32x, partially render screen if mid-frame vdp/palette changes

diff --git a/pico/32x/32x.c b/pico/32x/32x.c
index 46f5be1..b5ef091 100644 (file)
--- a/pico/32x/32x.c
+++ b/pico/32x/32x.c
@@ -243,34 +243,70 @@ void PicoReset32x(void)
   }
 }
 
-static void p32x_render_frame(void)
+static void Pico32xRenderSync(int lines)
 {
   if (Pico32xDrawMode != PDM32X_OFF && !PicoIn.skipFrame) {
-    int offs, lines;
+    int offs;
 
     pprof_start(draw);
 
-    offs = 8; lines = 224;
-    if (Pico.video.reg[1] & 8) {
+    offs = 8;
+    if (Pico.video.reg[1] & 8)
       offs = 0;
-      lines = 240;
-    }
 
     if ((Pico32x.vdp_regs[0] & P32XV_Mx) != 0 && // 32x not blanking
         (!(Pico.video.debug_p & PVD_KILL_32X)))
     {
       int md_bg = Pico.video.reg[7] & 0x3f;
 
-      // we draw full layer (not line-by-line)
-      PicoDraw32xLayer(offs, lines, md_bg);
+      // we draw lines up to the sync point (not line-by-line)
+      PicoDraw32xLayer(offs, lines-Pico32x.sync_line, md_bg);
     }
     else if (Pico32xDrawMode == PDM32X_BOTH)
-      PicoDraw32xLayerMdOnly(offs, lines);
+      PicoDraw32xLayerMdOnly(offs, lines-Pico32x.sync_line);
 
     pprof_end(draw);
   }
 }
 
+void Pico32xDrawSync(SH2 *sh2)
+{
+  if (sh2) {
+    unsigned int cycle = (sh2 ? sh2_cycles_done_m68k(sh2) : SekCyclesDone());
+    int line = ((cycle - Pico.t.m68c_frame_start) * (long long)((1LL<<32)/488.5)) >> 32;
+
+    if (Pico32x.sync_line < line && line < (Pico.video.reg[1] & 8 ? 240 : 224)) {
+      // make sure the MD image is also sync'ed to this line for merging
+      PicoDrawSync(line, 0, 0);
+
+      // pfff... need to save and restore some persistent data for MD
+      void *dest = Pico.est.DrawLineDest;
+      int incr = Pico.est.DrawLineDestIncr;
+      Pico32xRenderSync(line);
+      Pico.est.DrawLineDest = dest;
+      Pico.est.DrawLineDestIncr = incr;
+    }
+
+    // remember line we sync'ed to
+    Pico32x.sync_line = line;
+  }
+}
+
+static void p32x_render_frame(void)
+{
+  if (Pico32xDrawMode != PDM32X_OFF && !PicoIn.skipFrame) {
+    int lines;
+
+    pprof_start(draw);
+
+    lines = 224;
+    if (Pico.video.reg[1] & 8)
+      lines = 240;
+
+    Pico32xRenderSync(lines);
+  }
+}
+
 static void p32x_start_blank(void)
 {
   // enter vblank
@@ -616,6 +652,7 @@ void PicoFrame32x(void)
     pcd_prepare_frame();
 
   PicoFrameStart();
+  Pico32x.sync_line = 0;
   if (Pico32xDrawMode != PDM32X_BOTH)
     Pico.est.rendstatus |= PDRAW_SYNC_NEEDED;
   PicoFrameHints();

