buggy port of gfx/cd, silpheed started working?

git-svn-id: file:///home/notaz/opt/svn/PicoDrive@73 be3aeb3a-fb24-0410-a615-afba39da0efa

Pico/Area.c

@@ -52,6 +52,8 @@ int PicoAreaPackCpu(unsigned char *cpu, int is_sub)
   *(unsigned char *)(cpu+0x44)=(unsigned char)M68000_regs.ccr;
   *(unsigned char *)(cpu+0x45)=(unsigned char)M68000_regs.srh;
   *(unsigned int  *)(cpu+0x48)=M68000_regs.isp;
+  cpu[0x4c] = M68000_regs.irq;
+  // stop flag?
 #endif
 
 #ifdef EMU_C68K
@@ -60,6 +62,8 @@ int PicoAreaPackCpu(unsigned char *cpu, int is_sub)
   pc=context->pc-context->membase;
   *(unsigned int *)(cpu+0x44)=CycloneGetSr(context);
   *(unsigned int *)(cpu+0x48)=context->osp;
+  cpu[0x4c] = context->irq;
+  cpu[0x4d] = context->stopped;
 #endif
 
 #ifdef EMU_M68K
@@ -69,6 +73,8 @@ int PicoAreaPackCpu(unsigned char *cpu, int is_sub)
   pc=m68ki_cpu_p->pc;
   *(unsigned int  *)(cpu+0x44)=m68k_get_reg(NULL, M68K_REG_SR);
   *(unsigned int  *)(cpu+0x48)=m68ki_cpu_p->sp[0];
+  cpu[0x4c] = CPU_INT_LEVEL>>8;
+  cpu[0x4d] = CPU_STOPPED;
   m68k_set_context(oldcontext);
 #endif
 
@@ -84,6 +90,8 @@ int PicoAreaUnpackCpu(unsigned char *cpu, int is_sub)
   M68000_regs.ccr=*(unsigned char *)(cpu+0x44);
   M68000_regs.srh=*(unsigned char *)(cpu+0x45);
   M68000_regs.isp=*(unsigned int  *)(cpu+0x48);
+  M68000_regs.irq = cpu[0x4c];
+  // stop flag?
 #endif
 
 #ifdef EMU_C68K
@@ -93,6 +101,8 @@ int PicoAreaUnpackCpu(unsigned char *cpu, int is_sub)
   memcpy(context->d,cpu,0x40);
   context->membase=0;
   context->pc = context->checkpc(*(unsigned int *)(cpu+0x40)); // Base pc
+  context->irq = cpu[0x4c];
+  context->stopped = cpu[0x4d];
 #endif
 
 #ifdef EMU_M68K
@@ -102,6 +112,8 @@ int PicoAreaUnpackCpu(unsigned char *cpu, int is_sub)
   m68ki_cpu_p->pc=*(unsigned int *)(cpu+0x40);
   m68k_set_reg(M68K_REG_SR, *(unsigned int *)(cpu+0x44));
   m68ki_cpu_p->sp[0]=*(unsigned int *)(cpu+0x48);
+  CPU_INT_LEVEL = cpu[0x4c] << 8;
+  CPU_STOPPED = cpu[0x4d];
   m68k_set_context(oldcontext);
 #endif
   return 0;

Pico/cd/Memory.c

@@ -991,6 +991,7 @@ static u32 PicoReadS68k32(u32 a)
 #endif
 
 
+#ifndef _ASM_CD_MEMORY_C
 /* check: jaguar xj 220 (draws entire world using decode) */
 static void decode_write8(u32 a, u8 d, int r3)
 {
@@ -1043,7 +1044,7 @@ static void decode_write16(u32 a, u16 d, int r3)
 
   //dprintf("FIXME: decode");
 }
-
+#endif
 
 // -----------------------------------------------------------------
 

Pico/cd/Pico.c

@@ -90,7 +90,9 @@ static __inline void SekRunS68k(int cyc)
 #define PS_STEP_M68K ((488<<16)/20) // ~24
 //#define PS_STEP_S68K 13
 
-#ifndef _ASM_CD_PICO_C
+#ifdef _ASM_CD_PICO_C
+void SekRunPS(int cyc_m68k, int cyc_s68k);
+#else
 static __inline void SekRunPS(int cyc_m68k, int cyc_s68k)
 {
   int cycn, cycn_s68k, cyc_do;
@@ -133,8 +135,6 @@ static __inline void SekRunPS(int cyc_m68k, int cyc_s68k)
 
