Merge branch 'master' into fh/win32-winresize

core/hw/aica/aica.cpp

@@ -1,4 +1,5 @@
 #include "aica.h"
+#include "aica_if.h"
 #include "sgc_if.h"
 #include "aica_mem.h"
 #include <math.h>
@@ -180,6 +181,7 @@ template void WriteAicaReg<2>(u32 reg,u32 data);
 s32 libAICA_Init()
 {
 	init_mem();
+	aica_Init();
 
 	verify(sizeof(*CommonData)==0x508);
 	verify(sizeof(*DSPData)==0x15C8);
@@ -203,9 +205,12 @@ s32 libAICA_Init()
 	return rv_ok;
 }
 
-void libAICA_Reset(bool m)
+void libAICA_Reset(bool manual)
 {
+	if (!manual)
+		init_mem();
 	sgc_Init();
+	aica_Reset(manual);
 }
 
 void libAICA_Term()

core/hw/aica/aica_if.cpp

@@ -18,6 +18,7 @@ u32 VREG;//video reg =P
 u32 ARMRST;//arm reset reg
 u32 rtc_EN=0;
 int dma_sched_id;
+u32 RealTimeClock;
 
 u32 GetRTC_now()
 {
@@ -39,9 +40,9 @@ u32 ReadMem_aica_rtc(u32 addr,u32 sz)
 	switch( addr & 0xFF )
 	{
 	case 0:
-		return settings.dreamcast.RTC>>16;
+		return RealTimeClock>>16;
 	case 4:
-		return settings.dreamcast.RTC &0xFFFF;
+		return RealTimeClock &0xFFFF;
 	case 8:
 		return 0;
 	}
@@ -57,16 +58,16 @@ void WriteMem_aica_rtc(u32 addr,u32 data,u32 sz)
 	case 0:
 		if (rtc_EN)
 		{
-			settings.dreamcast.RTC&=0xFFFF;
+			RealTimeClock&=0xFFFF;
-			settings.dreamcast.RTC|=(data&0xFFFF)<<16;
+			RealTimeClock|=(data&0xFFFF)<<16;
 			rtc_EN=0;
 		}
 		return;
 	case 4:
 		if (rtc_EN)
 		{
-			settings.dreamcast.RTC&=0xFFFF0000;
+			RealTimeClock&=0xFFFF0000;
-			settings.dreamcast.RTC|= data&0xFFFF;
+			RealTimeClock|= data&0xFFFF;
 			//TODO: Clean the internal timer ?
 		}
 		return;
@@ -153,15 +154,12 @@ void WriteMem_aica_reg(u32 addr,u32 data,u32 sz)
 //Init/res/term
 void aica_Init()
 {
-	//mmnnn ? gotta fill it w/ something
+	RealTimeClock = GetRTC_now();
 }
 
 void aica_Reset(bool Manual)
 {
-	if (!Manual)
+	aica_Init();
-	{
-		aica_ram.Zero();
-	}
 }
 
 void aica_Term()

core/hw/aica/aica_if.h

@@ -3,6 +3,7 @@
 
 extern u32 VREG;
 extern VArray2 aica_ram;
+extern u32 RealTimeClock;
 u32 ReadMem_aica_rtc(u32 addr,u32 sz);
 void WriteMem_aica_rtc(u32 addr,u32 data,u32 sz);
 u32 ReadMem_aica_reg(u32 addr,u32 sz);
@@ -17,4 +18,4 @@ void aica_Term();
 
 void aica_sb_Init();
 void aica_sb_Reset(bool Manual);
-void aica_sb_Term();
+void aica_sb_Term();

core/hw/maple/maple_cfg.cpp

@@ -4,6 +4,7 @@
 #include "maple_devs.h"
 #include "maple_cfg.h"
 #include "cfg/cfg.h"
+#include "hw/naomi/naomi_cart.h"
 
 #define HAS_VMU
 /*
@@ -28,13 +29,31 @@ extern u16 kcode[4];
 extern u32 vks[4];
 extern s8 joyx[4],joyy[4];
 extern u8 rt[4],lt[4];
-extern bool naomi_test_button;
 
 u8 GetBtFromSgn(s8 val)
 {
 	return val+128;
 }
 
+u32 awave_button_mapping[] = {
+		AWAVE_SERVICE_KEY,	// DC_BTN_C
+		AWAVE_BTN1_KEY,		// DC_BTN_B
+		AWAVE_BTN0_KEY,		// DC_BTN_A
+		AWAVE_START_KEY,	// DC_BTN_START
+		AWAVE_UP_KEY,		// DC_DPAD_UP
+		AWAVE_DOWN_KEY,		// DC_DPAD_DOWN
+		AWAVE_LEFT_KEY,		// DC_DPAD_LEFT
+		AWAVE_RIGHT_KEY,	// DC_DPAD_RIGHT
+		AWAVE_TEST_KEY,		// DC_BTN_Z
+		AWAVE_BTN3_KEY,		// DC_BTN_Y
+		AWAVE_BTN2_KEY,		// DC_BTN_X
+		AWAVE_COIN_KEY,		// DC_BTN_D
+		// DC_DPAD2_UP
+		// DC_DPAD2_DOWN
+		// DC_DPAD2_LEFT
+		// DC_DPAD2_RIGHT
+};
+
 struct MapleConfigMap : IMapleConfigMap
 {
 	maple_device* dev;
@@ -59,17 +78,32 @@ struct MapleConfigMap : IMapleConfigMap
 
 		pjs->kcode=kcode[player_num];
 #if DC_PLATFORM == DC_PLATFORM_DREAMCAST
-		pjs->kcode |= 0xF901;
+		pjs->kcode |= 0xF901;		// mask off DPad2, C, D and Z
-#elif DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
-		if (naomi_test_button)
-			pjs->kcode &= ~(1 << 14);
-//		if (!(pjs->kcode & (1 << 9)))	// Hack (Y -> service btn)
-//			pjs->kcode &= ~(1 << 13);
-#endif
 		pjs->joy[PJAI_X1]=GetBtFromSgn(joyx[player_num]);
 		pjs->joy[PJAI_Y1]=GetBtFromSgn(joyy[player_num]);
 		pjs->trigger[PJTI_R]=rt[player_num];
 		pjs->trigger[PJTI_L]=lt[player_num];
+#elif DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
+		pjs->kcode = 0xFFFF;
+		for (int i = 0; i < 16; i++)
+		{
+			if ((kcode[player_num] & (1 << i)) == 0)
+				pjs->kcode &= ~awave_button_mapping[i];
+		}
+		pjs->joy[PJAI_X1] = GetBtFromSgn(joyx[player_num]);
+		if (NaomiGameInputs != NULL && NaomiGameInputs->axes[1].name != NULL && NaomiGameInputs->axes[1].type == Half)
+		{
+			// Driving games: put axis 2 on RT (accel) and axis 3 on LT (brake)
+			pjs->joy[PJAI_Y1] = rt[player_num];
+			pjs->joy[PJAI_X2] = lt[player_num];
+		}
+		else
+		{
+			pjs->joy[PJAI_Y1] = GetBtFromSgn(joyy[player_num]);
+			pjs->joy[PJAI_X2] = rt[player_num];
+			pjs->joy[PJAI_Y2] = lt[player_num];
+		}
+#endif
 	}
 	void SetImage(void* img)
 	{
@@ -77,6 +111,16 @@ struct MapleConfigMap : IMapleConfigMap
 	}
 };
 
+bool maple_atomiswave_coin_chute(int slot)
+{
+	for (int i = 0; i < 16; i++)
+	{
+		if ((kcode[slot] & (1 << i)) == 0 && awave_button_mapping[i] == AWAVE_COIN_KEY)
+			return true;
+	}
+	return false;
+}
+
 void mcfg_Create(MapleDeviceType type, u32 bus, u32 port, s32 player_num = -1)
 {
 	if (MapleDevices[bus][port] != NULL)
@@ -100,17 +144,32 @@ void mcfg_CreateAtomisWaveControllers()
 	// Then other devices on port 2 and 3 for analog axes, light guns, ...
 	mcfg_Create(MDT_SegaController, 0, 5);
 	mcfg_Create(MDT_SegaController, 1, 5);
-//	mcfg_Create(MDT_SegaController, 2, 5, 0);
+	if (NaomiGameInputs != NULL && NaomiGameInputs->axes[0].name != NULL)
-//	mcfg_Create(MDT_SegaController, 3, 5, 1);
+	{
-//	mcfg_Create(MDT_LightGun, 2, 5, 0);
+		// Game needs analog axes
-//	mcfg_Create(MDT_LightGun, 3, 5, 1);
+		mcfg_Create(MDT_SegaController, 2, 5, 0);
-//	mcfg_Create(MDT_Mouse, 2, 5, 0);
+		mcfg_Create(MDT_SegaController, 3, 5, 1);
-	// Guilty Gear Isuka (4P)		needs 4 std controllers
+		// Faster Than Speed			needs 1 std controller on port 0 (digital inputs) and one on port 2 (analog axes)
-	// Faster Than Speed			needs 1 std controller on port 0 (digital inputs) and one on port 2 (analog axes)
+		// Maximum Speed				same
-	// Maximum Speed				same
+	}
-	// Clay Challenge				needs 2 std controllers on port 0 & 1 (digital in) and light guns on port 2 & 3
+	else if (settings.input.JammaSetup == 1)
-	// Sports Shooting				same
+	{
-	// Sega Bass Fishing Challenge  needs a mouse (track-ball) on port 3
+		// 4 players
+		mcfg_Create(MDT_SegaController, 2, 5);
+		mcfg_Create(MDT_SegaController, 3, 5);
+	}
+	else if (settings.input.JammaSetup == 5)
+	{
+		// Clay Challenge				needs 2 std controllers on port 0 & 1 (digital in) and light guns on port 2 & 3
+		// Sports Shooting				same
+		mcfg_Create(MDT_LightGun, 2, 5, 0);
+		mcfg_Create(MDT_LightGun, 3, 5, 1);
+	}
+	else if (settings.input.JammaSetup == 3)
+	{
+		// Sega Bass Fishing Challenge  needs a mouse (track-ball) on port 2
+		mcfg_Create(MDT_Mouse, 2, 5, 0);
+	}
 }
 
 void mcfg_CreateDevices()

core/hw/maple/maple_cfg.h

@@ -68,3 +68,5 @@ void mcfg_CreateAtomisWaveControllers();
 void mcfg_DestroyDevices();
 void mcfg_SerializeDevices(void **data, unsigned int *total_size);
 void mcfg_UnserializeDevices(void **data, unsigned int *total_size);
+
+bool maple_atomiswave_coin_chute(int slot);

core/hw/maple/maple_devs.cpp

@@ -7,6 +7,7 @@
 #include "hw/naomi/naomi.h"
 #include "hw/naomi/naomi_cart.h"
 #include "hw/pvr/spg.h"
+#include "input/gamepad.h"
 #include <time.h>
 
 #include "deps/zlib/zlib.h"
@@ -198,6 +199,35 @@ struct maple_base: maple_device
 */
 struct maple_sega_controller: maple_base
 {
+	virtual u32 get_capabilities() {
+		// byte 0: 0  0  0  0  0  0  0  0
+		// byte 1: 0  0  a5 a4 a3 a2 a1 a0
+		// byte 2: R2 L2 D2 U2 D  X  Y  Z
+		// byte 3: R  L  D  U  St A  B  C
+
+		return 0xfe060f00;	// 4 analog axes (0-3) X Y A B Start U D L R
+	}
+
+	virtual u32 transform_kcode(u32 kcode) {
+		return kcode;
+	}
+
+	virtual u32 get_analog_axis(int index, const PlainJoystickState &pjs) {
+		switch (index)
+		{
+		case 0:
+			return pjs.trigger[PJTI_R];		// Right trigger
+		case 1:
+			return pjs.trigger[PJTI_L];		// Left trigger
+		case 2:
+			return pjs.joy[PJAI_X1];		// Stick X
+		case 3:
+			return pjs.joy[PJAI_Y1];		// Stick Y
+		default:
+			return 0x80;					// unused
+		}
+	}
+
 	virtual MapleDeviceType get_device_type()
 	{
 		return MDT_SegaController;
@@ -215,9 +245,9 @@ struct maple_sega_controller: maple_base
 
 			//struct data
 			//3*4
-			w32( 0xfe060f00);
+			w32(get_capabilities());
-			w32( 0);
+			w32(0);
-			w32( 0);
+			w32(0);
 
 			//1	area code
 			w8(0xFF);
@@ -250,26 +280,26 @@ struct maple_sega_controller: maple_base
 
 				//state data
 				//2 key code
-				w16(pjs.kcode);
+				w16(transform_kcode(pjs.kcode));
 
 				//triggers
 				//1 R
-				w8(pjs.trigger[PJTI_R]);
+				w8(get_analog_axis(0, pjs));
 				//1 L
-				w8(pjs.trigger[PJTI_L]);
+				w8(get_analog_axis(1, pjs));
 
 				//joyx
 				//1
-				w8(pjs.joy[PJAI_X1]);
+				w8(get_analog_axis(2, pjs));
 				//joyy
 				//1
-				w8(pjs.joy[PJAI_Y1]);
+				w8(get_analog_axis(3, pjs));
 
 				//not used
 				//1
-				w8(0x80);
+				w8(get_analog_axis(4, pjs));
 				//1
-				w8(0x80);
+				w8(get_analog_axis(5, pjs));
 			}
 
 			return MDRS_DataTransfer;
@@ -281,6 +311,38 @@ struct maple_sega_controller: maple_base
 	}
 };
 
+struct maple_atomiswave_controller: maple_sega_controller
+{
+	virtual u32 get_capabilities() override {
+		// byte 0: 0  0  0  0  0  0  0  0
+		// byte 1: 0  0  a5 a4 a3 a2 a1 a0
+		// byte 2: R2 L2 D2 U2 D  X  Y  Z
+		// byte 3: R  L  D  U  St A  B  C
+
+		return 0xff663f00;	// 6 analog axes, X Y L2/D2(?) A B C Start U D L R
+	}
+
+	virtual u32 transform_kcode(u32 kcode) override {
+		return kcode | AWAVE_TRIGGER_KEY;
+	}
+
+	virtual u32 get_analog_axis(int index, const PlainJoystickState &pjs) override {
+		switch (index)
+		{
+		case 2:
+			return pjs.joy[PJAI_X1];
+		case 3:
+			return pjs.joy[PJAI_Y1];
+		case 4:
+			return pjs.joy[PJAI_X2];
+		case 5:
+			return pjs.joy[PJAI_Y2];
+		default:
+			return 0x80;
+		}
+	}
+};
+
 /*
 	Sega Dreamcast Visual Memory Unit
 	This is pretty much done (?)
@@ -1267,6 +1329,10 @@ struct maple_mouse : maple_base
 
 struct maple_lightgun : maple_base
 {
+	virtual u32 transform_kcode(u32 kcode) {
+		return kcode | 0xFF01;
+	}
+
 	virtual MapleDeviceType get_device_type()
 	{
 		return MDT_LightGun;
@@ -1315,21 +1381,13 @@ struct maple_lightgun : maple_base
 			PlainJoystickState pjs;
 			config->GetInput(&pjs);
 
-			// Also use the mouse buttons
-			if (!(mo_buttons & 4))	// Left button
-				pjs.kcode &= ~4;	// A
-			if (!(mo_buttons & 2))	// Right button
-				pjs.kcode &= ~2;	// B
-			if (!(mo_buttons & 8))	// Wheel button
-				pjs.kcode &= ~8;	// Start
-
 			//caps
 			//4
 			w32(MFID_0_Input);
 
 			//state data
 			//2 key code
-			w16(pjs.kcode | 0xFF01);
+			w16(transform_kcode(pjs.kcode));
 
 			//not used
 			//2
@@ -1355,6 +1413,13 @@ struct maple_lightgun : maple_base
 	}
 };
 
+struct atomiswave_lightgun : maple_lightgun
+{
+	virtual u32 transform_kcode(u32 kcode) override {
+		return (kcode & AWAVE_TRIGGER_KEY) == 0 ? ~AWAVE_BTN0_KEY : ~0;
+	}
+};
+
 extern u16 kcode[4];
 extern s8 joyx[4],joyy[4];
 extern u8 rt[4], lt[4];
@@ -1381,20 +1446,29 @@ static u16 getRightTriggerAxis()
 	return rt[0] << 8;
 }
 
-NaomiInputMapping Naomi_Mapping = {
+u32 naomi_button_mapping[] = {
-	{ getJoystickXAxis, getJoystickYAxis, getRightTriggerAxis, getLeftTriggerAxis },
+		NAOMI_SERVICE_KEY,	// DC_BTN_C
-	{ 0,    0,    0,    0,    0,    0,    0,    0,    0, 1,    1,    0, 0 },
+		NAOMI_BTN1_KEY,		// DC_BTN_B
-	{ 0x40, 0x01, 0x02, 0x80, 0x20, 0x10, 0x08, 0x04, 0, 0x80, 0x40, 0, 0 },
+		NAOMI_BTN0_KEY,		// DC_BTN_A
-//  SERVICE BTN1  BTN0  START UP    DOWN  LEFT  RIGHT    BTN2  BTN3
+		NAOMI_START_KEY,	// DC_BTN_START
+		NAOMI_UP_KEY,		// DC_DPAD_UP
+		NAOMI_DOWN_KEY,		// DC_DPAD_DOWN
+		NAOMI_LEFT_KEY,		// DC_DPAD_LEFT
+		NAOMI_RIGHT_KEY,	// DC_DPAD_RIGHT
+		NAOMI_TEST_KEY,		// DC_BTN_Z
+		NAOMI_BTN3_KEY,		// DC_BTN_Y
+		NAOMI_BTN2_KEY,		// DC_BTN_X
+		NAOMI_COIN_KEY,		// DC_BTN_D
+		// DC_DPAD2_UP
+		// DC_DPAD2_DOWN
+		// DC_DPAD2_LEFT
+		// DC_DPAD2_RIGHT
 };
-
 /*
  * Sega JVS I/O board
 */
-bool coin_chute;
+static bool old_coin_chute[4];
-static bool old_coin_chute;
+static int coin_count[4];
-static int coin_count;
-bool naomi_test_button = false;
 
 struct maple_naomi_jamma;
 
@@ -2313,63 +2387,56 @@ u32 jvs_io_board::handle_jvs_message(u8 *buffer_in, u32 length_in, u8 *buffer_ou
 					{
 						JVS_STATUS1();	// report byte
 
-						LOGJVS("btns ");
+						u16 buttons[4] = { 0 };
-						JVS_OUT(naomi_test_button ? 0x80 : 0x00); // test, tilt1, tilt2, tilt3, unused, unused, unused, unused
+						for (int player = 0; player < buffer_in[cmdi + 1] && first_player + player < ARRAY_SIZE(kcode); player++)
-						// FIXME in-lst mapping
-						u8 buttons[8] = { 0 };
-						u32 keycode = ~kcode[0];
-						for (int i = 0; i < 16; i++)
-							if ((keycode & (1 << i)) != 0)
-							{
-								buttons[Naomi_Mapping.button_mapping_byte[i]] |= Naomi_Mapping.button_mapping_mask[i];
-							}
-						for (int player = 0; player < buffer_in[cmdi + 1]; player++)
 						{
-							u8 *cur_btns = &buttons[(first_player + player) * 2];
+							u32 keycode = ~kcode[first_player + player];
-							LOGJVS("P%d %02x ", player + 1 + first_player, cur_btns[0]);
+							for (int i = 0; i < 16; i++)
-							JVS_OUT(cur_btns[0]);
-							if (buffer_in[cmdi + 2] == 2)
 							{
-								LOGJVS("%02x ", cur_btns[1]);
+								if ((keycode & (1 << i)) != 0)
-								JVS_OUT(cur_btns[1]);
+									buttons[player] |= naomi_button_mapping[i];
+							}
+						}
+
+						LOGJVS("btns ");
+						JVS_OUT((buttons[0] & NAOMI_TEST_KEY) ? 0x80 : 0x00); // test, tilt1, tilt2, tilt3, unused, unused, unused, unused
+						for (int player = 0; player < buffer_in[cmdi + 1]; player++)
+						{
+							u16 cur_btns = first_player + player < ARRAY_SIZE(buttons) ? buttons[first_player + player] : 0;
+							LOGJVS("P%d %02x ", player + 1 + first_player, cur_btns >> 8);
+							JVS_OUT(cur_btns >> 8);
+							if (buffer_in[cmdi + 2] == 2)
+							{
+								LOGJVS("%02x ", cur_btns & 0xFF);
+								JVS_OUT(cur_btns);
 							}
 						}
-//								for (int player = 0; player < jvs_request[channel][cmdi + 1]; player++)
-//								{
-//									u32 keycode = ~kcode[player];
-//									if (keycode & DC_BTN_C)
-//										keycode |= 0xFFff;
-//
-//									if (jvs_request[channel][cmdi + 2] == 1)
-//										JVS_OUT(keycode);
-//									else
-//										w16(keycode);
-//								}
 						cmdi += 3;
 					}
 					break;
 
 				case 0x21:	// Read coins
 					{
-						if (coin_chute && !old_coin_chute)
-							coin_count++;
-						old_coin_chute = coin_chute;
 						JVS_STATUS1();	// report byte
 						LOGJVS("coins ");
 						for (int slot = 0; slot < buffer_in[cmdi + 1]; slot++)
 						{
-							if (slot == 0)
+							bool coin_chute = false;
+							u32 keycode = ~kcode[first_player + slot];
+							for (int i = 0; i < 16 && !coin_chute; i++)
 							{
-								LOGJVS("0:%d ", coin_count);
+								if (naomi_button_mapping[i] == NAOMI_COIN_KEY && (keycode & (1 << i)) != 0)
-								JVS_OUT((coin_count >> 8) & 0x3F);		// status (2 highest bits, 0: normal), coin count MSB
+									coin_chute = true;
-								JVS_OUT(coin_count);						// coin count LSB
-							}
-							else
-							{
-								LOGJVS("%d:0 ", slot);
-								JVS_OUT(0);
-								JVS_OUT(0);
 							}
+							if (coin_chute && !old_coin_chute[first_player + slot])
+								coin_count[first_player + slot] += 1;
+							old_coin_chute[first_player + slot] = coin_chute;
+
+							LOGJVS("%d:%d ", slot + 1 + first_player, coin_count[first_player + slot]);
+							// status (2 highest bits, 0: normal), coin count MSB
+							JVS_OUT((coin_count[first_player + slot] >> 8) & 0x3F);
+							// coin count LSB
+							JVS_OUT(coin_count[first_player + slot]);
 						}
 						cmdi += 2;
 					}
@@ -2393,34 +2460,50 @@ u32 jvs_io_board::handle_jvs_message(u8 *buffer_in, u32 length_in, u8 *buffer_ou
 							}
 							LOGJVS("x,y:%4x,%4x ", x, y);
 							JVS_OUT(x >> 8);		// X, MSB
-							JVS_OUT(x);			// X, LSB
+							JVS_OUT(x);				// X, LSB
 							JVS_OUT(y >> 8);		// Y, MSB
-							JVS_OUT(y);			// Y, LSB
+							JVS_OUT(y);				// Y, LSB
 							axis = 2;
 						}
 
