⎕io←0 throughout. ⎕io delenda est.

Stencil

A stencil operator ⌺ is available with Dyalog version 16.0. In brief, stencil is a dyadic operator f⌺s which applies f to (possibly overlapping) rectangles. The size of the rectangle and its movement are controlled by s. For example, enclosing 3-by-3 rectangles with default movements of 1:

   ⊢ a←4 5⍴⎕a
ABCDE
FGHIJ
KLMNO
PQRST
   {⊂⍵}⌺3 3 ⊢a
┌───┬───┬───┬───┬───┐
│   │   │   │   │   │
│ AB│ABC│BCD│CDE│DE │
│ FG│FGH│GHI│HIJ│IJ │
├───┼───┼───┼───┼───┤
│ AB│ABC│BCD│CDE│DE │
│ FG│FGH│GHI│HIJ│IJ │
│ KL│KLM│LMN│MNO│NO │
├───┼───┼───┼───┼───┤
│ FG│FGH│GHI│HIJ│IJ │
│ KL│KLM│LMN│MNO│NO │
│ PQ│PQR│QRS│RST│ST │
├───┼───┼───┼───┼───┤
│ KL│KLM│LMN│MNO│NO │
│ PQ│PQR│QRS│RST│ST │
│   │   │   │   │   │
└───┴───┴───┴───┴───┘

Stencil is also known as stencil code, tile, tessellation, and cut. It has applications in artificial neural networks, computational fluid dynamics, cellular automata, etc., and of course in Conway’s Game of Life.

The Rules of Life

Each cell of a boolean matrix has 8 neighbors adjacent to it horizontally, vertically, or diagonally. The Game of Life concerns the computation of the next generation boolean matrix.

(0) A 0-cell with 3 neighboring 1-cells becomes a 1.

(1) A 1-cell with 2 or 3 neighboring 1-cells remains at 1.

(2) All other cells remain or become a 0.

There are two main variations on the treatment of cells on the edges of the matrix: (a) the matrix is surrounded by a border of 0s; or (b) cells on an edge are adjacent to cells on the opposite edge, as on a torus.

There is an implementation of life in the dfns workspace and explained in a YouTube video. It assumes a toroidal topology.

   life←{↑1 ⍵∨.∧3 4=+/,¯1 0 1∘.⊖¯1 0 1∘.⌽⊂⍵}    ⍝ John Scholes

   ⊢ glider←5 5⍴0 0 1 0 0 1 0 1 0 0 0 1 1,12⍴0
0 0 1 0 0
1 0 1 0 0
0 1 1 0 0
0 0 0 0 0
0 0 0 0 0
   life glider
0 1 0 0 0
0 0 1 1 0
0 1 1 0 0
0 0 0 0 0
0 0 0 0 0

   {'.⍟'[⍵]}¨ (⍳8) {life⍣⍺⊢⍵}¨ ⊂glider
┌─────┬─────┬─────┬─────┬─────┬─────┬─────┬─────┐
│..⍟..│.⍟...│..⍟..│.....│.....│.....│.....│.....│
│⍟.⍟..│..⍟⍟.│...⍟.│.⍟.⍟.│...⍟.│..⍟..│...⍟.│.....│
│.⍟⍟..│.⍟⍟..│.⍟⍟⍟.│..⍟⍟.│.⍟.⍟.│...⍟⍟│....⍟│..⍟.⍟│
│.....│.....│.....│..⍟..│..⍟⍟.│..⍟⍟.│..⍟⍟⍟│...⍟⍟│
│.....│.....│.....│.....│.....│.....│.....│...⍟.│
└─────┴─────┴─────┴─────┴─────┴─────┴─────┴─────┘

Stencil Lives

It has long been known that stencil facilitates Game of Life computations. Eugene McDonnell explored the question in the APL88 paper Life: Nasty, Brutish, and Short [Ho51]. The shortest of the solutions derive as follows.

