accurate_sprites performance improvement, PSP is untested

[picodrive.git] / Pico / Draw.s

diff --git a/Pico/Draw.s b/Pico/Draw.s
index 5ddd5e6..3cafde9 100644 (file)
--- a/Pico/Draw.s
+++ b/Pico/Draw.s
@@ -896,7 +896,7 @@ DrawTilesFromCache:
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
 \r
-.global DrawSpritesFromCache @ int *hc, int sh\r
+.global DrawSpritesFromCache @ int *hc, int maxwidth, int prio, int sh\r
 \r
 DrawSpritesFromCache:\r
     stmfd   sp!, {r4-r11,lr}\r
@@ -911,7 +911,7 @@ DrawSpritesFromCache:
     mov     r12,#0xf\r
 .endif\r
     ldr     lr, =(Pico+0x10000) @ lr=Pico.vram\r
-    mov     r6, r1, lsl #31\r
+    mov     r6, r3, lsl #31\r
     orr     r6, r6, #1<<30\r
 \r
     mov     r10, r0\r
@@ -1032,11 +1032,12 @@ DrawSpritesFromCache:
 @ + 0  :    hhhhvvvv ab--hhvv yyyyyyyy yyyyyyyy // a: offscreen h, b: offs. v, h: horiz. size\r
 @ + 4  :    xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8\r
 \r
-.global DrawSprite @ unsigned int *sprite, int **hc, int sh\r
+.global DrawSprite @ unsigned int *sprite, int **hc, int sh, int acc_sprites\r
 \r
 DrawSprite:\r
     stmfd   sp!, {r4-r9,r11,lr}\r
 \r
+    orr     r8, r3, r2, lsl #4\r
     ldr     r3, [r0]        @ sprite[0]\r
     ldr     r7, =Scanline\r
     mov     r6, r3, lsr #28\r
@@ -1050,10 +1051,10 @@ DrawSprite:
     ldr     r9, [r0, #4]\r
     sub     r7, r7, r4, asr #16 @ r7=row=Scanline-sy\r
 \r
-    tst     r2, r2\r
     mov     r2, r9, asr #16 @ r2=sx\r
-    bic     r9, r9, #0xfe000000\r
-    orrne   r9, r9, #1<<31  @ r9=code|(sh<<31)\r
+    mov     r9, r9, lsl #16\r
+    mov     r9, r9, lsr #16\r
+    orr     r9, r9, r8, lsl #27 @ r9=code|sh[31]|as[27]\r
 \r
     tst     r9, #0x1000\r
     movne   r4, r5, lsl #3\r
@@ -1075,6 +1076,7 @@ DrawSprite:
     tst     r9, #0x8000\r
     bne     .dspr_cache       @ if(code&0x8000) // high priority - cache it\r
 \r
+.dspr_continue:\r
     @ cache some stuff to avoid mem access\r
 .if OVERRIDE_HIGHCOL\r
     ldr     r11,=HighCol\r
@@ -1089,11 +1091,10 @@ DrawSprite:
     mov     r5, r5, lsl #4     @ delta<<=4; // Delta of address\r
     and     r4, r9, #0x6000\r
     orr     r9, r9, r4, lsl #16\r
-    orr     r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)\r
+    orrs    r9, r9, #0x10000000 @ r9=scc1 a??? ... <code> (s=shadow/hilight, cc=pal, a=acc_spr)\r
 \r
-    tst     r9, #1<<31\r
     mov     r3, r4, lsr #9     @ r3=pal=((code>>9)&0x30);\r
-    orrne   r3, r3, #0x40      @ shadow by default\r
+    orrmi   r3, r3, #0x40      @ shadow by default\r
 \r
     add     r6, r6, #1         @ inc now\r
     adds    r0, r2, #0         @ mov sx to r0 and set ZV flags\r
@@ -1184,19 +1185,21 @@ DrawSprite:
     mov     r4, r8, lsl #16     @ tile\r
     tst     r9, #0x0800\r