+						int full_axis_count = 0;
+						int half_axis_count = 0;
 						for (; axis < buffer_in[cmdi + 1]; axis++)
 						{
-							// FIXME Need to know how many axes per player for proper mapping
 							u16 axis_value;
-							if (axis + first_player * 4 < 8 && Naomi_Mapping.axis[axis + first_player * 4] != NULL)
+							if (NaomiGameInputs != NULL
-								axis_value = Naomi_Mapping.axis[axis + first_player * 4]();
+									&& axis < ARRAY_SIZE(NaomiGameInputs->axes)
+									&& NaomiGameInputs->axes[axis].name != NULL
+									&& NaomiGameInputs->axes[axis].type == Half)
+							{
+								if (half_axis_count == 0)
+									axis_value = rt[first_player] << 8;
+								else if (half_axis_count == 1)
+									axis_value = lt[first_player] << 8;
+								else
+									axis_value = 0;
+								half_axis_count++;
+							}
 							else
 							{
-								switch (axis) {
+								switch (full_axis_count) {
 								case 0:
-									axis_value = (joyx[first_player + axis / 4] + 128) << 8;
+									axis_value = (joyx[first_player] + 128) << 8;
 									break;
 								case 1:
-									axis_value = (joyy[first_player + axis / 4] + 128) << 8;
+									axis_value = (joyy[first_player] + 128) << 8;
-									break;
-								case 2:
-									axis_value = rt[first_player + axis / 4] << 8;
-									break;
-								case 3:
-									axis_value = lt[first_player + axis / 4] << 8;
 									break;
+								// TODO right analog stick
+//								case 2:
+//									axis_value = (joyrx[first_player] + 128) << 8;
+//									break;
+//								case 3:
+//									axis_value = (joyry[first_player] + 128) << 8;
+//									break;
+								default:
+									axis_value = 128;
 								}
+								full_axis_count++;
 							}
 							LOGJVS("%d:%4x ", axis, axis_value);
 							JVS_OUT(axis_value >> 8);
@@ -2493,8 +2576,8 @@ u32 jvs_io_board::handle_jvs_message(u8 *buffer_in, u32 length_in, u8 *buffer_ou
 					break;
 
 				case 0x30:	// substract coin
-					if (buffer_in[cmdi + 1] == 1)
+					if (buffer_in[cmdi + 1] > 0 && first_player + buffer_in[cmdi + 1] - 1 < ARRAY_SIZE(coin_count))
-						coin_count -= (buffer_in[cmdi + 2] << 8) + buffer_in[cmdi + 3];
+						coin_count[first_player + buffer_in[cmdi + 1] - 1] -= (buffer_in[cmdi + 2] << 8) + buffer_in[cmdi + 3];
 					JVS_STATUS1();	// report byte
 					cmdi += 4;
 					break;
@@ -2542,7 +2625,11 @@ maple_device* maple_Create(MapleDeviceType type)
 	switch(type)
 	{
 	case MDT_SegaController:
-		rv=new maple_sega_controller();
+#if DC_PLATFORM != DC_PLATFORM_ATOMISWAVE
+		rv = new maple_sega_controller();
+#else
+		rv = new maple_atomiswave_controller();
+#endif
 		break;
 
 	case MDT_Microphone:
@@ -2566,7 +2653,11 @@ maple_device* maple_Create(MapleDeviceType type)
 		break;
 
 	case MDT_LightGun:
+#if DC_PLATFORM != DC_PLATFORM_ATOMISWAVE
 		rv = new maple_lightgun();
+#else
+		rv = new atomiswave_lightgun();
+#endif
 		break;
 
 	case MDT_NaomiJamma:

core/hw/naomi/naomi.cpp

@@ -6,6 +6,7 @@
 #include "hw/holly/sb.h"
 #include "hw/sh4/sh4_mem.h"
 #include "hw/holly/holly_intc.h"
+#include "hw/maple/maple_cfg.h"
 
 #include "naomi.h"
 #include "naomi_cart.h"
@@ -634,7 +635,6 @@ void Update_naomi()
 }
 
 static u8 aw_maple_devs;
-extern bool coin_chute;
 
 u32 libExtDevice_ReadMem_A0_006(u32 addr,u32 size) {
 	addr &= 0x7ff;
@@ -653,12 +653,13 @@ u32 libExtDevice_ReadMem_A0_006(u32 addr,u32 size) {
 			aw_ram_test_skipped = true;
 			return 0;
 		}
-		if (coin_chute)
 		{
-			// FIXME Coin Error if coin_chute is set for too long
+			u8 coin_input = 0xF;
-			return 0xE;
+			for (int slot = 0; slot < 4; slot++)
+				if (maple_atomiswave_coin_chute(slot))
+					coin_input &= ~(1 << slot);
+			return coin_input;
 		}
-		return 0xF;
 
 	case 0x284:		// Atomiswave maple devices
 		// ddcc0000 where cc/dd are the types of devices on maple bus 2 and 3:

core/hw/naomi/naomi_cart.cpp

@@ -32,7 +32,7 @@ fd_t*	RomCacheMap = NULL;
 u32		RomCacheMapCount;
 
 char naomi_game_id[33];
-InputDescriptors *naomi_game_inputs;
+InputDescriptors *NaomiGameInputs;
 u8 *naomi_default_eeprom;
 
 extern RomChip sys_rom;
@@ -246,7 +246,7 @@ static bool naomi_cart_LoadZip(char *filename)
 		break;
 	}
 	CurrentCartridge->SetKey(game->key);
-	naomi_game_inputs = game->inputs;
+	NaomiGameInputs = game->inputs;
 
 	for (int romid = 0; game->blobs[romid].filename != NULL; romid++)
 	{

core/hw/naomi/naomi_cart.h

@@ -110,6 +110,6 @@ struct InputDescriptors
    AxisDescriptor axes[8];
 };
 
-extern InputDescriptors *naomi_game_inputs;
+extern InputDescriptors *NaomiGameInputs;
 
 #endif //NAOMI_CART_H

core/hw/sh4/dyna/blockmanager.cpp

@@ -304,6 +304,8 @@ void bm_Rebuild()
 {
 	return;
 
+	die("this is broken in multiple levels, including compile options");
+
 	void RASDASD();
 	RASDASD();
 
@@ -321,7 +323,7 @@ void bm_Rebuild()
 			//constprop(all_blocks[i]);
 //#endif
 		}
-		ngen_Compile(all_blocks[i],false,false,all_blocks[i]->staging_runs>0,do_opts);
+		ngen_Compile(all_blocks[i],NoCheck,false,all_blocks[i]->staging_runs>0,do_opts);
 
 		blkmap.insert(all_blocks[i]);
 		verify(bm_GetBlock((RuntimeBlockInfo*)all_blocks[i]->code)==all_blocks[i]);

core/hw/sh4/dyna/driver.cpp

@@ -86,6 +86,15 @@ void recSh4_Run()
 
 	sh4_dyna_rcb=(u8*)&Sh4cntx + sizeof(Sh4cntx);
 	printf("cntx // fpcb offset: %td // pc offset: %td // pc %08X\n",(u8*)&sh4rcb.fpcb - sh4_dyna_rcb, (u8*)&sh4rcb.cntx.pc - sh4_dyna_rcb,sh4rcb.cntx.pc);
+
+	if (!settings.dynarec.safemode)
+		printf("Warning: Dynarec safe mode is off\n");
+
+	if (settings.dynarec.unstable_opt)
+		printf("Warning: Unstable optimizations is on\n");
+
+	if (settings.dynarec.SmcCheckLevel != FullCheck)
+		printf("Warning: SMC check mode is %d\n", settings.dynarec.SmcCheckLevel);
 	
 	verify(rcb_noffs(&next_pc)==-184);
 	ngen_mainloop(sh4_dyna_rcb);
@@ -119,34 +128,51 @@ u32 emit_FreeSpace()
 }
 
 
-bool DoCheck(u32 pc)
+SmcCheckEnum DoCheck(u32 pc)
 {
-	if (IsOnRam(pc))
-	{
-		if (!settings.dynarec.unstable_opt)
-			return true;
 
-		pc&=0xFFFFFF;
+	switch (settings.dynarec.SmcCheckLevel) {
-		switch(pc)
-		{
-			//DOA2LE
-			case 0x3DAFC6:
-			case 0x3C83F8:
 
-			//Shenmue 2
+		// Heuristic-elimintaed FastChecks
-			case 0x348000:
+		case NoCheck: {
-				
+			if (IsOnRam(pc))
-			//Shenmue
+			{
-			case 0x41860e:
+				pc&=0xFFFFFF;
-			
+				switch(pc)
+				{
+					//DOA2LE
+					case 0x3DAFC6:
+					case 0x3C83F8:
 
-				return true;
+					//Shenmue 2
+					case 0x348000:
+						
+					//Shenmue
+					case 0x41860e:
+					
 
-			default:
+						return FastCheck;
-				return false;
+
+					default:
+						return NoCheck;
+				}
+			}
+			return NoCheck;
 		}
+		break;
+
+		// Fast Check everything
+		case FastCheck:
+			return FastCheck;
+
+		// Full Check everything
+		case FullCheck:
+			return FullCheck;
+
+		default:
+			die("Unhandled settings.dynarec.SmcCheckLevel");
+			return FullCheck;
 	}
-	return false;
 }
 
 void AnalyseBlock(RuntimeBlockInfo* blk);

core/hw/sh4/dyna/ngen.h

@@ -85,7 +85,7 @@ u32 DYNACALL rdv_DoInterrupts_pc(u32 pc);
 void ngen_init();
 
 //Called to compile a block
-void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool staging,bool optimise);
+void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging,bool optimise);
 
 //Called when blocks are reseted
 void ngen_ResetBlocks();

core/hw/sh4/interpr/sh4_interpreter.cpp

@@ -232,7 +232,7 @@ int AicaUpdate(int tag, int c, int j)
 
 int DreamcastSecond(int tag, int c, int j)
 {
-	settings.dreamcast.RTC++;
+	RealTimeClock++;
 
 #if 1 //HOST_OS==OS_WINDOWS
 	prof_periodical();

core/hw/sh4/modules/ccn.cpp

@@ -12,6 +12,8 @@
 
 //Types
 
+#define printf_smc(...) // printf
+
 
 u32 CCN_QACR_TR[2];
 
@@ -72,13 +74,18 @@ void CCN_CCR_write(u32 addr, u32 value)
 	temp.reg_data=value;
 
 
-	//what is 0xAC13DBF8 from ?
+	if (temp.ICI) {
-	if (temp.ICI && curr_pc!=0xAC13DBF8)
+		printf_smc("Sh4: i-cache invalidation %08X\n",curr_pc);
-	{
+
-		//printf("Sh4: i-cache invalidation %08X\n",curr_pc);
+		if (settings.dynarec.SmcCheckLevel != FullCheck) {
-		// Shikigami No Shiro II sets ICI frequently
+			//TODO: Add skip/check vectors for Shikigami No Shiro II (uses ICI frequently)
-		// Any reason to flush the dynarec cache for this?
+			//which game is 0xAC13DBF8 from ?
-		//sh4_cpu.ResetCache();
+			if (curr_pc != 0xAC13DBF8)
+			{
+				printf("Sh4: code cache clear (ICI) pc: %08X\n",curr_pc);
+				sh4_cpu.ResetCache();
+			}
+		}
 	}
 
 	temp.ICI=0;

core/input/gamepad_device.cpp

@@ -51,7 +51,39 @@ bool GamepadDevice::gamepad_btn_input(u32 code, bool pressed)
 	if (key < 0x10000)
 	{
 		if (pressed)
+		{
 			kcode[_maple_port] &= ~(u16)key;
+			// Avoid two opposite dpad keys being pressed simultaneously
+			switch (key)
+			{
+			case DC_DPAD_UP:
+				kcode[_maple_port] |= (u16)DC_DPAD_DOWN;
+				break;
+			case DC_DPAD_DOWN:
+				kcode[_maple_port] |= (u16)DC_DPAD_UP;
+				break;
+			case DC_DPAD_LEFT:
+				kcode[_maple_port] |= (u16)DC_DPAD_RIGHT;
+				break;
+			case DC_DPAD_RIGHT:
+				kcode[_maple_port] |= (u16)DC_DPAD_LEFT;
+				break;
+			case DC_DPAD2_UP:
+				kcode[_maple_port] |= (u16)DC_DPAD2_DOWN;
+				break;
+			case DC_DPAD2_DOWN:
+				kcode[_maple_port] |= (u16)DC_DPAD2_UP;
+				break;
+			case DC_DPAD2_LEFT:
+				kcode[_maple_port] |= (u16)DC_DPAD2_RIGHT;
+				break;
+			case DC_DPAD2_RIGHT:
+				kcode[_maple_port] |= (u16)DC_DPAD2_LEFT;
+				break;
+			default:
+				break;
+			}
+		}
 		else
 			kcode[_maple_port] |= (u16)key;
 	}

core/linux-dist/x11.cpp

@@ -84,8 +84,6 @@ Atom wmDeleteMessage;
 void* x11_vis;
 
 extern bool dump_frame_switch;
-extern bool naomi_test_button;
-extern bool coin_chute;
 
 void dc_exit(void);
 
@@ -275,16 +273,6 @@ void input_x11_handle()
 							x11_fullscreen = !x11_fullscreen;
 							x11_window_set_fullscreen(x11_fullscreen);
 						}
-#if DC_PLATFORM == DC_PLATFORM_NAOMI || DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
-						else if (e.xkey.keycode == KEY_F8)
-						{
-							coin_chute = e.type == KeyPress;
-						}
-						else if (e.xkey.keycode == KEY_F7)
-						{
-							naomi_test_button = e.type == KeyPress;
-						}
-#endif
 					}
 				}
 				break;

core/nullDC.cpp

@@ -134,7 +134,7 @@ void LoadSpecialSettings()
 	safemode_game = false;
 	tr_poly_depth_mask_game = false;
 	extra_depth_game = false;
-	
+
 	if (reios_windows_ce)
 	{
 		printf("Enabling Extra depth scaling for Windows CE games\n");
@@ -190,7 +190,7 @@ void LoadSpecialSettings()
 	}
 #elif DC_PLATFORM == DC_PLATFORM_NAOMI || DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
 	printf("Game ID is [%s]\n", naomi_game_id);
-	
+
 	if (!strcmp("METAL SLUG 6", naomi_game_id) || !strcmp("WAVE RUNNER GP", naomi_game_id))
 	{
 		printf("Enabling Dynarec safe mode for game %s\n", naomi_game_id);
@@ -213,12 +213,14 @@ void LoadSpecialSettings()
 		printf("Enabling JVS rotary encoders for game %s\n", naomi_game_id);
 		settings.input.JammaSetup = 2;
 	}
-	else if (!strcmp("POWER STONE 2 JAPAN", naomi_game_id))
+	else if (!strcmp("POWER STONE 2 JAPAN", naomi_game_id)		// Naomi
+			|| !strcmp("GUILTY GEAR isuka", naomi_game_id))		// AW
 	{
 		printf("Enabling 4-player setup for game %s\n", naomi_game_id);
 		settings.input.JammaSetup = 1;
 	}
-	else if (!strcmp("SEGA MARINE FISHING JAPAN", naomi_game_id))
+	else if (!strcmp("SEGA MARINE FISHING JAPAN", naomi_game_id)
+				|| !strcmp(naomi_game_id, "BASS FISHING SIMULATOR VER.A"))	// AW
 	{
 		printf("Enabling specific JVS setup for game %s\n", naomi_game_id);
 		settings.input.JammaSetup = 3;
@@ -228,9 +230,11 @@ void LoadSpecialSettings()
 		printf("Enabling specific JVS setup for game %s\n", naomi_game_id);
 		settings.input.JammaSetup = 4;
 	}
-	else if (!strcmp("NINJA ASSAULT", naomi_game_id))
+	else if (!strcmp("NINJA ASSAULT", naomi_game_id)
+				|| !strcmp(naomi_game_id, "Sports Shooting USA")	// AW
+				|| !strcmp(naomi_game_id, "SEGA CLAY CHALLENGE"))	// AW
 	{
-		printf("Enabling specific JVS setup for game %s\n", naomi_game_id);
+		printf("Enabling lightgun setup for game %s\n", naomi_game_id);
 		settings.input.JammaSetup = 5;
 	}
 	else if (!strcmp(" BIOHAZARD  GUN SURVIVOR2", naomi_game_id))
@@ -309,11 +313,12 @@ int dc_start_game(const char *path)
 	{
 		InitSettings();
 		LoadSettings(false);
-		settings.dreamcast.RTC = GetRTC_now();	// FIXME This shouldn't be in settings anymore
 #if DC_PLATFORM == DC_PLATFORM_DREAMCAST
 		if (!settings.bios.UseReios)
 #endif
-			LoadRomFiles(get_readonly_data_path(DATA_PATH));
+			if (!LoadRomFiles(get_readonly_data_path(DATA_PATH)))
+				return -5;
+
 #if DC_PLATFORM == DC_PLATFORM_DREAMCAST
 		if (path == NULL)
 		{
@@ -345,7 +350,6 @@ int dc_start_game(const char *path)
 		return 0;
 	}
 
-	settings.dreamcast.RTC = GetRTC_now();	// FIXME This shouldn't be in settings anymore
 	if (settings.bios.UseReios || !LoadRomFiles(get_readonly_data_path(DATA_PATH)))
 	{
 #ifdef USE_REIOS
@@ -485,7 +489,6 @@ void dc_exit()
 
 void InitSettings()
 {
-	settings.dreamcast.RTC			= GetRTC_now();
 	settings.dynarec.Enable			= true;
 	settings.dynarec.idleskip		= true;
 	settings.dynarec.unstable_opt	= false;
@@ -495,9 +498,11 @@ void InitSettings()
 	settings.dreamcast.broadcast	= 4;	// default
 	settings.dreamcast.language     = 6;	// default
 	settings.dreamcast.FullMMU      = false;
+	settings.dynarec.SmcCheckLevel  = FullCheck;
 	settings.aica.LimitFPS			= true;
 	settings.aica.NoBatch			= false;	// This also controls the DSP. Disabled by default
     settings.aica.NoSound			= false;
+	settings.audio.backend 			= "auto";
 	settings.rend.UseMipmaps		= true;
 	settings.rend.WideScreen		= false;
 	settings.rend.ShowFPS			= false;
@@ -512,6 +517,10 @@ void InitSettings()
 	settings.rend.CustomTextures    = false;
 	settings.rend.DumpTextures      = false;
 	settings.rend.ScreenScaling     = 100;
+	settings.rend.ScreenStretching  = 100;
+	settings.rend.Fog				= true;
+	settings.rend.FloatVMUs			= false;
+	settings.rend.Rotate90			= false;
 
 	settings.pvr.ta_skip			= 0;
 	settings.pvr.rend				= 0;
@@ -558,11 +567,13 @@ void LoadSettings(bool game_specific)
 {
 	const char *config_section = game_specific ? cfgGetGameId() : "config";
 	const char *input_section = game_specific ? cfgGetGameId() : "input";
+	const char *audio_section = game_specific ? cfgGetGameId() : "audio";
 
 	settings.dynarec.Enable			= cfgLoadBool(config_section, "Dynarec.Enabled", settings.dynarec.Enable);
 	settings.dynarec.idleskip		= cfgLoadBool(config_section, "Dynarec.idleskip", settings.dynarec.idleskip);
 	settings.dynarec.unstable_opt	= cfgLoadBool(config_section, "Dynarec.unstable-opt", settings.dynarec.unstable_opt);
 	settings.dynarec.safemode		= cfgLoadBool(config_section, "Dynarec.safe-mode", settings.dynarec.safemode);
+	settings.dynarec.SmcCheckLevel  = (SmcCheckEnum)cfgLoadInt(config_section, "Dynarec.SmcCheckLevel", settings.dynarec.SmcCheckLevel);
 	//disable_nvmem can't be loaded, because nvmem init is before cfg load
 	settings.dreamcast.cable		= cfgLoadInt(config_section, "Dreamcast.Cable", settings.dreamcast.cable);
 	settings.dreamcast.region		= cfgLoadInt(config_section, "Dreamcast.Region", settings.dreamcast.region);
@@ -572,6 +583,7 @@ void LoadSettings(bool game_specific)
 	settings.aica.LimitFPS			= cfgLoadBool(config_section, "aica.LimitFPS", settings.aica.LimitFPS);
 	settings.aica.NoBatch			= cfgLoadBool(config_section, "aica.NoBatch", settings.aica.NoBatch);
     settings.aica.NoSound			= cfgLoadBool(config_section, "aica.NoSound", settings.aica.NoSound);
+    settings.audio.backend			= cfgLoadStr(audio_section, "backend", settings.audio.backend.c_str());
 	settings.rend.UseMipmaps		= cfgLoadBool(config_section, "rend.UseMipmaps", settings.rend.UseMipmaps);
 	settings.rend.WideScreen		= cfgLoadBool(config_section, "rend.WideScreen", settings.rend.WideScreen);
 	settings.rend.ShowFPS			= cfgLoadBool(config_section, "rend.ShowFPS", settings.rend.ShowFPS);
@@ -593,6 +605,10 @@ void LoadSettings(bool game_specific)
 	settings.rend.DumpTextures      = cfgLoadBool(config_section, "rend.DumpTextures", settings.rend.DumpTextures);
 	settings.rend.ScreenScaling     = cfgLoadInt(config_section, "rend.ScreenScaling", settings.rend.ScreenScaling);
 	settings.rend.ScreenScaling = min(max(1, settings.rend.ScreenScaling), 100);
+	settings.rend.ScreenStretching  = cfgLoadInt(config_section, "rend.ScreenStretching", settings.rend.ScreenStretching);
+	settings.rend.Fog				= cfgLoadBool(config_section, "rend.Fog", settings.rend.Fog);
+	settings.rend.FloatVMUs			= cfgLoadBool(config_section, "rend.FloatVMUs", settings.rend.FloatVMUs);
+	settings.rend.Rotate90			= cfgLoadBool(config_section, "rend.Rotate90", settings.rend.Rotate90);
 
 	settings.pvr.ta_skip			= cfgLoadInt(config_section, "ta.skip", settings.pvr.ta_skip);
 	settings.pvr.rend				= cfgLoadInt(config_section, "pvr.rend", settings.pvr.rend);
@@ -698,10 +714,13 @@ void SaveSettings()
 	cfgSaveBool("config", "Dynarec.unstable-opt", settings.dynarec.unstable_opt);
 	if (!safemode_game || !settings.dynarec.safemode)
 		cfgSaveBool("config", "Dynarec.safe-mode", settings.dynarec.safemode);
+	cfgSaveInt("config", "Dynarec.SmcCheckLevel", (int)settings.dynarec.SmcCheckLevel);
+
 	cfgSaveInt("config", "Dreamcast.Language", settings.dreamcast.language);
 	cfgSaveBool("config", "aica.LimitFPS", settings.aica.LimitFPS);
 	cfgSaveBool("config", "aica.NoBatch", settings.aica.NoBatch);
 	cfgSaveBool("config", "aica.NoSound", settings.aica.NoSound);
+	cfgSaveStr("audio", "backend", settings.audio.backend.c_str());
 	cfgSaveBool("config", "rend.WideScreen", settings.rend.WideScreen);
 	cfgSaveBool("config", "rend.ShowFPS", settings.rend.ShowFPS);
 	if (!rtt_to_buffer_game || !settings.rend.RenderToTextureBuffer)
@@ -714,6 +733,10 @@ void SaveSettings()
 	cfgSaveBool("config", "rend.CustomTextures", settings.rend.CustomTextures);
 	cfgSaveBool("config", "rend.DumpTextures", settings.rend.DumpTextures);
 	cfgSaveInt("config", "rend.ScreenScaling", settings.rend.ScreenScaling);
+	cfgSaveInt("config", "rend.ScreenStretching", settings.rend.ScreenStretching);
+	cfgSaveBool("config", "rend.Fog", settings.rend.Fog);
+	cfgSaveBool("config", "rend.FloatVMUs", settings.rend.FloatVMUs);
+	cfgSaveBool("config", "rend.Rotate90", settings.rend.Rotate90);
 	cfgSaveInt("config", "ta.skip", settings.pvr.ta_skip);
 	cfgSaveInt("config", "pvr.rend", settings.pvr.rend);
 

core/oslib/audiobackend_alsa.cpp

@@ -1,4 +1,4 @@
-#include "oslib/audiobackend_alsa.h"
+#include "oslib/audiostream.h"
 #if USE_ALSA
 #include <alsa/asoundlib.h>
 #include "cfg/cfg.h"
@@ -175,11 +175,13 @@ static void alsa_term()
 	snd_pcm_close(handle);
 }
 
