tfv: plug in the high speed multiply

This commit is contained in:
Vince Weaver 2017-11-25 19:37:02 -05:00
parent 7a7344f689
commit 382fe07bfd
4 changed files with 460 additions and 107 deletions

View File

@ -1,11 +1,23 @@
mode7 speed fixes:
+ Don't draw sky every frame, only if needed
+ faster multiply routine
+ move multiply vars to zero page
+ If result is AABBCCDD we only need BBCC for fixed point
result
+ re-arrange variables to better take advantage of self-modifying code
+ only doing the spacez calculation if it has changed
+ update the constants to be constants
+ move the screen width constant to own set of varaibles
(instead of temp)
+ leave one of multiply results in accumulator at end?
+ Skip key parsing if no key read
slower: move to 40x40 again
draw every other line. First with color/black
then xor in the second line
short term:
longer term:

View File

@ -19,6 +19,9 @@ flying_start:
jsr clear_screens
jsr set_gr_page0
jsr init_multiply_tables
;===============
; Init Variables
;===============

View File

@ -1,129 +1,338 @@
; http://www.llx.com/~nparker/a2/mult.html
; MULTIPLY NUM1H:NUM1L * NUM2H:NUM2L
; NUM2 is zeroed out
; result is in RESULT3:RESULT2:RESULT1:RESULT0
; Fast mutiply
;
; The old routine took around 700 cycles for a 16bitx16bit=32bit mutiply
; This routine, at an expense of 2kB of looku tables, takes around 250
; If you reuse a term the next time this drops closer to 200
;NUM1L: .byte 0
;NUM1H: .byte 0
;NUM2L: .byte 0
;NUM2H: .byte 0
;RESULT: .byte 0,0,0,0
;NEGATE: .byte 0
; If we have 2k to spare we should check out
; This routine was described by Stephen Judd and found
; in The Fridge and in the C=Hacking magazine
; http://codebase64.org/doku.php?id=base:seriously_fast_multiplication
multiply:
; The key thing to note is that
; (a+b)^2 (a-b)^2
; a*b = ------- - --------
; 4 4
; So if you have tables of the squares of 0..511 you can lookup and subtract
; instead of multiplying.
lda #$0 ; 2
sta NEGATE ; 3
; Table generation: I:0..511
; square1_lo = <((I*I)/4)
; square1_hi = >((I*I)/4)
; square2_lo = <(((I-255)*(I-255))/4)
; square2_hi = >(((I-255)*(I-255))/4)
; Handle Signed
lda NUM1H ; 3
bpl check_num2 ; 2nt/3
;==============
; 10
; Note: DOS3.3 starts at $9600
inc NEGATE ; 3
square1_lo EQU $8E00
square1_hi EQU $9000
square2_lo EQU $9200
square2_hi EQU $9400
clc ; 2s-complement NUM1H/NUM1L ; 2
lda NUM1L ; 3
eor #$ff ; 2
adc #$1 ; 2
sta NUM1L ; 3
; for(i=0;i<512;i++) {
; square1_lo[i]=((i*i)/4)&0xff;
; square1_hi[i]=(((i*i)/4)>>8)&0xff;
; square2_lo[i]=( ((i-255)*(i-255))/4)&0xff;
; square2_hi[i]=(( ((i-255)*(i-255))/4)>>8)&0xff;
; }
lda NUM1H ; 3
eor #$ff ; 2
adc #$0 ; 2
sta NUM1H ; 3
init_multiply_tables:
; Build the add tables
ldx #$00
txa
