stencil performance

[justin2004]

doing a stencil drag race between dyalog and april.
dyalog: 0.4 seconds
april: 68 seconds.

i wonder what kind of optimization dyalog is using?

      ⎕CY'dfns'
      kernel←3 3 ⍴ ¯1 2 ¯1
      img←?300 300 ⍴ 255
      result←({⍵×kernel}⌺3 3) time img
00.40
      ⍴result
300 300 3 3
⍝ and i don't think it is computing the result lazily because getting the sum of each 3x3 matrix is fast too
      ⍴+/+/ time result
00.01
300 300

APRIL> (april "⎕←kernel←3 3 ⍴ ¯1 2 ¯1 ")
¯1 2 ¯1
¯1 2 ¯1
¯1 2 ¯1
#2A((-1 2 -1) (-1 2 -1) (-1 2 -1))
APRIL> (april "img←?300 300 ⍴ 255 ")
...
APRIL> (time (april "result←({⍵×kernel}⌺3 3) img ⋄ 0"))
Evaluation took:
  68.986 seconds of real time
  334.412360 seconds of total run time (333.364623 user, 1.047737 system)
  [ Run times consist of 0.086 seconds GC time, and 334.327 seconds non-GC time. ]
  484.75% CPU
  158,368,755,265 processor cycles
  346,350,768 bytes consed
... 
APRIL> (april "⍴result")
#(300 300 3 3)

[justin2004]

since this is much quicker it might give some hints about what is costly in the previous expression.

APRIL> (time (april "result←({+/+/⍵×kernel}⌺3 3) img ⋄ 0"))
Evaluation took:
  6.518 seconds of real time
  47.692486 seconds of total run time (46.760758 user, 0.931728 system)
  [ Run times consist of 0.052 seconds GC time, and 47.641 seconds non-GC time. ]
  731.70% CPU
  14,963,064,023 processor cycles
  456,031,888 bytes consed

[justin2004]

this was mentioned in the orchard
https://www.dyalog.com/blog/2020/06/towards-improvements-to-stencil/