-audiobackend_t audiobackend_alsa = {
+static audiobackend_t audiobackend_alsa = {
     "alsa", // Slug
     "Advanced Linux Sound Architecture", // Name
     &alsa_init,
     &alsa_push,
     &alsa_term
 };
+
+static bool alsa = RegisterAudioBackend(&audiobackend_alsa);
 #endif

core/oslib/audiobackend_alsa.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_alsa;

core/oslib/audiobackend_android.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_android;

core/oslib/audiobackend_coreaudio.cpp

@@ -1,18 +1,18 @@
 /*
     Simple Core Audio backend for osx (and maybe ios?)
     Based off various audio core samples and dolphin's code
- 
+
     This is part of the Reicast project, please consult the
     LICENSE file for licensing & related information
- 
+
     This could do with some locking logic to avoid
     race conditions, and some variable length buffer
     logic to support chunk sizes other than 512 bytes
- 
+
     It does work on my macmini though
  */
 
-#include "oslib/audiobackend_coreaudio.h"
+#include "oslib/audiostream.h"
 
 #if HOST_OS == OS_DARWIN
 #include <atomic>
@@ -49,9 +49,9 @@ static OSStatus coreaudio_callback(void* ctx, AudioUnitRenderActionFlags* flags,
             samples_rptr = (samples_rptr + buf_size) % BUFSIZE;
         }
     }
-    
+
     bufferEmpty.Set();
-    
+
     return noErr;
 }
 
@@ -63,7 +63,7 @@ static void coreaudio_init()
     AudioStreamBasicDescription format;
     AudioComponentDescription desc;
     AudioComponent component;
-    
+
     desc.componentType = kAudioUnitType_Output;
 #if !defined(TARGET_IPHONE)
     desc.componentSubType = kAudioUnitSubType_DefaultOutput;
@@ -75,12 +75,12 @@ static void coreaudio_init()
     desc.componentFlagsMask = 0;
     desc.componentManufacturer = kAudioUnitManufacturer_Apple;
     component = AudioComponentFindNext(nullptr, &desc);
-    
+
     verify(component != nullptr);
-    
+
     err = AudioComponentInstanceNew(component, &audioUnit);
     verify(err == noErr);
-    
+
     FillOutASBDForLPCM(format, 44100,
                        2, 16, 16, false, false, false);
     err = AudioUnitSetProperty(audioUnit,
@@ -88,7 +88,7 @@ static void coreaudio_init()
                                kAudioUnitScope_Input, 0, &format,
                                sizeof(AudioStreamBasicDescription));
     verify(err == noErr);
-    
+
     callback_struct.inputProc = coreaudio_callback;
     callback_struct.inputProcRefCon = 0;
     err = AudioUnitSetProperty(audioUnit,
@@ -96,24 +96,24 @@ static void coreaudio_init()
                                kAudioUnitScope_Input, 0, &callback_struct,
                                sizeof callback_struct);
     verify(err == noErr);
-    
+
     /*
     err = AudioUnitSetParameter(audioUnit,
                                 kHALOutputParam_Volume,
                                 kAudioUnitParameterFlag_Output, 0,
                                 1, 0);
     verify(err == noErr);
-    
+
     */
-    
+
     err = AudioUnitInitialize(audioUnit);
-    
+
     verify(err == noErr);
-    
+
     err = AudioOutputUnitStart(audioUnit);
 
     verify(err == noErr);
-    
+
     bufferEmpty.Set();
 }
 
@@ -134,23 +134,23 @@ static u32 coreaudio_push(void* frame, u32 samples, bool wait)
         samples_wptr = (samples_wptr + byte_size) % BUFSIZE;
         break;
     }
-    
+
     return 1;
 }
 
 static void coreaudio_term()
 {
     OSStatus err;
-    
+
     err = AudioOutputUnitStop(audioUnit);
     verify(err == noErr);
-    
+
     err = AudioUnitUninitialize(audioUnit);
     verify(err == noErr);
-    
+
     err = AudioComponentInstanceDispose(audioUnit);
     verify(err == noErr);
-    
+
     bufferEmpty.Set();
 }
 
@@ -161,4 +161,7 @@ audiobackend_t audiobackend_coreaudio = {
     &coreaudio_push,
     &coreaudio_term
 };
+
+static bool core = RegisterAudioBackend(&audiobackend_coreaudio);
+
 #endif

core/oslib/audiobackend_coreaudio.h

@@ -1,5 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_coreaudio;
-

core/oslib/audiobackend_directsound.cpp

@@ -1,4 +1,4 @@
-#include "oslib/audiobackend_directsound.h"
+#include "oslib/audiostream.h"
 #if HOST_OS==OS_WINDOWS
 #include "oslib.h"
 #include <initguid.h>
@@ -19,31 +19,31 @@ static void directsound_init()
 	verifyc(dsound->SetCooperativeLevel((HWND)libPvr_GetRenderTarget(),DSSCL_PRIORITY));
 	IDirectSoundBuffer* buffer_;
 
-	WAVEFORMATEX wfx; 
+	WAVEFORMATEX wfx;
-	DSBUFFERDESC desc; 
+	DSBUFFERDESC desc;
 
-	// Set up WAV format structure. 
+	// Set up WAV format structure.
 
-	memset(&wfx, 0, sizeof(WAVEFORMATEX)); 
+	memset(&wfx, 0, sizeof(WAVEFORMATEX));
-	wfx.wFormatTag = WAVE_FORMAT_PCM; 
+	wfx.wFormatTag = WAVE_FORMAT_PCM;
-	wfx.nChannels = 2; 
+	wfx.nChannels = 2;
-	wfx.nSamplesPerSec = 44100; 
+	wfx.nSamplesPerSec = 44100;
-	wfx.nBlockAlign = 4; 
+	wfx.nBlockAlign = 4;
-	wfx.nAvgBytesPerSec = wfx.nSamplesPerSec * wfx.nBlockAlign; 
+	wfx.nAvgBytesPerSec = wfx.nSamplesPerSec * wfx.nBlockAlign;
-	wfx.wBitsPerSample = 16; 
+	wfx.wBitsPerSample = 16;
 
-	// Set up DSBUFFERDESC structure. 
+	// Set up DSBUFFERDESC structure.
 
 	ds_ring_size=8192*wfx.nBlockAlign;
 
-	memset(&desc, 0, sizeof(DSBUFFERDESC)); 
+	memset(&desc, 0, sizeof(DSBUFFERDESC));
-	desc.dwSize = sizeof(DSBUFFERDESC); 
+	desc.dwSize = sizeof(DSBUFFERDESC);
-	desc.dwFlags = DSBCAPS_GETCURRENTPOSITION2 | DSBCAPS_CTRLPOSITIONNOTIFY;// _CTRLPAN | DSBCAPS_CTRLVOLUME | DSBCAPS_CTRLFREQUENCY; 
+	desc.dwFlags = DSBCAPS_GETCURRENTPOSITION2 | DSBCAPS_CTRLPOSITIONNOTIFY;// _CTRLPAN | DSBCAPS_CTRLVOLUME | DSBCAPS_CTRLFREQUENCY;
-	
+
-	desc.dwBufferBytes = ds_ring_size; 
+	desc.dwBufferBytes = ds_ring_size;
-	desc.lpwfxFormat = &wfx; 
+	desc.lpwfxFormat = &wfx;
+
 
-	
 
 	if (settings.aica.HW_mixing==0)
 	{
@@ -71,7 +71,7 @@ static void directsound_init()
 
 	//Play the buffer !
 	verifyc(buffer->Play(0,0,DSBPLAY_LOOPING));
-	
+
 }
 
 
@@ -159,7 +159,7 @@ static u32 directsound_push(void* frame, u32 samples, bool wait)
 	wait &= w;
 	*/
 	int ffs=1;
-	
+
 	/*
 	while (directsound_IsAudioBufferedLots() && wait)
 		if (ffs == 0)
@@ -175,7 +175,7 @@ static u32 directsound_push(void* frame, u32 samples, bool wait)
 static void directsound_term()
 {
 	buffer->Stop();
-	
+
 	buffer->Release();
 	dsound->Release();
 }
@@ -187,4 +187,6 @@ audiobackend_t audiobackend_directsound = {
     &directsound_push,
     &directsound_term
 };
+
+static bool ds = RegisterAudioBackend(&audiobackend_directsound);
 #endif

core/oslib/audiobackend_directsound.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_directsound;

core/oslib/audiobackend_libao.cpp

@@ -1,4 +1,4 @@
-#include "oslib/audiobackend_libao.h"
+#include "oslib/audiostream.h"
 #ifdef USE_LIBAO
 
 #include <ao/ao.h>
@@ -10,12 +10,12 @@ static void libao_init()
 {
 	ao_initialize();
 	memset(&aoformat, 0, sizeof(aoformat));
-	
+
 	aoformat.bits = 16;
 	aoformat.channels = 2;
 	aoformat.rate = 44100;
 	aoformat.byte_format = AO_FMT_LITTLE;
-	
+
 	aodevice = ao_open_live(ao_default_driver_id(), &aoformat, NULL); // Live output
 	if (!aodevice)
 		aodevice = ao_open_live(ao_driver_id("null"), &aoformat, NULL);
@@ -23,13 +23,13 @@ static void libao_init()
 
 static u32 libao_push(void* frame, u32 samples, bool wait)
 {
-	if (aodevice) 
+	if (aodevice)
 		ao_play(aodevice, (char*)frame, samples * 4);
-	
+
 	return 1;
 }
 
-static void libao_term() 
+static void libao_term()
 {
 	if (aodevice)
 	{
@@ -46,4 +46,5 @@ audiobackend_t audiobackend_libao = {
 		&libao_term
 };
 
+static bool ao = RegisterAudioBackend(&audiobackend_libao);
 #endif

core/oslib/audiobackend_libao.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_libao;

core/oslib/audiobackend_omx.cpp

@@ -1,4 +1,4 @@
-#include "oslib/audiobackend_omx.h"
+#include "oslib/audiostream.h"
 #if USE_OMX
 
 #include <IL/OMX_Broadcom.h>
@@ -316,4 +316,5 @@ audiobackend_t audiobackend_omx = {
     &omx_term
 };
 
+static bool omx = RegisterAudioBackend(&audiobackend_omx);
 #endif

core/oslib/audiobackend_omx.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_omx;

core/oslib/audiobackend_oss.cpp

@@ -1,4 +1,4 @@
-#include "oslib/audiobackend_oss.h"
+#include "oslib/audiostream.h"
 #ifdef USE_OSS
 #include <sys/ioctl.h>
 #include <sys/fcntl.h>
@@ -51,4 +51,5 @@ audiobackend_t audiobackend_oss = {
 		&oss_term
 };
 
+static bool oss = RegisterAudioBackend(&audiobackend_oss);
 #endif

core/oslib/audiobackend_oss.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_oss;

core/oslib/audiobackend_pulseaudio.cpp

@@ -1,4 +1,4 @@
-#include "oslib/audiobackend_pulseaudio.h"
+#include "oslib/audiostream.h"
 #ifdef USE_PULSEAUDIO
 #include <sys/ioctl.h>
 #include <sys/fcntl.h>
@@ -45,4 +45,6 @@ audiobackend_t audiobackend_pulseaudio = {
 		&pulseaudio_push,
 		&pulseaudio_term
 };
+
+static bool pulse = RegisterAudioBackend(&audiobackend_pulseaudio);
 #endif

core/oslib/audiobackend_pulseaudio.h

@@ -1,4 +0,0 @@
-#pragma once
-#include "oslib/audiostream.h"
-
-extern audiobackend_t audiobackend_pulseaudio;

core/oslib/audiostream.cpp

@@ -2,14 +2,6 @@
 #include "cfg/cfg.h"
 #include "oslib/oslib.h"
 #include "audiostream.h"
-#include "oslib/audiobackend_directsound.h"
-#include "oslib/audiobackend_android.h"
-#include "oslib/audiobackend_alsa.h"
-#include "oslib/audiobackend_oss.h"
-#include "oslib/audiobackend_pulseaudio.h"
-#include "oslib/audiobackend_coreaudio.h"
-#include "oslib/audiobackend_omx.h"
-#include "oslib/audiobackend_libao.h"
 
 struct SoundFrame { s16 l;s16 r; };
 #define SAMPLE_COUNT 512
@@ -25,16 +17,27 @@ u32 gen_samples=0;
 
 double time_diff = 128/44100.0;
 double time_last;
+
 #ifdef LOG_SOUND
+// TODO Only works on Windows!
 WaveWriter rawout("d:\\aica_out.wav");
 #endif
 
-static bool audiobackends_registered = false;
 static unsigned int audiobackends_num_max = 1;
 static unsigned int audiobackends_num_registered = 0;
-static audiobackend_t **audiobackends = static_cast<audiobackend_t**>(calloc(audiobackends_num_max, sizeof(audiobackend_t*)));
+static audiobackend_t **audiobackends = NULL;
 static audiobackend_t *audiobackend_current = NULL;
 
+u32 GetAudioBackendCount()
+{
+	return audiobackends_num_registered;
+}
+
+audiobackend_t* GetAudioBackend(int num)
+{
+	return audiobackends[num];
+}
+
 bool RegisterAudioBackend(audiobackend_t *backend)
 {
 	/* This function announces the availability of an audio backend to reicast. */
@@ -44,10 +47,16 @@ bool RegisterAudioBackend(audiobackend_t *backend)
 		printf("ERROR: Tried to register invalid audio backend (NULL pointer).\n");
 		return false;
 	}
+
 	if (backend->slug == "auto" || backend->slug == "none") {
 		printf("ERROR: Tried to register invalid audio backend (slug \"%s\" is a reserved keyword).\n", backend->slug.c_str());
 		return false;
 	}
+
+	// First call to RegisterAudioBackend(), create the backend structure;
+	if (audiobackends == NULL)
+		audiobackends = static_cast<audiobackend_t**>(calloc(audiobackends_num_max, sizeof(audiobackend_t*)));
+
 	// Check if we need to allocate addition memory for storing the pointers and allocate if neccessary
 	if (audiobackends_num_registered == audiobackends_num_max)
 	{
@@ -67,46 +76,19 @@ bool RegisterAudioBackend(audiobackend_t *backend)
 		}
 		audiobackends = new_ptr;
 	}
+
 	audiobackends[audiobackends_num_registered] = backend;
 	audiobackends_num_registered++;
 	return true;
 }
 
-void RegisterAllAudioBackends() {
+audiobackend_t* GetAudioBackend(std::string slug)
-		#if HOST_OS==OS_WINDOWS
-		RegisterAudioBackend(&audiobackend_directsound);
-		#endif
-		#if ANDROID
-		RegisterAudioBackend(&audiobackend_android);
-		#endif
-		#if USE_OMX
-		RegisterAudioBackend(&audiobackend_omx);
-		#endif
-		#if USE_ALSA
-		RegisterAudioBackend(&audiobackend_alsa);
-		#endif
-		#if USE_OSS
-		RegisterAudioBackend(&audiobackend_oss);
-		#endif
-		#if USE_PULSEAUDIO
-		RegisterAudioBackend(&audiobackend_pulseaudio);
-		#endif
-		#if USE_LIBAO
-		RegisterAudioBackend(&audiobackend_libao);
-		#endif
-        #if HOST_OS == OS_DARWIN
-        RegisterAudioBackend(&audiobackend_coreaudio);
-        #endif
-		audiobackends_registered = true;
-}
-
-static audiobackend_t* GetAudioBackend(std::string slug)
 {
 	if (slug == "none")
 	{
 			printf("WARNING: Audio backend set to \"none\"!\n");
 	}
-	else if(audiobackends_num_registered > 0)
+	else if (audiobackends_num_registered > 0)
 	{
 		if (slug == "auto")
 		{
@@ -135,7 +117,8 @@ static audiobackend_t* GetAudioBackend(std::string slug)
 	return NULL;
 }
 
-u32 PushAudio(void* frame, u32 amt, bool wait) {
+u32 PushAudio(void* frame, u32 amt, bool wait)
+{
 	if (audiobackend_current != NULL) {
 		return audiobackend_current->push(frame, amt, wait);
 	}
@@ -151,6 +134,7 @@ u32 asRingUsedCount()
 	//s32 sz=(WritePtr+1)%RingBufferSampleCount-ReadPtr;
 	//return sz<0?sz+RingBufferSampleCount:sz;
 }
+
 u32 asRingFreeCount()
 {
 	return RingBufferSampleCount-asRingUsedCount();
@@ -169,6 +153,27 @@ void WriteSample(s16 r, s16 l)
 	}
 }
 
+static bool backends_sorted = false;
+void SortAudioBackends()
+{
+	if (backends_sorted)
+		return;
+
+	// Sort backends by slug
+	for (int n = audiobackends_num_registered; n > 0; n--)
+	{
+		for (int i = 0; i < n-1; i++)
+		{
+			if (audiobackends[i]->slug > audiobackends[i+1]->slug)
+			{
+				audiobackend_t* swap = audiobackends[i];
+				audiobackends[i] = audiobackends[i+1];
+				audiobackends[i+1] = swap;
+			}
+		}
+	}
+}
+
 void InitAudio()
 {
 	if (cfgLoadInt("audio", "disable", 0)) {
@@ -176,26 +181,24 @@ void InitAudio()
 		return;
 	}
 
-	cfgSaveInt("audio","disable",0);
+	cfgSaveInt("audio", "disable", 0);
-
-	if (!audiobackends_registered) {
-		//FIXME: There might some nicer way to do this.
-		RegisterAllAudioBackends();
-	}
 
 	if (audiobackend_current != NULL) {
 		printf("ERROR: The audio backend \"%s\" (%s) has already been initialized, you need to terminate it before you can call audio_init() again!\n", audiobackend_current->slug.c_str(), audiobackend_current->name.c_str());
 		return;
 	}
 
-	string audiobackend_slug = cfgLoadStr("audio", "backend", "auto"); // FIXME: This could be made a parameter
+	SortAudioBackends();
+
+	string audiobackend_slug = settings.audio.backend;
 	audiobackend_current = GetAudioBackend(audiobackend_slug);
 	if (audiobackend_current == NULL) {
 		printf("WARNING: Running without audio!\n");
 		return;
 	}
+
 	printf("Initializing audio backend \"%s\" (%s)...\n", audiobackend_current->slug.c_str(), audiobackend_current->name.c_str());
-	audiobackend_current->init();	
+	audiobackend_current->init();
 }
 
 void TermAudio()
@@ -204,5 +207,5 @@ void TermAudio()
 		audiobackend_current->term();
 		printf("Terminating audio backend \"%s\" (%s)...\n", audiobackend_current->slug.c_str(), audiobackend_current->name.c_str());
 		audiobackend_current = NULL;
-	}	
+	}
 }

core/oslib/audiostream.h

@@ -1,5 +1,6 @@
 #pragma once
 #include "types.h"
+#include <tuple>
 
 //Get used size in the ring buffer
 u32 asRingUsedCount();
@@ -24,3 +25,8 @@ extern bool RegisterAudioBackend(audiobackend_t* backend);
 extern void InitAudio();
 extern u32 PushAudio(void* frame, u32 amt, bool wait);
 extern void TermAudio();
+
+u32 GetAudioBackendCount();
+void SortAudioBackends();
+audiobackend_t* GetAudioBackend(int num);
+audiobackend_t* GetAudioBackend(std::string slug);

core/rec-ARM/rec_arm.cpp

@@ -2082,7 +2082,7 @@ __default:
 }
 
 
-void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool staging,bool optimise)
+void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging,bool optimise)
 {
 	//printf("Compile: %08X, %d, %d\n",block->addr,staging,optimise);
 	block->code=(DynarecCodeEntryPtr)EMIT_GET_PTR();
@@ -2114,39 +2114,68 @@ void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool st
 	reg.OpBegin(&block->oplist[0],0);
 
 	//scheduler
-	if (force_checks)
+	switch (smc_checks) {
-	{
+		case NoCheck:
-		s32 sz = block->sh4_code_size;
+			break;
-		u32 addr = block->addr;
-		MOV32(r0,addr);
 
-		while (sz > 0)
+		case FastCheck: {
-		{
+			MOV32(r0,block->addr);
-			if (sz > 2)
+			u32* ptr=(u32*)GetMemPtr(block->addr,4);
+			if (ptr != NULL)
 			{
-				u32* ptr=(u32*)GetMemPtr(addr,4);
 				MOV32(r2,(u32)ptr);
 				LDR(r2,r2,0);
 				MOV32(r1,*ptr);
 				CMP(r1,r2);
-
 				JUMP((u32)ngen_blockcheckfail, CC_NE);
-				addr += 4;
-				sz -= 4;
 			}
-			else
+		}
+		break;
+
+		case FullCheck: {
+			s32 sz = block->sh4_code_size;
+			u32 addr = block->addr;
+			MOV32(r0,addr);
+
+			while (sz > 0)
 			{
-				u16* ptr = (u16 *)GetMemPtr(addr, 2);
+				if (sz > 2)
-				MOV32(r2, (u32)ptr);
+				{
-				LDRH(r2, r2, 0, AL);
+					u32* ptr=(u32*)GetMemPtr(addr,4);
-				MOVW(r1, *ptr, AL);
+					if (ptr != NULL)
-				CMP(r1, r2);
+					{
+						MOV32(r2,(u32)ptr);
+						LDR(r2,r2,0);
+						MOV32(r1,*ptr);
+						CMP(r1,r2);
 
-				JUMP((u32)ngen_blockcheckfail, CC_NE);
+						JUMP((u32)ngen_blockcheckfail, CC_NE);
-				addr += 2;
+					}
-				sz -= 2;
+					addr += 4;
+					sz -= 4;
+				}
+				else
+				{
+					u16* ptr = (u16 *)GetMemPtr(addr, 2);
+					if (ptr != NULL)
+					{
+						MOV32(r2, (u32)ptr);
+						LDRH(r2, r2, 0, AL);
+						MOVW(r1, *ptr, AL);
+						CMP(r1, r2);
+
+						JUMP((u32)ngen_blockcheckfail, CC_NE);
+					}
+					addr += 2;
+					sz -= 2;
+				}
 			}
 		}
+		break;
+
+		default: {
+			die("unhandled smc_checks");
+		}
 	}
 
 	u32 cyc=block->guest_cycles;

core/rec-ARM64/rec_arm64.cpp

@@ -348,15 +348,15 @@ public:
 		return *ret_reg;
 	}
 
-	void ngen_Compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise)
+	void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging, bool optimise)
 	{
 		//printf("REC-ARM64 compiling %08x\n", block->addr);
 #ifdef PROFILING
 		SaveFramePointer();
 #endif
 		this->block = block;
-		if (force_checks)
+		
-			CheckBlock(block);
+		CheckBlock(smc_checks, block);
 
 		// run register allocator
 		regalloc.DoAlloc(block);
@@ -1292,49 +1292,72 @@ private:
 		verify (GetCursorAddress<Instruction *>() - start_instruction == code_size * kInstructionSize);
 	}
 
-	void CheckBlock(RuntimeBlockInfo* block)
+	void CheckBlock(SmcCheckEnum smc_checks, RuntimeBlockInfo* block)
 	{
-		s32 sz = block->sh4_code_size;
 
 		Label blockcheck_fail;
 		Label blockcheck_success;
 
-		u8* ptr = GetMemPtr(block->addr, sz);
+		switch (smc_checks) {
-		if (ptr == NULL)
+			case NoCheck:
-			// FIXME Can a block cross a RAM / non-RAM boundary??
+				return;
-			return;
 
