Performance annotation docs shows SIMD slows down code by 10x

[ggggggggg]

At http://docs.julialang.org/en/latest/manual/performance-tips/#performance-annotations the docs talk about performance annotations. This section has a few confusing bits, which I'll list in order of decreasing importance.

The timings from the @simd example show @simd slowing down the code by almost 10x.

GFlop        = 1.9467069505224963
GFlop (SIMD) = 17.578554163920018

That seems like a poor result for something intended to speed up code.

The following example also confused me a few times, despite clearly showing @fastmath speeding up code

$ julia wave.jl
elapsed time: 1.207814709 seconds (0 bytes allocated)
4.443986180758243

$ julia --math-mode=ieee wave.jl
elapsed time: 4.487083643 seconds (0 bytes allocated)
4.443986180758243

While scanning the documentation to re-read sections my eye was drawn the the results 4.44... rather than the timings. This caused me to think that @fastmath had no effect on the timing, until I slowed down to read it more carefully. I think it would be worth changing the example somewhat so that the result and the slower time don't both start with the same number.

[vtjnash]

flops are 1/time, so that's a 10x speedup, not slowdown

[eschnett]

The unit should probably be GFlop/sec instead of just GFlop, for clarity.

[eschnett]

Regarding @fastmath: We should probably just add a semicolon ; to the command, which discards the output of the (here uninteresting) result.

[StefanKarpinski]

So the fixes here seem to be:

• change "GFlop" to "GFlop/sec"
• note that larger is better for "GFlop/sec"
• don't print return value for the wave.jl example

[ggggggggg]

I see that misinterpreted the GFlop statement, hopefully my revealed ignorance will aid future readers.