Try #334:

Author: bors[bot]

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 1.0
 LLVM 0.9.14
 CUDAapi 0.4.0

src/CUDAnative.jl

@@ -31,6 +31,7 @@ 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 +1,40 @@
+##
+# 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
+
+Cassette.@context CUDACtx
+const cudactx = CUDACtx()
+
+# 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

@@ -172,7 +172,8 @@ 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
 """
@@ -202,7 +203,8 @@ 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)