-		Ldr(x9, reinterpret_cast<uintptr_t>(ptr));
+			case FastCheck: {
-
+				u8* ptr = GetMemPtr(block->addr, 4);
-		while (sz > 0)
+				if (ptr == NULL)
-		{
+					return;
-			if (sz >= 8)
+				Ldr(x9, reinterpret_cast<uintptr_t>(ptr));
-			{
+				Ldr(w10, MemOperand(x9));
-				Ldr(x10, MemOperand(x9, 8, PostIndex));
-				Ldr(x11, *(u64*)ptr);
-				Cmp(x10, x11);
-				sz -= 8;
-				ptr += 8;
-			}
-			else if (sz >= 4)
-			{
-				Ldr(w10, MemOperand(x9, 4, PostIndex));
 				Ldr(w11, *(u32*)ptr);
 				Cmp(w10, w11);
-				sz -= 4;
+				B(eq, &blockcheck_success);
-				ptr += 4;
 			}
-			else
+			break;
-			{
+
-				Ldrh(w10, MemOperand(x9, 2, PostIndex));
+			case FullCheck: {
-				Mov(w11, *(u16*)ptr);
+				s32 sz = block->sh4_code_size;
-				Cmp(w10, w11);
+
-				sz -= 2;
+				u8* ptr = GetMemPtr(block->addr, sz);
-				ptr += 2;
+				if (ptr == NULL)
+					return;
+
+				Ldr(x9, reinterpret_cast<uintptr_t>(ptr));
+
+				while (sz > 0)
+				{
+					if (sz >= 8)
+					{
+						Ldr(x10, MemOperand(x9, 8, PostIndex));
+						Ldr(x11, *(u64*)ptr);
+						Cmp(x10, x11);
+						sz -= 8;
+						ptr += 8;
+					}
+					else if (sz >= 4)
+					{
+						Ldr(w10, MemOperand(x9, 4, PostIndex));
+						Ldr(w11, *(u32*)ptr);
+						Cmp(w10, w11);
+						sz -= 4;
+						ptr += 4;
+					}
+					else
+					{
+						Ldrh(w10, MemOperand(x9, 2, PostIndex));
+						Mov(w11, *(u16*)ptr);
+						Cmp(w10, w11);
+						sz -= 2;
+						ptr += 2;
+					}
+					B(ne, &blockcheck_fail);
+				}
+				B(&blockcheck_success);
 			}
-			B(ne, &blockcheck_fail);
+			break;
+
+			default:
+				die("unhandled smc_checks");
 		}
-		B(&blockcheck_success);
 
 		Bind(&blockcheck_fail);
 		Ldr(w0, block->addr);
@@ -1404,13 +1427,13 @@ private:
 
 static Arm64Assembler* compiler;
 
-void ngen_Compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise)
+void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging, bool optimise)
 {
 	verify(emit_FreeSpace() >= 16 * 1024);
 
 	compiler = new Arm64Assembler();
 
-	compiler->ngen_Compile(block, force_checks, reset, staging, optimise);
+	compiler->ngen_Compile(block, smc_checks, reset, staging, optimise);
 
 	delete compiler;
 	compiler = NULL;

core/rec-cpp/rec_cpp.cpp

@@ -1190,10 +1190,10 @@ public:
 
 	size_t opcode_index;
 	opcodeExec** ptrsg;
-	void compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise) {
+	void compile(RuntimeBlockInfo* block,  SmcCheckEnum smc_checks, bool reset, bool staging, bool optimise) {
 		
 		//we need an extra one for the end opcode and optionally one more for block check
-		auto ptrs = fnnCtor_forreal(block->oplist.size() + 1 + (force_checks ? 1 : 0))(block->guest_cycles);
+		auto ptrs = fnnCtor_forreal(block->oplist.size() + 1 + (smc_checks != NoCheck ? 1 : 0))(block->guest_cycles);
 
 		ptrsg = ptrs.ptrs;
 
@@ -1207,9 +1207,16 @@ public:
 		}
 
 		size_t i = 0;
-		if (force_checks)
+		if (smc_checks != NoCheck)
 		{
+			verify (smc_checks == FastCheck || smc_checks == FullCheck)
 			opcodeExec* op;
+			int check_size = block->sh4_code_size;
+
+			if (smc_checks == FastCheck) {
+				check_size = 4;
+			}
+
 			switch (block->sh4_code_size)
 			{
 			case 4:
@@ -1227,6 +1234,7 @@ public:
 			}
 			ptrs.ptrs[i++] = op;
 		}
+
 		for (size_t opnum = 0; opnum < block->oplist.size(); opnum++, i++) {
 			opcode_index = i;
 			shil_opcode& op = block->oplist[opnum];
@@ -1551,14 +1559,14 @@ public:
 
 BlockCompiler* compiler;
 
-void ngen_Compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise)
+void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging, bool optimise)
 {
 	verify(emit_FreeSpace() >= 16 * 1024);
 
 	compiler = new BlockCompiler();
 
 
-	compiler->compile(block, force_checks, reset, staging, optimise);
+	compiler->compile(block, smc_checks, reset, staging, optimise);
 
 	delete compiler;
 }

core/rec-x64/rec_x64.cpp

@@ -244,12 +244,11 @@ public:
 		call_regsxmm.push_back(xmm3);
 	}
 
-	void compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise)
+	void compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging, bool optimise)
 	{
 		//printf("X86_64 compiling %08x to %p\n", block->addr, emit_GetCCPtr());
-		if (force_checks) {
+		CheckBlock(smc_checks, block);
-			CheckBlock(block);
+		
-		}
 		regalloc.DoAlloc(block);
 
 		sub(dword[rip + &cycle_counter], block->guest_cycles);
@@ -1091,41 +1090,66 @@ private:
 	typedef void (BlockCompiler::*X64BinaryOp)(const Xbyak::Operand&, const Xbyak::Operand&);
 	typedef void (BlockCompiler::*X64BinaryFOp)(const Xbyak::Xmm&, const Xbyak::Operand&);
 
-	void CheckBlock(RuntimeBlockInfo* block) {
+	void CheckBlock(SmcCheckEnum smc_checks, RuntimeBlockInfo* block) {
-		mov(call_regs[0], block->addr);
 
-		s32 sz=block->sh4_code_size;
+		switch (smc_checks) {
-		u32 sa=block->addr;
+			case NoCheck:
+				return;
 
-		while (sz > 0)
+		 	case FastCheck: {
-		{
+		 		void* ptr = (void*)GetMemPtr(block->addr, 4);
-			void* ptr = (void*)GetMemPtr(sa, sz > 8 ? 8 : sz);
+                if (ptr)
-			if (ptr)
+				{
-			{
+					mov(call_regs[0], block->addr);
-				mov(rax, reinterpret_cast<uintptr_t>(ptr));
+					mov(rax, reinterpret_cast<uintptr_t>(ptr));
-
-				if (sz >= 8) {
-					mov(rdx, *(u64*)ptr);
-					cmp(qword[rax], rdx);
-					sz -= 8;
-					sa += 8;
-				}
-				else if (sz >= 4) {
 					mov(edx, *(u32*)ptr);
 					cmp(dword[rax], edx);
-					sz -= 4;
+					jne(reinterpret_cast<const void*>(&ngen_blockcheckfail));		
-					sa += 4;
 				}
-				else {
+		 	}
-					mov(edx, *(u16*)ptr);
+		 	break;
-					cmp(word[rax],dx);
-					sz -= 2;
-					sa += 2;
-				}
-				jne(reinterpret_cast<const void*>(&ngen_blockcheckfail));
-			}
-		}
 
+		 	case FullCheck: {
+		 		s32 sz=block->sh4_code_size;
+				u32 sa=block->addr;
+
+				void* ptr = (void*)GetMemPtr(sa, sz > 8 ? 8 : sz);
+				if (ptr)
+				{
+					mov(call_regs[0], block->addr);
+
+					while (sz > 0)
+					{
+						mov(rax, reinterpret_cast<uintptr_t>(ptr));
+
+						if (sz >= 8) {
+							mov(rdx, *(u64*)ptr);
+							cmp(qword[rax], rdx);
+							sz -= 8;
+							sa += 8;
+						}
+						else if (sz >= 4) {
+							mov(edx, *(u32*)ptr);
+							cmp(dword[rax], edx);
+							sz -= 4;
+							sa += 4;
+						}
+						else {
+							mov(edx, *(u16*)ptr);
+							cmp(word[rax],dx);
+							sz -= 2;
+							sa += 2;
+						}
+						jne(reinterpret_cast<const void*>(&ngen_blockcheckfail));
+						ptr = (void*)GetMemPtr(sa, sz > 8 ? 8 : sz);
+					}
+		 		}
+		 	}
+		 	break;
+
+			default:
+				die("unhandled smc_checks");
+		}
 	}
 
 	void GenBinaryOp(const shil_opcode &op, X64BinaryOp natop)
@@ -1267,13 +1291,13 @@ void X64RegAlloc::Writeback_FPU(u32 reg, s8 nreg)
 
 static BlockCompiler* compiler;
 
-void ngen_Compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise)
+void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging, bool optimise)
 {
 	verify(emit_FreeSpace() >= 16 * 1024);
 
 	compiler = new BlockCompiler();
 	
-	compiler->compile(block, force_checks, reset, staging, optimise);
+	compiler->compile(block, smc_checks, reset, staging, optimise);
 
 	delete compiler;
 }

core/rec-x86/rec_x86_driver.cpp

@@ -229,29 +229,50 @@ u32 rdmt[6];
 extern u32 memops_t,memops_l;
 extern int mips_counter;
 
-void CheckBlock(RuntimeBlockInfo* block,x86_ptr_imm place)
+//TODO: Get back validating mode for this
+void CheckBlock(SmcCheckEnum smc_checks, RuntimeBlockInfo* block)
 {
-	s32 sz=block->sh4_code_size;
+	switch (smc_checks) {
-	u32 sa=block->addr;
+		case NoCheck:
-	while(sz>0)
+			break;
-	{
+
-		void* ptr=(void*)GetMemPtr(sa,4);
+		case FastCheck: {
-		if (ptr)
+			void* ptr = (void*)GetMemPtr(block->addr, 4);
-		{
+			if (ptr)
-			if (sz==2)
+			{
-				x86e->Emit(op_cmp16,ptr,*(u16*)ptr);
+				x86e->Emit(op_cmp32, ptr, *(u32*)ptr);
-			else
+				x86e->Emit(op_jne, x86_ptr_imm(ngen_blockcheckfail));
-				x86e->Emit(op_cmp32,ptr,*(u32*)ptr);
+			}
-			x86e->Emit(op_jne,place);
 		}
-		sz-=4;
+		break;
-		sa+=4;
+
+		case FullCheck: {
+			s32 sz=block->sh4_code_size;
+			u32 sa=block->addr;
+			while(sz>0)
+			{
+				void* ptr=(void*)GetMemPtr(sa,4);
+				if (ptr)
+				{
+					if (sz==2)
+						x86e->Emit(op_cmp16,ptr,*(u16*)ptr);
+					else
+						x86e->Emit(op_cmp32,ptr,*(u32*)ptr);
+					x86e->Emit(op_jne,x86_ptr_imm(ngen_blockcheckfail));
+				}
+				sz-=4;
+				sa+=4;
+			}
+		}
+		break;
+
+		default:
+			die("unhandled smc_checks");
 	}
-	
 }
 
 
-void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool staging,bool optimise)
+void ngen_Compile(RuntimeBlockInfo* block, SmcCheckEnum smc_checks, bool reset, bool staging,bool optimise)
 {
 	//initialise stuff
 	DetectCpuFeatures();
@@ -282,7 +303,7 @@ void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool st
 	//block invl. checks
 	x86e->Emit(op_mov32,ECX,block->addr);
 
-	CheckBlock(block,force_checks?x86_ptr_imm(ngen_blockcheckfail):x86_ptr_imm(ngen_blockcheckfail2));
+	CheckBlock(smc_checks, block);
 
 	//Scheduler
 	x86_Label* no_up=x86e->CreateLabel(false,8);

core/rend/gl4/abuffer.cpp

@@ -331,23 +331,23 @@ void initABuffer()
 	{
 		char source[16384];
 		sprintf(source, final_shader_source, 1);
-		gl4CompilePipelineShader(&g_abuffer_final_shader, source);
+		gl4CompilePipelineShader(&g_abuffer_final_shader, false, source);
 	}
 	if (g_abuffer_final_nosort_shader.program == 0)
 	{
 		char source[16384];
 		sprintf(source, final_shader_source, 0);
-		gl4CompilePipelineShader(&g_abuffer_final_nosort_shader, source);
+		gl4CompilePipelineShader(&g_abuffer_final_nosort_shader, false, source);
 	}
 	if (g_abuffer_clear_shader.program == 0)
-		gl4CompilePipelineShader(&g_abuffer_clear_shader, clear_shader_source);
+		gl4CompilePipelineShader(&g_abuffer_clear_shader, false, clear_shader_source);
 	if (g_abuffer_tr_modvol_shaders[0].program == 0)
 	{
 		char source[16384];
 		for (int mode = 0; mode < ModeCount; mode++)
 		{
 			sprintf(source, tr_modvol_shader_source, mode);
-			gl4CompilePipelineShader(&g_abuffer_tr_modvol_shaders[mode], source);
+			gl4CompilePipelineShader(&g_abuffer_tr_modvol_shaders[mode], false, source);
 		}
 	}
 
@@ -417,6 +417,17 @@ void termABuffer()
 		glDeleteBuffers(1, &g_quadBuffer);
 		g_quadBuffer = 0;
 	}
+	glcache.DeleteProgram(g_abuffer_final_shader.program);
+	g_abuffer_final_shader.program = 0;
+	glcache.DeleteProgram(g_abuffer_final_nosort_shader.program);
+	g_abuffer_final_nosort_shader.program = 0;
+	glcache.DeleteProgram(g_abuffer_clear_shader.program);
+	g_abuffer_clear_shader.program = 0;
+	for (int mode = 0; mode < ModeCount; mode++)
+	{
+		glcache.DeleteProgram(g_abuffer_tr_modvol_shaders[mode].program);
+		g_abuffer_tr_modvol_shaders[mode].program = 0;
+	}
 }
 
 void reshapeABuffer(int w, int h)

core/rend/gl4/gl4.h

@@ -1,6 +1,6 @@
 #pragma once
 #include "rend/gles/gles.h"
-#include <map>
+#include <unordered_map>
 
 void gl4DrawStrips(GLuint output_fbo);
 
@@ -44,7 +44,8 @@ struct gl4_ctx
 		GLuint extra_depth_scale;
 	} modvol_shader;
 
-	std::map<int, gl4PipelineShader *> shaders;
+	std::unordered_map<u32, gl4PipelineShader> shaders;
+	bool rotate90;
 
 	struct
 	{
@@ -53,16 +54,6 @@ struct gl4_ctx
 		GLuint modvol_vao;
 		GLuint tr_poly_params;
 	} vbo;
-
-	gl4PipelineShader *getShader(int programId) {
-		gl4PipelineShader *shader = shaders[programId];
-		if (shader == NULL) {
-			shader = new gl4PipelineShader();
-			shaders[programId] = shader;
-			shader->program = -1;
-		}
-		return shader;
-	}
 };
 
 extern gl4_ctx gl4;
@@ -76,7 +67,8 @@ bool gl4_render_output_framebuffer();
 void abufferDrawQuad(bool upsideDown = false, float x = 0.f, float y = 0.f, float w = 0.f, float h = 0.f);
 
 extern const char *gl4PixelPipelineShader;
-bool gl4CompilePipelineShader(gl4PipelineShader* s, const char *source = gl4PixelPipelineShader);
+bool gl4CompilePipelineShader(gl4PipelineShader* s, bool rotate_90, const char *source = gl4PixelPipelineShader);
+void gl4_delete_shaders();
 
 extern GLuint stencilTexId;
 extern GLuint depthTexId;

core/rend/gl4/gldraw.cpp

@@ -45,10 +45,15 @@ static GLuint texSamplers[2];
 static GLuint depth_fbo;
 GLuint depthSaveTexId;
 
-static int gl4GetProgramID(u32 cp_AlphaTest, u32 pp_ClipTestMode,
+static gl4PipelineShader *gl4GetProgram(u32 cp_AlphaTest, u32 pp_ClipTestMode,
 							u32 pp_Texture, u32 pp_UseAlpha, u32 pp_IgnoreTexA, u32 pp_ShadInstr, u32 pp_Offset,
 							u32 pp_FogCtrl, bool pp_TwoVolumes, u32 pp_DepthFunc, bool pp_Gouraud, bool pp_BumpMap, bool fog_clamping, int pass)
 {
+	if (settings.rend.Rotate90 != gl4.rotate90)
+	{
+		gl4_delete_shaders();
+		gl4.rotate90 = settings.rend.Rotate90;
+	}
 	u32 rv=0;
 
 	rv|=pp_ClipTestMode;
@@ -66,45 +71,27 @@ static int gl4GetProgramID(u32 cp_AlphaTest, u32 pp_ClipTestMode,
 	rv <<= 1; rv |= fog_clamping;
 	rv <<= 2; rv |= pass;
 
-	return rv;
+	gl4PipelineShader *shader = &gl4.shaders[rv];
-}
+	if (shader->program == 0)
-
+	{
-static void setCurrentShader(u32 cp_AlphaTest, u32 pp_ClipTestMode,
+		shader->cp_AlphaTest = cp_AlphaTest;
-							u32 pp_Texture, u32 pp_UseAlpha, u32 pp_IgnoreTexA, u32 pp_ShadInstr, u32 pp_Offset,
+		shader->pp_ClipTestMode = pp_ClipTestMode;
-							u32 pp_FogCtrl, bool pp_TwoVolumes, u32 pp_DepthFunc, bool pp_Gouraud, bool pp_BumpMap, bool fog_clamping, int pass)
+		shader->pp_Texture = pp_Texture;
-{
+		shader->pp_UseAlpha = pp_UseAlpha;
-	int shaderId = gl4GetProgramID(cp_AlphaTest,
+		shader->pp_IgnoreTexA = pp_IgnoreTexA;
-				pp_ClipTestMode + 1,
+		shader->pp_ShadInstr = pp_ShadInstr;
-				pp_Texture,
+		shader->pp_Offset = pp_Offset;
-				pp_UseAlpha,
+		shader->pp_FogCtrl = pp_FogCtrl;
-				pp_IgnoreTexA,
+		shader->pp_TwoVolumes = pp_TwoVolumes;
-				pp_ShadInstr,
+		shader->pp_DepthFunc = pp_DepthFunc;
-				pp_Offset,
+		shader->pp_Gouraud = pp_Gouraud;
-				pp_FogCtrl,
+		shader->pp_BumpMap = pp_BumpMap;
-				pp_TwoVolumes,
+		shader->fog_clamping = fog_clamping;
-				pp_DepthFunc,
+		shader->pass = pass;
-				pp_Gouraud,
+		gl4CompilePipelineShader(shader, settings.rend.Rotate90);
-				pp_BumpMap,
-				fog_clamping,
-				pass);
-	CurrentShader = gl4.getShader(shaderId);
-	if (CurrentShader->program == -1) {
-		CurrentShader->cp_AlphaTest = cp_AlphaTest;
-		CurrentShader->pp_ClipTestMode = pp_ClipTestMode;
-		CurrentShader->pp_Texture = pp_Texture;
-		CurrentShader->pp_UseAlpha = pp_UseAlpha;
-		CurrentShader->pp_IgnoreTexA = pp_IgnoreTexA;
-		CurrentShader->pp_ShadInstr = pp_ShadInstr;
-		CurrentShader->pp_Offset = pp_Offset;
-		CurrentShader->pp_FogCtrl = pp_FogCtrl;
-		CurrentShader->pp_TwoVolumes = pp_TwoVolumes;
-		CurrentShader->pp_DepthFunc = pp_DepthFunc;
-		CurrentShader->pp_Gouraud = pp_Gouraud;
-		CurrentShader->pp_BumpMap = pp_BumpMap;
-		CurrentShader->fog_clamping = fog_clamping;
-		CurrentShader->pass = pass;
-		gl4CompilePipelineShader(CurrentShader);
 	}
+
+	return shader;
 }
 
 static void SetTextureRepeatMode(int index, GLuint dir, u32 clamp, u32 mirror)
@@ -132,7 +119,7 @@ template <u32 Type, bool SortingEnabled>
 
 	if (pass == 0)
 	{
-		setCurrentShader(Type == ListType_Punch_Through ? 1 : 0,
+		CurrentShader = gl4GetProgram(Type == ListType_Punch_Through ? 1 : 0,
 				clipping,
 				Type == ListType_Punch_Through ? gp->pcw.Texture : 0,
 				1,
@@ -153,6 +140,8 @@ template <u32 Type, bool SortingEnabled>
 		bool two_volumes_mode = (gp->tsp1.full != -1) && Type != ListType_Translucent;
 		bool color_clamp = gp->tsp.ColorClamp && (pvrrc.fog_clamp_min != 0 || pvrrc.fog_clamp_max != 0xffffffff);
 
+		int fog_ctrl = settings.rend.Fog ? gp->tsp.FogCtrl : 2;
+
 		int depth_func = 0;
 		if (Type == ListType_Translucent)
 		{
@@ -162,14 +151,14 @@ template <u32 Type, bool SortingEnabled>
 				depth_func = gp->isp.DepthMode;
 		}
 
-		setCurrentShader(Type == ListType_Punch_Through ? 1 : 0,
+		CurrentShader = gl4GetProgram(Type == ListType_Punch_Through ? 1 : 0,
 				clipping,
 				gp->pcw.Texture,
 				gp->tsp.UseAlpha,
 				gp->tsp.IgnoreTexA,
 				gp->tsp.ShadInstr,
 				gp->pcw.Offset,
-				gp->tsp.FogCtrl,
+				fog_ctrl,
 				two_volumes_mode,
 				depth_func,
 				gp->pcw.Gouraud,
@@ -681,7 +670,7 @@ static void gl4_draw_quad_texture(GLuint texture, bool upsideDown, float x = 0.f
 
 	ShaderUniforms.trilinear_alpha = 1.0;
 
-	setCurrentShader(0,
+	CurrentShader = gl4GetProgram(0,
 				0,
 				1,
 				0,
@@ -713,11 +702,15 @@ void gl4DrawFramebuffer(float w, float h)
 
 bool gl4_render_output_framebuffer()
 {
-	glBindFramebuffer(GL_FRAMEBUFFER, 0);
 	glViewport(0, 0, screen_width, screen_height);
-	if (gl.ofbo.tex == 0)
+	glcache.Disable(GL_SCISSOR_TEST);
+	if (gl.ofbo.fbo == 0)
 		return false;
-
+	glBindFramebuffer(GL_READ_FRAMEBUFFER, gl.ofbo.fbo);
-	gl4_draw_quad_texture(gl.ofbo.tex, true);
+	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+	glBlitFramebuffer(0, 0, gl.ofbo.width, gl.ofbo.height,
+			0, 0, screen_width, screen_height,
+			GL_COLOR_BUFFER_BIT, GL_LINEAR);
+	glBindFramebuffer(GL_FRAMEBUFFER, 0);
 	return true;
 }

core/rend/gl4/gles.cpp

@@ -14,6 +14,7 @@ static const char* VertexShaderSource =
 "\
 #version 140 \n\
 #define pp_Gouraud %d \n\
+#define ROTATE_90 %d \n\
  \n\
 #if pp_Gouraud == 0 \n\
 #define INTERPOLATION flat \n\
@@ -56,6 +57,9 @@ void main() \n\
 	\n\
 	vpos.w = extra_depth_scale / vpos.z; \n\
 	vpos.z = vpos.w; \n\
+#if ROTATE_90 == 1 \n\
+	vpos.xy = vec2(vpos.y, -vpos.x);  \n\
+#endif \n\
 	vpos.xy=vpos.xy*scale.xy-scale.zw;  \n\
 	vpos.xy*=vpos.w;  \n\
 	gl_Position = vpos; \n\
@@ -393,11 +397,11 @@ gl4_ctx gl4;
 
 struct gl4ShaderUniforms_t gl4ShaderUniforms;
 
-bool gl4CompilePipelineShader(	gl4PipelineShader* s, const char *source /* = PixelPipelineShader */)
+bool gl4CompilePipelineShader(	gl4PipelineShader* s, bool rotate_90, const char *source /* = PixelPipelineShader */)
 {
 	char vshader[16384];
 
-	sprintf(vshader, VertexShaderSource, s->pp_Gouraud);
+	sprintf(vshader, VertexShaderSource, s->pp_Gouraud, rotate_90);
 
 	char pshader[16384];
 
