From: kub <derkub@gmail.com>
Date: Thu, 21 Jan 2021 18:09:17 +0000 (+0100)
Subject: vdp renderer, partial sync 8bit fast ARM asm with C code
X-Git-Tag: v2.00~608
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=959ea39b232ee01a24a016e8844fc4536254e368;p=picodrive.git

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

diff --git a/pico/draw2.c b/pico/draw2.c
index 9d65ea54..1e27af3d 100644
--- a/pico/draw2.c
+++ b/pico/draw2.c
@@ -34,7 +34,7 @@ void (*PicoPrepareCram)()=0;            // prepares PicoCramHigh for renderer to
 
 // stuff available in asm:
 #ifdef _ASM_DRAW_C
-void BackFillFull(void *dst, int reg7);
+void BackFillFull(void *dst, int reg7, int lwidth);
 void DrawLayerFull(int plane, u32 *hcache, int planestart, int planeend,
                    struct PicoEState *est);
 void DrawTilesFromCacheF(u32 *hc, struct PicoEState *est);
@@ -183,7 +183,7 @@ static void DrawWindowFull(int start, int end, int prio, struct PicoEState *est)
 	nametab += nametab_step*(start-scrstart);
 
 	// check priority
-	code=PicoMem.vram[nametab+tile_start];
+	code=est->PicoMem_vram[nametab+tile_start];
 	if ((code>>15) != prio) return; // hack: just assume that whole window uses same priority
 
 	scrpos+=8*est->Draw2Width+8;
@@ -197,7 +197,7 @@ static void DrawWindowFull(int start, int end, int prio, struct PicoEState *est)
 //			unsigned short *pal=NULL;
 			unsigned char pal;
 
-			code=PicoMem.vram[nametab+tilex];
+			code=est->PicoMem_vram[nametab+tilex];
 			if (code==blank) continue;
 
 			// Get tile address/2:
@@ -250,7 +250,7 @@ static void DrawLayerFull(int plane, u32 *hcache, int planestart, int planeend,
 
 	if(!(pvid->reg[11]&3)) { // full screen scroll
 		// Get horizontal scroll value
-		hscroll=PicoMem.vram[htab&0x7fff];
+		hscroll=est->PicoMem_vram[htab&0x7fff];
 		htab = 0; // this marks that we don't have to update scroll value
 	}
 
@@ -297,7 +297,7 @@ static void DrawLayerFull(int plane, u32 *hcache, int planestart, int planeend,
 		if(htab) {
 			int htaddr=htab+(trow<<4);
 			if(trow) htaddr-=(vscroll&7)<<1;
-			hscroll=PicoMem.vram[htaddr&0x7fff];
+			hscroll=est->PicoMem_vram[htaddr&0x7fff];
 		}
 
 		// Draw tiles across screen:
@@ -323,7 +323,7 @@ static void DrawLayerFull(int plane, u32 *hcache, int planestart, int planeend,
 #endif
 			vsidx++;
 
-			code=PicoMem.vram[nametab_row+(tilex&xmask)];
+			code=est->PicoMem_vram[nametab_row+(tilex&xmask)];
 			if (code==blank) continue;
 
 			if (code>>15) { // high priority tile
@@ -442,18 +442,19 @@ static void DrawSpriteFull(unsigned int *sprite, struct PicoEState *est)
 	pal=(unsigned char)((code>>9)&0x30);
 
 	// goto first vertically visible tile
-	while(sy <= scrstart*8) { sy+=8; tile+=tdeltay; height--; }
+	sy -= scrstart*8;
+	while(sy <= 0) { sy+=8; tile+=tdeltay; height--; }
 
 	scrpos = est->Draw2FB;
 	if (est->rendstatus&PDRAW_BORDER_32)
 		scrpos += 32;
-	scrpos+=(sy-scrstart*8)*est->Draw2Width;
+	scrpos+=sy*est->Draw2Width;
 
 	for (; height > 0; height--, sy+=8, tile+=tdeltay)
 	{
 		int w = width, x=sx, t=tile, s;
 
-		if((sy-scrstart*8) >= END_ROW*8+8) return; // offscreen
+		if(sy >= END_ROW*8+8) return; // offscreen
 
 		for (; w; w--,x+=8,t+=tdeltax)
 		{
@@ -502,7 +503,7 @@ static void DrawAllSpritesFull(int prio, int maxwidth, struct PicoEState *est)
 		unsigned int *sprite=NULL;
 		int code, code2, sx, sy, height;
 