.byte $c9 ; CMP #immediate - skip TYA and clear carry flag
lb1: tya
adc #$00 ; 0
ml1: sta square1_hi,x ; square1_hi[0]=0
tay ; y=0
cmp #$40 ; subtract 64 and update flags (c=0)
txa ; a=0
ror ; rotate
ml9: adc #$00 ; add 0
sta ml9+1 ; update add value
inx ; x=1
ml0: sta square1_lo,x ; square1_lo[0]=1
bne lb1 ; if not zero, loop
inc ml0+2 ; increment values
inc ml1+2 ; increment values
clc ; c=0
iny ; y=1
bne lb1 ; loop
; Build the subtract tables based on the existing one
ldx #$00
ldy #$ff
second_table:
lda square1_hi+1,x
sta square2_hi+$100,x
lda square1_hi,x
sta square2_hi,y
lda square1_lo+1,x
sta square2_lo+$100,x
lda square1_lo,x
sta square2_lo,y
dey
inx
bne second_table
rts
; Fast 16x16 bit unsigned multiplication, 32-bit result
; Input: NUM1H:NUM1L * NUM2H:NUM2L
; Result: RESULT3:RESULT2:RESULT1:RESULT0
;
; Does self-modifying code to hard-code NUM1H:NUM1L into the code
; carry=0: re-use previous NUM1H:NUM1L
; carry=1: reload NUM1H:NUM1L (58 cycles slower)
;
; clobbered: RESULT, X, A, C
; Allocation setup: T1,T2 and RESULT preferably on Zero-page.
;
; NUM1H (x_i), NUM1L (x_f)
; NUM2H (y_i), NUM2L (y_f)
; NUM1L * NUM2L = AAaa
; NUM1L * NUM2H = BBbb
; NUM1H * NUM2L = CCcc
; NUM1H * NUM2H = DDdd
;
; AAaa
; BBbb
; CCcc
; + DDdd
; ----------
; RESULT
fixed_16x16_mul_unsigned:
sec ; FIXME-remove when we implement this
bcc num1_same_as_last_time ; 2nt/3
;============================
; Set up self-modifying code
; this changes the code to be hard-coded to multiply by NUM1H:NUM1L
;============================
lda NUM1L ; load the low byte ; 3
sta sm1a+1 ; 3
sta sm3a+1 ; 3
sta sm5a+1 ; 3
sta sm7a+1 ; 3
eor #$ff ; invert the bits for subtracting ; 2
sta sm2a+1 ; 3
sta sm4a+1 ; 3
sta sm6a+1 ; 3
sta sm8a+1 ; 3
lda NUM1H ; load the high byte ; 3
sta sm1b+1 ; 3
sta sm3b+1 ; 3
sta sm5b+1 ; 3
sta sm7b+1 ; 3
eor #$ff ; invert the bits for subtractin ; 2
sta sm2b+1 ; 3
sta sm4b+1 ; 3
sta sm6b+1 ; 3
sta sm8b+1 ; 3
;===========
; 25
check_num2:
lda NUM2H ; 3
bpl unsigned_multiply ; 2nt/3
;==============
; 6
; 58
inc NEGATE ; 3
num1_same_as_last_time:
clc ; 2
lda NUM2L ; 3
eor #$ff ; 2
adc #$1 ; 2
sta NUM2L ; 3
;==========================
; Perform NUM1L * NUM2L = AAaa
;==========================
lda NUM2H ; 3
eor #$ff ; 2
adc #$0 ; 2
sta NUM2H ; 3
;=============
; 25
unsigned_multiply:
ldx NUM2L ; (low le) ; 3
sec ; 2
sm1a:
lda square1_lo,x ; 4
sm2a:
sbc square2_lo,x ; 4
lda #0 ; Initialize RESULT to 0 ; 2
sta RESULT+2 ; 3
ldx #16 ; 16x16 multiply ; 2
;============
; 7
multiply_mainloop:
lsr NUM2H ; Shift right 16-bit NUM2 ; 5
ror NUM2L ; low bit goes into carry ; 5
bcc shift_output ; 0 or 1? ; 2nt/3
;============
; 13
; a is _aa
tay ; If 1, add NUM1 (hi byte RESULT in A) ; 2
clc ; 2
lda NUM1L ; 3
adc RESULT+2 ; 3