@@ -478,28 +482,45 @@ bool gl4CompilePipelineShader(	gl4PipelineShader* s, const char *source /* = Pix
 
 void gl_term();
 
+void gl4_delete_shaders()
+{
+	for (auto it : gl4.shaders)
+	{
+		if (it.second.program != 0)
+			glcache.DeleteProgram(it.second.program);
+	}
+	gl4.shaders.clear();
+	glcache.DeleteProgram(gl4.modvol_shader.program);
+	gl4.modvol_shader.program = 0;
+}
+
 static void gles_term(void)
 {
-	glDeleteProgram(gl4.modvol_shader.program);
 	glDeleteBuffers(1, &gl4.vbo.geometry);
 	gl4.vbo.geometry = 0;
 	glDeleteBuffers(1, &gl4.vbo.modvols);
 	glDeleteBuffers(1, &gl4.vbo.idxs);
 	glDeleteBuffers(1, &gl4.vbo.idxs2);
 	glDeleteBuffers(1, &gl4.vbo.tr_poly_params);
-	for (auto it = gl4.shaders.begin(); it != gl4.shaders.end(); it++)
+	gl4_delete_shaders();
-	{
-		if (it->second->program != -1)
-			glDeleteProgram(it->second->program);
-		delete it->second;
-	}
-	gl4.shaders.clear();
 	glDeleteVertexArrays(1, &gl4.vbo.main_vao);
 	glDeleteVertexArrays(1, &gl4.vbo.modvol_vao);
 
 	gl_term();
 }
 
+static void create_modvol_shader()
+{
+	if (gl4.modvol_shader.program != 0)
+		return;
+	char vshader[16384];
+	sprintf(vshader, VertexShaderSource, 1, settings.rend.Rotate90);
+
+	gl4.modvol_shader.program=gl_CompileAndLink(vshader, ModifierVolumeShader);
+	gl4.modvol_shader.scale          = glGetUniformLocation(gl4.modvol_shader.program, "scale");
+	gl4.modvol_shader.extra_depth_scale = glGetUniformLocation(gl4.modvol_shader.program, "extra_depth_scale");
+}
+
 static bool gl_create_resources()
 {
 	if (gl4.vbo.geometry != 0)
@@ -521,12 +542,7 @@ static bool gl_create_resources()
 	gl4SetupMainVBO();
 	gl4SetupModvolVBO();
 
-	char vshader[16384];
+	create_modvol_shader();
-	sprintf(vshader, VertexShaderSource, 1);
-
-	gl4.modvol_shader.program=gl_CompileAndLink(vshader, ModifierVolumeShader);
-	gl4.modvol_shader.scale          = glGetUniformLocation(gl4.modvol_shader.program, "scale");
-	gl4.modvol_shader.extra_depth_scale = glGetUniformLocation(gl4.modvol_shader.program, "extra_depth_scale");
 
 	gl_load_osd_resources();
 
@@ -604,6 +620,7 @@ static bool RenderFrame()
 		old_screen_scaling = settings.rend.ScreenScaling;
 	}
 	DoCleanup();
+	create_modvol_shader();
 
 	bool is_rtt=pvrrc.isRTT;
 
@@ -662,16 +679,41 @@ static bool RenderFrame()
 	/*
 		Handle Dc to screen scaling
 	*/
-	float screen_scaling = is_rtt ? 1.f : settings.rend.ScreenScaling / 100.f;
+	float screen_scaling = settings.rend.ScreenScaling / 100.f;
-	float dc2s_scale_h = is_rtt ? (screen_width / dc_width) : (screen_height / 480.0);
+	float screen_stretching = settings.rend.ScreenStretching / 100.f;
-	dc2s_scale_h *=  screen_scaling;
-	float ds2s_offs_x =  is_rtt ? 0 : (((screen_width * screen_scaling) - dc2s_scale_h * 640.0) / 2);
 
-	//-1 -> too much to left
+	float dc2s_scale_h;
-	gl4ShaderUniforms.scale_coefs[0] = 2.0f / (screen_width * screen_scaling / dc2s_scale_h * scale_x);
+	float ds2s_offs_x;
-	gl4ShaderUniforms.scale_coefs[1] = (is_rtt ? 2 : -2) / dc_height;		// FIXME CT2 needs 480 here instead of dc_height=512
+
-	gl4ShaderUniforms.scale_coefs[2] = 1 - 2 * ds2s_offs_x / (screen_width * screen_scaling);
+	if (is_rtt)
-	gl4ShaderUniforms.scale_coefs[3] = (is_rtt ? 1 : -1);
+	{
+		gl4ShaderUniforms.scale_coefs[0] = 2.0f / dc_width;
+		gl4ShaderUniforms.scale_coefs[1] = 2.0f / dc_height;	// FIXME CT2 needs 480 here instead of dc_height=512
+		gl4ShaderUniforms.scale_coefs[2] = 1;
+		gl4ShaderUniforms.scale_coefs[3] = 1;
+	}
+	else
+	{
+		if (settings.rend.Rotate90)
+		{
+			dc2s_scale_h = screen_height / 640.0;
+			ds2s_offs_x =  (screen_width - dc2s_scale_h * 480.0 * screen_stretching) / 2;
+			gl4ShaderUniforms.scale_coefs[0] = 2.0f / (screen_width / dc2s_scale_h * scale_x) * screen_stretching;
+			gl4ShaderUniforms.scale_coefs[1] = -2.0f / dc_width;
+			gl4ShaderUniforms.scale_coefs[2] = 1 - 2 * ds2s_offs_x / screen_width;
+			gl4ShaderUniforms.scale_coefs[3] = 1;
+		}
+		else
+		{
+			dc2s_scale_h = screen_height / 480.0;
+			ds2s_offs_x =  (screen_width - dc2s_scale_h * 640.0 * screen_stretching) / 2;
+			//-1 -> too much to left
+			gl4ShaderUniforms.scale_coefs[0] = 2.0f / (screen_width / dc2s_scale_h * scale_x) * screen_stretching;
+			gl4ShaderUniforms.scale_coefs[1] = -2.0f / dc_height;
+			gl4ShaderUniforms.scale_coefs[2] = 1 - 2 * ds2s_offs_x / screen_width;
+			gl4ShaderUniforms.scale_coefs[3] = -1;
+		}
+	}
 
 	gl4ShaderUniforms.extra_depth_scale = settings.rend.ExtraDepthScale;
 
@@ -704,7 +746,7 @@ static bool RenderFrame()
 	gl4ShaderUniforms.fog_clamp_max[2] = ((pvrrc.fog_clamp_max >> 0) & 0xFF) / 255.0f;
 	gl4ShaderUniforms.fog_clamp_max[3] = ((pvrrc.fog_clamp_max >> 24) & 0xFF) / 255.0f;
 	
-	if (fog_needs_update)
+	if (fog_needs_update && settings.rend.Fog)
 	{
 		fog_needs_update = false;
 		UpdateFogTexture((u8 *)FOG_TABLE, GL_TEXTURE5, GL_RED);
@@ -764,7 +806,7 @@ static bool RenderFrame()
 	{
 		if (settings.rend.ScreenScaling != 100 || gl.swap_buffer_not_preserved)
 		{
-			output_fbo = init_output_framebuffer(screen_width * screen_scaling, screen_height * screen_scaling);
+			output_fbo = init_output_framebuffer(screen_width * screen_scaling + 0.5f, screen_height * screen_scaling + 0.5f);
 		}
 		else
 		{
@@ -826,22 +868,31 @@ static bool RenderFrame()
 			float min_y = pvrrc.fb_Y_CLIP.min / scale_y;
 			if (!is_rtt)
 			{
+				if (settings.rend.Rotate90)
+				{
+					float t = width;
+					width = height;
+					height = t;
+					t = min_x;
+					min_x = min_y;
+					min_y = 640 - t - height;
+				}
 				// Add x offset for aspect ratio > 4/3
-            	min_x = min_x * dc2s_scale_h + ds2s_offs_x;
+            	min_x = (min_x * dc2s_scale_h * screen_stretching + ds2s_offs_x) * screen_scaling;
 				// Invert y coordinates when rendering to screen
-				min_y = screen_height * screen_scaling - (min_y + height) * dc2s_scale_h;
+				min_y = (screen_height - (min_y + height) * dc2s_scale_h) * screen_scaling;
-				width *= dc2s_scale_h;
+				width *= dc2s_scale_h * screen_stretching * screen_scaling;
-				height *= dc2s_scale_h;
+				height *= dc2s_scale_h * screen_scaling;
 
 				if (ds2s_offs_x > 0)
 				{
-					float rounded_offs_x = ds2s_offs_x + 0.5f;
+					float scaled_offs_x = ds2s_offs_x * screen_scaling;
 
 					glcache.ClearColor(0.f, 0.f, 0.f, 0.f);
 					glcache.Enable(GL_SCISSOR_TEST);
-					glScissor(0, 0, rounded_offs_x, screen_height);
+					glScissor(0, 0, scaled_offs_x + 0.5f, screen_height * screen_scaling + 0.5f);
 					glClear(GL_COLOR_BUFFER_BIT);
-					glScissor(screen_width - rounded_offs_x, 0, rounded_offs_x, screen_height);
+					glScissor(screen_width * screen_scaling - scaled_offs_x + 0.5f, 0, scaled_offs_x + 1.f, screen_height * screen_scaling + 0.5f);
 					glClear(GL_COLOR_BUFFER_BIT);
 				}
 			}

core/rend/gles/CustomTexture.cpp

@@ -50,23 +50,28 @@ void CustomTexture::LoaderThread()
 			
 			if (texture != NULL)
 			{
-				// FIXME texture may have been deleted. Need to detect this.
 				texture->ComputeHash();
-				int width, height;
+				if (texture->custom_image_data != NULL)
-				u8 *image_data = LoadCustomTexture(texture->texture_hash, width, height);
-				if (image_data == NULL)
 				{
-					image_data = LoadCustomTexture(texture->old_texture_hash, width, height);
+					delete [] texture->custom_image_data;
+					texture->custom_image_data = NULL;
 				}
-				if (image_data != NULL)
+				if (!texture->dirty)
 				{
-					if (texture->custom_image_data != NULL)
+					int width, height;
-						delete [] texture->custom_image_data;
+					u8 *image_data = LoadCustomTexture(texture->texture_hash, width, height);
-					texture->custom_width = width;
+					if (image_data == NULL)
-					texture->custom_height = height;
+					{
-					texture->custom_image_data = image_data;
+						image_data = LoadCustomTexture(texture->old_texture_hash, width, height);
+					}
+					if (image_data != NULL)
+					{
+						texture->custom_width = width;
+						texture->custom_height = height;
+						texture->custom_image_data = image_data;
+					}
 				}
-				texture->custom_load_in_progress = false;
+				texture->custom_load_in_progress--;
 			}
 
 		} while (texture != NULL);
@@ -144,10 +149,9 @@ u8* CustomTexture::LoadCustomTexture(u32 hash, int& width, int& height)
 void CustomTexture::LoadCustomTextureAsync(TextureCacheData *texture_data)
 {
 	if (!Init())
-	{
-		texture_data->custom_load_in_progress = false;
 		return;
-	}
+
+	texture_data->custom_load_in_progress++;
 	work_queue_mutex.Lock();
 	work_queue.insert(work_queue.begin(), texture_data);
 	work_queue_mutex.Unlock();

core/rend/gles/CustomTexture.h

@@ -53,7 +53,7 @@ private:
 	cThread loader_thread;
 #endif
 	cResetEvent wakeup_thread;
-	std::vector<struct TextureCacheData *> work_queue;
+	std::vector<TextureCacheData *> work_queue;
 	cMutex work_queue_mutex;
 };
 

core/rend/gles/glcache.h

@@ -149,6 +149,19 @@ public:
 		return _texture_ids[--_texture_cache_size];
 	}
 
+	void DeleteProgram(GLuint program)
+	{
+		GLsizei shader_count;
+		GLuint shaders[2];
+		glGetAttachedShaders(program, ARRAY_SIZE(shaders), &shader_count, shaders);
+		for (int i = 0; i < shader_count; i++)
+			glDeleteShader(shaders[i]);
+
+		glDeleteProgram(program);
+		if (_program == program)
+			_program = 0;
+	}
+
 	void Reset() {
 		_texture = 0xFFFFFFFFu;
 		_src_blend_factor = 0xFFFFFFFFu;
@@ -179,7 +192,7 @@ public:
 	void DisableCache() { _disable_cache = true; }
 	void EnableCache()
 	{
-	   _disable_cache = true;
+	   _disable_cache = false;
 	   Reset();
 	}
 

core/rend/gles/gldraw.cpp

@@ -114,13 +114,30 @@ s32 SetTileClip(u32 val, GLint uniform)
 			csy /= scale_y;
 			cex /= scale_x;
 			cey /= scale_y;
-			float t = cey;
+			float dc2s_scale_h;
-			cey = 480 - csy;
+			float ds2s_offs_x;
-			csy = 480 - t;
+			float screen_stretching = settings.rend.ScreenStretching / 100.f;
-			float dc2s_scale_h = screen_height / 480.0f;
+
-			float ds2s_offs_x = (screen_width - dc2s_scale_h * 640) / 2;
+			if (settings.rend.Rotate90)
-			csx = csx * dc2s_scale_h + ds2s_offs_x;
+			{
-			cex = cex * dc2s_scale_h + ds2s_offs_x;
+				float t = cex;
+				cex = cey;
+				cey = 640 - csx;
+				csx = csy;
+				csy = 640 - t;
+				dc2s_scale_h = screen_height / 640.0f;
+				ds2s_offs_x =  (screen_width - dc2s_scale_h * 480.0 * screen_stretching) / 2;
+			}
+			else
+			{
+				float t = cey;
+				cey = 480 - csy;
+				csy = 480 - t;
+				dc2s_scale_h = screen_height / 480.0f;
+				ds2s_offs_x =  (screen_width - dc2s_scale_h * 640.0 * screen_stretching) / 2;
+			}
+			csx = csx * dc2s_scale_h * screen_stretching + ds2s_offs_x;
+			cex = cex * dc2s_scale_h * screen_stretching + ds2s_offs_x;
 			csy = csy * dc2s_scale_h;
 			cey = cey * dc2s_scale_h;
 		}
@@ -173,28 +190,25 @@ __forceinline
 		ShaderUniforms.trilinear_alpha = 1.f;
 
 	bool color_clamp = gp->tsp.ColorClamp && (pvrrc.fog_clamp_min != 0 || pvrrc.fog_clamp_max != 0xffffffff);
+	int fog_ctrl = settings.rend.Fog ? gp->tsp.FogCtrl : 2;
 
-	CurrentShader = &gl.pogram_table[
+	CurrentShader = GetProgram(Type == ListType_Punch_Through ? 1 : 0,
-									 GetProgramID(Type == ListType_Punch_Through ? 1 : 0,
+								  SetTileClip(gp->tileclip, -1) + 1,
-											 	  SetTileClip(gp->tileclip, -1) + 1,
+								  gp->pcw.Texture,
-												  gp->pcw.Texture,
+								  gp->tsp.UseAlpha,
-												  gp->tsp.UseAlpha,
+								  gp->tsp.IgnoreTexA,
-												  gp->tsp.IgnoreTexA,
+								  gp->tsp.ShadInstr,
-												  gp->tsp.ShadInstr,
+								  gp->pcw.Offset,
-												  gp->pcw.Offset,
+								  fog_ctrl,
-												  gp->tsp.FogCtrl,
+								  gp->pcw.Gouraud,
-												  gp->pcw.Gouraud,
+								  gp->tcw.PixelFmt == PixelBumpMap,
-												  gp->tcw.PixelFmt == PixelBumpMap,
+								  color_clamp,
-												  color_clamp,
+								  ShaderUniforms.trilinear_alpha != 1.f);
-												  ShaderUniforms.trilinear_alpha != 1.f)];
 	
-	if (CurrentShader->program == -1)
+	glcache.UseProgram(CurrentShader->program);
-		CompilePipelineShader(CurrentShader);
+	if (CurrentShader->trilinear_alpha != -1)
-	else
+		glUniform1f(CurrentShader->trilinear_alpha, ShaderUniforms.trilinear_alpha);
-	{
+
-		glcache.UseProgram(CurrentShader->program);
-		ShaderUniforms.Set(CurrentShader);
-	}
 	SetTileClip(gp->tileclip, CurrentShader->pp_ClipTest);
 
 	//This bit control which pixels are affected
@@ -1122,14 +1136,8 @@ static void DrawQuad(GLuint texId, float x, float y, float w, float h, float u0,
 
 	ShaderUniforms.trilinear_alpha = 1.0;
 
-	PipelineShader *shader = &gl.pogram_table[GetProgramID(0, 1, 1, 0, 1, 0, 0, 2, false, false, false, false)];
+	PipelineShader *shader = GetProgram(0, 1, 1, 0, 1, 0, 0, 2, false, false, false, false);
-	if (shader->program == -1)
+	glcache.UseProgram(shader->program);
-		CompilePipelineShader(shader);
-	else
-	{
-		glcache.UseProgram(shader->program);
-		ShaderUniforms.Set(shader);
-	}
 
 	glActiveTexture(GL_TEXTURE0);
 	glcache.BindTexture(GL_TEXTURE_2D, texId);
@@ -1150,17 +1158,27 @@ void DrawFramebuffer(float w, float h)
 
 bool render_output_framebuffer()
 {
-#if HOST_OS != OS_DARWIN
+	glcache.Disable(GL_SCISSOR_TEST);
-	//Fix this in a proper way
+	if (gl.gl_major < 3)
-	glBindFramebuffer(GL_FRAMEBUFFER, 0);
+	{
-#endif
+		glViewport(0, 0, screen_width, screen_height);
-	glViewport(0, 0, screen_width, screen_height);
+		if (gl.ofbo.tex == 0)
-	if (gl.ofbo.tex == 0)
+			return false;
-		return false;
+		glBindFramebuffer(GL_FRAMEBUFFER, 0);
-
+		float scl = 480.f / screen_height;
-    float scl = 480.f / screen_height;
+		float tx = (screen_width * scl - 640.f) / 2;
-    float tx = (screen_width * scl - 640.f) / 2;
+		DrawQuad(gl.ofbo.tex, -tx, 0, 640.f + tx * 2, 480.f, 0, 1, 1, 0);
-	DrawQuad(gl.ofbo.tex, -tx, 0, 640.f + tx * 2, 480.f, 0, 1, 1, 0);
+	}
-
+	else
+	{
+		if (gl.ofbo.fbo == 0)
+			return false;
+		glBindFramebuffer(GL_READ_FRAMEBUFFER, gl.ofbo.fbo);
+		glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+		glBlitFramebuffer(0, 0, gl.ofbo.width, gl.ofbo.height,
+				0, 0, screen_width, screen_height,
+				GL_COLOR_BUFFER_BIT, GL_LINEAR);
+    	glBindFramebuffer(GL_FRAMEBUFFER, 0);
+	}
 	return true;
 }

core/rend/gles/gles.cpp

@@ -79,6 +79,7 @@ const char* VertexShaderSource =
 %s \n\
 #define TARGET_GL %s \n\
 #define pp_Gouraud %d \n\
+#define ROTATE_90 %d \n\
  \n\
 #define GLES2 0 \n\
 #define GLES3 1 \n\
@@ -136,6 +137,9 @@ void main() \n\
 	vpos.z = vpos.w; \n\
 #else \n\
 	vpos.z=depth_scale.x+depth_scale.y*vpos.w;  \n\
+#endif \n\
+#if ROTATE_90 == 1 \n\
+	vpos.xy = vec2(vpos.y, -vpos.x);  \n\
 #endif \n\
 	vpos.xy=vpos.xy*scale.xy-scale.zw;  \n\
 	vpos.xy*=vpos.w;  \n\
@@ -850,9 +854,20 @@ GLuint fogTextureId;
 extern void gl_term();
 #endif
 
+static void gl_delete_shaders()
+{
+	for (auto it : gl.shaders)
+	{
+		if (it.second.program != 0)
+			glcache.DeleteProgram(it.second.program);
+	}
+	gl.shaders.clear();
+	glcache.DeleteProgram(gl.modvol_shader.program);
+	gl.modvol_shader.program = 0;
+}
+
 static void gles_term()
 {
-	glDeleteProgram(gl.modvol_shader.program);
 	glDeleteBuffers(1, &gl.vbo.geometry);
 	gl.vbo.geometry = 0;
 	glDeleteBuffers(1, &gl.vbo.modvols);
@@ -865,7 +880,7 @@ static void gles_term()
 	gl_free_osd_resources();
 	free_output_framebuffer();
 
-	memset(gl.pogram_table, 0, sizeof(gl.pogram_table));
+	gl_delete_shaders();
 	gl_term();
 }
 
@@ -1014,10 +1029,15 @@ GLuint gl_CompileAndLink(const char* VertexShader, const char* FragmentShader)
 	return program;
 }
 
-int GetProgramID(u32 cp_AlphaTest, u32 pp_ClipTestMode,
+PipelineShader *GetProgram(u32 cp_AlphaTest, u32 pp_ClipTestMode,
 							u32 pp_Texture, u32 pp_UseAlpha, u32 pp_IgnoreTexA, u32 pp_ShadInstr, u32 pp_Offset,
 							u32 pp_FogCtrl, bool pp_Gouraud, bool pp_BumpMap, bool fog_clamping, bool trilinear)
 {
+	if (settings.rend.Rotate90 != gl.rotate90)
+	{
+		gl_delete_shaders();
+		gl.rotate90 = settings.rend.Rotate90;
+	}
 	u32 rv=0;
 
 	rv|=pp_ClipTestMode;
@@ -1033,14 +1053,32 @@ int GetProgramID(u32 cp_AlphaTest, u32 pp_ClipTestMode,
 	rv<<=1; rv|=fog_clamping;
 	rv<<=1; rv|=trilinear;
 
-	return rv;
+	PipelineShader *shader = &gl.shaders[rv];
+	if (shader->program == 0)
+	{
+		shader->cp_AlphaTest = cp_AlphaTest;
+		shader->pp_ClipTestMode = pp_ClipTestMode-1;
+		shader->pp_Texture = pp_Texture;
+		shader->pp_UseAlpha = pp_UseAlpha;
+		shader->pp_IgnoreTexA = pp_IgnoreTexA;
+		shader->pp_ShadInstr = pp_ShadInstr;
+		shader->pp_Offset = pp_Offset;
+		shader->pp_FogCtrl = pp_FogCtrl;
+		shader->pp_Gouraud = pp_Gouraud;
+		shader->pp_BumpMap = pp_BumpMap;
+		shader->fog_clamping = fog_clamping;
+		shader->trilinear = trilinear;
+		CompilePipelineShader(shader);
+	}
+
+	return shader;
 }
 
 bool CompilePipelineShader(	PipelineShader* s)
 {
 	char vshader[8192];
 
-	sprintf(vshader, VertexShaderSource, gl.glsl_version_header, gl.gl_version, s->pp_Gouraud);
+	sprintf(vshader, VertexShaderSource, gl.glsl_version_header, gl.gl_version, s->pp_Gouraud, settings.rend.Rotate90);
 
 	char pshader[8192];
 
@@ -1126,13 +1164,30 @@ void gl_load_osd_resources()
 
 void gl_free_osd_resources()
 {
-	glDeleteProgram(gl.OSD_SHADER.program);
+	glcache.DeleteProgram(gl.OSD_SHADER.program);
 
     if (osd_tex != 0) {
         glcache.DeleteTextures(1, &osd_tex);
         osd_tex = 0;
     }
 }
+
+static void create_modvol_shader()
+{
+	if (gl.modvol_shader.program != 0)
+		return;
+	char vshader[8192];
+	sprintf(vshader, VertexShaderSource, gl.glsl_version_header, gl.gl_version, 1, settings.rend.Rotate90);
+	char fshader[8192];
+	sprintf(fshader, ModifierVolumeShader, gl.glsl_version_header, gl.gl_version);
+
+	gl.modvol_shader.program=gl_CompileAndLink(vshader, fshader);
+	gl.modvol_shader.scale          = glGetUniformLocation(gl.modvol_shader.program, "scale");
+	gl.modvol_shader.sp_ShaderColor = glGetUniformLocation(gl.modvol_shader.program, "sp_ShaderColor");
+	gl.modvol_shader.depth_scale    = glGetUniformLocation(gl.modvol_shader.program, "depth_scale");
+	gl.modvol_shader.extra_depth_scale = glGetUniformLocation(gl.modvol_shader.program, "extra_depth_scale");
+}
+
 bool gl_create_resources()
 {
 	if (gl.vbo.geometry != 0)
@@ -1156,84 +1211,7 @@ bool gl_create_resources()
 	glGenBuffers(1, &gl.vbo.idxs);
 	glGenBuffers(1, &gl.vbo.idxs2);
 