By hook or by crook, find all the 3-by-3 boolean matrices U which lead to a middle 1. A succinct Game of Life then obtains.

   B ← {1 1⌷life 3 3⍴(9⍴2)⊤⍵}¨ ⍳2*9
   U ← {3 3⍴(9⍴2)⊤⍵}¨ ⍸B  ⍝ ⍸ ←→ {⍵/⍳⍴⍵}

   life1 ← {U ∊⍨ {⊂⍵}⌺3 3⊢⍵}    ⍝ Eugene McDonnell

Comparing life and life1, and also illustrating that the toroidal and 0-border computations can be expressed one with the other.

   b←1=?97 103⍴3

   x←1 1↓¯1 ¯1↓ life 0,0,⍨0⍪0⍪⍨b
   y←life1 b
   x≡y
1

   g←{(¯1↑⍵)⍪⍵⍪1↑⍵}
   p←life b
   q←1 1↓¯1 ¯1↓ life1 (g b[;102]),(g b),(g b[;0])
   p≡q
1

Adám Brudzewsky points out that life can be terser as a train (fork):

   life1a ← U ∊⍨ ⊢∘⊂⌺3 3    ⍝ Adám Brudzewsky
   life1b ← U ∊⍨ {⊂⍵}⌺3 3

   (life1 ≡ life1a) b
1
   (life1 ≡ life1b) b
1

life1 is an example of implementing a calculation by look-up rather than by a more conventional computation, discussed in a recent blog post. There is a variation which is more efficient because the look-up is effected with integers rather than boolean matrices:

   A←3 3⍴2*⌽⍳9
   life2 ← {B[{+/,A×⍵}⌺3 3⊢⍵]}

   (life1 ≡ life1b) b
1

Jay Foad offers another stencil life, translating an algorithm in k by Arthur Whitney:

   life3 ← {3=s-⍵∧4=s←{+/,⍵}⌺3 3⊢⍵}    ⍝ Jay Foad

   (life1 ≡ life3) b
1

The algorithm combines the life rules into a single expression, wherein s←{+/,⍵}⌺3 3 ⊢⍵

(0) for 0-cells s is the number of neighbors; and
(1) for 1-cells s is the number of neighbors plus 1, and the plus 1 only matters if s is 4.

The same idea can be retrofit into the toroidal life:

   lifea←{3=s-⍵∧4=s←⊃+/,¯1 0 1∘.⊖¯1 0 1∘.⌽⊂⍵}

   (life ≡ lifea) b
1

Collected Definitions and Timings

life   ← {↑1 ⍵∨.∧3 4=+/,¯1 0 1∘.⊖¯1 0 1∘.⌽⊂⍵}
lifea  ← {3=s-⍵∧4=s←⊃+/,¯1 0 1∘.⊖¯1 0 1∘.⌽⊂⍵}

  B←{1 1⌷life 3 3⍴(9⍴2)⊤⍵}¨ ⍳2*9
  U←{3 3⍴(9⍴2)⊤⍵}¨ ⍸ B
  A←3 3⍴2*⌽⍳9

life1  ← {U ∊⍨ {⊂⍵}⌺3 3⊢⍵}
life1a ← U ∊⍨ ⊢∘⊂⌺3 3
life1b ← U ∊⍨ {⊂⍵}⌺3 3
life2  ← {B[{+/,A×⍵}⌺3 3⊢⍵]}
life3  ← {3=s-⍵∧4=s←{+/,⍵}⌺3 3⊢⍵}

   cmpx (⊂'life') ,¨ '1' '1a' '1b' '2' '3' '' 'a' ,¨ ⊂' b'
  life1 b  → 2.98E¯3 |    0% ⎕⎕⎕⎕⎕
  life1a b → 1.97E¯2 | +561% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
  life1b b → 2.99E¯3 |    0% ⎕⎕⎕⎕⎕
  life2 b  → 2.71E¯4 |  -91%
  life3 b  → 6.05E¯5 |  -98%
* life b   → 1.50E¯4 |  -95%
* lifea b  → 1.41E¯4 |  -96%

The * indicate that life and lifea give a different result (toroidal v 0-border).

life1a is much slower than the others because ⊢∘⊂⌺ is not implemented by special code.

