Support @fastmath min

[Seelengrab]

Now that we have @fastmath min in Base ( JuliaLang/julia#47972), how do I make SIMD.jl work as well? Just using @fastmath min(a, b) with a and b being Vec of Float32 complains about a missing method:

julia> using SIMD, Base.Cartesian

julia> t = Vec(@ntuple 8 _ -> -Inf32)
<8 x Float32>[-Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf]

julia> @fastmath min(t,t)
ERROR: MethodError: no method matching minnum(::NTuple{8, VecElement{Float32}}, ::NTuple{8, VecElement{Float32}}, ::SIMD.Intrinsics.FastMathFlags{128})

Closest candidates are:
  minnum(::T, ::T) where T<:(Union{Tuple{Vararg{VecElement{var"#s6"}, var"#s1"}} where var"#s1", var"#s6"} where var"#s6"<:Union{Float32, Float64})
   @ SIMD ~/.julia/packages/SIMD/Ls1Up/src/LLVM_intrinsics.jl:265

Stacktrace:
 [1] min_fast(x::Vec{8, Float32}, y::Vec{8, Float32})
   @ SIMD ~/.julia/packages/SIMD/Ls1Up/src/simdvec.jl:259
 [2] top-level scope
   @ REPL[5]:1

This should ideally directly map to vminps on x86 with AVX, but right now it also has an additional vmpcunordps when not using @fastmath:

julia> @code_native debuginfo=:none syntax=:intel min(t,t)
	.text
	.file	"min"
	.globl	julia_min_346                   # -- Begin function julia_min_346
	.p2align	4, 0x90
	.type	julia_min_346,@function
julia_min_346:                          # @julia_min_346
	.cfi_startproc
# %bb.0:                                # %top
	#APP
	mov	rcx, qword ptr fs:[0]
	#NO_APP
	mov	rcx, qword ptr [rcx - 8]
	mov	rax, rdi
	mov	rcx, qword ptr [rcx + 16]
	mov	rcx, qword ptr [rcx + 16]
	#MEMBARRIER
	mov	rcx, qword ptr [rcx]
	#MEMBARRIER
	vmovups	ymm0, ymmword ptr [rsi]
	vmovups	ymm1, ymmword ptr [rdx]
	vminps	ymm2, ymm1, ymm0
	vcmpunordps	k1, ymm0, ymm0
	vmovaps	ymm2 {k1}, ymm1
	vmovaps	ymmword ptr [rdi], ymm2
	vzeroupper
	ret
.Lfunc_end0:
	.size	julia_min_346, .Lfunc_end0-julia_min_346
	.cfi_endproc
                                        # -- End function
	.section	".note.GNU-stack","",@progbits

Compare to this workaround for length 8:

julia> my_min(a, b) = Vec(@ntuple 8 i -> @fastmath min(a[i], b[i]))
my_min (generic function with 1 method)

julia> @code_native debuginfo=:none syntax=:intel my_min(t,t)
	.text
	.file	"my_min"
	.globl	julia_my_min_295                # -- Begin function julia_my_min_295
	.p2align	4, 0x90
	.type	julia_my_min_295,@function
julia_my_min_295:                       # @julia_my_min_295
	.cfi_startproc
# %bb.0:                                # %top
	#APP
	mov	rcx, qword ptr fs:[0]
	#NO_APP
	mov	rcx, qword ptr [rcx - 8]
	mov	rax, rdi
	mov	rcx, qword ptr [rcx + 16]
	mov	rcx, qword ptr [rcx + 16]
	#MEMBARRIER
	mov	rcx, qword ptr [rcx]
	#MEMBARRIER
	vmovups	ymm0, ymmword ptr [rsi]
	vminps	ymm0, ymm0, ymmword ptr [rdx]
	vmovaps	ymmword ptr [rdi], ymm0
	vzeroupper
	ret
.Lfunc_end0:
	.size	julia_my_min_295, .Lfunc_end0-julia_my_min_295
	.cfi_endproc
                                        # -- End function
	.section	".note.GNU-stack","",@progbits

[KristofferC]

Did you try:

function my_min(a,b)
    mask = @fastmath a < b
    vifelse(mask, a, b)
end

?

We could make that the fastmath min in SIMD .

[Seelengrab]

I have not - I'm too used to @ntuple all the things :)