Apple-2-SW/dos33fsprogs
1
0
Fork 0
mirror of https://github.com/deater/dos33fsprogs.git synced 2024-12-27 02:31:00 +00:00
Code Issues Projects Releases Wiki Activity

hgr: sier: more optimizations

Browse Source
This commit is contained in:
Vince Weaver 2021-04-12 16:39:28 -04:00
parent 6044cf6a1e
commit 8ecf1031f4
3 changed files with 301 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
graphics/hgr/tiny_triangles/Makefile
View File

@ -8,7 +8,7 @@ EMPTY_DISK = ../../../empty_disk/empty.dsk
all: triangles.dsk
triangles.dsk: HELLO TINY_TRIANGLES XOR MOD9 MOD9_64 WRITING MOVE_RIGHT \
SIER_HGR SIER_HPLOT SIER_MOVERIGHT RAINBOW_SQUARES
SIER_HGR SIER_HPLOT SIER_MOVERIGHT SIER_INLINE RAINBOW_SQUARES
cp $(EMPTY_DISK) triangles.dsk
$(DOS33) -y triangles.dsk SAVE A HELLO
$(DOS33) -y triangles.dsk BSAVE -a 0x070 TINY_TRIANGLES
@ -18,6 +18,7 @@ triangles.dsk: HELLO TINY_TRIANGLES XOR MOD9 MOD9_64 WRITING MOVE_RIGHT \
$(DOS33) -y triangles.dsk BSAVE -a 0x300 WRITING
$(DOS33) -y triangles.dsk BSAVE -a 0x300 MOVE_RIGHT
$(DOS33) -y triangles.dsk BSAVE -a 0x300 SIER_HGR
$(DOS33) -y triangles.dsk BSAVE -a 0x300 SIER_INLINE
$(DOS33) -y triangles.dsk BSAVE -a 0x300 SIER_MOVERIGHT
$(DOS33) -y triangles.dsk BSAVE -a 0x300 SIER_HPLOT
$(DOS33) -y triangles.dsk BSAVE -a 0x300 RAINBOW_SQUARES
@ -61,6 +62,15 @@ sier_hgr.o: sier_hgr.s
###
SIER_INLINE: sier_inline.o
ld65 -o SIER_INLINE sier_inline.o -C $(LINKER_SCRIPTS)/apple2_300.inc
sier_inline.o: sier_inline.s
ca65 -o sier_inline.o sier_inline.s -l sier_inline.lst
###
SIER_HPLOT: sier_hplot.o
ld65 -o SIER_HPLOT sier_hplot.o -C $(LINKER_SCRIPTS)/apple2_300.inc
@ -119,4 +129,5 @@ move_right.o: move_right.s
clean:
rm -f *~ *.o *.lst HELLO TINY_TRIANGLES XOR MOD9 MOD9_64 WRITING \
MOVE_RIGHT SIER_HGR SIER_HPLOT SIER_MOVERIGHT \
SIER_INLINE \
RAINBOW_SQUARES