-	memset(gl.pogram_table,0,sizeof(gl.pogram_table));
+	create_modvol_shader();
-
-	PipelineShader* dshader=0;
-	u32 compile=0;
-#define forl(name,max) for(u32 name=0;name<=max;name++)
-	forl(cp_AlphaTest,1)
-	{
-		forl(pp_ClipTestMode,2)
-		{
-			forl(pp_UseAlpha,1)
-			{
-				forl(pp_Texture,1)
-				{
-					forl(pp_FogCtrl,3)
-					{
-						forl(pp_IgnoreTexA,1)
-						{
-							forl(pp_ShadInstr,3)
-							{
-								forl(pp_Offset,1)
-								{
-									forl(pp_Gouraud,1)
-									{
-										forl(pp_BumpMap,1)
-										{
-											forl(fog_clamping,1)
-											{
-												forl(trilinear,1)
-												{
-													dshader=&gl.pogram_table[GetProgramID(cp_AlphaTest,pp_ClipTestMode,pp_Texture,pp_UseAlpha,pp_IgnoreTexA,
-																			pp_ShadInstr,pp_Offset,pp_FogCtrl, (bool)pp_Gouraud, (bool)pp_BumpMap, (bool)fog_clamping,
-																			(bool)trilinear)];
-
-														dshader->cp_AlphaTest = cp_AlphaTest;
-														dshader->pp_ClipTestMode = pp_ClipTestMode-1;
-														dshader->pp_Texture = pp_Texture;
-														dshader->pp_UseAlpha = pp_UseAlpha;
-														dshader->pp_IgnoreTexA = pp_IgnoreTexA;
-														dshader->pp_ShadInstr = pp_ShadInstr;
-														dshader->pp_Offset = pp_Offset;
-														dshader->pp_FogCtrl = pp_FogCtrl;
-														dshader->pp_Gouraud = pp_Gouraud;
-														dshader->pp_BumpMap = pp_BumpMap;
-														dshader->fog_clamping = fog_clamping;
-														dshader->trilinear = trilinear;
-														dshader->program = -1;
-												}
-											}
-										}
-									}
-								}
-							}
-						}
-					}
-				}
-			}
-		}
-	}
-	
-	char vshader[8192];
-	sprintf(vshader, VertexShaderSource, gl.glsl_version_header, gl.gl_version, 1);
-	char fshader[8192];
-	sprintf(fshader, ModifierVolumeShader, gl.glsl_version_header, gl.gl_version);
-
-	gl.modvol_shader.program=gl_CompileAndLink(vshader, fshader);
-	gl.modvol_shader.scale          = glGetUniformLocation(gl.modvol_shader.program, "scale");
-	gl.modvol_shader.sp_ShaderColor = glGetUniformLocation(gl.modvol_shader.program, "sp_ShaderColor");
-	gl.modvol_shader.depth_scale    = glGetUniformLocation(gl.modvol_shader.program, "depth_scale");
-	gl.modvol_shader.extra_depth_scale = glGetUniformLocation(gl.modvol_shader.program, "extra_depth_scale");
-
-	//#define PRECOMPILE_SHADERS
-	#ifdef PRECOMPILE_SHADERS
-	for (u32 i=0;i<sizeof(gl.pogram_table)/sizeof(gl.pogram_table[0]);i++)
-	{
-		if (!CompilePipelineShader(	&gl.pogram_table[i] ))
-			return false;
-	}
-	#endif
 
 	gl_load_osd_resources();
 
@@ -1553,6 +1531,7 @@ static void upload_vertex_indices()
 bool RenderFrame()
 {
 	DoCleanup();
+	create_modvol_shader();
 
 	bool is_rtt=pvrrc.isRTT;
 
@@ -1719,17 +1698,41 @@ bool RenderFrame()
 	/*
 		Handle Dc to screen scaling
 	*/
-	float screen_scaling = is_rtt ? 1.f : settings.rend.ScreenScaling / 100.f;
+	float screen_stretching = settings.rend.ScreenStretching / 100.f;
-	float dc2s_scale_h = is_rtt ? (screen_width / dc_width) : (screen_height / 480.0);
+	float screen_scaling = settings.rend.ScreenScaling / 100.f;
-	dc2s_scale_h *=  screen_scaling;
-	float ds2s_offs_x =  is_rtt ? 0 : (((screen_width * screen_scaling) - dc2s_scale_h * 640.0) / 2);
 
-	//-1 -> too much to left
+	float dc2s_scale_h;
-	ShaderUniforms.scale_coefs[0] = 2.0f / (screen_width * screen_scaling / dc2s_scale_h * scale_x);
+	float ds2s_offs_x;
-	ShaderUniforms.scale_coefs[1]= (is_rtt ? 2 : -2) / dc_height;		// FIXME CT2 needs 480 here instead of dc_height=512
-	ShaderUniforms.scale_coefs[2]= 1 - 2 * ds2s_offs_x / (screen_width * screen_scaling);
-	ShaderUniforms.scale_coefs[3]= (is_rtt ? 1 : -1);
 
+	if (is_rtt)
+	{
+		ShaderUniforms.scale_coefs[0] = 2.0f / dc_width;
+		ShaderUniforms.scale_coefs[1] = 2.0f / dc_height;	// FIXME CT2 needs 480 here instead of dc_height=512
+		ShaderUniforms.scale_coefs[2] = 1;
+		ShaderUniforms.scale_coefs[3] = 1;
+	}
+	else
+	{
+		if (settings.rend.Rotate90)
+		{
+			dc2s_scale_h = screen_height / 640.0f;
+			ds2s_offs_x =  (screen_width - dc2s_scale_h * 480.0f * screen_stretching) / 2;
+			ShaderUniforms.scale_coefs[0] = 2.0f / (screen_width / dc2s_scale_h * scale_x) * screen_stretching;
+			ShaderUniforms.scale_coefs[1] = -2.0f / dc_width;
+			ShaderUniforms.scale_coefs[2] = 1 - 2 * ds2s_offs_x / screen_width;
+			ShaderUniforms.scale_coefs[3] = 1;
+		}
+		else
+		{
+			dc2s_scale_h = screen_height / 480.0f;
+			ds2s_offs_x =  (screen_width - dc2s_scale_h * 640.0f * screen_stretching) / 2;
+			ShaderUniforms.scale_coefs[0] = 2.0f / (screen_width / dc2s_scale_h * scale_x) * screen_stretching;
+			ShaderUniforms.scale_coefs[1] = -2.0f / dc_height;
+			ShaderUniforms.scale_coefs[2] = 1 - 2 * ds2s_offs_x / screen_width;
+			ShaderUniforms.scale_coefs[3] = -1;
+		}
+		//-1 -> too much to left
+	}
 
 	ShaderUniforms.depth_coefs[0]=2/(vtx_max_fZ-vtx_min_fZ);
 	ShaderUniforms.depth_coefs[1]=-vtx_min_fZ-1;
@@ -1768,7 +1771,7 @@ bool RenderFrame()
 	ShaderUniforms.fog_clamp_max[2] = ((pvrrc.fog_clamp_max >> 0) & 0xFF) / 255.0f;
 	ShaderUniforms.fog_clamp_max[3] = ((pvrrc.fog_clamp_max >> 24) & 0xFF) / 255.0f;
 	
-	if (fog_needs_update)
+	if (fog_needs_update && settings.rend.Fog)
 	{
 		fog_needs_update = false;
 		UpdateFogTexture((u8 *)FOG_TABLE, GL_TEXTURE1, gl.fog_image_format);
@@ -1782,16 +1785,12 @@ bool RenderFrame()
 
 	ShaderUniforms.PT_ALPHA=(PT_ALPHA_REF&0xFF)/255.0f;
 
-//	for (u32 i=0;i<sizeof(gl.pogram_table)/sizeof(gl.pogram_table[0]);i++)
+	for (auto it : gl.shaders)
-//	{
+	{
-//		PipelineShader* s=&gl.pogram_table[i];
+		glcache.UseProgram(it.second.program);
-//		if (s->program == -1)
+		ShaderUniforms.Set(&it.second);
-//			continue;
+	}
-//
+
-//		glcache.UseProgram(s->program);
-//
-//		ShaderUniforms.Set(s);
-//	}
 	//setup render target first
 	if (is_rtt)
 	{
@@ -1836,7 +1835,7 @@ bool RenderFrame()
 	{
 		if (settings.rend.ScreenScaling != 100 || gl.swap_buffer_not_preserved)
 		{
-			init_output_framebuffer(screen_width * screen_scaling, screen_height * screen_scaling);
+			init_output_framebuffer(screen_width * screen_scaling + 0.5f, screen_height * screen_scaling + 0.5f);
 		}
 		else
 		{
@@ -1883,9 +1882,6 @@ bool RenderFrame()
 			glBufferData(GL_ARRAY_BUFFER,pvrrc.modtrig.bytes(),pvrrc.modtrig.head(),GL_STREAM_DRAW); glCheck();
 		}
 
-		int offs_x=ds2s_offs_x+0.5f;
-		//this needs to be scaled
-
 		//not all scaling affects pixel operations, scale to adjust for that
 		scale_x *= scissoring_scale_x;
 
@@ -1904,22 +1900,31 @@ bool RenderFrame()
 			float min_y = pvrrc.fb_Y_CLIP.min / scale_y;
 			if (!is_rtt)
 			{
+				if (settings.rend.Rotate90)
+				{
+					float t = width;
+					width = height;
+					height = t;
+					t = min_x;
+					min_x = min_y;
+					min_y = 640 - t - height;
+				}
 				// Add x offset for aspect ratio > 4/3
-				min_x = min_x * dc2s_scale_h + ds2s_offs_x;
+				min_x = (min_x * dc2s_scale_h * screen_stretching + ds2s_offs_x) * screen_scaling;
 				// Invert y coordinates when rendering to screen
-				min_y = screen_height * screen_scaling - (min_y + height) * dc2s_scale_h;
+				min_y = (screen_height - (min_y + height) * dc2s_scale_h) * screen_scaling;
-				width *= dc2s_scale_h;
+				width *= dc2s_scale_h * screen_stretching * screen_scaling;
-				height *= dc2s_scale_h;
+				height *= dc2s_scale_h * screen_scaling;
 
 				if (ds2s_offs_x > 0)
 				{
-					float rounded_offs_x = ds2s_offs_x + 0.5f;
+					float scaled_offs_x = ds2s_offs_x * screen_scaling;
 
 					glcache.ClearColor(0.f, 0.f, 0.f, 0.f);
 					glcache.Enable(GL_SCISSOR_TEST);
-					glScissor(0, 0, rounded_offs_x, screen_height);
+					glScissor(0, 0, scaled_offs_x + 0.5f, screen_height * screen_scaling + 0.5f);
 					glClear(GL_COLOR_BUFFER_BIT);
-					glScissor(screen_width - rounded_offs_x, 0, rounded_offs_x, screen_height);
+					glScissor(screen_width * screen_scaling - scaled_offs_x + 0.5f, 0, scaled_offs_x + 1.f, screen_height * screen_scaling + 0.5f);
 					glClear(GL_COLOR_BUFFER_BIT);
 				}
 			}

core/rend/gles/gles.h

@@ -1,4 +1,6 @@
 #pragma once
+#include <unordered_map>
+#include <atomic>
 #include "rend/rend.h"
 
 #if (defined(GLES) && !defined(TARGET_NACL32) && HOST_OS != OS_DARWIN && !defined(USE_SDL)) || defined(_ANDROID)
@@ -93,7 +95,9 @@ struct gl_ctx
 
 	} modvol_shader;
 
-	PipelineShader pogram_table[24576];
+	std::unordered_map<u32, PipelineShader> shaders;
+	bool rotate90;
+
 	struct
 	{
 		GLuint program;
@@ -117,6 +121,7 @@ struct gl_ctx
 	struct
 	{
 		GLuint depthb;
+		GLuint colorb;
 		GLuint tex;
 		GLuint fbo;
 		int width;
@@ -176,7 +181,7 @@ void free_output_framebuffer();
 
 void HideOSD();
 void OSD_DRAW(bool clear_screen);
-int GetProgramID(u32 cp_AlphaTest, u32 pp_ClipTestMode,
+PipelineShader *GetProgram(u32 cp_AlphaTest, u32 pp_ClipTestMode,
 							u32 pp_Texture, u32 pp_UseAlpha, u32 pp_IgnoreTexA, u32 pp_ShadInstr, u32 pp_Offset,
 							u32 pp_FogCtrl, bool pp_Gouraud, bool pp_BumpMap, bool fog_clamping, bool trilinear);
 
@@ -223,9 +228,6 @@ extern struct ShaderUniforms_t
 		if (s->sp_FOG_COL_VERT!=-1)
 			glUniform3fv( s->sp_FOG_COL_VERT, 1, ps_FOG_COL_VERT);
 
-		if (s->trilinear_alpha != -1)
-			glUniform1f(s->trilinear_alpha, trilinear_alpha);
-		
 		if (s->fog_clamp_min != -1)
 			glUniform4fv(s->fog_clamp_min, 1, fog_clamp_min);
 		if (s->fog_clamp_max != -1)
@@ -272,10 +274,10 @@ struct TextureCacheData
 	//a texture can't be both VQ and PAL at the same time
 	u32 texture_hash;			// xxhash of texture data, used for custom textures
 	u32 old_texture_hash;		// legacy hash
-	u8* custom_image_data;		// loaded custom image data
+	u8* volatile custom_image_data;		// loaded custom image data
-	u32 custom_width;
+	volatile u32 custom_width;
-	u32 custom_height;
+	volatile u32 custom_height;
-	bool custom_load_in_progress;
+	std::atomic_int custom_load_in_progress;
 	
 	void PrintTextureName();
 	

core/rend/gles/gltex.cpp

@@ -296,10 +296,7 @@ void TextureCacheData::Update()
 		}
 	}
 	if (settings.rend.CustomTextures)
-	{
-		custom_load_in_progress = true;
 		custom_texture.LoadCustomTextureAsync(this);
-	}
 
 	void *temp_tex_buffer = NULL;
 	u32 upscaled_w = w;
@@ -428,7 +425,7 @@ void TextureCacheData::UploadToGPU(GLuint textype, int width, int height, u8 *te
 
 void TextureCacheData::CheckCustomTexture()
 {
-	if (custom_image_data != NULL)
+	if (custom_load_in_progress == 0 && custom_image_data != NULL)
 	{
 		UploadToGPU(GL_UNSIGNED_BYTE, custom_width, custom_height, custom_image_data);
 		delete [] custom_image_data;
@@ -446,7 +443,7 @@ bool TextureCacheData::NeedsUpdate() {
 	
 bool TextureCacheData::Delete()
 {
-	if (custom_load_in_progress)
+	if (custom_load_in_progress > 0)
 		return false;
 	
 	if (pData) {
@@ -736,11 +733,7 @@ TextureCacheData *getTextureCacheData(TSP tsp, TCW tcw) {
 	}
 	else //create if not existing
 	{
-		TextureCacheData tfc={0};
+		tf=&TexCache[key];
-		TexCache[key] = tfc;
-
-		tx=TexCache.find(key);
-		tf=&tx->second;
 
 		tf->tsp = tsp;
 		tf->tcw = tcw;
@@ -800,11 +793,7 @@ text_info raw_GetTexture(TSP tsp, TCW tcw)
 	}
 	else //create if not existing
 	{
-		TextureCacheData tfc = { 0 };
+		tf = &TexCache[key];
-		TexCache[key] = tfc;
-
-		tx = TexCache.find(key);
-		tf = &tx->second;
 
 		tf->tsp = tsp;
 		tf->tcw = tcw;
@@ -1007,13 +996,7 @@ GLuint init_output_framebuffer(int width, int height)
 {
 	if (width != gl.ofbo.width || height != gl.ofbo.height)
 	{
-		if (gl.ofbo.fbo != 0)
+		free_output_framebuffer();
-		{
-			glDeleteFramebuffers(1, &gl.ofbo.fbo);
-			gl.ofbo.fbo = 0;
-			glDeleteRenderbuffers(1, &gl.ofbo.depthb);
-			glcache.DeleteTextures(1, &gl.ofbo.tex);
-		}
 		gl.ofbo.width = width;
 		gl.ofbo.height = height;
 	}
@@ -1037,15 +1020,25 @@ GLuint init_output_framebuffer(int width, int height)
 		else
 			glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH24_STENCIL8, width, height);
 
-		// Create a texture for rendering to
+		if (gl.gl_major < 3)
-		gl.ofbo.tex = glcache.GenTexture();
+		{
-		glcache.BindTexture(GL_TEXTURE_2D, gl.ofbo.tex);
+			// Create a texture for rendering to
+			gl.ofbo.tex = glcache.GenTexture();
+			glcache.BindTexture(GL_TEXTURE_2D, gl.ofbo.tex);
 
-		glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, 0);
+			glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, 0);
-		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+			glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+			glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
-		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+			glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+			glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+		}
+		else
+		{
+			// Use a renderbuffer and glBlitFramebuffer
+			glGenRenderbuffers(1, &gl.ofbo.colorb);
+			glBindRenderbuffer(GL_RENDERBUFFER, gl.ofbo.colorb);
+			glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, width, height);
+		}
 
 		// Create the framebuffer
 		glGenFramebuffers(1, &gl.ofbo.fbo);
@@ -1057,13 +1050,20 @@ GLuint init_output_framebuffer(int width, int height)
 		if (!gl.is_gles || gl.GL_OES_packed_depth_stencil_supported)
 			glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_RENDERBUFFER, gl.ofbo.depthb);
 
-		// Attach the texture to the FBO
+		// Attach the texture/renderbuffer to the FBO
-		glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, gl.ofbo.tex, 0);
+		if (gl.gl_major < 3)
+			glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, gl.ofbo.tex, 0);
+		else
+			glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, gl.ofbo.colorb);
 
 		// Check that our FBO creation was successful
 		GLuint uStatus = glCheckFramebufferStatus(GL_FRAMEBUFFER);
 
 		verify(uStatus == GL_FRAMEBUFFER_COMPLETE);
+
+		glcache.Disable(GL_SCISSOR_TEST);
+		glcache.ClearColor(0.f, 0.f, 0.f, 0.f);
+		glClear(GL_COLOR_BUFFER_BIT);
 	}
 	else
 		glBindFramebuffer(GL_FRAMEBUFFER, gl.ofbo.fbo);
@@ -1082,7 +1082,15 @@ void free_output_framebuffer()
 		gl.ofbo.fbo = 0;
 		glDeleteRenderbuffers(1, &gl.ofbo.depthb);
 		gl.ofbo.depthb = 0;
-		glcache.DeleteTextures(1, &gl.ofbo.tex);
+		if (gl.ofbo.tex != 0)
-		gl.ofbo.tex = 0;
+		{
+			glcache.DeleteTextures(1, &gl.ofbo.tex);
+			gl.ofbo.tex = 0;
+		}
+		if (gl.ofbo.colorb != 0)
+		{
+			glDeleteRenderbuffers(1, &gl.ofbo.colorb);
+			gl.ofbo.colorb = 0;
+		}
 	}
 }

core/rend/gles/imgui_impl_opengl3.cpp

@@ -67,6 +67,7 @@
 #endif
 
 #include "gles.h"
+#include "glcache.h"
 
 // OpenGL Data
 static char         g_GlslVersionString[32] = "";
@@ -127,28 +128,7 @@ void    ImGui_ImplOpenGL3_RenderDrawData(ImDrawData* draw_data, bool save_backgr
     draw_data->ScaleClipRects(io.DisplayFramebufferScale);
 
     // Backup GL state
-    GLenum last_active_texture; glGetIntegerv(GL_ACTIVE_TEXTURE, (GLint*)&last_active_texture);
     glActiveTexture(GL_TEXTURE0);
-    GLint last_program; glGetIntegerv(GL_CURRENT_PROGRAM, &last_program);
-    GLint last_texture; glGetIntegerv(GL_TEXTURE_BINDING_2D, &last_texture);
-    GLint last_sampler; glGetIntegerv(GL_SAMPLER_BINDING, &last_sampler);
-    GLint last_array_buffer; glGetIntegerv(GL_ARRAY_BUFFER_BINDING, &last_array_buffer);
-    GLint last_vertex_array; glGetIntegerv(GL_VERTEX_ARRAY_BINDING, &last_vertex_array);
-#ifdef GL_POLYGON_MODE
-    GLint last_polygon_mode[2]; glGetIntegerv(GL_POLYGON_MODE, last_polygon_mode);
-#endif
-    GLint last_viewport[4]; glGetIntegerv(GL_VIEWPORT, last_viewport);
-    GLint last_scissor_box[4]; glGetIntegerv(GL_SCISSOR_BOX, last_scissor_box);
-    GLenum last_blend_src_rgb; glGetIntegerv(GL_BLEND_SRC_RGB, (GLint*)&last_blend_src_rgb);
-    GLenum last_blend_dst_rgb; glGetIntegerv(GL_BLEND_DST_RGB, (GLint*)&last_blend_dst_rgb);
-    GLenum last_blend_src_alpha; glGetIntegerv(GL_BLEND_SRC_ALPHA, (GLint*)&last_blend_src_alpha);
-    GLenum last_blend_dst_alpha; glGetIntegerv(GL_BLEND_DST_ALPHA, (GLint*)&last_blend_dst_alpha);
-    GLenum last_blend_equation_rgb; glGetIntegerv(GL_BLEND_EQUATION_RGB, (GLint*)&last_blend_equation_rgb);
-    GLenum last_blend_equation_alpha; glGetIntegerv(GL_BLEND_EQUATION_ALPHA, (GLint*)&last_blend_equation_alpha);
-    GLboolean last_enable_blend = glIsEnabled(GL_BLEND);
-    GLboolean last_enable_cull_face = glIsEnabled(GL_CULL_FACE);
-    GLboolean last_enable_depth_test = glIsEnabled(GL_DEPTH_TEST);
-    GLboolean last_enable_scissor_test = glIsEnabled(GL_SCISSOR_TEST);
     bool clip_origin_lower_left = true;
 #ifdef GL_CLIP_ORIGIN
     if (gl.gl_major >= 4 && glClipControl != NULL)
@@ -166,13 +146,13 @@ void    ImGui_ImplOpenGL3_RenderDrawData(ImDrawData* draw_data, bool save_backgr
 
 		// (Re-)create the background texture and reserve space for it
 		if (g_BackgroundTexture != 0)
-	    	glDeleteTextures(1, &g_BackgroundTexture);
+			glcache.DeleteTextures(1, &g_BackgroundTexture);
-		glGenTextures(1, &g_BackgroundTexture);
+		g_BackgroundTexture = glcache.GenTexture();
-		glBindTexture(GL_TEXTURE_2D, g_BackgroundTexture);
+		glcache.BindTexture(GL_TEXTURE_2D, g_BackgroundTexture);
-		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+		glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 		glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB, fb_width, fb_height, 0, GL_RGB, GL_UNSIGNED_BYTE, (GLvoid*)NULL);
 
 		// Copy the current framebuffer into it
@@ -180,12 +160,12 @@ void    ImGui_ImplOpenGL3_RenderDrawData(ImDrawData* draw_data, bool save_backgr
     }
 
