Merge pull request #403 from JuliaGPU/tb/latency

Latency improvements

Manifest.toml

@@ -65,7 +65,9 @@ version = "5.1.0"
 
 [[GPUCompiler]]
 deps = ["DataStructures", "InteractiveUtils", "LLVM", "Libdl", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "c24978f4722e39510d2225db8823be74db4f7ecb"
+git-tree-sha1 = "3d8cec72696d81556689e4dd131d4b714abbb47c"
+repo-rev = "177bff3"
+repo-url = "https://github.com/JuliaGPU/GPUCompiler.jl.git"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
 version = "0.7.1"
 

src/CUDA.jl

@@ -90,4 +90,7 @@ export NVML, has_nvml
 
 include("deprecated.jl")
 
+include("precompile.jl")
+_precompile_()
+
 end

src/precompile.jl

@@ -0,0 +1,9 @@
+function _precompile_()
+    ccall(:jl_generating_output, Cint, ()) == 1 || return nothing
+    @assert precompile(Tuple{CUDA.HostKernel{identity, Tuple{Nothing}},Nothing})
+    @assert precompile(Tuple{Type{CuModule},String,Dict{CUDA.CUjit_option_enum, Any}})
+    @assert precompile(Tuple{typeof(CUDA.prepare_cuda_call)})
+    @assert precompile(Tuple{typeof(GPUCompiler.load_runtime),GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget, CUDA.CUDACompilerParams},LLVM.Context})
+    @assert precompile(Tuple{typeof(cufunction),typeof(identity),Type{Tuple{Nothing}}})
+    @assert precompile(Tuple{typeof(which(CUDA.pack_arguments,(Function,Vararg{Any, N} where N,)).generator.gen),Any,Any,Any})
+end