gpuify_cthulhu.patch

gistfile1.txt

diff --git a/Project.toml b/Project.toml
index 104d043..d15c4ad 100644
--- a/Project.toml
+++ b/Project.toml
@@ -5,6 +5,7 @@ version = "0.2.8"
 
 [deps]
 Cassette = "7057c7e9-c182-5462-911a-8362d720325c"
+Cthulhu = "f68482b8-f384-11e8-15f7-abe071a5a75f"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 
diff --git a/src/GPUifyLoops.jl b/src/GPUifyLoops.jl
index b8c67bc..335cce7 100644
--- a/src/GPUifyLoops.jl
+++ b/src/GPUifyLoops.jl
@@ -19,6 +19,7 @@ export CPU, CUDA, Device
 
 using StaticArrays
 using Requires
+using Cthulhu
 
 export @setup, @loop, @synchronize
 export @scratch, @shmem
@@ -85,6 +86,7 @@ function split_kwargs(kwargs)
     call_kws     = [:blocks, :threads, :shmem, :stream, :config]
     compiler_kwargs = []
     call_kwargs = []
+    descend = false
     for kwarg in kwargs
         key, val = kwarg
         if isa(key, Symbol)
@@ -92,6 +94,8 @@ function split_kwargs(kwargs)
                 push!(compiler_kwargs, kwarg)
             elseif key in call_kws
                 push!(call_kwargs, kwarg)
+            elseif key == :descend
+              descend = true
             else
                 throw(ArgumentError("unknown keyword argument '$key'"))
             end
@@ -99,7 +103,7 @@ function split_kwargs(kwargs)
             throw(ArgumentError("non-symbolic keyword '$key'"))
         end
     end
-    return compiler_kwargs, call_kwargs
+    return compiler_kwargs, call_kwargs, descend
 end
 
 @init @require CUDAnative="be33ccc6-a3ff-5ff2-a52e-74243cff1e17" begin
@@ -116,7 +120,7 @@ end
     global const CUDANativeVersion = version_check()
 
     function launch(::CUDA, f::F, args...; kwargs...) where F
-        compiler_kwargs, call_kwargs = split_kwargs(kwargs)
+        compiler_kwargs, call_kwargs, descend_into = split_kwargs(kwargs)
         args = (ctx, f, args...)
         GC.@preserve args begin
             kernel_args = map(cudaconvert, args)
@@ -130,7 +134,11 @@ end
             maxthreads = CUDAnative.maxthreads(kernel)
             config = launch_config(f, maxthreads, args...; call_kwargs...)
 
-            kernel(kernel_args...; config...)
+            if descend_into
+              @descend kernel(kernel_args...; config...)
+            else
+              kernel(kernel_args...; config...)
+            end
         end
         return nothing
     end