{+/,⍵}⌺ is the fastest of the special codes because the computation has mathematical properties absent from {⊂⍵}⌺ and {+/,A×⍵}⌺.

The effect of special code v not, can be observed (for example) by use of redundant parentheses:

   cmpx '{+/,⍵}⌺3 3⊢b' '{+/,(⍵)}⌺3 3⊢b'
  {+/,⍵}⌺3 3⊢b   → 3.13E¯5  |      0%
  {+/,(⍵)}⌺3 3⊢b → 2.63E¯2  | +83900% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

   cmpx '{+/,A×⍵}⌺3 3⊢b' '{+/,A×(⍵)}⌺3 3⊢b'
  {+/,A×⍵}⌺3 3⊢b   → 2.42E¯4 |      0%
  {+/,A×(⍵)}⌺3 3⊢b → 2.98E¯2 | +12216% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

   ⍝ no special code in either of the following expressions
   cmpx '{+/,2×⍵}⌺3 3⊢b' '{+/,2×(⍵)}⌺3 3⊢b'
  {+/,2×⍵}⌺3 3⊢b   → 2.92E¯2 |      0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕ 
  {+/,2×(⍵)}⌺3 3⊢b → 3.03E¯2 |     +3% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

That life and lifea do not use stencil and yet are competitive, illustrates the efficacy of boolean operations and of letting primitives “see” large arguments.

Finally, if you wish to play with stencil a description and a hi-fi model of it can be found here.

Follow

Introduction

The e-mail arrived in the early afternoon from Morten, in Finland attending the FinnAPL Forest Seminar.

Background

⌺ is stencil, a new dyadic operator which will be available in version 16.0. In brief, f⌺s applies the operand function f to windows of size s. The window is moved over the argument centered over every possible position. The size is commonly odd and the movement commonly 1. For example:

   {⊂⍵}⌺5 3⊢6 5⍴⍳30
┌───────┬────────┬────────┬────────┬───────┐
│0  0  0│ 0  0  0│ 0  0  0│ 0  0  0│ 0  0 0│
│0  0  0│ 0  0  0│ 0  0  0│ 0  0  0│ 0  0 0│
│0  1  2│ 1  2  3│ 2  3  4│ 3  4  5│ 4  5 0│
│0  6  7│ 6  7  8│ 7  8  9│ 8  9 10│ 9 10 0│
│0 11 12│11 12 13│12 13 14│13 14 15│14 15 0│
├───────┼────────┼────────┼────────┼───────┤
│0  0  0│ 0  0  0│ 0  0  0│ 0  0  0│ 0  0 0│
│0  1  2│ 1  2  3│ 2  3  4│ 3  4  5│ 4  5 0│
│0  6  7│ 6  7  8│ 7  8  9│ 8  9 10│ 9 10 0│
│0 11 12│11 12 13│12 13 14│13 14 15│14 15 0│
│0 16 17│16 17 18│17 18 19│18 19 20│19 20 0│
├───────┼────────┼────────┼────────┼───────┤
│0  1  2│ 1  2  3│ 2  3  4│ 3  4  5│ 4  5 0│
│0  6  7│ 6  7  8│ 7  8  9│ 8  9 10│ 9 10 0│
│0 11 12│11 12 13│12 13 14│13 14 15│14 15 0│
│0 16 17│16 17 18│17 18 19│18 19 20│19 20 0│
│0 21 22│21 22 23│22 23 24│23 24 25│24 25 0│
├───────┼────────┼────────┼────────┼───────┤
│0  6  7│ 6  7  8│ 7  8  9│ 8  9 10│ 9 10 0│
│0 11 12│11 12 13│12 13 14│13 14 15│14 15 0│
│0 16 17│16 17 18│17 18 19│18 19 20│19 20 0│
│0 21 22│21 22 23│22 23 24│23 24 25│24 25 0│
│0 26 27│26 27 28│27 28 29│28 29 30│29 30 0│
├───────┼────────┼────────┼────────┼───────┤
│0 11 12│11 12 13│12 13 14│13 14 15│14 15 0│
│0 16 17│16 17 18│17 18 19│18 19 20│19 20 0│
│0 21 22│21 22 23│22 23 24│23 24 25│24 25 0│
│0 26 27│26 27 28│27 28 29│28 29 30│29 30 0│
│0  0  0│ 0  0  0│ 0  0  0│ 0  0  0│ 0  0 0│
├───────┼────────┼────────┼────────┼───────┤
│0 16 17│16 17 18│17 18 19│18 19 20│19 20 0│
│0 21 22│21 22 23│22 23 24│23 24 25│24 25 0│
│0 26 27│26 27 28│27 28 29│28 29 30│29 30 0│
│0  0  0│ 0  0  0│ 0  0  0│ 0  0  0│ 0  0 0│
│0  0  0│ 0  0  0│ 0  0  0│ 0  0  0│ 0  0 0│
└───────┴────────┴────────┴────────┴───────┘
   {+/,⍵}⌺5 3⊢6 5⍴⍳30
 39  63  72  81  57
 72 114 126 138  96
115 180 195 210 145
165 255 270 285 195
152 234 246 258 176
129 198 207 216 147

