Re: More on compiler optimization

;; 35.5 μs, 0.66 bytes/FFT alloc 1M iter on 1024 pt FFT
(defun tst-fft (&key (nel 1024) (niter 1_000_000.))
  #F
  (let* ((cx (make-cx-sep-array nel))
         (nr (map 'vector 'dfloat (vm:unoise nel :limit 1.0)))
         (ni (map 'vector 'dfloat (vm:unoise nel :limit 1.0))))
    (with-sep-arrays (rr ii) cx
      (replace rr nr)
      (replace ii ni))
    ;; test only the speed of the FFT
    (time
     (loop repeat niter do
             (fft! cx)
           ))
    ))


> On Feb 12, 2025, at 18:03, David McClain (as dbm at refined-audiometrics dot com) <[email protected]> wrote:
> 
> 
>> Looking at my own FFT implementation, which does not primarily address GC issues, I have a timing of 46.7 μs/1024 pt FFT on M1 iMac.
>> 
>> Starts in Line 437. Looks like it allocs about 3 kB / 1024 pt FFT.
> 
>> 
> <cx-arrays.lisp>
>> 
>> 
>> 
>