workaround for all-tiles-hi-priority performance issue

[picodrive.git] / Pico / Draw.s

diff --git a/Pico/Draw.s b/Pico/Draw.s
index cb1a6fa..91a647c 100644 (file)
--- a/Pico/Draw.s
+++ b/Pico/Draw.s
@@ -306,7 +306,7 @@ DrawLayer:
 \r
     tst     r9, #1<<31\r
     mov     r3, #0\r
-    orrne   r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|ty)\r
+    orrne   r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty)\r
     movne   r3, #0x40       @ default to shadowed pal on sh mode\r
 \r
     mvn     r9, #0          @ r9=prevcode=-1\r
@@ -342,6 +342,7 @@ DrawLayer:
     beq     .DrawStrip_samecode @ we know stuff about this tile already\r
 \r
     mov     r9, r7          @ remember code\r
+    orr     r10, r10, #1<<21 @ seen non hi-prio tile\r
 \r
     movs    r2, r9, lsl #20 @ if (code&0x1000)\r
     mov     r2, r2, lsl #1\r
@@ -386,18 +387,6 @@ DrawLayer:
     strneb  r4, [r1], #1       @ have a remaining unaligned pixel?\r
     b       .dsloop_subr1\r
 \r
-.DrawStrip_hiprio:\r
-    tst     r10, #0x00c00000\r
-    beq     .DrawStrip_hiprio_maybempt\r
-    sub     r0, r1, r11\r
-    orr     r7, r7, r0,  lsl #16\r
-    orr     r7, r7, r10, lsl #25 @ (ty<<25)\r
-    tst     r7, #0x1000\r
-    eorne   r7, r7, #7<<26  @ if(code&0x1000) cval^=7<<26;\r
-    str     r7, [r6], #4    @ cache hi priority tile\r
-    mov     r0, #0xf\r
-    b       .dsloop\r
-\r
 .DrawStrip_hiprio_maybempt:\r
     cmp     r7, r9\r
     beq     .dsloop         @ must've been empty, otherwise we wouldn't get here\r
@@ -409,13 +398,29 @@ DrawLayer:
     ldr     r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r
     mov     r9, r7          @ remember code\r
     tst     r2, r2\r
-    orrne   r10, r10, #1<<22\r
-    bne     .DrawStrip_hiprio\r
+    beq     .dsloop\r
+    orr     r10, r10, #1<<22\r
+\r
+.DrawStrip_hiprio:\r
+    tst     r10, #0x00c00000\r
+    beq     .DrawStrip_hiprio_maybempt\r
+    sub     r0, r1, r11\r
+    orr     r7, r7, r0,  lsl #16\r
+    orr     r7, r7, r10, lsl #25 @ (ty<<25)\r
+    tst     r7, #0x1000\r
+    eorne   r7, r7, #7<<26  @ if(code&0x1000) cval^=7<<26;\r
+    str     r7, [r6], #4    @ cache hi priority tile\r
+    mov     r0, #0xf\r
     b       .dsloop\r
 \r
 .dsloop_exit:\r
+    tst     r10, #1<<21 @ seen non hi-prio tile\r
+    ldreq   r1, =rendstatus\r
     mov     r0, #0\r
+    ldreq   r2, [r1]\r
     str     r0, [r6]    @ terminate the cache list\r
+    orreq   r2, r2, #0x40 @ had a layer with all hi-prio tiles\r
+    streq   r2, [r1]\r
 \r
     ldmfd   sp!, {r4-r11,lr}\r
     bx      lr\r
@@ -426,7 +431,7 @@ DrawLayer:
     rsb     r8, r3, #0\r
     mov     r8, r8, lsr #3        @ r8=tilex=(-ts->hscroll)>>3\r
     bic     r8, r8, #0xff000000\r
-    orr     r8, r8, r5, lsl #25   @ r8=(xmask[31:25]|tilex[15:0])\r
+    orr     r8, r8, r5, lsl #25   @ r8=(xmask[31:25]|had_output[24]|tilex[15:0])\r
 \r
     ldr     r4, =Scanline\r
     orr     r5, r1, r10, lsl #24\r