In addition, for matrix right arguments with movement 1, special code is provided for the following operand functions:

{∧/,⍵}   {∨/,⍵}   {=/,⍵}   {≠/,⍵} 

{⍵}      {,⍵}     {⊂⍵}     {+/,⍵}

{+/,A×⍵}     {E<+/,A×⍵}       
{+/⍪A×⍤2⊢⍵}  {E<+/⍪A×⍤2⊢⍵}

The comparison < can be one of < ≤ = ≠ ≥ >.

The variables r g b in the problem are integer matrices with shape 227 316 having values between 0 and 255. For present purposes we can initialize them to random values:

   r←¯1+?227 316⍴256
   g←¯1+?227 316⍴256
   b←¯1+?227 316⍴256

Opening Moves

I had other obligations and did not attend to the problem until 7 PM. A low-hanging fruit was immediately apparent: {+/,⍵×A}⌺s is not implemented by special code but {+/,A×⍵}⌺s is. The difference in performance is significant:

   edges1←{⍺←0.7 ⋄ 1 1↓¯1 ¯1↓⍺<(|EdgeDetect apply1 ⍵)÷1⌈(+⌿÷≢),⍵}
   apply1←{stencil←⍺ ⋄ {+/,stencil×⍵}⌺(⍴stencil)⊢⍵}

   cmpx 'e←⊃∨/0.2 edges1¨r g b' 'e←⊃∨/0.2 edges ¨r g b'
  e←⊃∨/0.2 edges1¨r g b → 8.0E¯3 |     0% ⎕
  e←⊃∨/0.2 edges ¨r g b → 6.1E¯1 | +7559% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

I fired off an e-mail reporting this good result, then turned to other more urgent matters. An example of the good being an enemy of the better, I suppose.

When I returned to the problem at 11 PM, the smug good feelings have largely dissipated:

• Why should {+/,A×⍵}⌺s be so much faster than {+/,⍵×A}⌺s? That is, why is there special code for the former but not the latter? (Answer: all the special codes end with … ⍵}.)
• I can not win: If the factor is small, then why isn’t it larger; if it is large, it’s only because the original code wasn’t very good.
• The large factor is because, for this problem, C (special code) is still so much faster than APL (magic function).
• If the absolute value can somehow be replaced, the operand function can then be in the form {E<+/,A×⍵}⌺s, already implemented by special code. Alternatively, {E<|+/,A×⍵}⌺s can be implemented by new special code.