89
graphics/hgr/tiny_triangles/sier_hgr.s
View File

@ -14,10 +14,13 @@
; a series of 16-bit 8.8 fixed point adds
; TODO:
; HPLOT timing
; MOVERIGHT timing
; MOVERIGHT MOVEDOWN timing
; LOOKUP TABLE timing
; HPLOT: roughly 30s / frame
; MOVERIGHT: roughly 14s / frame
; MOVERIGHT NO COLORSHIFT:roughly 11s / frame
; MOVERIGHT MOVEDOWN roughly 11s / frame
; INLINE HPLOT roughly 9s / frame
; INLINE EVERYTHING roughly 7s / frame
; XT/YT lookup tables roughly 6s / frame
; zero page
@ -65,6 +68,8 @@ MOVE_DOWN = $F504 ; clc at f504, needed?
COLORTBL = $F6F6
XT_LOOKUP_TABLE = $1000
YT_LOOKUP_TABLE = $1100
;.zeropage
;.globalzp T_L,T_H
@ -75,46 +80,55 @@ COLORTBL = $F6F6
sier:
jsr HGR2 ; set FULLGR, sets A=0
; lda #0 ; start with multiplier 0
sta T_L
sta T_L ; start with multiplier 0
sta T_H
sier_outer:
lda #$40
lda #$40 ; start on page2 ($4000)
sta GBASH
lda #0 ; YY starts at 0
sta YY
sta GBASL
ldx #0 ; get X 0 for later
stx YY ; YY starts at 0
stx GBASL ; GBASL is $00
sta YY_TL
sta YY_TH
; create XX_T lookup table
; note, same as YY_T lookup table?
sta YY
stx XX_TL ; start at 0
stx XX_TH
; calc XX*T
; only really care about XX_TH
xt_table_loop:
clc
lda XX_TL ; 3
tl_smc:
adc T_L ; 2
sta XX_TL ; 3
lda XX_TH ; 3
th_smc:
adc T_H ; 3
sta XX_TH ; 3
sta YT_LOOKUP_TABLE,X ; 5
eor #$ff ; negate, as we subtract ; 2
sta XT_LOOKUP_TABLE,X ; 5
inx ; 2
bne xt_table_loop ; 3/2
sier_yloop:
lda #$C0 ; 192
lda #$C0 ; 192 reset hmask at begin of line
sta HGR_HMASK
; calc YY_T (8.8 fixed point add)
; save space by skipping clc as it's only a slight variation w/o
; clc
lda YY_TL
adc T_L
sta YY_TL
lda YY_TH
adc T_H
sta YY_TH
ldx YY ; 3
lda YT_LOOKUP_TABLE,X ; 4
sta yy_th_smc+1 ; 4
; reset XX to 0
ldy #0 ; y is x/7
ldx #0 ; XX
stx XX_TL
stx XX_TH
sier_xloop:
@ -124,29 +138,20 @@ sier_xloop:
; SAVED = XX+(Y*T)
; clc
txa ; XX ; 2
adc YY_TH ; 3
yy_th_smc:
adc #00 ; 2
sta SAVED ; 3
; calc XX*T
; clc
lda XX_TL ; 3
adc T_L ; 3
sta XX_TL ; 3
lda XX_TH ; 3
adc T_H ; 3
sta XX_TH ; 3
; calc (YY-X_T)
eor #$ff ; 2
lda XT_LOOKUP_TABLE,X ; ~(XX*T) ; 4
; calc (YY-XX*T)
sec ; 2
adc YY ; 3
; want (YY-(XX*T)) & (XX+(YY*T)
and SAVED ; 3
;============
; 36
; 19
; and #$f8
@ -199,8 +204,8 @@ lr_1:
;=================
; total roughly 36+4+19+16+5 = 80
; 49152 per inside *80 = 3,932,160
; total roughly 19+4+19+16+5 = 63
; 49152 per inside *80 = 3,145,728
; apple II cyles/frame = 17,030
; 1FPS = 1,021,800

