(inds_t[dims]) && return output
 
-    f = cufunc(f)
-
     # iteration domain across the main dimension
     Rdim = CartesianIndices((size(input, dims),))
 

src/broadcast.jl

@@ -14,99 +14,6 @@ Base.similar(bc::Broadcasted{CuArrayStyle{N}}, ::Type{T}) where {N,T} =
 Base.similar(bc::Broadcasted{CuArrayStyle{N}}, ::Type{T}, dims) where {N,T} =
     CuArray{T}(undef, dims)
 
-
-## replace base functions with libdevice alternatives
-
-cufunc(f) = f
-cufunc(::Type{T}) where T = (x...) -> T(x...) # broadcasting type ctors isn't GPU compatible
-
-Broadcast.broadcasted(::CuArrayStyle{N}, f, args...) where {N} =
-  Broadcasted{CuArrayStyle{N}}(cufunc(f), args, nothing)
-
-const device_intrinsics = :[
-  cos, cospi, sin, sinpi, tan, acos, asin, atan,
-  cosh, sinh, tanh, acosh, asinh, atanh, angle,
-  log, log10, log1p, log2, logb, ilogb,
-  exp, exp2, exp10, expm1, ldexp,
-  erf, erfinv, erfc, erfcinv, erfcx,
-  brev, clz, ffs, byte_perm, popc,
-  isfinite, isinf, isnan, nearbyint,
-  nextafter, signbit, copysign, abs,
-  sqrt, rsqrt, cbrt, rcbrt, pow,
-  ceil, floor, saturate,
-  lgamma, tgamma,
-  j0, j1, jn, y0, y1, yn,
-  normcdf, normcdfinv, hypot,
-  fma, sad, dim, mul24, mul64hi, hadd, rhadd, scalbn].args
-
-for f in device_intrinsics
-  isdefined(Base, f) || continue
-  @eval cufunc(::typeof(Base.$f)) = $f
-end
-
-# broadcast ^
-
-culiteral_pow(::typeof(^), x::T, ::Val{0}) where {T<:Real} = one(x)
-culiteral_pow(::typeof(^), x::T, ::Val{1}) where {T<:Real} = x
-culiteral_pow(::typeof(^), x::T, ::Val{2}) where {T<:Real} = x * x
-culiteral_pow(::typeof(^), x::T, ::Val{3}) where {T<:Real} = x * x * x
-culiteral_pow(::typeof(^), x::T, ::Val{p}) where {T<:Real,p} = pow(x, Int32(p))
-
-cufunc(::typeof(Base.literal_pow)) = culiteral_pow
-cufunc(::typeof(Base.:(^))) = pow
-
-using MacroTools
-
-const _cufuncs = [copy(device_intrinsics); :^]
-cufuncs() = (global _cufuncs; _cufuncs)
-
-_cuint(x::Int) = Int32(x)
-_cuint(x::Expr) = x.head == :call && x.args[1] == :Int32 && x.args[2] isa Int ? Int32(x.args[2]) : x
-_cuint(x) = x
-
-function _cupowliteral(x::Expr)
-  if x.head == :call && x.args[1] == :(CUDA.cufunc(^)) && x.args[3] isa Int32
-    num = x.args[3]
-    if 0 <= num <= 3
-      sym = gensym(:x)
-      new_x = Expr(:block, :($sym = $(x.args[2])))
-
-      if iszero(num)
-        push!(new_x.args, :(one($sym)))
-      else
-        unroll = Expr(:call, :*)
-        for x = one(num):num
-          push!(unroll.args, sym)
-        end
-        push!(new_x.args, unroll)
-      end
-
-      x = new_x
-    end
-  end
-  x
-end
-_cupowliteral(x) = x
-
-function replace_device(ex)
-  global _cufuncs
-  MacroTools.postwalk(ex) do x
-    x = x in _cufuncs ? :(CUDA.cufunc($x)) : x
-    x = _cuint(x)
-    x = _cupowliteral(x)
-    x
-  end
-end
-
-macro cufunc(ex)
-  global _cufuncs
-  def = MacroTools.splitdef(ex)
-  f = def[:name]
-  def[:name] = Symbol(:cu, f)
-  def[:body] = replace_device(def[:body])
-  push!(_cufuncs, f)
-  quote
-    $(esc(MacroTools.combinedef(def)))
-    CUDA.cufunc(::typeof($(esc(f)))) = $(esc(def[:name]))
-  end
-end
+# broadcasting type ctors isn't GPU compatible
+Broadcast.broadcasted(::CuArrayStyle{N}, f::Type{T}, args...) where {N, T} =
+    Broadcasted{CuArrayStyle{N}}((x...) -> T(x...), args, nothing)

src/compiler/gpucompiler.jl

@@ -39,3 +39,7 @@ function GPUCompiler.link_libraries!(job::CUDACompilerJob, mod::LLVM.Module,
            job, mod, undefined_fns)
     link_libdevice!(mod, job.target.cap, undefined_fns)
 end
+
+GPUCompiler.ci_cache(::CUDACompilerJob) = ci_cache
+
+GPUCompiler.method_table(::CUDACompilerJob) = method_table

src/device/intrinsics.jl

@@ -1,5 +1,34 @@
 # wrappers for functionality provided by the CUDA toolkit
 
+const overrides = quote end
+
+macro device_override(ex)
+    code = quote
+        $GPUCompiler.@override($method_table, $ex)
+    end
+    if VERSION >= v"1.7-"
+        return esc(code)
+    else
+        push!(overrides.args, code)
+        return
+    end
+end
+
+macro device_function(ex)
+    ex = macroexpand(__module__, ex)
+    def = splitdef(ex)
+
+    # generate a function that errors
+    def[:body] = quote
+        error("This function is not intended for use on the CPU")
+    end
+
+    esc(quote
+        $(combinedef(def))
+        @device_override $ex
+    end)
+end
+
 # extensions to the C language
 include("intrinsics/memory_shared.jl")
 include("intrinsics/indexing.jl")