# Feedback on benchmark

**URL:** <https://discourse.julialang.org/t/feedback-on-benchmark/33186>\
**Category:** General Usage\
**Created:** [January 10, 2020, 9:23am UTC](https://discourse.julialang.org/t/feedback-on-benchmark/33186 "2020-01-10T09:23:06Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![jw3126](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/jw3126/32/3086_2.png) [@jw3126](https://discourse.julialang.org/u/jw3126)\
**Post date:** [January 10, 2020, 9:23am UTC](https://discourse.julialang.org/t/feedback-on-benchmark/33186/1 "2020-01-10T09:23:06Z")

</div>

It is my understanding that hardware operations do not always take the same amount of cycles on a given type. For instance it seems that `fma` is slower on subnormal floats then on normal floats.  
I wanted to observe this effect on a hot simd loop. Here is my benchmark:

```julia

function _reduce(op, v)
    ret = first(v)
    @inbounds @simd for x in v
        ret = op(ret,x,)
    end
    ret
end

function printop(io::IO, op, args...)
    print(io, op, "(")
    for arg in args
        print(io, arg, ", ")
    end
    println(io, ")")
end
printop(op, args...) = printop(stdout, op, args...)
function reducebench(op, N::Integer, arg)
    v = fill(arg, N)
    _reduce(op, v)
    printop(op, arg)
    @time _reduce(op, v)
end

fma112(a,b) = fma(a,a,b)
N = 10^7
for T in [Float32]
    for op in [+, *, min, fma112]
        println('*'^20, " $op(::$T, ::$T) ", '*'^20)
        for arg in [nextfloat(zero(T)), one(T), T(NaN), T(Inf)]
            reducebench(op, N, arg)
        end
    end
end 

```

So it seems only `fma` on subnormal is slow, there are no other bad combinations. Does that sound right? Is this benchmark sane or am I measureing something wrong? Are there other interesting cases of operations being slow on certain arguments?

```julia
********************+(::Float32, ::Float32)********************
+(1.0e-45, )
  0.009499 seconds (1 allocation: 16 bytes)
+(1.0, )
  0.009120 seconds (1 allocation: 16 bytes)
+(NaN, )
  0.010403 seconds (1 allocation: 16 bytes)
+(Inf, )
  0.009389 seconds (1 allocation: 16 bytes)
*********************(::Float32, ::Float32)********************
*(1.0e-45, )
  0.014134 seconds (1 allocation: 16 bytes)
*(1.0, )
  0.013916 seconds (1 allocation: 16 bytes)
*(NaN, )
  0.014362 seconds (1 allocation: 16 bytes)
*(Inf, )
  0.013973 seconds (1 allocation: 16 bytes)
********************min(::Float32, ::Float32)********************
min(1.0e-45, )
  0.019460 seconds (1 allocation: 16 bytes)
min(1.0, )
  0.019750 seconds (1 allocation: 16 bytes)
min(NaN, )
  0.022992 seconds (1 allocation: 16 bytes)
min(Inf, )
  0.019744 seconds (1 allocation: 16 bytes)
********************fma112(::Float32, ::Float32)********************
fma112(1.0e-45, )
  0.411726 seconds (1 allocation: 16 bytes)
fma112(1.0, )
  0.013925 seconds (1 allocation: 16 bytes)
fma112(NaN, )
  0.013971 seconds (1 allocation: 16 bytes)
fma112(Inf, )
  0.013883 seconds (1 allocation: 16 bytes)

```