Muladd fallback

[ChrisRackauckas]

u0 = GPUArray(rand(Float32, 32, 32))
k1 = similar(u0)
uprev = similar(u0)
muladd.(2,k1,uprev)

MethodError: no method matching clintrinsic(::Tuple{Expr,DataType})
Closest candidates are:
  clintrinsic(!Matched::Base.#getindex, !Matched::Type{Tuple{Tuple{Vararg{T,N}},I<:Integer}}) where {N, T, I<:Integer} at C:\Users\Chris\.julia\v0.6\Transpiler\src\clike/opencl\intrinsics.jl:172
  clintrinsic(!Matched::Base.#getindex, !Matched::Type{Tuple{T,I<:Union{Int32, Int64, UInt64}}}) where {T, I<:Union{Int32, Int64, UInt64}} at C:\Users\Chris\.julia\v0.6\Transpiler\src\clike/opencl\intrinsics.jl:178
  clintrinsic(!Matched::Base.#setindex!, !Matched::Type{Tuple{T<:Union{Transpiler.CLTranspiler.CLIntrinsics.CLArray, Transpiler.CLTranspiler.CLIntrinsics.LocalMemory},Val,I<:Integer}}) where {T<:Union{Transpiler.CLTranspiler.CLIntrinsics.CLArray, Transpiler.CLTranspiler.CLIntrinsics.LocalMemory}, Val, I<:Integer} at C:\Users\Chris\.julia\v0.6\Transpiler\src\clike/opencl\intrinsics.jl:183
  ...
isintrinsic(::Sugar.LazyMethod{:CL}) at intrinsics.jl:157
_dependencies!(::DataStructures.OrderedSet{Sugar.LazyMethod}, ::Sugar.LazyMethod{:CL}) at methods.jl:330
_dependencies!(::Sugar.LazyMethod{:CL}, ::Sugar.LazyMethod{:CL}) at methods.jl:323
_dependencies!(::DataStructures.OrderedSet{Sugar.LazyMethod}, ::Sugar.LazyMethod{:CL}) at methods.jl:331
_dependencies!(::Sugar.LazyMethod{:CL}, ::Sugar.LazyMethod{:CL}) at methods.jl:323
_dependencies!(::DataStructures.OrderedSet{Sugar.LazyMethod}, ::Sugar.LazyMethod{:CL}) at methods.jl:331
_dependencies!(::Sugar.LazyMethod{:CL}, ::Sugar.LazyMethod{:CL}) at methods.jl:323
_dependencies!(::DataStructures.OrderedSet{Sugar.LazyMethod}, ::Sugar.LazyMethod{:CL}) at methods.jl:331
dependencies!(::Sugar.LazyMethod{:CL}, ::Bool) at methods.jl:310
(::Transpiler.CLTranspiler.##27#28{Tuple{GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},##13#14,Tuple{Int32,Int32},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}},GPUArrays.CLBackend.#broadcast_kernel,OpenCL.cl.CmdQueue,OpenCL.cl.Context,NTuple{5,DataType}})() at compilation.jl:84
get!(::Transpiler.CLTranspiler.##27#28{Tuple{GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},##13#14,Tuple{Int32,Int32},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}},GPUArrays.CLBackend.#broadcast_kernel,OpenCL.cl.CmdQueue,OpenCL.cl.Context,NTuple{5,DataType}}, ::Dict{Any,Transpiler.CLTranspiler.CLFunction}, ::Tuple{GPUArrays.CLBackend.#broadcast_kernel,NTuple{5,DataType}}) at dict.jl:449
Transpiler.CLTranspiler.CLFunction(::Function, ::Tuple{GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},##13#14,Tuple{Int32,Int32},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}}, ::OpenCL.cl.CmdQueue) at compilation.jl:79
acc_broadcast!(::##13#14, ::GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}, ::Tuple{GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext},GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}}) at opencl.jl:168
broadcast(::Function, ::GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}, ::GPUArrays.GPUArray{Float32,2,OpenCL.cl.Buffer{Float32},GPUArrays.CLBackend.CLContext}) at abstractarray.jl:196
include_string(::String, ::String) at loading.jl:515
eval(::Module, ::Any) at boot.jl:235
(::Atom.##61#64)() at eval.jl:102
withpath(::Atom.##61#64, ::Void) at utils.jl:30
withpath(::Function, ::Void) at eval.jl:38
macro expansion at eval.jl:101 [inlined]
(::Atom.##60#63{Dict{String,Any}})() at task.jl:80

Would a simple fallback would be fine?

uprev .+ 2.*k1

That works. If it can FMA, that's better.

[SimonDanisch]

I think I should just add fma as an intrinsic... Allthough when I look at the implementation of muladd, it ultimately relies on float_muladd, which will make it hard to transpile!
We could make those equal to fma in the transpiler

[ChrisRackauckas]

We could make those equal to fma in the transpiler

That's probably a good idea. muladd makes sense for pure-Julia codes because of potential CPU limitations, but if you're going to the GPU or other accelerators, 99.999% of the time you should be using FMA.

[SimonDanisch]

added this as part of my transpiler rewrite branch... Will try to merge it soon!

[SimonDanisch]

Solved by #37