     orrne   r4, r4, #0x10000    @ code&0x0800\r
-    mov     r2, r2, lsl #22\r
-    orr     r4, r4, r2, lsr #16 @ (sx<<6)&0x0000ffc0\r
-    and     r2, r9, #0x6000\r
-    orr     r4, r4, r2, lsr #9  @ (code>>9)&0x30\r
+    mov     r0, r2, lsl #22\r
+    orr     r4, r4, r0, lsr #16 @ (sx<<6)&0x0000ffc0\r
+    and     r0, r9, #0x6000\r
+    orr     r4, r4, r0, lsr #9  @ (code>>9)&0x30\r
     mov     r3, r3, lsl #12\r
-    ldr     r2, [r1]\r
+    ldr     r0, [r1]\r
     orr     r4, r4, r3, lsr #28 @ (sprite[0]>>24)&0xf\r
 \r
-    str     r4, [r2], #4\r
-    str     r2, [r1]\r
+    str     r4, [r0], #4\r
+    str     r0, [r1]\r
 \r
-    ldmfd   sp!, {r4-r9,r11,lr}\r
-    bx      lr\r
+    tst     r9, #(1<<27)\r
+    ldmeqfd sp!, {r4-r9,r11,lr}\r
+    bne     .dspr_continue      @ draw anyway if accurate sprites enabled\r
+    bxeq    lr\r
 \r
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
@@ -1452,23 +1455,29 @@ FinalizeLineBGR444:
     bne     .fl_loopcpBGR444_hi\r
 \r
     sub     r3, r4, #0x40*3*2\r
+    mov     r6, #1\r
 \r
 \r
 .fl_noshBGR444:\r
+    ldr     r12,=rendstatus\r
+    eors    r6, r6, #1          @ sh is 0\r
+    ldr     r12,[r12]\r
+    mov     lr, #0xff\r
+    tstne   r12,#(1<<2)         @ and PDRAW_ACC_SPRITES\r
+\r
 .if OVERRIDE_HIGHCOL\r
     ldr     r1, =HighCol\r
-    mov     lr, #0xff\r
+    movne   lr, #0x3f\r
     ldr     r1, [r1]\r
     mov     lr, lr, lsl #1\r
     add     r1, r1, #8\r
 .else\r
     ldr     r1, =(HighCol+8)\r
-    mov     lr, #0xff\r
+    movne   lr, #0x3f\r
     mov     lr, lr, lsl #1\r
 .endif\r
 \r
 .fl_loopBGR444:\r
-\r
     ldr     r12, [r1], #4\r
     subs    r2, r2, #1\r
 \r
@@ -1478,11 +1487,10 @@ FinalizeLineBGR444:
     ldrh    r5, [r3, r5]\r
     and     r6, lr, r12, lsr #15\r
     ldrh    r6, [r3, r6]\r
+    and     r12,lr, r12, lsr #23\r
+    ldrh    r12,[r3, r12]              @ 1c.i.\r
     orr     r4, r4, r5, lsl #16\r
-\r
-    and     r5, lr, r12, lsr #23\r
-    ldrh    r5, [r3, r5]              @ 2c.i.\r
-    orr     r5, r6, r5, lsl #16\r
+    orr     r5, r6, r12,lsl #16\r
 \r
     stmia   r0!, {r4,r5}\r
     bne     .fl_loopBGR444\r
@@ -1617,8 +1625,16 @@ FinalizeLineRGB555:
     bne     .fl_loopcpRGB555_hi\r
 \r
     sub     r3, r3, #0x40*2\r
+    mov     r6, #1\r
 \r
 .fl_noshRGB555:\r
+    ldr     r12,=rendstatus\r
+    eors    r6, r6, #1          @ sh is 0\r
+    ldr     r12,[r12]\r
+    mov     lr, #0xff\r
+    tstne   r12,#(1<<2)         @ and PDRAW_ACC_SPRITES\r
+    movne   lr, #0x3f\r
+\r
 .if OVERRIDE_HIGHCOL\r
     ldr     r1, =HighCol\r
     ldr     r0, =DrawLineDest\r
@@ -1632,7 +1648,6 @@ FinalizeLineRGB555:
 .endif\r
 \r
     ldrb    r12, [r8, #12]\r
-    mov     lr, #0xff\r
     mov     lr, lr, lsl #1\r
 \r
     tst     r12, #1\r