     // Setup render state: alpha-blending enabled, no face culling, no depth testing, scissor enabled, polygon fill
-    glEnable(GL_BLEND);
+    glcache.Enable(GL_BLEND);
     glBlendEquation(GL_FUNC_ADD);
-    glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
+    glcache.BlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
-    glDisable(GL_CULL_FACE);
+    glcache.Disable(GL_CULL_FACE);
-    glDisable(GL_DEPTH_TEST);
+    glcache.Disable(GL_DEPTH_TEST);
-    glEnable(GL_SCISSOR_TEST);
+    glcache.Enable(GL_SCISSOR_TEST);
 #ifdef GL_POLYGON_MODE
     if (glPolygonMode != NULL)
     	glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
@@ -205,7 +185,7 @@ void    ImGui_ImplOpenGL3_RenderDrawData(ImDrawData* draw_data, bool save_backgr
         { 0.0f,         0.0f,        -1.0f,   0.0f },
         { (R+L)/(L-R),  (T+B)/(B-T),  0.0f,   1.0f },
     };
-    glUseProgram(g_ShaderHandle);
+    glcache.UseProgram(g_ShaderHandle);
     glUniform1i(g_AttribLocationTex, 0);
     glUniformMatrix4fv(g_AttribLocationProjMtx, 1, GL_FALSE, &ortho_projection[0][0]);
     if (gl.gl_major >= 3 && glBindSampler != NULL)
@@ -260,7 +240,7 @@ void    ImGui_ImplOpenGL3_RenderDrawData(ImDrawData* draw_data, bool save_backgr
                         glScissor((int)clip_rect.x, (int)clip_rect.y, (int)clip_rect.z, (int)clip_rect.w); // Support for GL 4.5's glClipControl(GL_UPPER_LEFT)
 
                     // Bind texture, Draw
-                    glBindTexture(GL_TEXTURE_2D, (GLuint)(intptr_t)pcmd->TextureId);
+                    glcache.BindTexture(GL_TEXTURE_2D, (GLuint)(intptr_t)pcmd->TextureId);
                     glDrawElements(GL_TRIANGLES, (GLsizei)pcmd->ElemCount, sizeof(ImDrawIdx) == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT, idx_buffer_offset);
                 }
             }
@@ -269,28 +249,6 @@ void    ImGui_ImplOpenGL3_RenderDrawData(ImDrawData* draw_data, bool save_backgr
     }
     if (vao_handle != 0)
     	glDeleteVertexArrays(1, &vao_handle);
-
-    // Restore modified GL state
-    glUseProgram(last_program);
-    glBindTexture(GL_TEXTURE_2D, last_texture);
-    if (gl.gl_major >= 3 && glBindSampler != NULL)
-    	glBindSampler(0, last_sampler);
-    glActiveTexture(last_active_texture);
-    glBindBuffer(GL_ARRAY_BUFFER, last_array_buffer);
-    if (gl.gl_major >= 3)
-    	glBindVertexArray(last_vertex_array);
-    glBlendEquationSeparate(last_blend_equation_rgb, last_blend_equation_alpha);
-    glBlendFuncSeparate(last_blend_src_rgb, last_blend_dst_rgb, last_blend_src_alpha, last_blend_dst_alpha);
-    if (last_enable_blend) glEnable(GL_BLEND); else glDisable(GL_BLEND);
-    if (last_enable_cull_face) glEnable(GL_CULL_FACE); else glDisable(GL_CULL_FACE);
-    if (last_enable_depth_test) glEnable(GL_DEPTH_TEST); else glDisable(GL_DEPTH_TEST);
-    if (last_enable_scissor_test) glEnable(GL_SCISSOR_TEST); else glDisable(GL_SCISSOR_TEST);
-#ifdef GL_POLYGON_MODE
-    if (glPolygonMode != NULL)
-    	glPolygonMode(GL_FRONT_AND_BACK, (GLenum)last_polygon_mode[0]);
-#endif
-    glViewport(last_viewport[0], last_viewport[1], (GLsizei)last_viewport[2], (GLsizei)last_viewport[3]);
-    glScissor(last_scissor_box[0], last_scissor_box[1], (GLsizei)last_scissor_box[2], (GLsizei)last_scissor_box[3]);
 }
 
 bool ImGui_ImplOpenGL3_CreateFontsTexture()
@@ -302,21 +260,16 @@ bool ImGui_ImplOpenGL3_CreateFontsTexture()
     io.Fonts->GetTexDataAsRGBA32(&pixels, &width, &height);   // Load as RGBA 32-bits (75% of the memory is wasted, but default font is so small) because it is more likely to be compatible with user's existing shaders. If your ImTextureId represent a higher-level concept than just a GL texture id, consider calling GetTexDataAsAlpha8() instead to save on GPU memory.
 
     // Upload texture to graphics system
-    GLint last_texture;
+    g_FontTexture = glcache.GenTexture();
-    glGetIntegerv(GL_TEXTURE_BINDING_2D, &last_texture);
+    glcache.BindTexture(GL_TEXTURE_2D, g_FontTexture);
-    glGenTextures(1, &g_FontTexture);
+    glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-    glBindTexture(GL_TEXTURE_2D, g_FontTexture);
+    glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
     glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, pixels);
 
     // Store our identifier
     io.Fonts->TexID = (ImTextureID)(intptr_t)g_FontTexture;
 
-    // Restore state
-    glBindTexture(GL_TEXTURE_2D, last_texture);
-
     return true;
 }
 
@@ -325,7 +278,7 @@ void ImGui_ImplOpenGL3_DestroyFontsTexture()
     if (g_FontTexture)
     {
         ImGuiIO& io = ImGui::GetIO();
-        glDeleteTextures(1, &g_FontTexture);
+        glcache.DeleteTextures(1, &g_FontTexture);
         io.Fonts->TexID = 0;
         g_FontTexture = 0;
     }
@@ -369,12 +322,6 @@ static bool CheckProgram(GLuint handle, const char* desc)
 
 bool    ImGui_ImplOpenGL3_CreateDeviceObjects()
 {
-    // Backup GL state
-    GLint last_texture, last_array_buffer, last_vertex_array;
-    glGetIntegerv(GL_TEXTURE_BINDING_2D, &last_texture);
-    glGetIntegerv(GL_ARRAY_BUFFER_BINDING, &last_array_buffer);
-    glGetIntegerv(GL_VERTEX_ARRAY_BINDING, &last_vertex_array);
-
     // Parse GLSL version string
     int glsl_version = 130;
     sscanf(g_GlslVersionString, "#version %d", &glsl_version);
@@ -534,12 +481,6 @@ bool    ImGui_ImplOpenGL3_CreateDeviceObjects()
 
     ImGui_ImplOpenGL3_CreateFontsTexture();
 
-    // Restore modified GL state
-    glBindTexture(GL_TEXTURE_2D, last_texture);
-    glBindBuffer(GL_ARRAY_BUFFER, last_array_buffer);
-    if (gl.gl_major >= 3)
-    	glBindVertexArray(last_vertex_array);
-
     return true;
 }
 
@@ -549,26 +490,23 @@ void ImGui_ImplOpenGL3_DestroyDeviceObjects()
     if (g_ElementsHandle) glDeleteBuffers(1, &g_ElementsHandle);
     g_VboHandle = g_ElementsHandle = 0;
 
-    if (g_ShaderHandle && g_VertHandle) glDetachShader(g_ShaderHandle, g_VertHandle);
+    glcache.DeleteProgram(g_ShaderHandle);
-    if (g_VertHandle) glDeleteShader(g_VertHandle);
     g_VertHandle = 0;
-
-    if (g_ShaderHandle && g_FragHandle) glDetachShader(g_ShaderHandle, g_FragHandle);
-    if (g_FragHandle) glDeleteShader(g_FragHandle);
     g_FragHandle = 0;
-
-    if (g_ShaderHandle) glDeleteProgram(g_ShaderHandle);
     g_ShaderHandle = 0;
 
     ImGui_ImplOpenGL3_DestroyFontsTexture();
 
     if (g_BackgroundTexture != 0)
-    	glDeleteTextures(1, &g_BackgroundTexture);
+    	glcache.DeleteTextures(1, &g_BackgroundTexture);
     g_BackgroundTexture = 0;
 }
 
 void ImGui_ImplOpenGL3_DrawBackground()
 {
+	glcache.Disable(GL_SCISSOR_TEST);
+	glcache.ClearColor(0, 0, 0, 0);
+	glClear(GL_COLOR_BUFFER_BIT);
 	if (g_BackgroundTexture != 0)
 	{
 		ImGuiIO& io = ImGui::GetIO();
@@ -579,20 +517,14 @@ void ImGui_ImplOpenGL3_DrawBackground()
 		ImGui::GetWindowDrawList()->AddImage((ImTextureID)(uintptr_t)g_BackgroundTexture, ImVec2(0, 0), io.DisplaySize, ImVec2(0, 1), ImVec2(1, 0), 0xffffffff);
 		ImGui::End();
 	}
-	else
-	{
-		glClearColor(0, 0, 0, 0);
-		glClear(GL_COLOR_BUFFER_BIT);
-	}
 }
 
 ImTextureID ImGui_ImplOpenGL3_CreateVmuTexture(const unsigned int *data)
 {
-	GLuint tex_id;
+	GLuint tex_id = glcache.GenTexture();
-    glGenTextures(1, &tex_id);
+    glcache.BindTexture(GL_TEXTURE_2D, tex_id);
-    glBindTexture(GL_TEXTURE_2D, tex_id);
+    glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+    glcache.TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, 48, 32, 0, GL_RGBA, GL_UNSIGNED_BYTE, data);
 
     return reinterpret_cast<ImTextureID>(tex_id);
@@ -600,5 +532,5 @@ ImTextureID ImGui_ImplOpenGL3_CreateVmuTexture(const unsigned int *data)
 
 void ImGui_ImplOpenGL3_DeleteVmuTexture(ImTextureID tex_id)
 {
-	glDeleteTextures(1, &(GLuint &)tex_id);
+	glcache.DeleteTextures(1, &(GLuint &)tex_id);
 }

core/rend/gui.cpp

@@ -39,7 +39,9 @@
 #include "linux-dist/main.h"	// FIXME for kcode[]
 #include "gui_util.h"
 #include "gui_android.h"
-#include "version.h"
+#include "version/version.h"
+#include "oslib/audiostream.h"
+
 
 extern void dc_loadstate();
 extern void dc_savestate();
@@ -299,7 +301,9 @@ static void gui_display_commands()
     if (!settings_opening)
     	ImGui_ImplOpenGL3_DrawBackground();
 
-	display_vmus();
+    if (!settings.rend.FloatVMUs)
+    	// If floating VMUs, they are already visible on the background
+    	display_vmus();
 
     ImGui::SetNextWindowPos(ImVec2(screen_width / 2.f, screen_height / 2.f), ImGuiCond_Always, ImVec2(0.5f, 0.5f));
     ImGui::SetNextWindowSize(ImVec2(330 * scaling, 0));
@@ -622,6 +626,8 @@ void directory_selected_callback(bool cancelled, std::string selection)
 
 static void gui_display_settings()
 {
+	static bool maple_devices_changed;
+
 	ImGui_Impl_NewFrame();
     ImGui::NewFrame();
 
@@ -644,10 +650,14 @@ static void gui_display_settings()
     		gui_state = Commands;
     	else
     		gui_state = Main;
+    	if (maple_devices_changed)
+    	{
+    		maple_devices_changed = false;
 #if DC_PLATFORM == DC_PLATFORM_DREAMCAST
-    	maple_ReconnectDevices();
+    		maple_ReconnectDevices();
-    	reset_vmus();
+    		reset_vmus();
 #endif
+    	}
        	SaveSettings();
     }
 	if (game_started)
@@ -801,13 +811,14 @@ static void gui_display_settings()
 		if (ImGui::BeginTabItem("Controls"))
 		{
 			ImGui::PushStyleVar(ImGuiStyleVar_FramePadding, normal_padding);
-#if DC_PLATFORM == DC_PLATFORM_DREAMCAST
+#if DC_PLATFORM == DC_PLATFORM_DREAMCAST || DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
 		    if (ImGui::CollapsingHeader("Dreamcast Devices", ImGuiTreeNodeFlags_DefaultOpen))
 		    {
 				for (int bus = 0; bus < MAPLE_PORTS; bus++)
 				{
 					ImGui::Text("Device %c", bus + 'A');
 					ImGui::SameLine();
+#if DC_PLATFORM == DC_PLATFORM_DREAMCAST
 					char device_name[32];
 					sprintf(device_name, "##device%d", bus);
 					float w = ImGui::CalcItemWidth() / 3;
@@ -818,7 +829,10 @@ static void gui_display_settings()
 						{
 							bool is_selected = settings.input.maple_devices[bus] == maple_device_type_from_index(i);
 							if (ImGui::Selectable(maple_device_types[i], &is_selected))
+							{
 								settings.input.maple_devices[bus] = maple_device_type_from_index(i);
+								maple_devices_changed = true;
+							}
 							if (is_selected)
 								ImGui::SetItemDefaultFocus();
 						}
@@ -836,7 +850,10 @@ static void gui_display_settings()
 							{
 								bool is_selected = settings.input.maple_expansion_devices[bus][port] == maple_expansion_device_type_from_index(i);
 								if (ImGui::Selectable(maple_expansion_device_types[i], &is_selected))
+								{
 									settings.input.maple_expansion_devices[bus][port] = maple_expansion_device_type_from_index(i);
+									maple_devices_changed = true;
+								}
 								if (is_selected)
 									ImGui::SetItemDefaultFocus();
 							}
@@ -845,6 +862,10 @@ static void gui_display_settings()
 						ImGui::PopID();
 					}
 					ImGui::PopItemWidth();
+#elif DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
+					if (MapleDevices[bus][5] != NULL)
+						ImGui::Text("%s", maple_device_name(MapleDevices[bus][5]->get_device_type()));
+#endif
 				}
 				ImGui::Spacing();
 		    }
@@ -945,15 +966,27 @@ static void gui_display_settings()
 		    	ImGui::Checkbox("Shadows", &settings.rend.ModifierVolumes);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Enable modifier volumes, usually used for shadows");
+		    	ImGui::Checkbox("Fog", &settings.rend.Fog);
+	            ImGui::SameLine();
+	            ShowHelpMarker("Enable fog effects");
 		    	ImGui::Checkbox("Widescreen", &settings.rend.WideScreen);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Draw geometry outside of the normal 4:3 aspect ratio. May produce graphical glitches in the revealed areas");
 		    	ImGui::Checkbox("Show FPS Counter", &settings.rend.ShowFPS);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Show on-screen frame/sec counter");
+		    	ImGui::Checkbox("Show VMU in game", &settings.rend.FloatVMUs);
+	            ImGui::SameLine();
+	            ShowHelpMarker("Show the VMU LCD screens while in game");
+		    	ImGui::Checkbox("Rotate screen 90°", &settings.rend.Rotate90);
+	            ImGui::SameLine();
+	            ShowHelpMarker("Rotate the screen 90° counterclockwise");
 		    	ImGui::SliderInt("Scaling", (int *)&settings.rend.ScreenScaling, 1, 100);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Downscaling factor relative to native screen resolution. Higher is better");
+		    	ImGui::SliderInt("Horizontal Stretching", (int *)&settings.rend.ScreenStretching, 100, 150);
+	            ImGui::SameLine();
+	            ShowHelpMarker("Stretch the screen horizontally");
 		    	ImGui::SliderInt("Frame Skipping", (int *)&settings.pvr.ta_skip, 0, 6);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Number of frames to skip between two actually rendered frames");
@@ -991,6 +1024,48 @@ static void gui_display_settings()
 			ImGui::Checkbox("Disable Sound", &settings.aica.NoSound);
             ImGui::SameLine();
             ShowHelpMarker("Disable the emulator sound output");
+
+			audiobackend_t* backend = NULL;;
+			std::string backend_name = settings.audio.backend;
+			if (backend_name != "auto" && backend_name != "none")
+			{
+				backend = GetAudioBackend(settings.audio.backend);
+				if (backend != NULL)
+					backend_name = backend->slug;
+			}
+
+			SortAudioBackends();
+			if (ImGui::BeginCombo("Audio Backend", backend_name.c_str(), ImGuiComboFlags_None))
+			{
+				bool is_selected = (settings.audio.backend == "auto");
+				if (ImGui::Selectable("auto", &is_selected))
+					settings.audio.backend = "auto";
+				ImGui::SameLine(); ImGui::Text("-");
+				ImGui::SameLine(); ImGui::Text("Autoselect audio backend");
+
+				is_selected = (settings.audio.backend == "none");
+				if (ImGui::Selectable("none", &is_selected))
+					settings.audio.backend = "none";
+				ImGui::SameLine(); ImGui::Text("-");
+				ImGui::SameLine(); ImGui::Text("No audio backend");
+
+				for (int i = 0; i < GetAudioBackendCount(); i++)
+				{
+					audiobackend_t* backend = GetAudioBackend(i);
+					is_selected = (settings.audio.backend == backend->slug);
+
+					if (ImGui::Selectable(backend->slug.c_str(), &is_selected))
+						settings.audio.backend = backend->slug;
+					ImGui::SameLine(); ImGui::Text("-");
+					ImGui::SameLine(); ImGui::Text(backend->name.c_str());
+	                if (is_selected)
+	                    ImGui::SetItemDefaultFocus();
+				}
+				ImGui::EndCombo();
+			}
+            ImGui::SameLine();
+            ShowHelpMarker("The audio backend to use");
+
 			ImGui::Checkbox("Enable DSP", &settings.aica.NoBatch);
             ImGui::SameLine();
             ShowHelpMarker("Enable the Dreamcast Digital Sound Processor. Only recommended on fast and arm64 platforms");
@@ -1022,10 +1097,33 @@ static void gui_display_settings()
 	            ShowHelpMarker("Do not optimize integer division. Recommended");
 		    	ImGui::Checkbox("Unstable Optimizations", &settings.dynarec.unstable_opt);
 	            ImGui::SameLine();
-	            ShowHelpMarker("Enable unsafe optimizations. May cause crash or environmental disaster");
+	            ShowHelpMarker("Enable unsafe optimizations. Will cause crash or environmental disaster");
 		    	ImGui::Checkbox("Idle Skip", &settings.dynarec.idleskip);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Skip wait loops. Recommended");
+				ImGui::PushItemWidth(ImGui::CalcTextSize("Largeenough").x);
+	            const char *preview = settings.dynarec.SmcCheckLevel == NoCheck ? "Faster" : settings.dynarec.SmcCheckLevel == FastCheck ? "Fast" : "Full";
+				if (ImGui::BeginCombo("SMC Checks", preview	, ImGuiComboFlags_None))
+				{
+					bool is_selected = settings.dynarec.SmcCheckLevel == NoCheck;
+					if (ImGui::Selectable("Faster", &is_selected))
+						settings.dynarec.SmcCheckLevel = NoCheck;
+					if (is_selected)
+						ImGui::SetItemDefaultFocus();
+					is_selected = settings.dynarec.SmcCheckLevel == FastCheck;
+					if (ImGui::Selectable("Fast", &is_selected))
+						settings.dynarec.SmcCheckLevel = FastCheck;
+					if (is_selected)
+						ImGui::SetItemDefaultFocus();
+					is_selected = settings.dynarec.SmcCheckLevel == FullCheck;
+					if (ImGui::Selectable("Full", &is_selected))
+						settings.dynarec.SmcCheckLevel = FullCheck;
+					if (is_selected)
+						ImGui::SetItemDefaultFocus();
+					ImGui::EndCombo();
+				}
+	            ImGui::SameLine();
+	            ShowHelpMarker("How to detect self-modifying code. Full check recommended");
 		    }
 		    if (ImGui::CollapsingHeader("Other", ImGuiTreeNodeFlags_DefaultOpen))
 		    {
@@ -1230,6 +1328,7 @@ static void gui_start_game(const std::string& path)
 	{
 		gui_state = Main;
 		game_started = false;
+		cfgSetVirtual("config", "image", "");
 		switch (rc) {
 		case -3:
 			error_msg = "Audio/video initialization failed";
@@ -1387,24 +1486,6 @@ void gui_display_ui()
 		gui_state = Closed;
 }
 
-void gui_display_fps(const char *string)
-{
-	ImGui_Impl_NewFrame();
-    ImGui::NewFrame();
-
-    ImGui::SetNextWindowBgAlpha(0);
-    ImGui::SetNextWindowPos(ImVec2(0, screen_height), ImGuiCond_Always, ImVec2(0.f, 1.f));	// Lower left corner
-
-    ImGui::Begin("##fps", NULL, ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoNav
-    		| ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoBackground);
-    ImGui::SetWindowFontScale(2);
-    ImGui::TextColored(ImVec4(1, 1, 0, 0.7), "%s", string);
-    ImGui::End();
-
-    ImGui::Render();
-    ImGui_ImplOpenGL3_RenderDrawData(ImGui::GetDrawData());
-}
-
 static float LastFPSTime;
 static int lastFrameCount = 0;
 static float fps = -1;
@@ -1452,26 +1533,32 @@ void gui_display_osd()
 	if (osd_message.empty())
 	{
 		message = getFPSNotification();
-		if (message.empty())
-			return;
 	}
 	else
 		message = osd_message;
 
-	ImGui_Impl_NewFrame();
+	if (!message.empty() || settings.rend.FloatVMUs)
-    ImGui::NewFrame();
+	{
+		ImGui_Impl_NewFrame();
+		ImGui::NewFrame();
 
-    ImGui::SetNextWindowBgAlpha(0);
+		if (!message.empty())
-    ImGui::SetNextWindowPos(ImVec2(0, screen_height), ImGuiCond_Always, ImVec2(0.f, 1.f));	// Lower left corner
+		{
+			ImGui::SetNextWindowBgAlpha(0);
+			ImGui::SetNextWindowPos(ImVec2(0, screen_height), ImGuiCond_Always, ImVec2(0.f, 1.f));	// Lower left corner
 
-    ImGui::Begin("##osd", NULL, ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoNav
+			ImGui::Begin("##osd", NULL, ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoNav
-    		| ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoBackground);
+					| ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoBackground);
-    ImGui::SetWindowFontScale(1.5);
+			ImGui::SetWindowFontScale(1.5);
-    ImGui::TextColored(ImVec4(1, 1, 0, 0.7), "%s", message.c_str());
+			ImGui::TextColored(ImVec4(1, 1, 0, 0.7), "%s", message.c_str());
-    ImGui::End();
+			ImGui::End();
+		}
+		if (settings.rend.FloatVMUs)
+			display_vmus();
 
-    ImGui::Render();
+		ImGui::Render();
-    ImGui_ImplOpenGL3_RenderDrawData(ImGui::GetDrawData());
+		ImGui_ImplOpenGL3_RenderDrawData(ImGui::GetDrawData());
+	}
 }
 
 void gui_open_onboarding()

core/types.h

@@ -39,7 +39,7 @@
 #undef _CRT_SECURE_NO_DEPRECATE
 #endif
 
-#define _CRT_SECURE_NO_DEPRECATE 
+#define _CRT_SECURE_NO_DEPRECATE
 //unnamed struncts/unions
 #pragma warning( disable : 4201)
 