diff --git a/pico/32x/draw.c b/pico/32x/draw.c
index a119e6b..6b32fe5 100644 (file)
--- a/pico/32x/draw.c
+++ b/pico/32x/draw.c
@@ -173,19 +173,19 @@ void FinalizeLine32xRGB555(int sh, int line, struct PicoEState *est)
 #define make_do_loop(name, pre_code, post_code, md_code)        \
 /* Direct Color Mode */                                         \
 static void do_loop_dc##name(unsigned short *dst,               \
-    unsigned short *dram, int lines_sft_offs, int mdbg)         \
+    unsigned short *dram, unsigned lines_sft_offs, int mdbg)    \
 {                                                               \
   int inv_bit = (Pico32x.vdp_regs[0] & P32XV_PRI) ? 0x8000 : 0; \
   unsigned char  *pmd = Pico.est.Draw2FB +                      \
                           328 * (lines_sft_offs & 0xff) + 8;    \
   unsigned short *palmd = Pico.est.HighPal;                     \
   unsigned short *p32x;                                         \
-  int lines = lines_sft_offs >> 16;                             \
+  int lines = (lines_sft_offs >> 16) & 0xff;                    \
   int l;                                                        \
   (void)palmd;                                                  \
   for (l = 0; l < lines; l++, pmd += 8) {                       \
     pre_code;                                                   \
-    p32x = dram + dram[l];                                      \
+    p32x = dram + dram[l + (lines_sft_offs >> 24)];             \
     do_line_dc(dst, p32x, pmd, inv_bit, md_code);               \
     post_code;                                                  \
     dst += DrawLineDestIncrement32x/2 - 320;                    \
@@ -194,19 +194,19 @@ static void do_loop_dc##name(unsigned short *dst,               \
                                                                 \
 /* Packed Pixel Mode */                                         \
 static void do_loop_pp##name(unsigned short *dst,               \
-    unsigned short *dram, int lines_sft_offs, int mdbg)         \
+    unsigned short *dram, unsigned lines_sft_offs, int mdbg)    \
 {                                                               \
   unsigned short *pal = Pico32xMem->pal_native;                 \
   unsigned char  *pmd = Pico.est.Draw2FB +                      \
                           328 * (lines_sft_offs & 0xff) + 8;    \
   unsigned short *palmd = Pico.est.HighPal;                     \
   unsigned char  *p32x;                                         \
-  int lines = lines_sft_offs >> 16;                             \
+  int lines = (lines_sft_offs >> 16) & 0xff;                    \
   int l;                                                        \
   (void)palmd;                                                  \
   for (l = 0; l < lines; l++, pmd += 8) {                       \
     pre_code;                                                   \
-    p32x = (void *)(dram + dram[l]);                            \
+    p32x = (void *)(dram + dram[l + (lines_sft_offs >> 24)]);   \
     p32x += (lines_sft_offs >> 8) & 1;                          \
     do_line_pp(dst, p32x, pmd, md_code);                        \
     post_code;                                                  \
@@ -216,19 +216,19 @@ static void do_loop_pp##name(unsigned short *dst,               \
                                                                 \
 /* Run Length Mode */                                           \
 static void do_loop_rl##name(unsigned short *dst,               \
-    unsigned short *dram, int lines_sft_offs, int mdbg)         \
+    unsigned short *dram, unsigned lines_sft_offs, int mdbg)    \
 {                                                               \
   unsigned short *pal = Pico32xMem->pal_native;                 \
   unsigned char  *pmd = Pico.est.Draw2FB +                      \
                           328 * (lines_sft_offs & 0xff) + 8;    \
   unsigned short *palmd = Pico.est.HighPal;                     \
   unsigned short *p32x;                                         \
-  int lines = lines_sft_offs >> 16;                             \
+  int lines = (lines_sft_offs >> 16) & 0xff;                    \
   int l;                                                        \
   (void)palmd;                                                  \
   for (l = 0; l < lines; l++, pmd += 8) {                       \
     pre_code;                                                   \
-    p32x = dram + dram[l];                                      \
+    p32x = dram + dram[l + (lines_sft_offs >> 24)];             \
     do_line_rl(dst, p32x, pmd, md_code);                        \
     post_code;                                                  \
     dst += DrawLineDestIncrement32x/2 - 320;                    \
@@ -239,11 +239,11 @@ static void do_loop_rl##name(unsigned short *dst,               \
 #undef make_do_loop
 #define make_do_loop(name, pre_code, post_code, md_code) \
 extern void do_loop_dc##name(unsigned short *dst,        \
-    unsigned short *dram, int lines_offs, int mdbg);     \
+    unsigned short *dram, unsigned lines_offs, int mdbg);\
 extern void do_loop_pp##name(unsigned short *dst,        \
-    unsigned short *dram, int lines_offs, int mdbg);     \
+    unsigned short *dram, unsigned lines_offs, int mdbg);\
 extern void do_loop_rl##name(unsigned short *dst,        \
-    unsigned short *dram, int lines_offs, int mdbg);
+    unsigned short *dram, unsigned lines_offs, int mdbg);
 #endif
 
 make_do_loop(,,,)
@@ -251,7 +251,7 @@ make_do_loop(_md, , , MD_LAYER_CODE)
 make_do_loop(_scan, PICOSCAN_PRE, PICOSCAN_POST, )
 make_do_loop(_scan_md, PICOSCAN_PRE, PICOSCAN_POST, MD_LAYER_CODE)
 
-typedef void (*do_loop_func)(unsigned short *dst, unsigned short *dram, int lines, int mdbg);
+typedef void (*do_loop_func)(unsigned short *dst, unsigned short *dram, unsigned lines, int mdbg);
 enum { DO_LOOP, DO_LOOP_MD, DO_LOOP_SCAN, DO_LOOP_MD_SCAN };
 
 static const do_loop_func do_loop_dc_f[] = { do_loop_dc, do_loop_dc_md, do_loop_dc_scan, do_loop_dc_scan_md };
@@ -266,6 +266,8 @@ void PicoDraw32xLayer(int offs, int lines, int md_bg)
   int lines_sft_offs;
   int which_func;
 
