Not all of fastmath is supported?

[macd]

It seems that not all of the fast math primitives are supported? Besides div_fast, I also noticed that conj_fast doesn't work either, but maybe not important until LoopVectorization support complex arithmetic.

julia @v1.5> a, b = rand(100), rand(100);

julia @v1.5> @fastmath @avx for i = 1:100
             b[i] = a[i] / 33
             end
ERROR: UndefVarError: div_fast not defined
Stacktrace:
 [1] top-level scope at REPL[38]:1
 [2] eval(::Module, ::Any) at ./boot.jl:331
 [3] eval_user_input(::Any, ::REPL.REPLBackend) at /home/macd/julia-versions/julia/usr/share/julia/stdlib/v1.5/REPL/src/REPL.jl:130
 [4] run_backend(::REPL.REPLBackend) at /home/macd/julia-versions/dot_julia/dev/Revise/src/Revise.jl:1070
 [5] top-level scope at none:1

julia @v1.5> 

[chriselrod]

LoopVectorization assumes that it's allowed to reorder your operations, so FastMath is the default.

julia> using LoopVectorization

julia> function div33!(b, a)
           @avx for i = 1:100
                b[i] = a[i] / 33
           end
       end
div33! (generic function with 1 method)

julia> a = rand(100); b = similar(a);