Regarding the last point, the performance improvement potentially can be:

   edges2←{⍺←0.7 ⋄ 1 1↓¯1 ¯1↓(⍺ EdgeDetect)apply2 ⍵}
   apply2←{threshold stencil←⍺ ⋄ 
                    {threshold<+/,stencil×⍵}⌺(⍴stencil)⊢⍵}

   cmpx (⊂'e←⊃∨/0.2 edges'),¨'21 ',¨⊂'¨r g b'
  e←⊃∨/0.2 edges2¨r g b → 4.8E¯3 |      0%
* e←⊃∨/0.2 edges1¨r g b → 7.5E¯3 |    +57%
* e←⊃∨/0.2 edges ¨r g b → 7.4E¯1 | +15489% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

The * in the result of cmpx indicates that the expressions don’t give the same results. That is expected; edges2 is not meant to give the same results but is to get a sense of the performance difference. (Here, an additional factor of 1.57.)

edges1 has the expression ⍺<matrix÷1⌈(+⌿÷≢),⍵ where ⍺ and the divisor are both scalars. Reordering the expression eliminates one matrix operation: (⍺×1⌈(+⌿÷≢),⍵)<matrix. Thus:

   edges1a←{⍺←0.7 ⋄ 1 1↓¯1 ¯1↓(⍺×1⌈(+⌿÷≢),⍵)<|EdgeDetect apply1 ⍵}

   cmpx (⊂'e←⊃∨/0.2 edges'),¨('1a' '1 ' '  '),¨⊂'¨r g b'
  e←⊃∨/0.2 edges1a¨r g b → 6.3E¯3 |      0%
  e←⊃∨/0.2 edges1 ¨r g b → 7.6E¯3 |    +22%
  e←⊃∨/0.2 edges  ¨r g b → 6.5E¯1 | +10232% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

The time was almost 1 AM. To bed.

Further Maneuvers

Fresh and promising ideas came with the morning. The following discussion applies to the operand function {+/,A×⍵}.

(0) Scalar multiple: If all the elements of A are equal, then {+/,A×⍵}⌺(⍴A)⊢r ←→ (⊃A)×{+/,⍵}⌺(⍴A)⊢r.

   A←3 3⍴?17
   ({+/,A×⍵}⌺(⍴A)⊢r) ≡ (⊃A)×{+/,⍵}⌺(⍴A)⊢r
1

(1) Sum v inner product: {+/,⍵}⌺(⍴A)⊢r is significantly faster than {+/,A×⍵}⌺(⍴A)⊢r because the former exploits mathematical properties absent from the latter.

   A←?3 3⍴17
   cmpx '{+/,⍵}⌺(⍴A)⊢r' '{+/,A×⍵}⌺(⍴A)⊢r'
  {+/,⍵}⌺(⍴A)⊢r   → 1.4E¯4 |    0% ⎕⎕⎕
* {+/,A×⍵}⌺(⍴A)⊢r → 1.3E¯3 | +828% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

(The * in the result of cmpx is expected.)

(2) Linearity: the stencil of the sum equals the sum of the stencils.

   A←?3 3⍴17
   B←?3 3⍴17
   ({+/,(A+B)×⍵}⌺(⍴A)⊢r) ≡ ({+/,A×⍵}⌺(⍴A)⊢r) + {+/,B×⍵}⌺(⍴A)⊢r 
1

(3) Middle: If B is zero everywhere except the middle, then {+/,B×⍵}⌺(⍴B)⊢r ←→ mid×r where mid is the middle value.

   B←(⍴A)⍴0 0 0 0 9
   B
0 0 0
0 9 0
0 0 0
   ({+/,B×⍵}⌺(⍴B)⊢r) ≡ 9×r
1

(4) A faster solution.