+  offs += Pico32x.sync_line;
+
   Pico.est.DrawLineDest = (char *)DrawLineDestBase32x + offs * DrawLineDestIncrement32x;
   Pico.est.DrawLineDestIncr = DrawLineDestIncrement32x;
   dram = Pico32xMem->dram[Pico32x.vdp_regs[0x0a/2] & P32XV_FS];
@@ -299,7 +301,7 @@ do_it:
     which_func = have_scan ? DO_LOOP_MD_SCAN : DO_LOOP_MD;
   else
     which_func = have_scan ? DO_LOOP_SCAN : DO_LOOP;
-  lines_sft_offs = (lines << 16) | offs;
+  lines_sft_offs = (Pico32x.sync_line << 24) | (lines << 16) | offs;
   if (Pico32x.vdp_regs[2 / 2] & P32XV_SFT)
     lines_sft_offs |= 1 << 8;
 
@@ -313,16 +315,18 @@ void PicoDraw32xLayerMdOnly(int offs, int lines)
   unsigned short *dst = (void *)((char *)DrawLineDestBase32x + offs * DrawLineDestIncrement32x);
   unsigned char  *pmd = Pico.est.Draw2FB + 328 * offs + 8;
   unsigned short *pal = Pico.est.HighPal;
-  int poffs = 0, plen = 320;
+  int plen = 320;
   int l, p;
 
   PicoDrawUpdateHighPal();
 
-  dst += poffs;
+  offs += Pico32x.sync_line;
+  dst += Pico32x.sync_line * DrawLineDestIncrement32x;
+
   for (l = 0; l < lines; l++) {
     if (have_scan) {
       PicoScan32xBegin(l + offs);
-      dst = (unsigned short *)Pico.est.DrawLineDest + poffs;
+      dst = (unsigned short *)Pico.est.DrawLineDest;
     }
     for (p = 0; p < plen; p += 4) {
       dst[p + 0] = pal[*pmd++];

diff --git a/pico/32x/draw_arm.S b/pico/32x/draw_arm.S
index f2b992f..f78f68c 100644 (file)
--- a/pico/32x/draw_arm.S
+++ b/pico/32x/draw_arm.S
@@ -97,13 +97,15 @@
     sub     r0, r0, #320*2
     add     r0, r0, r12
     add     r4, r4, #1
-    cmp     r4, r2, lsr #16
+    and     r12, r2, #0xff0000
+    cmp     r4, r12, lsr #16
     call_scan_fin_ge \call_scan
     ldmgefd sp!, {r4-r11,pc}
 
 1: @ loop_outer_entry:
     call_scan_begin \call_scan
-    mov     r12,r4, lsl #1
+    add     r12,r4, r2, lsr #24
+    mov     r12,r12,lsl #1
     ldrh    r12,[r1, r12]
     add     r11,r11,#8
     mov     r6, #320/2
@@ -277,13 +279,15 @@
     sub     r0, r0, #320*2
     add     r0, r0, r12
     add     r4, r4, #1
-    cmp     r4, r2, lsr #16
+    and     r12, r2, #0xff0000
+    cmp     r4, r12, lsr #16
     call_scan_fin_ge \call_scan
     ldmgefd sp!, {r4-r11,pc}
 
 1: @ loop_outer_entry:
     call_scan_begin \call_scan
-    mov     r12,r4, lsl #1
+    add     r12,r4, r2, lsr #24
+    mov     r12,r12,lsl #1
     ldrh    r12,[r1, r12]
     add     r11,r11,#8
     mov     r6, #320/2
@@ -454,13 +458,15 @@
     sub     r0, r0, #320*2
     add     r0, r0, r12
     add     r4, r4, #1
-    cmp     r4, r2, lsr #16
+    and     r12, r2, #0xff0000
+    cmp     r4, r12, lsr #16
     call_scan_fin_ge \call_scan
     ldmgefd sp!, {r4-r11,pc}
 
 1: @ loop_outer_entry:
     call_scan_begin \call_scan
-    mov     r12,r4, lsl #1
+    add     r12,r4, r2, lsr #24
+    mov     r12,r12,lsl #1
     ldrh    r12,[r1, r12]
     add     r11,r11,#8
     mov     r6, #320

diff --git a/pico/32x/memory.c b/pico/32x/memory.c
index 86c8cd7..12c3f73 100644 (file)
--- a/pico/32x/memory.c
+++ b/pico/32x/memory.c
@@ -675,7 +675,7 @@ static u32 p32x_vdp_read16(u32 a)
   return d;
 }
 