@@ -463,7 +468,7 @@ DrawLayer:
     add     r10,r10, #0x01000000\r
     and     r4, r10, #0x003f0000\r
     cmp     r4, r10, asr #8\r
-    ble     .dsloop_exit\r
+    ble     .dsloop_vs_exit\r
 \r
     @ calc offset and read tileline code to r7, also calc ty\r
     add     r7, lr, #0x012000\r
@@ -500,6 +505,7 @@ DrawLayer:
     beq     .DrawStrip_vs_samecode @ we know stuff about this tile already\r
 \r
     mov     r9, r7          @ remember code\r
+    orr     r8, r8, #1<<24  @ seen non hi-prio tile\r
 \r
     movs    r2, r9, lsl #20 @ if (code&0x1000)\r
     mov     r2, r2, lsl #1\r
@@ -571,6 +577,18 @@ DrawLayer:
     bne     .DrawStrip_vs_hiprio\r
     b       .dsloop_vs\r
 \r
+.dsloop_vs_exit:\r
+    tst     r8, #1<<24 @ seen non hi-prio tile\r
+    ldreq   r1, =rendstatus\r
+    mov     r0, #0\r
+    ldreq   r2, [r1]\r
+    str     r0, [r6]    @ terminate the cache list\r
+    orreq   r2, r2, #0x40 @ had a layer with all hi-prio tiles\r
+    streq   r2, [r1]\r
+\r
+    ldmfd   sp!, {r4-r11,lr}\r
+    bx      lr\r
+\r
 \r
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
@@ -650,14 +668,15 @@ BackFill:
 DrawTilesFromCache:\r
     stmfd   sp!, {r4-r8,r11,lr}\r
 \r
-    mvn     r5, #0         @ r5=prevcode=-1\r
-    mov     r8, r1\r
-\r
     @ cache some stuff to avoid mem access\r
     ldr     r11,=HighCol\r
     ldr     lr, =(Pico+0x10000) @ lr=Pico.vram\r
     mov     r12,#0xf\r
 \r
+    mvn     r5, #0         @ r5=prevcode=-1\r
+    movs    r8, r1\r
+    bne     .dtfc_check_rendflags\r
+\r
     @ scratch: r4, r7\r
 .dtfc_loop:\r
     ldr     r6, [r0], #4    @ read code\r
@@ -769,6 +788,45 @@ DrawTilesFromCache:
     mov     r12, #0xf\r
     b       .dtfc_loop\r
 \r
+@ check if we have detected layer covered with hi-prio tiles:\r
+.dtfc_check_rendflags:\r
+    ldr     r1, =rendstatus\r
+    ldr     r2, [r1]\r
+    tst     r2, #0xc0\r
+    beq     .dtfc_loop\r
+    mov     r8, #0          @ sh/hi mode off\r
+    tst     r2, #0x80\r
+    bne     .dtfc_loop      @ already processed\r
+    orr     r2, r2, #0x80\r
+    str     r2, [r1]\r
+\r
+    add     r1, r11,#8\r
+    mov     r3, #320/4\r
+    mov     r7, #0x80\r
+    orr     r7, r7, r7, lsl #8\r
+    orr     r7, r7, r7, lsl #16\r
+    mov     r6, #0x3f\r
+    orr     r6, r6, r6, lsl #8\r
+    orr     r6, r6, r6, lsl #16\r
+.dtfc_loop_shprep:\r
+    subs    r3, r3, #1\r
+    bmi     .dtfc_loop      @ done\r
+    ldr     r2, [r1]\r
+    tst     r2, r7\r
+    andeq   r2, r2, r6\r
+    streq   r2, [r1], #4\r
+    beq     .dtfc_loop_shprep\r
+    tst     r2,     #0x80000000\r
+    biceq   r2, r2, #0xc0000000\r
+    tst     r2,     #0x00800000\r
+    biceq   r2, r2, #0x00c00000\r
+    tst     r2,     #0x00008000\r
+    biceq   r2, r2, #0x0000c000\r
+    tst     r2,     #0x00000080\r
+    biceq   r2, r2, #0x000000c0\r
+    str     r2, [r1], #4\r
+    b       .dtfc_loop_shprep\r
+\r
 .pool\r
 \r
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r