Optimized _filt_fir! for DF2TFilter

[wheeheee]

Also adjusts cutoff, uses LazyString for errors.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.90%. Comparing base (b3f9da6) to head (32fdfc3).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #569      +/-   ##
==========================================
- Coverage   97.90%   97.90%   -0.01%     
==========================================
  Files          19       19              
  Lines        3252     3249       -3     
==========================================
- Hits         3184     3181       -3     
  Misses         68       68              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[martinholters]

I don't understand the SMALL_FILT_VECT_CUTOFF dance, but apart from that LGTM.

[wheeheee]

I don't understand the SMALL_FILT_VECT_CUTOFF dance, but apart from that LGTM.

I increased it to 19 because (for me, on 1.11.1) it's a small improvement.

julia> x = rand(10_000); out = similar(x);

julia> @btime filt!($out, b, a, $x, si) setup=((a, b, si) = (1., rand(19), zeros(18)));
  20.700 μs (0 allocations: 0 bytes) # SMALL_FILT_VECT_CUTOFF = 18
  18.600 μs (0 allocations: 0 bytes) # SMALL_FILT_VECT_CUTOFF = 19

Or if your question is what it's for, removing inbounds below the cutoff prevents it from vectorizing in some non-optimal way (as it was in 0.7.9).

julia> @btime filt!($out, b, a, $x, si) setup=((a, b, si) = (1., rand(15), zeros(14)));
  12.900 μs (0 allocations: 0 bytes) # SMALL_FILT_VECT_CUTOFF = 19
  17.200 μs (0 allocations: 0 bytes) # SMALL_FILT_VECT_CUTOFF = 12

And if it returns an NTuple, sometimes (I think for x = rand(10_000, 1)) this happens nevertheless, so I specialized it for DF2TFilter

[martinholters]

Ah, so the @inbounds is semantically correct for any N, but whether it helps or hinders performance depends on N? That's... ugh...

Anyway, from my point of view, this is good to go.

[wheeheee]

Ah, so the @inbounds is semantically correct for any N, but whether it helps or hinders performance depends on N?

Yup. And sometimes also on whether a view is passed to _filt_fir!, instead of a Vector...