RaySollium99/picodrive
1
0
Fork 0
mirror of https://github.com/RaySollium99/picodrive.git synced 2025-09-05 07:17:45 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

vdp renderer, partial sync 8bit fast ARM asm with C code

Browse source
This commit is contained in:
kub 2021-01-21 19:09:17 +01:00
parent 1cc774814d
commit 959ea39b23
4 changed files with 59 additions and 38 deletions
Download patch file Download diff file Expand all files Collapse all files

54
pico/draw2_arm.S
View file

@ -6,10 +6,17 @@
* See COPYING file in the top-level directory.
*
* this is highly specialized, be careful if changing related C code!
*
* NB: this only deals with buffers having line width at 328
*/
#include "pico_int_offs.h"
.equ PDRAW_INTERLACE, (1<<3)
.equ PDRAW_32_COLS, (1<<8)
.equ PDRAW_BORDER_32, (1<<9)
.equ PDRAW_30_ROWS, (1<<11)
@ define these constants in your include file:
@ .equiv START_ROW, 1
@ .equiv END_ROW, 27
@ -24,14 +31,17 @@
.text
.align 2
@ void BackFillFull(void *dst, int reg7)
@ void BackFillFull(unsigned char *dst, int reg7, int lwidth)
.global BackFillFull
BackFillFull:
stmfd sp!, {r4-r9,lr}
stmfd sp!, {r4-r10,lr}
sub r10,r2, #320 @ unused bytes in a line
add lr, r0, #8 @ 8 px overlap area at start of line
add lr, lr, r2, lsl #3 @ 8 lines overlap area at top
add lr, r0, #328*8
mov r0, r1, lsl #26
mov r0, r0, lsr #26
@ -52,7 +62,6 @@ BackFillFull:
@ go go go!
.bff_loop:
add lr, lr, #8
subs r12, r12, #1
stmia lr!, {r0-r9} @ 10*4*8
@ -64,9 +73,10 @@ BackFillFull:
stmia lr!, {r0-r9}
stmia lr!, {r0-r9}
add lr, lr, r10 @ skip unused rest of line
bne .bff_loop
ldmfd sp!, {r4-r9,lr}
ldmfd sp!, {r4-r10,lr}
bx lr
.pool
@ -413,14 +423,15 @@ DrawLayerFull:
orr lr, lr, r4, lsl #13 @ lr|=nametab_bits{3}<<13
ldr r11,[sp, #9*4] @ est
sub r4, r9, #(START_ROW<<24)
ldr r4, [r11, #OFS_EST_Draw2Start]
ldr r7, [r11, #OFS_EST_rendstatus]
ldr r11, [r11, #OFS_EST_Draw2FB]
tst r7, #0x100 @ H32 border mode?
sub r4, r9, r4, lsl #24
tst r7, #PDRAW_BORDER_32 @ H32 border mode?
addne r11, r11, #32
mov r4, r4, asr #24
mov r7, #328*8
mla r11, r4, r7, r11 @ scrpos+=8*328*(planestart-START_ROW);
mla r11, r4, r7, r11 @ scrpos+=8*328*(planestart-Draw2Start);
@ Get vertical scroll value:
add_c24 r7, r10, (OFS_PMEM_vsram-OFS_PMEM_vram)
@ -588,15 +599,16 @@ DrawLayerFull:
.global DrawTilesFromCacheF
DrawTilesFromCacheF:
stmfd sp!, {r4-r10,lr}
stmfd sp!, {r4-r11,lr}
mov r9, #0xff000000 @ r9=prevcode=-1
mvn r6, #0 @ r6=prevy=-1
ldr r7, [r1, #OFS_EST_rendstatus]
ldr r4, [r1, #OFS_EST_Draw2FB]
ldr r11,[r1, #OFS_EST_Draw2Start]
ldr r2, [r0], #4 @ read y offset
tst r7, #0x100 @ H32 border mode?
tst r7, #PDRAW_BORDER_32 @ H32 border mode?
addne r4, r4, #32
mov r7, #328
mla r2, r7, r2, r4
@ -612,13 +624,14 @@ DrawTilesFromCacheF:
.dtfcf_loop:
ldr r7, [r8], #4 @ read code
movs r1, r7, lsr #16 @ r1=dx;
ldmeqfd sp!, {r4-r10,pc} @ dx is never zero, this must be a terminator, return
ldmeqfd sp!, {r4-r11,pc} @ dx is never zero, this must be a terminator, return
@ row changed?
cmp r6, r7, lsr #27
movne r6, r7, lsr #27
subne r6, r6, r11
movne r4, #328*8
mlane r5, r4, r6, r12 @ r5=pd = scrpos + prevy*328*8
mlane r5, r4, r6, r12 @ r5=pd = scrpos + (prevy-Draw2Start)*328*8
bic r1, r1, #0xf800
add r1, r5, r1 @ r1=pdest (halfwords)
@ -695,6 +708,7 @@ DrawWindowFull:
ldr r4, [r11, #OFS_Pico_video_reg+12]
mov r5, #1 @ nametab_step
ldr r11, [r3, #OFS_EST_Draw2FB]
ldr r6, [r3, #OFS_EST_Draw2Start]
tst r4, #1 @ 40 cell mode?
andne r12, r12, #0xf000 @ 0x3c<<10
movne r5, r5, lsl #7
@ -702,11 +716,12 @@ DrawWindowFull:
ldr r7, [r3, #OFS_EST_rendstatus]
and r12, r12, #0xf800
mov r5, r5, lsl #6 @ nametab_step
tst r7, #0x100
tst r7, #PDRAW_BORDER_32
addne r11, r11, #32 @ center screen in H32 mode
0: and r4, r0, #0xff
mla r12, r5, r4, r12 @ nametab += nametab_step*start;
sub r4, r4, r6
mla r12, r5, r4, r12 @ nametab += nametab_step*(start-Draw2Start];
ldr r10, [r3, #OFS_EST_PicoMem_vram]
mov r4, r0, lsr #16 @ r4=start_cell_h
@ -728,11 +743,11 @@ DrawWindowFull:
and r4, r0, #0xff
add r11, r11, #328*8
sub r4, r4, #START_ROW
sub r4, r4, r6
add r11, r11, #8
mov r7, #328*8
mla r11, r7, r4, r11 @ scrpos+=8*328*(start-START_ROW);
mla r11, r7, r4, r11 @ scrpos+=8*328*(start-Draw2Start);
mov r0, #0xf
.dwfloop_outer:
@ -927,12 +942,13 @@ DrawSpriteFull:
ldr r0, [r1, #OFS_EST_rendstatus]
ldr r11, [r1, #OFS_EST_Draw2FB]
ldr r2, [r1, #OFS_EST_Draw2Start]
ldr r10, [r1, #OFS_EST_PicoMem_vram]
tst r0, #0x100 @ H32 border mode?
tst r0, #PDRAW_BORDER_32 @ H32 border mode?
addne r11, r11, #32
sub r1, r12, #(START_ROW*8)
sub r12, r12, r2, lsl #3
mov r0, #328
mla r11, r1, r0, r11 @ scrpos+=(sy-START_ROW*8)*328;
mla r11, r12, r0, r11 @ scrpos+=(sy-Draw2Start*8)*328;
orr r5, r5, r5, lsl #16 @
orr r5, r6, r5, lsl #8 @ r5=width|(height<<8)|(height<<24)