sta RESULT+2 ; 3
tya ; 2
adc NUM1H ; 3
;============
; 18
shift_output:
ror A ; "Stairstep" shift ; 2
ror RESULT+2 ; 5
ror RESULT+1 ; 5
ror RESULT ; 5
dex ; 2
bne multiply_mainloop ; 2nt/3
;=============
; 22
sta RESULT+3 ; 3
;; Negate if necessary
lda NEGATE ; 3
and #$1 ; 2
beq positive ; 2nt/3
;==============
; 11
clc ; 2
lda RESULT+0 ; 3
eor #$ff ; 2
adc #$1 ; 2
sta RESULT+0 ; 3
lda RESULT+1 ; 3
eor #$ff ; 2
adc #$0 ; 2
sta RESULT+1 ; 3
sm3a:
lda square1_hi,x ; 4
sm4a:
sbc square2_hi,x ; 4
; a is _AA
sta _AA+1 ; 3
;===========
; 27
lda RESULT+2 ; 3
eor #$ff ; 2
adc #$0 ; 2
sta RESULT+2 ; 3
; Perform NUM1H * NUM2L = CCcc
sec ; 2
sm1b:
lda square1_lo,x ; 4
sm2b:
sbc square2_lo,x ; 4
; a is _cc
sta _cc+1 ; 3
sm3b:
lda square1_hi,x ; 4
sm4b:
sbc square2_hi,x ; 4
; a is _CC
sta _CC+1 ; 3
;===========
; 24
lda RESULT+3 ; 3
eor #$ff ; 2
adc #$0 ; 2
;==========================
; Perform NUM1L * NUM2H = BBbb
;==========================
ldx NUM2H ; 3
sec ; 2
sm5a:
lda square1_lo,x ; 4
sm6a:
sbc square2_lo,x ; 4
; a is _bb
sta _bb+1 ; 3
sm7a:
lda square1_hi,x ; 4
sm8a:
sbc square2_hi,x ; 4
; a is _BB
sta _BB+1 ; 3
;===========
; 27
;==========================
; Perform NUM1H * NUM2H = DDdd
;==========================
sec ; 2
sm5b:
lda square1_lo,x ; 4
sm6b:
sbc square2_lo,x ; 4
; a is _dd
sta _dd+1 ; 3
sm7b:
lda square1_hi,x ; 4
sm8b:
sbc square2_hi,x ; 4
; a = _DD
sta RESULT+3 ; 3
;===========
; 42
positive:
; 24
;===========================================
; Add the separate multiplications together
;===========================================
clc ; 2
_AA:
lda #0 ; loading _AA ; 2
_bb:
adc #0 ; adding in _bb ; 2
sta RESULT+1 ; 3
; product[2]=_BB+_CC+c
_BB:
lda #0 ; loading _BB ; 2
_CC:
adc #0 ; adding in _CC ; 2
sta RESULT+2 ; 3
;===========
; 19
; product[3]=_DD+c
bcc dd_no_carry1 ; ^2nt/3
inc RESULT+3 ; 5
clc ; 2
;=============
; 6
dd_no_carry1:
; product[1]=_AA+_bb+_cc
_cc:
lda #0 ; load _cc ; 2
adc RESULT+1 ; 3
sta RESULT+1 ; 3
; product[2]=_BB+_CC+_dd+c
_dd:
lda #0 ; load _dd ; 2
adc RESULT+2 ; 3
sta RESULT+2 ; 3
;===========
; 19
; product[3]=_DD+c
bcc dd_no_carry2 ; ^2nt/3
inc RESULT+3 ; 5
;=============
; 4
dd_no_carry2:
; *z_i=product[1];
; *z_f=product[0];
rts ; 6
;=================
; Signed multiply
;=================
multiply:
jsr fixed_16x16_mul_unsigned ; 6
lda NUM1H ; x_i ; 3
;===========
; 12
bpl x_positive ;^3/2nt
sec ; 2
lda RESULT+2 ; 3
sbc NUM2L ; 3
sta RESULT+2 ; 3
lda RESULT+3 ; 3
sbc NUM2H ; 3
sta RESULT+3 ; 3
;============
; 19
x_positive:
lda NUM2H ; y_i ; 3
;============
; ; 6
bpl y_positive ; 3/2nt
sec ; 2
lda RESULT+2 ; 3
sbc NUM1L ; 3
sta RESULT+2 ; 3
lda RESULT+3 ; 3
sbc NUM1H ; 3
sta RESULT+3 ; 3
;===========
; 19
y_positive:
; *z_i=product[2];
; *z_f=product[1];
rts ; 6