242
graphics/hgr/tiny_triangles/sier_inline.s Normal file
View File

@ -0,0 +1,242 @@
; sierpinski-like demo
; based on the code from Hellmood's Memories demo
; by Vince `deater` Weaver <vince@deater.net>
; the simple sierpinski you more or less just plot
; X AND Y
; Hellmood's you plot something more or less like
; COLOR = ( (Y-(X*T)) & (X+(Y*T) ) & 0xf0
; where T is an incrementing frame value
; to get speed on 6502/Apple II we change the multiplies to
; a series of 16-bit 8.8 fixed point adds
; TODO:
; HPLOT: roughly 30s / frame
; MOVERIGHT: roughly 14s / frame
; MOVERIGHT NO COLORSHIFT:roughly 11s / frame
; MOVERIGHT MOVEDOWN roughly 11s / frame
; INLINE HPLOT roughly 9s / frame
; INLINE EVERYTHING roughly 7s / frame
; XT/YT lookup tables roughly 6s / frame
; zero page
HGR_BITS = $1C
GBASL = $26
GBASH = $27
MASK = $2E
COLOR = $30
HGR_HMASK = $30
HGR_X = $E0
HGR_Y = $E2
HGR_COLOR = $E4
HGR_HORIZ = $E5
SAVEX = $F7
XX_TH = $F8
XX_TL = $F9
YY = $FA
YY_TH = $FB
YY_TL = $FC
T_L = $FD
T_H = $FE
SAVED = $FF
; Soft switches
FULLGR = $C052
PAGE1 = $C054
PAGE2 = $C055
LORES = $C056 ; Enable LORES graphics
; ROM routines
HGR = $F3E2
HGR2 = $F3D8
HPOSN = $F411
HPLOT0 = $F457
HPLOT1 = $F45A ; skip the HPOSN call
COLOR_SHIFT = $F47E ; shift color for odd/even Y (0..7 or 7..13)
MOVE_RIGHT = $F48A ; move next plot location one to the right
MOVE_DOWN = $F504 ; clc at f504, needed?
; f505 is applesoft entry point but assumes c?
; move next plot location one to the right
; note moveright/movedown respect HGR_PAGE
COLORTBL = $F6F6
XT_LOOKUP_TABLE = $1000
YT_LOOKUP_TABLE = $1100
;.zeropage
;.globalzp T_L,T_H
;================================
; Clear screen and setup graphics
;================================
sier:
jsr HGR2 ; set FULLGR, sets A=0
sta T_L ; start with multiplier 0
sta T_H
sier_outer:
lda #$40 ; start on page2 ($4000)
sta GBASH
ldx #0 ; get X 0 for later
stx YY ; YY starts at 0
stx GBASL ; GBASL is $00
; create XX_T lookup table
; note, same as YY_T lookup table?
stx XX_TL ; start at 0
stx XX_TH
; calc XX*T
; only really care about XX_TH
xt_table_loop:
clc
lda XX_TL ; 3
tl_smc:
adc T_L ; 2
sta XX_TL ; 3
lda XX_TH ; 3
th_smc:
adc T_H ; 3
sta XX_TH ; 3
sta YT_LOOKUP_TABLE,X ; 5
eor #$ff ; negate, as we subtract ; 2
sta XT_LOOKUP_TABLE,X ; 5
inx ; 2
bne xt_table_loop ; 3/2
sier_yloop:
lda #$C0 ; 192 reset hmask at begin of line
sta HGR_HMASK
ldx YY ; 3
lda YT_LOOKUP_TABLE,X ; 4
sta yy_th_smc+1 ; 4
; reset XX to 0
ldy #0 ; y is x/7
ldx #0 ; XX
sier_xloop:
; want (YY-(XX*T)) & (XX+(YY*T)
; SAVED = XX+(Y*T)
; clc
txa ; XX ; 2
yy_th_smc:
adc #00 ; 2
sta SAVED ; 3
lda XT_LOOKUP_TABLE,X ; ~(XX*T) ; 4
; calc (YY-XX*T)
sec ; 2
adc YY ; 3
; want (YY-(XX*T)) & (XX+(YY*T)
and SAVED ; 3
;============
; 19
; and #$f8
beq black ; 2/3
white:
lda #$ff ; white ; 2
; .byte $2C ; bit trick
black:
;=====
; 4?
color_ready:
; sta HGR_BITS
no_shift:
; inline HPLOT1 (starting at $F45C)
eor (GBASL),Y ; 5+
and HGR_HMASK ; 3
eor (GBASL),Y ; 5+
sta (GBASL),Y ; 6
;=======
; 19
; inline MOVE_RIGHT
lda HGR_HMASK ; get mask ; 3
asl ; adjust ; 2
eor #$80 ; toggle top bit ; 2
bmi lr_1 ; if set, done? ; 3/2
lda #$81 ; otherwise set to $81 ; 2
iny ; and move to next mult of 7 ; 2
; no need to check for
; right boundary
; as we do that separately
lr_1:
sta HGR_HMASK ; 3
;======
; 16
;==================================
inx ; 2
bne sier_xloop ; 3/2
;=================
; total roughly 19+4+19+16+5 = 63
; 49152 per inside *80 = 3,145,728
; apple II cyles/frame = 17,030
; 1FPS = 1,021,800
;==================================
jsr MOVE_DOWN ; X/Y left alone
inc YY ; repeat until Y=192
ldy YY
cpy #192
bne sier_yloop
; inc T
; clc
lda T_L
blah_smc:
adc #1
sta T_L
bcc no_carry
inc T_H
no_carry:
; speed up the zoom as it goes
inc blah_smc+1
;flip_pages:
; TODO if frame rate ever gets fast enough
jmp sier_outer ; what can we branch on?