Cos() allocates and is slow in fastmath?

General Usage
1

in v1.6.1, when --math-mode=fast:

julia> @btime cos($1.2);
  217.234 ns (8 allocations: 576 bytes)
julia> @btime sin($1.2);
  175.154 ns (6 allocations: 544 bytes)
julia> @btime tan($1.2);
  188.837 ns (6 allocations: 544 bytes)

while in normal math mode:

julia> @btime cos($1.2);
  47.322 ns (2 allocations: 32 bytes)
julia> @btime sin($1.2);
  17.283 ns (0 allocations: 0 bytes)
julia> @btime tan($1.2);
  29.544 ns (0 allocations: 0 bytes)

two problems here:

  1. cos() allocates in either math mode, while sin() and tan() does not in normal math. Supposed that elementary function like cos() should not allocate, right?
  2. “fast” math is much slower? perhaps because of the allocation?

thanks.

2

Something very strange is going on in your case–I can’t reproduce any of this in Julia 1.6.1 on Linux:

Normal math mode:

julia> @btime cos($1.2)
  7.360 ns (0 allocations: 0 bytes)
0.3623577544766736

julia> @btime sin($1.2)
  7.587 ns (0 allocations: 0 bytes)
0.9320390859672263

julia> @btime tan($1.2)
  11.403 ns (0 allocations: 0 bytes)
2.5721516221263188

--math-mode=fast

julia> @btime cos($1.2)
  5.883 ns (0 allocations: 0 bytes)
0.3623577544766736

julia> @btime sin($1.2)
  5.632 ns (0 allocations: 0 bytes)
0.9320390859672264

julia> @btime tan($1.2)
  8.081 ns (0 allocations: 0 bytes)
2.5721516221263188

What is your platform and how did you install Julia?

2 Likes
3

Mac. Just downloaded the .dmg from julialang and simply installed it.

4

I know the problem: I have overrided Base.isnan().

Taking away my own Base.isnan() and everything back to normal.

lessons learned: functions in Base could only be extended, not to be overrided… :cold_sweat:

thanks.

3 Likes