-		sprite=(u32 *)(PicoMem.vram+((table+(link<<2))&0x7ffc)); // Find sprite
+		sprite=(u32 *)(est->PicoMem_vram+((table+(link<<2))&0x7ffc)); // Find sprite
 
 		// get sprite info
 		code = sprite[0];
@@ -559,16 +560,18 @@ static void DrawAllSpritesFull(int prio, int maxwidth, struct PicoEState *est)
 }
 
 #ifndef _ASM_DRAW_C
-static void BackFillFull(void *dst, int reg7)
+static void BackFillFull(unsigned char *dst, int reg7, int lwidth)
 {
 	unsigned int back;
+	int i;
 
 	// Start with a background color:
 	back=reg7&0x3f;
 	back|=back<<8;
 	back|=back<<16;
 
-	memset32(dst, back, Pico.est.Draw2Width*(8+(END_ROW-START_ROW)*8)/4);
+	for (i = 0, dst += 8*lwidth; i < (END_ROW-START_ROW)*8; i++, dst += lwidth)
+		memset32(dst+8, back, 320/4);
 }
 #endif
 
@@ -589,19 +592,19 @@ static void DrawDisplayFull(void)
 	}
 	if(est->rendstatus & PDRAW_30_ROWS) {
 		// In 240 line mode, the top and bottom 8 lines are omitted
-		// since this renderer always renderers 224 lines
+		// since this renderer always renders 224 lines
 		scrstart ++, scrend ++;
 	}
 	est->Draw2Start = scrstart;
 
-	planestart = scrstart, planeend = scrend;
-	winstart = scrstart, winend = scrend;
-
 	// 32C border for centering? (for asm)
 	est->rendstatus &= ~PDRAW_BORDER_32;
 	if ((est->rendstatus&PDRAW_32_COLS) && !(PicoIn.opt&POPT_DIS_32C_BORDER))
 		est->rendstatus |= PDRAW_BORDER_32;
 
+	planestart = scrstart, planeend = scrend;
+	winstart = scrstart, winend = scrend;
+
 	// horizontal window?
 	if ((win=pvid->reg[0x12]))
 	{
@@ -716,7 +719,7 @@ PICO_INTERNAL void PicoFrameFull()
 	if (PicoPrepareCram) PicoPrepareCram();
 
 	// Draw screen:
-	BackFillFull(Pico.est.Draw2FB, Pico.video.reg[7]);
+	BackFillFull(Pico.est.Draw2FB, Pico.video.reg[7], Pico.est.Draw2Width);
 	if (Pico.video.reg[1] & 0x40)
 		DrawDisplayFull();
 
diff --git a/pico/draw2_arm.S b/pico/draw2_arm.S
index ded0d5a5..3101a823 100644
--- a/pico/draw2_arm.S
+++ b/pico/draw2_arm.S
@@ -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)
diff --git a/pico/pico_int.h b/pico/pico_int.h
index d66b67e2..26fb60a1 100644
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -355,8 +355,8 @@ struct PicoEState
   unsigned char *HighCol;
   u32 *HighPreSpr;
   struct Pico *Pico;
-  void *PicoMem_vram;
-  void *PicoMem_cram;
+  unsigned short *PicoMem_vram;
+  unsigned short *PicoMem_cram;
   unsigned int  *PicoOpt;
   unsigned char *Draw2FB;
   int Draw2Width;
diff --git a/tools/mkoffsets.sh b/tools/mkoffsets.sh
index c347587c..e45567f5 100755
--- a/tools/mkoffsets.sh
+++ b/tools/mkoffsets.sh
@@ -144,6 +144,8 @@ get_define OFS_EST_ PicoEState PicoMem_vram	; echo "$line" >>$fn
 get_define OFS_EST_ PicoEState PicoMem_cram	; echo "$line" >>$fn
 get_define OFS_EST_ PicoEState PicoOpt		; echo "$line" >>$fn
 get_define OFS_EST_ PicoEState Draw2FB		; echo "$line" >>$fn
+get_define OFS_EST_ PicoEState Draw2Width	; echo "$line" >>$fn
+get_define OFS_EST_ PicoEState Draw2Start	; echo "$line" >>$fn
 get_define OFS_EST_ PicoEState HighPal		; echo "$line" >>$fn
 
 get_define OFS_PMEM_ PicoMem vram		; echo "$line" >>$fn