 //fprintf(stderr, "==    end %3i/%3i    ==\n", d_cm, d_cs);
 }
-#else
-void SekRunPS(int cyc_m68k, int cyc_s68k);
 #endif
 
 

Pico/cd/gfx_cd.c

@@ -44,14 +44,14 @@ typedef struct
 	unsigned int Stamp_Map_Adr;
 	unsigned int Buffer_Adr;
 	unsigned int Vector_Adr;
-	unsigned int Jmp_Adr;
+	unsigned int Function;	// Jmp_Adr;
 	unsigned int Float_Part;
 	unsigned int Draw_Speed;
 
 	unsigned int XS;
 	unsigned int YS;
-	unsigned int DXS;
+	/*unsigned*/ int DXS;
-	unsigned int DYS;
+	/*unsigned*/ int DYS;
 	unsigned int XD;
 	unsigned int YD;
 	unsigned int XD_Mul;
@@ -65,13 +65,34 @@ static void gfx_cd_start(void)
 
 	dprintf("gfx_cd_start()");
 
+	rot_comp.XD_Mul = ((rot_comp.Reg_5C & 0x1f) + 1) * 4;
+	rot_comp.Function = (rot_comp.Reg_58 & 7) | (Pico_mcd->s68k_regs[3] & 0x18);	// Jmp_Adr
+	rot_comp.Buffer_Adr = (rot_comp.Reg_5E & 0xfff8) << 2;
+	rot_comp.YD = (rot_comp.Reg_60 >> 3) & 7;
+	rot_comp.Vector_Adr = (rot_comp.Reg_66 & 0xfffe) << 2;
+
 	upd_len = (rot_comp.Reg_62 >> 3) & 0x3f;
 	upd_len = Table_Rot_Time[upd_len];
-
 	rot_comp.Draw_Speed = rot_comp.Float_Part = upd_len;
 
 	rot_comp.Reg_58 |= 0x8000;	// Stamp_Size,  we start a new GFX operation
 
+	switch (rot_comp.Reg_58 & 6)	// Scr_16?
+	{
+		case 0:	// ?
+			rot_comp.Stamp_Map_Adr = (rot_comp.Reg_5A & 0xff80) << 2;
+			break;
+		case 2: // .Dot_32
+			rot_comp.Stamp_Map_Adr = (rot_comp.Reg_5A & 0xffe0) << 2;
+			break;
+		case 4: // .Scr_16
+			rot_comp.Stamp_Map_Adr = 0x20000;
+			break;
+		case 6: // .Scr_16_Dot_32
+			rot_comp.Stamp_Map_Adr = (rot_comp.Reg_5A & 0xe000) << 2;
+			break;
+	}
+
 	gfx_cd_update();
 }
 
