Try #334:

Project.toml

@@ -5,6 +5,7 @@ uuid = "be33ccc6-a3ff-5ff2-a52e-74243cff1e17"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 CUDAapi = "3895d2a7-ec45-59b8-82bb-cfc6a382f9b3"
 CUDAdrv = "c5f51814-7f29-56b8-a69c-e4d8f6be1fde"
+Cassette = "7057c7e9-c182-5462-911a-8362d720325c"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
@@ -20,3 +21,6 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
 test = ["Test", "BenchmarkTools", "SpecialFunctions"]
+
+[compat]
+julia = ">= 1.1"

REQUIRE

@@ -1,4 +1,4 @@
-julia 1.0
+julia 1.1
 CUDAdrv 1.1
 LLVM 0.9.14
 CUDAapi 0.4.0

src/CUDAnative.jl

@@ -23,6 +23,7 @@ end
 include("utils.jl")
 
 # needs to be loaded _before_ the compiler infrastructure, because of generated functions
+isdevice() = false
 include(joinpath("device", "tools.jl"))
 include(joinpath("device", "pointer.jl"))
 include(joinpath("device", "array.jl"))
@@ -31,6 +32,7 @@ include(joinpath("device", "cuda_intrinsics.jl"))
 include(joinpath("device", "runtime_intrinsics.jl"))
 
 include("compiler.jl")
+include("context.jl")
 include("execution.jl")
 include("reflection.jl")
 

src/context.jl

@@ -0,0 +1,52 @@
+##
+# Implements contextual dispatch through Cassette.jl
+# Goals:
+# - Rewrite common CPU functions to appropriate GPU intrinsics
+# 
+# TODO:
+# - error (erf, ...)
+# - pow
+# - min, max
+# - mod, rem
+# - gamma
+# - bessel
+# - distributions
+# - unsorted
+
+using Cassette
+
+function transform(ctx, ref)
+ ci = ref.code_info
+ noinline = any(@nospecialize(x) -> Core.Compiler.isexpr(x, :meta) && x.args[1] == :noinline, ci.code)
+ if !noinline
+ ci.inlineable = true
+ end
+ return ci
+end
+const InlinePass = Cassette.@pass transform
+
+Cassette.@context CUDACtx
+const cudactx = Cassette.disablehooks(CUDACtx(pass = InlinePass))
+
+Cassette.overdub(::CUDACtx, ::typeof(datatype_align), ::Type{T}) where {T} = datatype_align(T) 
+Cassette.overdub(ctx::CUDACtx, ::typeof(isdevice)) = true
+
+# libdevice.jl
+for f in (:cos, :cospi, :sin, :sinpi, :tan,
+ :acos, :asin, :atan, 
+ :cosh, :sinh, :tanh,
+ :acosh, :asinh, :atanh, 
+ :log, :log10, :log1p, :log2,
+ :exp, :exp2, :exp10, :expm1, :ldexp,
+ :isfinite, :isinf, :isnan,
+ :signbit, :abs,
+ :sqrt, :cbrt,
+ :ceil, :floor,)
+ @eval function Cassette.overdub(ctx::CUDACtx, ::typeof(Base.$f), x::Union{Float32, Float64})
+ @Base._inline_meta
+ return CUDAnative.$f(x)
+ end
+end
+
+contextualize(f::F) where F = (args...) -> Cassette.overdub(cudactx, f, args...)
+

src/execution.jl

@@ -175,7 +175,8 @@ kernel to determine the launch configuration:
  GC.@preserve args begin
  kernel_args = cudaconvert.(args)
  kernel_tt = Tuple{Core.Typeof.(kernel_args)...}
- kernel = cufunction(f, kernel_tt; compilation_kwargs)
+ kernel_f = contextualize(f)
+ kernel = cufunction(kernel_f, kernel_tt; compilation_kwargs)
  kernel(kernel_args...; launch_kwargs)
  end
 """
@@ -205,7 +206,8 @@ macro cuda(ex...)
  GC.@preserve $(vars...) begin
  local kernel_args = cudaconvert.(($(var_exprs...),))
  local kernel_tt = Tuple{Core.Typeof.(kernel_args)...}
- local kernel = cufunction($(esc(f)), kernel_tt; $(map(esc, compiler_kwargs)...))
+ local kernel_f = contextualize($(esc(f)))
+ local kernel = cufunction(kernel_f, kernel_tt; $(map(esc, compiler_kwargs)...))
  kernel(kernel_args...; $(map(esc, call_kwargs)...))
  end
  end)