@@ -123,7 +123,7 @@ enum HollyInterruptType
 };
 
 enum HollyInterruptID
-{		
+{
 		// asic9a /sh4 external holly normal [internal]
 		holly_RENDER_DONE_vd = holly_nrm | 0,	//bit 0 = End of Render interrupt : Video
 		holly_RENDER_DONE_isp = holly_nrm | 1,	//bit 1 = End of Render interrupt : ISP
@@ -132,11 +132,11 @@ enum HollyInterruptID
 		holly_SCANINT1 = holly_nrm | 3,			//bit 3 = V Blank-in interrupt
 		holly_SCANINT2 = holly_nrm | 4,			//bit 4 = V Blank-out interrupt
 		holly_HBLank = holly_nrm | 5,			//bit 5 = H Blank-in interrupt
-												
+
 		holly_YUV_DMA = holly_nrm | 6,			//bit 6 = End of Transferring interrupt : YUV
 		holly_OPAQUE = holly_nrm | 7,			//bit 7 = End of Transferring interrupt : Opaque List
 		holly_OPAQUEMOD = holly_nrm | 8,		//bit 8 = End of Transferring interrupt : Opaque Modifier Volume List
-		
+
 		holly_TRANS = holly_nrm | 9,			//bit 9 = End of Transferring interrupt : Translucent List
 		holly_TRANSMOD = holly_nrm | 10,		//bit 10 = End of Transferring interrupt : Translucent Modifier Volume List
 		holly_PVR_DMA = holly_nrm | 11,			//bit 11 = End of DMA interrupt : PVR-DMA
@@ -145,12 +145,12 @@ enum HollyInterruptID
 		holly_MAPLE_VBOI = holly_nrm | 13,		//bit 13 = Maple V blank over interrupt
 		holly_GDROM_DMA = holly_nrm | 14,		//bit 14 = End of DMA interrupt : GD-DMA
 		holly_SPU_DMA = holly_nrm | 15,			//bit 15 = End of DMA interrupt : AICA-DMA
-		
+
 		holly_EXT_DMA1 = holly_nrm | 16,		//bit 16 = End of DMA interrupt : Ext-DMA1(External 1)
 		holly_EXT_DMA2 = holly_nrm | 17,		//bit 17 = End of DMA interrupt : Ext-DMA2(External 2)
 		holly_DEV_DMA = holly_nrm | 18,			//bit 18 = End of DMA interrupt : Dev-DMA(Development tool DMA)
-		
+
-		holly_CH2_DMA = holly_nrm | 19,			//bit 19 = End of DMA interrupt : ch2-DMA 
+		holly_CH2_DMA = holly_nrm | 19,			//bit 19 = End of DMA interrupt : ch2-DMA
 		holly_PVR_SortDMA = holly_nrm | 20,		//bit 20 = End of DMA interrupt : Sort-DMA (Transferring for alpha sorting)
 		holly_PUNCHTHRU = holly_nrm | 21,		//bit 21 = End of Transferring interrupt : Punch Through List
 
@@ -188,8 +188,8 @@ enum HollyInterruptID
 		//bit 23 = G2 : AICA-DMA Time out
 		//bit 24 = G2 : Ext-DMA1 Time out
 		//bit 25 = G2 : Ext-DMA2 Time out
-		//bit 26 = G2 : Dev-DMA Time out 
+		//bit 26 = G2 : Dev-DMA Time out
-		//bit 27 = G2 : Time out in CPU accessing 	
+		//bit 27 = G2 : Time out in CPU accessing
 };
 
 
@@ -200,7 +200,7 @@ struct vram_block
 	u32 end;
 	u32 len;
 	u32 type;
- 
+
 	void* userdata;
 };
 
@@ -229,7 +229,7 @@ struct NDC_WINDOW_RECT
 //******************************************************
 //*********************** PowerVR **********************
 //******************************************************
- 
+
 void libCore_vramlock_Unlock_block  (vram_block* block);
 void libCore_vramlock_Unlock_block_wb  (vram_block* block);
 vram_block* libCore_vramlock_Lock(u32 start_offset,u32 end_offset,void* userdata);
@@ -246,7 +246,7 @@ enum DiscType
 	CdRom_XA=0x20,
 	CdRom_Extra=0x30,
 	CdRom_CDI=0x40,
-	GdRom=0x80,		
+	GdRom=0x80,
 
 	NoDisk=0x1,			//These are a bit hacky .. but work for now ...
 	Open=0x2,			//tray is open :)
@@ -576,7 +576,7 @@ enum RegIO
 	RIO_RO = REG_RO | REG_WF,
 	RIO_RO_FUNC = REG_RO | REG_RF | REG_WF,
 	RIO_CONST = REG_RO | REG_WF,
-	RIO_WO_FUNC = REG_WF | REG_RF | REG_WO, 
+	RIO_WO_FUNC = REG_WF | REG_RF | REG_WO,
 	RIO_NO_ACCESS = REG_WF | REG_RF | REG_NO_ACCESS
 };
 
@@ -601,6 +601,11 @@ struct RegisterStruct
 	u32 flags;					//Access flags !
 };
 
+enum SmcCheckEnum {
+	FullCheck = 0,
+	FastCheck = 1,
+	NoCheck = 2
+};
 
 struct settings_t
 {
@@ -627,7 +632,11 @@ struct settings_t
 		f32 ExtraDepthScale;
 		bool CustomTextures;
 		bool DumpTextures;
-		int ScreenScaling;	// in percent. 50 means half the native resolution
+		int ScreenScaling;		// in percent. 50 means half the native resolution
+		int ScreenStretching;	// in percent. 150 means stretch from 4/3 to 6/3
+		bool Fog;
+		bool FloatVMUs;
+		bool Rotate90;			// Rotate the screen 90 deg CC
 	} rend;
 
 	struct
@@ -637,8 +646,9 @@ struct settings_t
 		bool unstable_opt;
 		bool safemode;
 		bool disable_nvmem;
+		SmcCheckEnum SmcCheckLevel;
 	} dynarec;
-	
+
 	struct
 	{
 		u32 run_counts;
@@ -647,7 +657,6 @@ struct settings_t
 	struct
 	{
 		u32 cable;			// 0 -> VGA, 1 -> VGA, 2 -> RGB, 3 -> TV
-		u32 RTC;
 		u32 region;			// 0 -> JP, 1 -> USA, 2 -> EU, 3 -> default
 		u32 broadcast;		// 0 -> NTSC, 1 -> PAL, 2 -> PAL/M, 3 -> PAL/N, 4 -> default
 		u32 language;		// 0 -> JP, 1 -> EN, 2 -> DE, 3 -> FR, 4 -> SP, 5 -> IT, 6 -> default
@@ -670,6 +679,9 @@ struct settings_t
 		bool NoSound;
 	} aica;
 
+	struct{
+		std::string backend;
+	} audio;
 #if USE_OMX
 	struct
 	{
@@ -699,7 +711,7 @@ struct settings_t
 	{
 		u32 ta_skip;
 		u32 rend;
-		
+
 		u32 MaxThreads;
 		bool SynchronousRender;
 	} pvr;
@@ -742,7 +754,7 @@ static inline void do_nada(...) { }
 #ifdef _ANDROID
 #include <android/log.h>
 
-#ifdef printf 
+#ifdef printf
 #undef printf
 #endif
 
@@ -849,7 +861,7 @@ void libARM_Update(u32 cycles);
 			else if (sz==2)							\
 				return *(u16*)&arr[addr];			\
 			else if (sz==4)							\
-				return *(u32*)&arr[addr];}	
+				return *(u32*)&arr[addr];}
 
 #define WriteMemArr(arr,addr,data,sz)				\
 			{if(sz==1)								\
@@ -857,7 +869,7 @@ void libARM_Update(u32 cycles);
 			else if (sz==2)							\
 				{*(u16*)&arr[addr]=(u16)data;}		\
 			else if (sz==4)							\
-			{*(u32*)&arr[addr]=data;}}	
+			{*(u32*)&arr[addr]=data;}}
 
 #define WriteMemArrRet(arr,addr,data,sz)				\
 			{if(sz==1)								\
@@ -865,7 +877,7 @@ void libARM_Update(u32 cycles);
 			else if (sz==2)							\
 				{*(u16*)&arr[addr]=(u16)data;return;}		\
 			else if (sz==4)							\
-			{*(u32*)&arr[addr]=data;return;}}	
+			{*(u32*)&arr[addr]=data;return;}}
 
 struct OnLoad
 {

core/windows/winmain.cpp

@@ -188,8 +188,6 @@ u16 kcode[4] = { 0xffff, 0xffff, 0xffff, 0xffff };
 u32 vks[4];
 s8 joyx[4],joyy[4];
 u8 rt[4],lt[4];
-extern bool coin_chute;
-extern bool naomi_test_button;
 // Mouse
 extern s32 mo_x_abs;
 extern s32 mo_y_abs;
@@ -220,12 +218,6 @@ void UpdateInputState(u32 port)
 	std::shared_ptr<XInputGamepadDevice> gamepad = XInputGamepadDevice::GetXInputDevice(port);
 	if (gamepad != NULL)
 		gamepad->ReadInput();
-
-#if DC_PLATFORM == DC_PLATFORM_NAOMI || DC_PLATFORM == DC_PLATFORM_ATOMISWAVE
-		// FIXME
-		coin_chute = GetAsyncKeyState(VK_F8);
-		naomi_test_button = GetAsyncKeyState(VK_F7);
-#endif
 }
 
 // Windows class name to register
@@ -772,6 +764,7 @@ cThread::cThread(ThreadEntryFP* function,void* prm)
 
 void cThread::Start()
 {
+	verify(hThread == NULL);
 	hThread=CreateThread(NULL,0,(LPTHREAD_START_ROUTINE)Entry,param,0,NULL);
 	ResumeThread(hThread);
 }
@@ -779,6 +772,8 @@ void cThread::Start()
 void cThread::WaitToEnd()
 {
 	WaitForSingleObject(hThread,INFINITE);
+	CloseHandle(hThread);
+	hThread = NULL;
 }
 //End thread class
 

shell/android-studio/reicast/src/dreamcast/AndroidManifest.xml

@@ -52,5 +52,56 @@
 		            android:scheme="file" />
 		    </intent-filter>
 		</activity>
+		<activity-alias
+			android:name="com.reicast.emulator.MainActivity"
+			android:targetActivity="com.reicast.emulator.NativeGLActivity">
+			<intent-filter>
+				<action android:name="android.intent.action.VIEW" />
+
+				<category android:name="android.intent.category.DEFAULT" />
+				<category android:name="android.intent.category.BROWSABLE" />
+
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.GDI"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.gdi"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.CHD"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.chd"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.CDI"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.cdi"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.CUE"
+					android:scheme="file" />
+				<data
+					android:host="*"
+					android:mimeType="*/*"
+					android:pathPattern=".*\\.cue"
+					android:scheme="file" />
+			</intent-filter>
+		</activity-alias>
 	</application>
 </manifest>

shell/android-studio/reicast/src/main/AndroidManifest.xml

@@ -61,6 +61,16 @@
                 <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
             </intent-filter>
         </activity>
+        <activity-alias
+            android:name="com.reicast.emulator.MainActivity"
+            android:targetActivity="com.reicast.emulator.NativeGLActivity">
+            <intent-filter>
+                <action android:name="android.intent.action.MAIN" />
+                <category android:name="android.intent.category.LAUNCHER" />
+                <category android:name="tv.ouya.intent.category.GAME" />
+                <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
+            </intent-filter>
+        </activity-alias>
         <provider
             android:name="android.support.v4.content.FileProvider"
             android:authorities="${applicationId}.provider"

shell/android-studio/reicast/src/main/java/com/reicast/emulator/BaseGLActivity.java

@@ -228,22 +228,24 @@ public abstract class BaseGLActivity extends Activity implements ActivityCompat.
 
     @Override
     public boolean onKeyDown(int keyCode, KeyEvent event) {
-        if (keyCode == KeyEvent.KEYCODE_BACK) {
+        if (event.getRepeatCount() == 0) {
-            if (!JNIdc.guiIsOpen()) {
+            if (keyCode == KeyEvent.KEYCODE_BACK) {
-                showMenu();
+                if (!JNIdc.guiIsOpen()) {
-                return true;
+                    showMenu();
+                    return true;
+                }
+                else if (JNIdc.guiIsContentBrowser()) {
+                    finish();
+                    return true;
+                }
             }
-            else if (JNIdc.guiIsContentBrowser()) {
+            if (InputDeviceManager.getInstance().joystickButtonEvent(event.getDeviceId(), keyCode, true))
-                finish();
                 return true;
-            }
-        }
-        if (InputDeviceManager.getInstance().joystickButtonEvent(event.getDeviceId(), keyCode, true))
-            return true;
 
-        if (ViewConfiguration.get(this).hasPermanentMenuKey()) {
+            if (ViewConfiguration.get(this).hasPermanentMenuKey()) {
-            if (keyCode == KeyEvent.KEYCODE_MENU) {
+                if (keyCode == KeyEvent.KEYCODE_MENU) {
-                return showMenu();
+                    return showMenu();
+                }
             }
         }
         return super.onKeyDown(keyCode, event);
@@ -296,7 +298,7 @@ public abstract class BaseGLActivity extends Activity implements ActivityCompat.
     }
 
     // Called from native code
-    private void generateErrorLog() {
+    protected void generateErrorLog() {
         try {
             new GenerateLogs(this).execute(getFilesDir().getAbsolutePath());
 		} catch (RuntimeException e) {

shell/android-studio/reicast/src/main/java/com/reicast/emulator/emu/JNIdc.java

@@ -21,8 +21,7 @@ public final class JNIdc
 	public static native int send(int cmd, int opt);
 	public static native int data(int cmd, byte[] data);
 
-	public static native void rendinitNative(Surface surface, int w, int h);
+	public static native void rendinitNative(Surface surface);
-	public static native boolean rendframeNative();
 	public static native void rendinitJava(int w, int h);
 	public static native boolean rendframeJava();
 	public static native void rendtermJava();

shell/android-studio/reicast/src/main/java/com/reicast/emulator/emu/NativeGLView.java

@@ -19,8 +19,6 @@ import com.reicast.emulator.NativeGLActivity;
 import com.reicast.emulator.config.Config;
 
 public class NativeGLView extends SurfaceView implements SurfaceHolder.Callback {
-    private Handler handler = new Handler();
-
     private boolean surfaceReady = false;
     private boolean paused = false;
     VirtualJoystickDelegate vjoyDelegate;
@@ -66,23 +64,6 @@ public class NativeGLView extends SurfaceView implements SurfaceHolder.Callback
 
         if (NativeGLActivity.syms != null)
             JNIdc.data(1, NativeGLActivity.syms);
-
-        startRendering();
-    }
-
-    private void startRendering() {
-        // Continuously render frames
-        handler.removeCallbacksAndMessages(null);
-        handler.postAtTime(new Runnable() {
-            @Override
-            public void run() {
-                if (!paused)
-                {
-                    JNIdc.rendframeNative();
-                    handler.post(this);
-                }
-            }
-        }, SystemClock.uptimeMillis() + 500);
     }
 
     @Override
@@ -111,7 +92,7 @@ public class NativeGLView extends SurfaceView implements SurfaceHolder.Callback
     public void surfaceChanged(SurfaceHolder surfaceHolder, int format, int w, int h) {
         //Log.i("reicast", "NativeGLView.surfaceChanged: " + w + "x" + h);
         surfaceReady = true;
-        JNIdc.rendinitNative(surfaceHolder.getSurface(), w, h);
+        JNIdc.rendinitNative(surfaceHolder.getSurface());
         Emulator.getCurrentActivity().handleStateChange(false);
     }
 
@@ -119,7 +100,7 @@ public class NativeGLView extends SurfaceView implements SurfaceHolder.Callback
     public void surfaceDestroyed(SurfaceHolder surfaceHolder) {
         //Log.i("reicast", "NativeGLView.surfaceDestroyed");
         surfaceReady = false;
-        JNIdc.rendinitNative(null, 0, 0);
+        JNIdc.rendinitNative(null);
         Emulator.getCurrentActivity().handleStateChange(true);
     }
 
@@ -142,7 +123,6 @@ public class NativeGLView extends SurfaceView implements SurfaceHolder.Callback
             requestFocus();
             JNIdc.resume();
         }
-        startRendering();
     }
 
     @TargetApi(19)

shell/android-studio/reicast/src/main/java/com/reicast/emulator/emu/VirtualJoystickDelegate.java

@@ -15,7 +15,7 @@ import com.reicast.emulator.periph.InputDeviceManager;
 import com.reicast.emulator.periph.VJoy;
 
 public class VirtualJoystickDelegate {
-    private Vibrator vib;
+    private VibratorThread vibratorThread;
 
     private boolean editVjoyMode = false;
     private int selectedVjoyElement = -1;
@@ -39,7 +39,10 @@ public class VirtualJoystickDelegate {
     public VirtualJoystickDelegate(View view) {
         this.view = view;
         this.context = view.getContext();
-        vib = (Vibrator)context.getSystemService(Context.VIBRATOR_SERVICE);
+
+        vibratorThread = new VibratorThread(context);
+        vibratorThread.start();
+
         readCustomVjoyValues();
         scaleGestureDetector = new ScaleGestureDetector(context, new OscOnScaleGestureListener());
     }
@@ -224,8 +227,9 @@ public class VirtualJoystickDelegate {
                             if (y > vjoy[j][1] && y <= (vjoy[j][1] + vjoy[j][3])) {
                                 if (vjoy[j][4] >= -2) {
                                     if (vjoy[j][5] == 0)
-                                        if (!editVjoyMode && Emulator.vibrationDuration > 0)
+                                        if (!editVjoyMode) {
-                                            vib.vibrate(Emulator.vibrationDuration);
+                                            vibratorThread.vibrate();
+                                        }
                                     vjoy[j][5] = 2;
                                 }
 
@@ -397,4 +401,51 @@ public class VirtualJoystickDelegate {
             selectedVjoyElement = -1;
         }
     }
+
+    private class VibratorThread extends Thread
+    {
+        private Vibrator vibrator;
+        private boolean vibrate = false;
+        private boolean stopping = false;
+
+        VibratorThread(Context context) {
+            vibrator = (Vibrator)context.getSystemService(Context.VIBRATOR_SERVICE);
+        }
+
+        @Override
+        public void run() {
+            while (!stopping) {
+                boolean doVibrate;
+                synchronized (this) {
+                    doVibrate = false;
+                    try {
+                        this.wait();
+                    } catch (InterruptedException e) {
+                    }
+                    if (vibrate) {
+                        doVibrate = true;
+                        vibrate = false;
+                    }
+                }
+                if (doVibrate)
+                    vibrator.vibrate(Emulator.vibrationDuration);
+            }
+        }
+
+        public void stopVibrator() {
+            synchronized (this) {
+                stopping = true;
+                notify();
+            }
+        }
+
+        public void vibrate() {
+            if (Emulator.vibrationDuration > 0) {
+                synchronized (this) {
+                    vibrate = true;
+                    notify();
+                }
+            }
+        }
+    }
 }

shell/android-studio/reicast/src/main/jni/src/Android.cpp

@@ -20,7 +20,7 @@
 #include "hw/maple/maple_devs.h"
 #include "hw/maple/maple_if.h"
 #include "hw/naomi/naomi_cart.h"
-#include "oslib/audiobackend_android.h"
+#include "oslib/audiostream.h"
 #include "imgread/common.h"
 #include "rend/gui.h"
 #include "cfg/cfg.h"
@@ -91,8 +91,7 @@ JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_destroy(JNIEnv *env,j
 JNIEXPORT jint JNICALL Java_com_reicast_emulator_emu_JNIdc_send(JNIEnv *env,jobject obj,jint id, jint v)  __attribute__((visibility("default")));
 JNIEXPORT jint JNICALL Java_com_reicast_emulator_emu_JNIdc_data(JNIEnv *env,jobject obj,jint id, jbyteArray d)  __attribute__((visibility("default")));
 
-JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendinitNative(JNIEnv *env, jobject obj, jobject surface, jint w, jint h)  __attribute__((visibility("default")));
+JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendinitNative(JNIEnv *env, jobject obj, jobject surface)  __attribute__((visibility("default")));
-JNIEXPORT jboolean JNICALL Java_com_reicast_emulator_emu_JNIdc_rendframeNative(JNIEnv *env,jobject obj)  __attribute__((visibility("default")));
 JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendinitJava(JNIEnv *env, jobject obj, jint w, jint h)  __attribute__((visibility("default")));
 JNIEXPORT jboolean JNICALL Java_com_reicast_emulator_emu_JNIdc_rendframeJava(JNIEnv *env, jobject obj)  __attribute__((visibility("default")));
 JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendtermJava(JNIEnv *env, jobject obj)  __attribute__((visibility("default")));
@@ -404,35 +403,57 @@ JNIEXPORT jint JNICALL Java_com_reicast_emulator_emu_JNIdc_data(JNIEnv *env, job
 
 extern void gl_swap();
 extern void egl_stealcntx();
+volatile static bool render_running;
+volatile static bool render_reinit;
 
-JNIEXPORT jboolean JNICALL Java_com_reicast_emulator_emu_JNIdc_rendframeNative(JNIEnv *env,jobject obj)
+void *render_thread_func(void *)
 {
-    if (g_window == NULL)
+	render_running = true;
-        return false;
+
-    if (!egl_makecurrent())
+	rend_init_renderer();
-        return false;
+
-    jboolean ret = (jboolean)rend_single_frame();
+    while (render_running) {
-    if (ret)
+        if (render_reinit)
-        gl_swap();
+        {
-    return ret;
+        	render_reinit = false;
+        	rend_init_renderer();
+        }
+        else
+            if (!egl_makecurrent())
+                break;;
+
+        bool ret = rend_single_frame();
+        if (ret)
+            gl_swap();
+    }
+    egl_makecurrent();
+    rend_term_renderer();
+    ANativeWindow_release(g_window);
+    g_window = NULL;
+	render_running = false;
+
+    return NULL;
 }
 
-JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendinitNative(JNIEnv * env, jobject obj, jobject surface, jint width, jint height)
+static cThread render_thread(render_thread_func, NULL);
+
+JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendinitNative(JNIEnv * env, jobject obj, jobject surface)
 {
-    if (g_window != NULL)
+	if (render_thread.hThread != NULL)
-    {
+	{
-        egl_makecurrent();
+		if (surface == NULL)
-        rend_term_renderer();
+		{
-        ANativeWindow_release(g_window);
+			render_running = false;
-        g_window = NULL;
+	        render_thread.WaitToEnd();
-    }
+		}
-    if (surface != NULL)
+		else
-    {
+			render_reinit = true;
+	}
+	else if (surface != NULL)
+	{
         g_window = ANativeWindow_fromSurface(env, surface);
-        rend_init_renderer();
+        render_thread.Start();
-        screen_width = width;
+	}
-        screen_height = height;
-    }
 }
 
 JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_JNIdc_rendinitJava(JNIEnv * env, jobject obj, jint width, jint height)
@@ -529,6 +550,9 @@ audiobackend_t audiobackend_android = {
         &androidaudio_term
 };
 
+static bool android = RegisterAudioBackend(&audiobackend_android);
+
+
 JNIEXPORT void JNICALL Java_com_reicast_emulator_emu_AudioBackend_setInstance(JNIEnv *env, jobject obj, jobject instance)
 {
     if (g_audioBackend != NULL)
@@ -568,7 +592,7 @@ void os_DebugBreak()
 
 	raise(SIGABRT);
     //pthread_exit(NULL);
-	
+
     // Attach debugger here to figure out what went wrong
     for(;;) ;
 }

shell/android-studio/reicast/src/naomi/AndroidManifest.xml

@@ -3,7 +3,7 @@
         android:name="com.reicast.emulator.Emulator">
 
         <activity
-            android:name="com.reicast.emulator.MainActivity">
+            android:name="com.reicast.emulator.NativeGLActivity">
             <intent-filter>
                 <action android:name="android.intent.action.VIEW" />
 
@@ -62,5 +62,66 @@
                     android:scheme="file" />
             </intent-filter>
         </activity>
+        <activity-alias
+            android:name="com.reicast.emulator.MainActivity"
+            android:targetActivity="com.reicast.emulator.NativeGLActivity">
+            <intent-filter>
+                <action android:name="android.intent.action.VIEW" />
+
+                <category android:name="android.intent.category.DEFAULT" />
+                <category android:name="android.intent.category.BROWSABLE" />
+
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.LST"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.lst"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.BIN"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.bin"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.DAT"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.dat"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.ZIP"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.zip"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.7Z"
+                    android:scheme="file" />
+                <data
+                    android:host="*"
+                    android:mimeType="*/*"
+                    android:pathPattern=".*\\.7z"
+                    android:scheme="file" />
+            </intent-filter>
+        </activity-alias>
     </application>
 </manifest>