julia> div33!(b, a); b ≈ (a ./ 33)
true

# julia> @code_llvm debuginfo=:none div33!(b, a)

define nonnull %jl_value_t addrspace(10)* @"japi1_div33!_2500"(%jl_value_t addrspace(10)*, %jl_value_t addrspace(10)**, i32) #0 {
L40:
  %3 = alloca %jl_value_t addrspace(10)**, align 8
  store volatile %jl_value_t addrspace(10)** %1, %jl_value_t addrspace(10)*** %3, align 8
  %4 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %1, align 8
  %5 = getelementptr inbounds %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %1, i64 1
  %6 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %5, align 8
  %7 = addrspacecast %jl_value_t addrspace(10)* %4 to %jl_value_t addrspace(11)*
  %8 = addrspacecast %jl_value_t addrspace(11)* %7 to %jl_value_t*
  %9 = bitcast %jl_value_t* %8 to double**
  %10 = load double*, double** %9, align 8
  %11 = addrspacecast %jl_value_t addrspace(10)* %6 to %jl_value_t addrspace(11)*
  %12 = addrspacecast %jl_value_t addrspace(11)* %11 to %jl_value_t*
  %13 = bitcast %jl_value_t* %12 to double**
  %14 = load double*, double** %13, align 8
  %ptr.i = bitcast double* %14 to <8 x double>*
  %res.i = load <8 x double>, <8 x double>* %ptr.i, align 8
  %res.i18 = fmul fast <8 x double> %res.i, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %ptr.i17 = bitcast double* %10 to <8 x double>*
  store <8 x double> %res.i18, <8 x double>* %ptr.i17, align 8
  %offsetptr.i.1 = getelementptr inbounds double, double* %14, i64 8
  %ptr.i.1 = bitcast double* %offsetptr.i.1 to <8 x double>*
  %res.i.1 = load <8 x double>, <8 x double>* %ptr.i.1, align 8
  %res.i18.1 = fmul fast <8 x double> %res.i.1, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.1 = getelementptr inbounds double, double* %10, i64 8
  %ptr.i17.1 = bitcast double* %offsetptr.i16.1 to <8 x double>*
  store <8 x double> %res.i18.1, <8 x double>* %ptr.i17.1, align 8
  %offsetptr.i.2 = getelementptr inbounds double, double* %14, i64 16
  %ptr.i.2 = bitcast double* %offsetptr.i.2 to <8 x double>*
  %res.i.2 = load <8 x double>, <8 x double>* %ptr.i.2, align 8
  %res.i18.2 = fmul fast <8 x double> %res.i.2, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.2 = getelementptr inbounds double, double* %10, i64 16
  %ptr.i17.2 = bitcast double* %offsetptr.i16.2 to <8 x double>*
  store <8 x double> %res.i18.2, <8 x double>* %ptr.i17.2, align 8
  %offsetptr.i.3 = getelementptr inbounds double, double* %14, i64 24
  %ptr.i.3 = bitcast double* %offsetptr.i.3 to <8 x double>*
  %res.i.3 = load <8 x double>, <8 x double>* %ptr.i.3, align 8
  %res.i18.3 = fmul fast <8 x double> %res.i.3, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.3 = getelementptr inbounds double, double* %10, i64 24
  %ptr.i17.3 = bitcast double* %offsetptr.i16.3 to <8 x double>*
  store <8 x double> %res.i18.3, <8 x double>* %ptr.i17.3, align 8
  %offsetptr.i.4 = getelementptr inbounds double, double* %14, i64 32
  %ptr.i.4 = bitcast double* %offsetptr.i.4 to <8 x double>*
  %res.i.4 = load <8 x double>, <8 x double>* %ptr.i.4, align 8
  %res.i18.4 = fmul fast <8 x double> %res.i.4, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.4 = getelementptr inbounds double, double* %10, i64 32
  %ptr.i17.4 = bitcast double* %offsetptr.i16.4 to <8 x double>*
  store <8 x double> %res.i18.4, <8 x double>* %ptr.i17.4, align 8
  %offsetptr.i.5 = getelementptr inbounds double, double* %14, i64 40
  %ptr.i.5 = bitcast double* %offsetptr.i.5 to <8 x double>*
  %res.i.5 = load <8 x double>, <8 x double>* %ptr.i.5, align 8
  %res.i18.5 = fmul fast <8 x double> %res.i.5, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.5 = getelementptr inbounds double, double* %10, i64 40
  %ptr.i17.5 = bitcast double* %offsetptr.i16.5 to <8 x double>*
  store <8 x double> %res.i18.5, <8 x double>* %ptr.i17.5, align 8
  %offsetptr.i.6 = getelementptr inbounds double, double* %14, i64 48
  %ptr.i.6 = bitcast double* %offsetptr.i.6 to <8 x double>*
  %res.i.6 = load <8 x double>, <8 x double>* %ptr.i.6, align 8
  %res.i18.6 = fmul fast <8 x double> %res.i.6, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.6 = getelementptr inbounds double, double* %10, i64 48
  %ptr.i17.6 = bitcast double* %offsetptr.i16.6 to <8 x double>*
  store <8 x double> %res.i18.6, <8 x double>* %ptr.i17.6, align 8
  %offsetptr.i.7 = getelementptr inbounds double, double* %14, i64 56
  %ptr.i.7 = bitcast double* %offsetptr.i.7 to <8 x double>*
  %res.i.7 = load <8 x double>, <8 x double>* %ptr.i.7, align 8
  %res.i18.7 = fmul fast <8 x double> %res.i.7, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.7 = getelementptr inbounds double, double* %10, i64 56
  %ptr.i17.7 = bitcast double* %offsetptr.i16.7 to <8 x double>*
  store <8 x double> %res.i18.7, <8 x double>* %ptr.i17.7, align 8
  %offsetptr.i.8 = getelementptr inbounds double, double* %14, i64 64
  %ptr.i.8 = bitcast double* %offsetptr.i.8 to <8 x double>*
  %res.i.8 = load <8 x double>, <8 x double>* %ptr.i.8, align 8
  %res.i18.8 = fmul fast <8 x double> %res.i.8, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.8 = getelementptr inbounds double, double* %10, i64 64
  %ptr.i17.8 = bitcast double* %offsetptr.i16.8 to <8 x double>*
  store <8 x double> %res.i18.8, <8 x double>* %ptr.i17.8, align 8
  %offsetptr.i.9 = getelementptr inbounds double, double* %14, i64 72
  %ptr.i.9 = bitcast double* %offsetptr.i.9 to <8 x double>*
  %res.i.9 = load <8 x double>, <8 x double>* %ptr.i.9, align 8
  %res.i18.9 = fmul fast <8 x double> %res.i.9, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.9 = getelementptr inbounds double, double* %10, i64 72
  %ptr.i17.9 = bitcast double* %offsetptr.i16.9 to <8 x double>*
  store <8 x double> %res.i18.9, <8 x double>* %ptr.i17.9, align 8
  %offsetptr.i.10 = getelementptr inbounds double, double* %14, i64 80
  %ptr.i.10 = bitcast double* %offsetptr.i.10 to <8 x double>*
  %res.i.10 = load <8 x double>, <8 x double>* %ptr.i.10, align 8
  %res.i18.10 = fmul fast <8 x double> %res.i.10, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.10 = getelementptr inbounds double, double* %10, i64 80
  %ptr.i17.10 = bitcast double* %offsetptr.i16.10 to <8 x double>*
  store <8 x double> %res.i18.10, <8 x double>* %ptr.i17.10, align 8
  %offsetptr.i.11 = getelementptr inbounds double, double* %14, i64 88
  %ptr.i.11 = bitcast double* %offsetptr.i.11 to <8 x double>*
  %res.i.11 = load <8 x double>, <8 x double>* %ptr.i.11, align 8
  %res.i18.11 = fmul fast <8 x double> %res.i.11, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08>
  %offsetptr.i16.11 = getelementptr inbounds double, double* %10, i64 88
  %ptr.i17.11 = bitcast double* %offsetptr.i16.11 to <8 x double>*
  store <8 x double> %res.i18.11, <8 x double>* %ptr.i17.11, align 8
  %offsetptr.i12 = getelementptr inbounds double, double* %14, i64 96
  %ptr.i13 = bitcast double* %offsetptr.i12 to <8 x double>*
  %res.i14 = call <8 x double> @llvm.masked.load.v8f64.p0v8f64(<8 x double>* nonnull %ptr.i13, i32 8, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 false, i1 false, i1 false, i1 false>, <8 x double> undef)
  %res.i10 = fmul fast <8 x double> %res.i14, <double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double 0x3F9F07C1F07C1F08, double undef, double undef, double undef, double undef>
  %offsetptr.i8 = getelementptr inbounds double, double* %10, i64 96
  %ptr.i9 = bitcast double* %offsetptr.i8 to <8 x double>*
  call void @llvm.masked.store.v8f64.p0v8f64(<8 x double> %res.i10, <8 x double>* %ptr.i9, i32 8, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 false, i1 false, i1 false, i1 false>)
  ret %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 139821213853376 to %jl_value_t*) to %jl_value_t addrspace(10)*)
}

Notice that it uses fmul fast instead of fdiv (multiplication is much faster than division), and that

julia> reinterpret(Float64, 0x3F9F07C1F07C1F08)
0.030303030303030304

julia> 1 / 33
0.030303030303030304

If you want to avoid FastMath, you can use evfdiv:

function ediv33!(x, y)
    @avx for i = 1:100
        x[i] = evfdiv(y[i], 33)
    end
end

[macd]

Ahhh, interesting. I had missed that fact and when writing some experimental code, I had remembered seeing a fastmath macro. But what's confusing is that using @fastmath with @avx seems to "work" on some operations. Maybe "work" it too strong a characterization, to be replaced by "not complain". Thanks for the explanation.