A project I'm working on needs a fast square root but I couldn't find anything suitable online. After implementing several versions of the bit-by-bit algorithm I discovered the following code is particularly efficient when unrolled:

```
/* Return the square root of numb */
int isqrt( numb )
{
int root = 0, bit = 04000h;
while ( bit != 0 )
{
if ( numb >= root + bit )
{
numb = numb - root - bit;
root = root / 2 + bit;
}
else
root = root / 2;
bit = bit / 4;
}
return root;
}
```

### First Make It Work

The looping version is small but clunky. It spends almost 400 t-states shifting bits. We'll be able to eliminate most of the shifting by hard-coding the bit positions when the loop is unrolled:

```
; 16-bit integer square root
; 34 bytes, 1005-1101 cycles (average 1053)
; call with de = number to square root
; returns hl = square root
; corrupts bc, de
ld bc,08000h
ld h,c
ld l,c
sqrloop:
srl b
rr c
add hl,bc
ex de,hl
sbc hl,de
jr c,sqrbit
ex de,hl
add hl,bc
jr sqrfi
sqrbit:
add hl,de
ex de,hl
or a
sbc hl,bc
sqrfi:
srl h
rr l
srl b
rr c
jr nc,sqrloop
```

### Then Make It Work Faster

First the loop is unrolled. The first 4 iterations are optimized to work on 8-bit values and bit positions are hard-coded. The first and last iteration are optimized as a special case, we work with the bitwise complement of the root instead of the root and small jumps are replace with overlapping code. The final code finds the root in an average of 362 t-states:

```
; fast 16-bit integer square root
; 92 bytes, 344-379 cycles (average 362)
; v2 - 3 t-state optimization spotted by Russ McNulty
; call with hl = number to square root
; returns a = square root
; corrupts hl, de
ld a,h
ld de,0B0C0h
add a,e
jr c,sq7
ld a,h
ld d,0F0h
sq7:
; ----------
add a,d
jr nc,sq6
res 5,d
db 254
sq6:
sub d
sra d
; ----------
set 2,d
add a,d
jr nc,sq5
res 3,d
db 254
sq5:
sub d
sra d
; ----------
inc d
add a,d
jr nc,sq4
res 1,d
db 254
sq4:
sub d
sra d
ld h,a
; ----------
add hl,de
jr nc,sq3
ld e,040h
db 210
sq3:
sbc hl,de
sra d
ld a,e
rra
; ----------
or 010h
ld e,a
add hl,de
jr nc,sq2
and 0DFh
db 218
sq2:
sbc hl,de
sra d
rra
; ----------
or 04h
ld e,a
add hl,de
jr nc,sq1
and 0F7h
db 218
sq1:
sbc hl,de
sra d
rra
; ----------
inc a
ld e,a
add hl,de
jr nc,sq0
and 0FDh
sq0:
sra d
rra
cpl
```

This is clever and elegant! It was bothering me all night, so I'm glad I stumbled upon this-- it's easily 30% faster than my fastest. When I actually get some time, I plan to dissect this to extend the precision by 8 bits. I'm using a 16→8.8 square root routine as a first approximation for 80-bit floats, then an iteration of Newton's method with 32-bit result, then a final iteration with 64-bit result to obtain the mantissa. Currently it's roughly 28000cc.

ReplyDeleteI can't believe I never thought to use a conditional jump to skip over the SBC instructions like that! I've abused tricks like that before in division to skip parts of an iteration (and I didn't miss that CP, either-- that's snazzy).

Really very well written and executed.

ReplyDelete