   A←EdgeDetect
   B←(⍴A)⍴0 0 0 0 9
   C←(⍴A)⍴¯1
   A B C
┌────────┬─────┬────────┐
│¯1 ¯1 ¯1│0 0 0│¯1 ¯1 ¯1│
│¯1  8 ¯1│0 9 0│¯1 ¯1 ¯1│
│¯1 ¯1 ¯1│0 0 0│¯1 ¯1 ¯1│
└────────┴─────┴────────┘
   A ≡ B+C
1

Whence:

   ({+/,A×⍵}⌺(⍴A)⊢r) ≡ ({+/,B×⍵}⌺(⍴A)⊢r) + {+/,C×⍵}⌺(⍴A)⊢r ⍝ (2)
1
   ({+/,A×⍵}⌺(⍴A)⊢r) ≡ ({+/,B×⍵}⌺(⍴A)⊢r) - {+/,⍵}⌺(⍴A)⊢r   ⍝ (0)
1
   ({+/,A×⍵}⌺(⍴A)⊢r) ≡ (9×r) - {+/,⍵}⌺(⍴A)⊢r                ⍝ (3)
1

Putting it all together:

edges3←{
  ⍺←0.7
  mid←⊃EdgeDetect↓⍨⌊2÷⍨⍴EdgeDetect
  1 1↓¯1 ¯1↓(⍺×1⌈(+⌿÷≢),⍵)<|(⍵×1+mid)-{+/,⍵}⌺(⍴EdgeDetect)⊢⍵
}

Comparing the various edges:

   x←(⊂'e←⊃∨/0.2 edges'),¨('3 ' '2 ' '1a' '1 ' '  '),¨⊂'¨r g b'
   cmpx x
  e←⊃∨/0.2 edges3 ¨r g b → 3.4E¯3 |      0%
* e←⊃∨/0.2 edges2 ¨r g b → 4.3E¯3 |    +25%
  e←⊃∨/0.2 edges1a¨r g b → 6.4E¯3 |    +88%
  e←⊃∨/0.2 edges1 ¨r g b → 7.5E¯3 |   +122%
  e←⊃∨/0.2 edges  ¨r g b → 6.5E¯1 | +19022% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

Fin

I don’t know if the Finns would be impressed. The exercise has been amusing in any case.

Follow

(⎕io←0 and timings are done in Dyalog version 15.0.)

Table Look-Up

Some functions can be computed faster by table look-up than a more traditional and more conventional calculation. For example:

      b←?1e6⍴2

      cmpx '*b' '(*0 1)[b]'
 *b        → 1.32E¯2 |   0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
 (*0 1)[b] → 1.23E¯3 | -91% ⎕⎕⎕

Some observations about this benchmark:

(a) The advantage of table look-up depends on the time to calculate the function directly, versus the time to do indexing, ⍺[⍵] or ⍵⌷⍺, plus a small amount of time to calculate the function on a (much) smaller representative set.

Indexing is particularly efficient when the indices are boolean:

      bb←b,1
      bi←b,2
      cmpx '2 3[bb]' '2 3 3[bi]'
 2 3[bb]   → 1.12E¯4 |    0% ⎕⎕⎕⎕⎕
 2 3 3[bi] → 7.39E¯4 | +558% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

bi is the same as bb except that the last element of 2 forces it to be 1-byte integers (for bi) instead of 1-bit booleans (for bb).

(b) Although the faster expression works best with 0-origin, the timing for 1-origin is similar.

      cmpx '*b' '{⎕io←0 ⋄ (*0 1)[⍵]}b'
 *b                   → 1.32E¯2 |   0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
 {⎕io←0 ⋄ (*0 1)[⍵]}b → 1.22E¯3 | -91% ⎕⎕⎕

That is, if the current value of ⎕io is 1, the implementation can use a local ⎕io setting of 0.