-static void p32x_vdp_write8(u32 a, u32 d)
+static void p32x_vdp_write8(u32 a, u32 d, SH2 *sh2)
 {
   u16 *r = Pico32x.vdp_regs;
   a &= 0x0f;
@@ -684,11 +684,15 @@ static void p32x_vdp_write8(u32 a, u32 d)
   switch (a) {
     case 0x01:
       // priority inversion is handled in palette
-      if ((r[0] ^ d) & P32XV_PRI)
+      if ((r[0] ^ d) & P32XV_PRI) {
+        Pico32xDrawSync(sh2);
         Pico32x.dirty_pal = 1;
+      }
       r[0] = (r[0] & P32XV_nPAL) | (d & 0xff);
       break;
     case 0x03: // shift (for pp mode)
+      if ((r[2 / 2] ^ d) & P32XV_SFT)
+        Pico32xDrawSync(sh2);
       r[2 / 2] = d & 1;
       break;
     case 0x05: // fill len
@@ -733,7 +737,7 @@ static void p32x_vdp_write16(u32 a, u32 d, SH2 *sh2)
     return;
   }
 
-  p32x_vdp_write8(a | 1, d);
+  p32x_vdp_write8(a | 1, d, sh2);
 }
 
 // ------------------------------------------------------------------
@@ -1088,13 +1092,15 @@ static void PicoWrite8_32x_on(u32 a, u32 d)
 
   if (!(Pico32x.regs[0] & P32XS_FM)) {
     if ((a & 0xfff0) == 0x5180) { // a15180
-      p32x_vdp_write8(a, d);
+      p32x_vdp_write8(a, d, NULL);
       return;
     }
 
     // TODO: verify
     if ((a & 0xfe00) == 0x5200) { // a15200
       elprintf(EL_32X|EL_ANOMALY, "m68k 32x PAL w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
+      if (((u8 *)Pico32xMem->pal)[MEM_BE2(a & 0x1ff)] != d)
+        Pico32xDrawSync(NULL);
       ((u8 *)Pico32xMem->pal)[MEM_BE2(a & 0x1ff)] = d;
       Pico32x.dirty_pal = 1;
       return;
@@ -1147,6 +1153,8 @@ static void PicoWrite16_32x_on(u32 a, u32 d)
     }
 
     if ((a & 0xfe00) == 0x5200) { // a15200
+      if (Pico32xMem->pal[(a & 0x1ff) / 2] != d)
+        Pico32xDrawSync(NULL);
       Pico32xMem->pal[(a & 0x1ff) / 2] = d;
       Pico32x.dirty_pal = 1;
       return;
@@ -1683,12 +1691,14 @@ static void REGPARM(3) sh2_write8_cs0(u32 a, u32 d, SH2 *sh2)
   if (Pico32x.regs[0] & P32XS_FM) {
     if ((a & 0x3fff0) == 0x4100) {
       sh2->poll_cnt = 0;
-      p32x_vdp_write8(a, d);
+      p32x_vdp_write8(a, d, sh2);
       goto out;
     }
 
     if ((a & 0x3fe00) == 0x4200) {
       sh2->poll_cnt = 0;
+      if (((u8 *)Pico32xMem->pal)[MEM_BE2(a & 0x1ff)] != d)
+        Pico32xDrawSync(sh2);
       ((u8 *)Pico32xMem->pal)[MEM_BE2(a & 0x1ff)] = d;
       Pico32x.dirty_pal = 1;
       goto out;
@@ -1763,6 +1773,8 @@ static void REGPARM(3) sh2_write16_cs0(u32 a, u32 d, SH2 *sh2)
 
     if ((a & 0x3fe00) == 0x4200) {
       sh2->poll_cnt = 0;
+      if (Pico32xMem->pal[(a & 0x1ff) / 2] != d)
+        Pico32xDrawSync(sh2);
       Pico32xMem->pal[(a & 0x1ff) / 2] = d;
       Pico32x.dirty_pal = 1;
       goto out;

diff --git a/pico/pico_int.h b/pico/pico_int.h
index 8d9eef9..2678ca6 100644 (file)
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -657,7 +657,8 @@ struct Pico32x
   unsigned short pwm_p[2];       // pwm pos in fifo\r
   unsigned int pwm_cycle_p;      // pwm play cursor (32x cycles)\r
   unsigned int hint_counter;\r
-  unsigned int reserved[5];\r
+  unsigned int sync_line;\r
+  unsigned int reserved[4];\r
 };\r
 \r
 struct Pico32xMem\r
@@ -1051,6 +1052,7 @@ void Pico32xStartup(void);
 void Pico32xShutdown(void);\r
 void PicoUnload32x(void);\r
 void PicoFrame32x(void);\r
+void Pico32xDrawSync(SH2 *sh2);\r
 void Pico32xStateLoaded(int is_early);\r
 void Pico32xPrepare(void);\r
 void p32x_sync_sh2s(unsigned int m68k_target);\r