@@ -88,9 +109,249 @@ static void gfx_completed(void)
 }
 
 
-//static void gfx_do(void)
+static void gfx_do(void)
-//{
+{
-//}
+	unsigned int eax, ebx, ecx, edx, esi, edi, pixel = 0;
+	unsigned int func = rot_comp.Function;
+	unsigned short *ptrs;
+
+	rot_comp.XD = rot_comp.Reg_60 & 7;
+	rot_comp.Buffer_Adr = ((rot_comp.Reg_5E & 0xfff8) + rot_comp.YD) << 2;
+	rot_comp.H_Dot = rot_comp.Reg_62 & 0x1ff;
+	ecx = *(unsigned int *)(Pico_mcd->word_ram2M + rot_comp.Vector_Adr);
+	edx = ecx >> 16;
+	ecx = (ecx & 0xffff) << 8;
+	edx <<= 8;
+	rot_comp.DXS = *(int *)(Pico_mcd->word_ram2M + rot_comp.Vector_Adr + 4);
+	rot_comp.DYS =  rot_comp.DXS >> 16;
+	rot_comp.DXS = (rot_comp.DXS << 16) >> 16; // sign extend
+	rot_comp.Vector_Adr += 8;
+	//if ((rot_comp.H_Dot & 0x1ff) == 0)
+	//	goto nothing_to_draw;
+
+	// MAKE_IMAGE_LINE
+	while (rot_comp.H_Dot)
+	{
+		if (func & 2)		// mode 32x32 dot
+		{
+			if (func & 4)	// 16x16 screen
+			{
+				eax = (ecx >> (11+5)) & 0x7f;
+				ebx = (edx >> (11-2)) & 0x3f80;
+			}
+			else		// 1x1 screen
+			{
+				eax = (ecx >> (11+5)) & 0x0f;
+				ebx = (edx >> (11+2)) & 0x38;
+			}
+		}
+		else			// mode 16x16 dot
+		{
+			if (func & 4)	// 16x16 screen
+			{
+				eax = (ecx >> (11+4)) & 0xff;
+				ebx = (edx >> (11-4)) & 0xff00;
+			}
+			else		// 1x1 screen
+			{
+				eax = (ecx >> (11+4)) & 0x0f;
+				ebx = (edx >> (11+0)) & 0xf0;
+			}
+		}
+		ebx += eax;
+
+		esi = rot_comp.Stamp_Map_Adr;
+		ptrs = (unsigned short *) (Pico_mcd->word_ram2M + rot_comp.Stamp_Map_Adr + ebx*2);
+		edi = ptrs[0] | (ptrs[1] << 16);
+
+		// MAKE_IMAGE_PIXEL
+		if (!(func & 1))	// NOT TILED
+		{
+			if (func & 4)	// 16x16 screen
+			{
+				if ((edx & 0x00800000) || (ecx & 0x00800000)) goto Pixel_Out;
+			}
+			else		// 1x1 screen
+			{
+				if ((edx & 0x00f80000) || (ecx & 0x00f80000)) goto Pixel_Out;
+			}
+		}
+		esi = edi;
+		edi >>= (11+1);
+		esi = (esi & 0x7ff) << 7;
+		if (!esi) goto Pixel_Out;
+		edi &= (0x1c>>1);
+		eax = ecx;
+		ebx = edx;
+		if (func & 2) edi |= 1;	// 32 dots?
+		switch (edi)
+		{
+			case 0x00:	// No_Flip_0, 16x16 dots
+				ebx = (ebx >> 9) & 0x3c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x1000;		// bswap
+				eax = ((eax >> 8) & 0x40) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x01:	// No_Flip_0, 32x32 dots
+				ebx = (ebx >> 9) & 0x7c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x1000;		// bswap
+				eax = ((eax >> 7) & 0x180) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x02:	// No_Flip_90, 16x16 dots
+				eax = (eax >> 9) & 0x3c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x2800;		// bswap
+				eax += ((ebx >> 8) & 0x40) ^ 0x40;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x03:	// No_Flip_90, 32x32 dots
+				eax = (eax >> 9) & 0x7c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x2800;		// bswap
+				eax += (ebx >> 7) & 0x180;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x04:	// No_Flip_180, 16x16 dots
+				ebx = ((ebx >> 9) & 0x3c) ^ 0x3c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x2800;		// bswap and flip
+				eax = (((eax >> 8) & 0x40) ^ 0x40) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x05:	// No_Flip_180, 32x32 dots
+				ebx = ((ebx >> 9) & 0x7c) ^ 0x7c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x2800;		// bswap and flip