129
tfv/tfv_multiply_slow.s Normal file
View File

@ -0,0 +1,129 @@
; http://www.llx.com/~nparker/a2/mult.html
; MULTIPLY NUM1H:NUM1L * NUM2H:NUM2L
; NUM2 is zeroed out
; result is in RESULT3:RESULT2:RESULT1:RESULT0
;NUM1L: .byte 0
;NUM1H: .byte 0
;NUM2L: .byte 0
;NUM2H: .byte 0
;RESULT: .byte 0,0,0,0
;NEGATE: .byte 0
; If we have 2k to spare we should check out
; http://codebase64.org/doku.php?id=base:seriously_fast_multiplication
multiply:
lda #$0 ; 2
sta NEGATE ; 3
; Handle Signed
lda NUM1H ; 3
bpl check_num2 ; 2nt/3
;==============
; 10
inc NEGATE ; 3
clc ; 2s-complement NUM1H/NUM1L ; 2
lda NUM1L ; 3
eor #$ff ; 2
adc #$1 ; 2
sta NUM1L ; 3
lda NUM1H ; 3
eor #$ff ; 2
adc #$0 ; 2
sta NUM1H ; 3
;===========
; 25
check_num2:
lda NUM2H ; 3
bpl unsigned_multiply ; 2nt/3
;==============
; 6
inc NEGATE ; 3
clc ; 2
lda NUM2L ; 3
eor #$ff ; 2
adc #$1 ; 2
sta NUM2L ; 3
lda NUM2H ; 3
eor #$ff ; 2
adc #$0 ; 2
sta NUM2H ; 3
;=============
; 25
unsigned_multiply:
lda #0 ; Initialize RESULT to 0 ; 2
sta RESULT+2 ; 3
ldx #16 ; 16x16 multiply ; 2
;============
; 7
multiply_mainloop:
lsr NUM2H ; Shift right 16-bit NUM2 ; 5
ror NUM2L ; low bit goes into carry ; 5
bcc shift_output ; 0 or 1? ; 2nt/3
;============
; 13
tay ; If 1, add NUM1 (hi byte RESULT in A) ; 2
clc ; 2
lda NUM1L ; 3
adc RESULT+2 ; 3
sta RESULT+2 ; 3
tya ; 2
adc NUM1H ; 3
;============
; 18
shift_output:
ror A ; "Stairstep" shift ; 2
ror RESULT+2 ; 5
ror RESULT+1 ; 5
ror RESULT ; 5
dex ; 2
bne multiply_mainloop ; 2nt/3
;=============
; 22
sta RESULT+3 ; 3
;; Negate if necessary
lda NEGATE ; 3
and #$1 ; 2
beq positive ; 2nt/3
;==============
; 11
clc ; 2
lda RESULT+0 ; 3
eor #$ff ; 2
adc #$1 ; 2
sta RESULT+0 ; 3
lda RESULT+1 ; 3
eor #$ff ; 2
adc #$0 ; 2
sta RESULT+1 ; 3
lda RESULT+2 ; 3
eor #$ff ; 2
adc #$0 ; 2
sta RESULT+2 ; 3
lda RESULT+3 ; 3
eor #$ff ; 2
adc #$0 ; 2
sta RESULT+3 ; 3
;===========
; 42
positive:
rts ; 6