(c) The fact that *b is currently slower than (*0 1)[b] represents a judgment by the implementers that *b is not a sufficiently common computation to warrant writing faster code for it. Other similar functions already embody the table look-up technique:

      cmpx '⍕b' '¯1↓,(2 2⍴''0 1 '')[b;]'
 ⍕b                   → 6.95E¯4 |  0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
 ¯1↓,(2 2⍴'0 1 ')[b;] → 6.92E¯4 | -1% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

Small Domains

Table look-up works best for booleans, but it also works for other small domains such as 1-byte integers:

      i1←?1e6⍴128    ⍝ 1-byte integers

      cmpx '*i1' '(*⍳128)[i1]'
 *i1         → 1.48E¯2 |   0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
 (*⍳128)[i1] → 9.30E¯4 | -94% ⎕⎕

      cmpx '○i1' '(○⍳128)[i1]'
 ○i1         → 2.78E¯3 |   0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
 (○⍳128)[i1] → 9.25E¯4 | -67% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

In general, a table for 1-byte integers needs to have values for ¯128+⍳256. Here ⍳128 is used to simulate that in C indexing can be done on 1-byte indices without extra computation on the indices. In APL, general 1-byte indices are used as table[i1+128]; in C, the expression is (table+offset)[i].

Domains considered “small” get bigger all the time as machines come equipped with ever larger memories.

Larger Domains

Table look-up is applicable when the argument is not a substantial subset of a large domain and there is a fast way to detect that it is not.

      i4s←1e7+?1e6⍴1e5    ⍝ small-range 4-byte integers

      G←{(⊂u⍳⍵)⌷⍺⍺ u←∪⍵}

      cmpx '⍟i4s' '⍟G i4s'
 ⍟i4s   → 1.44E¯2 |   0% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕
 ⍟G i4s → 9.59E¯3 | -34% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

The definition of G implements that the argument is not used directly as indices (as in (⊂⍵)⌷⍺⍺ u) but needed to be looked up, the u⍳⍵ in (⊂u⍳⍵)⌷⍺⍺ u. Thus both index-of and indexing are key enablers for making table look-up competitive.

§4.3 of Notation as a Tool of Thought presents a different way to distribute the results of a calculation on the “nub” (unique items). But it takes more time and space and is limited to numeric scalar results.

      H←{(⍺⍺ u)+.×(u←∪⍵)∘.=⍵}

      i4a←1e7+?1e4⍴1e3    ⍝ small-range 4-byte integers

      cmpx '⍟i4a' '⍟G i4a' '⍟H i4a'
 ⍟i4a   → 1.56E¯4 |       0%
 ⍟G i4a → 6.25E¯5 |     -60%
 ⍟H i4a → 1.78E¯1 | +113500% ⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕⎕

The benchmark is done on the smaller i4a as a benchmark on i4s founders on the ≢∪⍵ by ≢⍵ boolean matrix (for i4s, 12.5 GB = ×/ 0.125 1e5 1e6) created by H.

Mathematical Background

“New Math” teaches that a function is a set of ordered pairs whose first items are unique:

*⍵: {(⍵,*⍵)|⍵∊C}      The set of all (⍵,*⍵) where ⍵ is a complex number
⍟⍵: {(⍵,⍟⍵)|⍵∊C~{0}}  The set of all (⍵,⍟⍵) where ⍵ is a non-zero complex number
○⍵: {(⍵,○⍵)|⍵∊C}      The set of all (⍵,○⍵) where ⍵ is a complex number
⍕⍵: {(⍵,⍕⍵)|0=⍴⍴⍵}    The set of all (⍵,⍕⍵) where ⍵ is a scalar

When ⍵ is restricted (for example) to the boolean domain, the functions, the sets of ordered pairs, are more simply represented by enumerating all the possibilities:

*⍵: {(0,1), (1,2.71828)}
⍟⍵: {(0,Err), (1,0)}
○⍵: {(0,0), (1,3.14159)}
⍕⍵: {(0,'0'), (1,'1')}

Realized and Potential Performance Improvements

realized (available no later than 15.0)

potential (non-exhaustive list)

a is a scalar; f and g are primitive scalar dyadic functions.

Follow