+				eax = (((eax >> 7) & 0x180) ^ 0x180) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x06:	// No_Flip_270, 16x16 dots
+				eax = ((eax >> 9) & 0x3c) ^ 0x3c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x1000;		// bswap
+				eax += (ebx >> 8) & 0x40;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x07:	// No_Flip_270, 32x32 dots
+				eax = ((eax >> 9) & 0x7c) ^ 0x7c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x1000;		// bswap
+				eax += (ebx >> 7) & 0x180;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x08:	// Flip_0, 16x16 dots
+				ebx = ((ebx >> 9) & 0x3c) ^ 0x3c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x2800;		// bswap, flip
+				eax = (((eax >> 8) & 0x40) ^ 0x40) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x09:	// Flip_0, 32x32 dots
+				ebx = ((ebx >> 9) & 0x7c) ^ 0x7c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x2800;		// bswap, flip
+				eax = (((eax >> 7) & 0x180) ^ 0x180) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x0a:	// Flip_90, 16x16 dots
+				eax = ((eax >> 9) & 0x3c) ^ 0x3c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x2800;		// bswap, flip
+				eax += ((ebx >> 8) & 0x40) ^ 0x40;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x0b:	// Flip_90, 32x32 dots
+				eax = ((eax >> 9) & 0x7c) ^ 0x7c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x2800;		// bswap, flip
+				eax += ((ebx >> 7) & 0x180) ^ 0x180;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x0c:	// Flip_180, 16x16 dots
+				ebx = ((ebx >> 9) & 0x3c) ^ 0x3c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x1000;		// bswap
+				eax = ((eax >> 8) & 0x40) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x0d:	// Flip_180, 32x32 dots
+				ebx = ((ebx >> 9) & 0x7c) ^ 0x7c;
+				ebx += esi;
+				edi = (eax & 0x3800) ^ 0x1000;		// bswap
+				eax = ((eax >> 7) & 0x180) + ebx;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x0e:	// Flip_270, 16x16 dots
+				eax = (eax >> 9) & 0x3c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x1000;		// bswap, flip
+				eax += (ebx >> 8) & 0x40;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+			case 0x0f:	// Flip_270, 32x32 dots
+				eax = (eax >> 9) & 0x7c;
+				eax += esi;
+				edi = (ebx & 0x3800) ^ 0x1000;		// bswap, flip
+				eax += (ebx >> 7) & 0x180;
+				pixel = *(Pico_mcd->word_ram2M + (edi >> 12) + eax);
+				if (!(edi & 0x800)) pixel >>= 4;
+				break;
+		}
+
+Pixel_Out:
+//		pixel = 0;
+//Finish:
+		pixel &= 0x0f;
+		if (!pixel && (func & 0x18)) goto Next_Pixel;
+		esi = rot_comp.Buffer_Adr + ((rot_comp.XD>>1)^1);	// pixel addr
+		eax = *(Pico_mcd->word_ram2M + esi);			// old pixel
+		if (rot_comp.XD & 1)
+		{
+			if ((eax & 0x0f) && (func & 0x18) == 0x08) goto Next_Pixel; // underwrite
+			*(Pico_mcd->word_ram2M + esi) = pixel | (eax & 0xf0);
+		}
+		else
+		{
+			if ((eax & 0xf0) && (func & 0x18) == 0x08) goto Next_Pixel; // underwrite
+			*(Pico_mcd->word_ram2M + esi) = (pixel << 4) | (eax & 0xf);
+		}
+
+
+Next_Pixel:
+		ecx += rot_comp.DXS;
+		edx += rot_comp.DYS;
+		rot_comp.XD++;
+		if (rot_comp.XD >= 8)
+		{
+			rot_comp.Buffer_Adr += ((rot_comp.Reg_5C & 0x1f) + 1) << 5;
+			rot_comp.XD = 0;
+		}
+		rot_comp.H_Dot--;
+	}
+
+
+//nothing_to_draw:
+	rot_comp.YD++;
+	// rot_comp.V_Dot--; // will be done by caller
+}
 
