Fine-grained fast-math flags

[lcw]

Is your feature request related to a problem? Please describe.

To get kernel performance matching clang we have had to add fast-math flags such as contract (which clang and nvcc do by default). Currently, we do this by an ugly-hack, see for example

CUDA.jl/perf/volumerhs.jl

Lines 21 to 57 in bb37b50

	# HACK: module-local versions of core arithmetic; needed to get FMA
	for (jlf, f) in zip((:+, :*, :-), (:add, :mul, :sub))
	for (T, llvmT) in ((:Float32, "float"), (:Float64, "double"))
	ir = """
	%x = f$f contract nsz $llvmT %0, %1
	ret $llvmT %x
	"""
	@eval begin
	# the @pure is necessary so that we can constant propagate.
	@inline Base.@pure function $jlf(a::$T, b::$T)
	Base.llvmcall($ir, $T, Tuple{$T, $T}, a, b)
	end
	end
	end
	@eval function $jlf(args...)
	Base.$jlf(args...)
	end
	end
	let (jlf, f) = (:div_arcp, :div)
	for (T, llvmT) in ((:Float32, "float"), (:Float64, "double"))
	ir = """
	%x = f$f fast $llvmT %0, %1
	ret $llvmT %x
	"""
	@eval begin
	# the @pure is necessary so that we can constant propagate.
	@inline Base.@pure function $jlf(a::$T, b::$T)
	Base.llvmcall($ir, $T, Tuple{$T, $T}, a, b)
	end
	end
	end
	@eval function $jlf(args...)
	Base.$jlf(args...)
	end
	end
	rcp(x) = div_arcp(one(x), x) # still leads to rcp.rn which is also a function call

Describe the solution you'd like

I would like a macro like @fastmath that had fine-grained control over the fast-math flags.

Describe alternatives you've considered

KernelAbstractions used to do this with https://github.com/JuliaLabs/Cassette.jl and other people use macros (although it opens up less optimization and thus not desired) https://github.com/JuliaLabs/Cassette.jl. I don't know if https://github.com/JuliaDebug/CassetteOverlay.jl can be used with kernels but it might be a possible way to implement this.

It would be nice if this functionality eventually got added to base julia.

[maleadt]

It would be nice if this functionality eventually got added to base julia.

I agree, so better file this on the Julia repository?

[lcw]

Looks like there already is at least one JuliaLang/julia#49890.

[maleadt]

I think we can close this issue then?

[lcw]

Sure.

[vchuravy]

I was thinking we could do @cuda math=(:contract, :reassoc) and then use a overlay table to switch the implementation.

[lcw]

I like that idea. So all of the code in the kernel (even within function calls) would use contract and reassoc?

[vchuravy]

Yeah, kinda inspired by JuliaLang/julia#50239, I think we could solve this with stacked OverlayMethodTables

[maleadt]

I think it would be better to prototype this in an external package, and have CUDA.jl use that package's overlay table. That way the functionality wouldn't be locked into the CUDA.jl ecosystem either.

[lcw]

Something akin to https://github.com/JuliaSIMD/LLVMLoopInfo.jl? That would be great. Is the idea to use CassetteOverlay to create some standard passes for each fast-math flag and the use those in the kernels via macros? I am not sure how to stack these for combining fast-math flags.

[vchuravy]

No more like https://github.com/vchuravy/FastmathOverlay.jl

I don't have a good solution for combining flags... yet.