 
 void gfx_cd_update(void)
@@ -102,7 +363,6 @@ void gfx_cd_update(void)
 
 	if (!*V_Dot)
 	{
-		// ...
 		gfx_completed();
 		return;
 	}
@@ -120,7 +380,9 @@ void gfx_cd_update(void)
 
 	while (jobs--)
 	{
-		// jmp [Jmp_Adr]:
+		if (PicoOpt & 0x1000)
+			gfx_do();	// jmp [Jmp_Adr]:
+
 		(*V_Dot)--;		// dec byte [V_Dot]
 
 		if (!*V_Dot)

Pico/cd/gfx_cd.h

@@ -16,7 +16,7 @@ typedef struct
 	unsigned int Stamp_Map_Adr;
 	unsigned int Buffer_Adr;
 	unsigned int Vector_Adr;
-	unsigned int Jmp_Adr;
+	unsigned int Function;		// Jmp_Adr;
 	unsigned int Float_Part;
 	unsigned int Draw_Speed;
 

platform/gp2x/menu.c

@@ -744,8 +744,9 @@ static void draw_cd_menu_options(int menu_sel, char *b_us, char *b_eu, char *b_j
 	gp2x_text_out8(tl_x, (y+=10), "CD LEDs                    %s", (currentConfig.EmuOpt &0x0400)?"ON":"OFF"); // 3
 	gp2x_text_out8(tl_x, (y+=10), "CDDA audio (using mp3s)    %s", (currentConfig.PicoOpt&0x0800)?"ON":"OFF"); // 4
 	gp2x_text_out8(tl_x, (y+=10), "PCM audio                  %s", (currentConfig.PicoOpt&0x0400)?"ON":"OFF"); // 5
-	gp2x_text_out8(tl_x, (y+=10), "Better sync (slow)         %s", (currentConfig.PicoOpt&0x2000)?"ON":"OFF"); // 6
+	gp2x_text_out8(tl_x, (y+=10), "ReadAhead buffer      %s", ra_buff); // 6
-	gp2x_text_out8(tl_x, (y+=10), "ReadAhead buffer      %s", ra_buff); // 7
+	gp2x_text_out8(tl_x, (y+=10), "Scale/Rot. fx (buggy,slow) %s", (currentConfig.PicoOpt&0x1000)?"ON":"OFF"); // 7
+	gp2x_text_out8(tl_x, (y+=10), "Better sync (slow)         %s", (currentConfig.PicoOpt&0x2000)?"ON":"OFF"); // 8
 	gp2x_text_out8(tl_x, (y+=10), "Done");
 
 	// draw cursor
@@ -761,7 +762,7 @@ static void draw_cd_menu_options(int menu_sel, char *b_us, char *b_eu, char *b_j
 
 static void cd_menu_loop_options(void)
 {
-	int menu_sel = 0, menu_sel_max = 8;
+	int menu_sel = 0, menu_sel_max = 9;
 	unsigned long inp = 0;
 	char bios_us[32], bios_eu[32], bios_jp[32], *bios, *p;
 
@@ -791,8 +792,7 @@ static void cd_menu_loop_options(void)
 				case  3: currentConfig.EmuOpt ^=0x0400; break;
 				case  4: currentConfig.PicoOpt^=0x0800; break;
 				case  5: currentConfig.PicoOpt^=0x0400; break;
-				case  6: currentConfig.PicoOpt^=0x2000; break;
+				case  6:
-				case  7:
 					if (inp & GP2X_LEFT) {
 						PicoCDBuffers >>= 1;
 						if (PicoCDBuffers < 64) PicoCDBuffers = 0;
@@ -802,7 +802,9 @@ static void cd_menu_loop_options(void)
 						if (PicoCDBuffers > 4096) PicoCDBuffers = 4096;
 					}
 					break;
-				case  8: return;
+				case  7: currentConfig.PicoOpt^=0x1000; break;
+				case  8: currentConfig.PicoOpt^=0x2000; break;
+				case  9: return;
 			}
 		}
 		if(inp & (GP2X_X|GP2X_A)) return;

platform/linux/Makefile

@@ -66,7 +66,7 @@ tidy:
 
 PicoDrive : $(OBJS)
 	@echo $@
-	@$(GCC) $(COPT) $(OBJS) ../gp2x/helix/helix_mp3_x86.a $(LDFLAGS) -lm -o $@
+	@$(GCC) $(COPT) $(OBJS) ../gp2x/helix/helix_mp3_x86.a $(LDFLAGS) -lm -Wl,-Map=PicoDrive.map -o $@
 
 
 ../../cpu/mz80/mz80.o : ../../cpu/mz80/mz80.asm

platform/linux/port_config.h

@@ -12,8 +12,8 @@
 // pico.c
 #define CAN_HANDLE_240_LINES	1
 
-#define dprintf(f,...) printf("%05i:%03i: " f "\n",Pico.m.frame_count,Pico.m.scanline,##__VA_ARGS__)
+//#define dprintf(f,...) printf("%05i:%03i: " f "\n",Pico.m.frame_count,Pico.m.scanline,##__VA_ARGS__)
-//#define dprintf(x...)
+#define dprintf(x...)
 
 #endif //PORT_CONFIG_H
 

platform/readme.txt

@@ -221,6 +221,7 @@ Changelog
   * Fixed a bug related to game specific config saving.
   * Frameskipper was skipping sound processing, what caused some audio desyncs. Fixed.
   * Fixed reset not working for some games.
+  + New assembly optimized memory handlers for CD (gives at least a few fps).
 
 1.201
   + Added basic cheat support (GameGenie and Genecyst patches).