diff --git a/src/include/86box/machine.h b/src/include/86box/machine.h
index 1af0ba709..b3c875925 100644
--- a/src/include/86box/machine.h
+++ b/src/include/86box/machine.h
@@ -690,6 +690,7 @@ extern int	machine_xt_pc4i_init(const machine_t *);
 extern int	machine_xt_mpc1600_init(const machine_t *);
 extern int	machine_xt_pcspirit_init(const machine_t *);
 extern int	machine_xt_pc700_init(const machine_t *);
+extern int	machine_xt_multitechpc500_init(const machine_t *);
 
 extern int	machine_xt_iskra3104_init(const machine_t *);
 
diff --git a/src/machine/m_xt.c b/src/machine/m_xt.c
index e356d1cbb..ac50b2310 100644
--- a/src/machine/m_xt.c
+++ b/src/machine/m_xt.c
@@ -402,3 +402,22 @@ machine_xt_pc700_init(const machine_t *model)
 
     return ret;
 }
+
+
+int
+machine_xt_multitechpc500_init(const machine_t* model)
+{
+    int ret;
+
+    ret = bios_load_linear("roms/machines/multitech_pc500/rom404.bin",
+			   0x000f8000, 32768, 0);
+
+    if (bios_only || !ret)
+	return ret;
+
+    device_add(&keyboard_pc_device);
+
+    machine_xt_common_init(model);
+
+    return ret;
+}
diff --git a/src/sio/sio_f82c710.c b/src/sio/sio_f82c710.c
index 9d52e399b..6b866c094 100644
--- a/src/sio/sio_f82c710.c
+++ b/src/sio/sio_f82c710.c
@@ -43,306 +43,318 @@
 #include <86box/nvr.h>
 #include <86box/sio.h>
 
+
 typedef struct upc_t
 {
-	uint32_t local;
-	int configuration_state; /* state of algorithm to enter configuration mode */
-	int configuration_mode;
-	uint16_t cri_addr; /* cri = configuration index register, addr is even */
-	uint16_t cap_addr; /* cap = configuration access port, addr is odd and is cri_addr + 1 */
-	uint8_t cri; /* currently indexed register */
+    uint32_t	local;
+    int		configuration_state;	/* state of algorithm to enter configuration mode */
+    int		configuration_mode;
+    uint16_t	cri_addr;		/* cri = configuration index register, addr is even */
+    uint16_t	cap_addr;		/* cap = configuration access port, addr is odd and is cri_addr + 1 */
+    uint8_t	cri;			/* currently indexed register */
+    uint8_t	last_write;
 
-        /* these regs are not affected by reset */
-        uint8_t regs[15]; /* there are 16 indexes, but there is no need to store the last one which is: R = cri_addr / 4, W = exit config mode */
-	fdc_t *fdc;
-	nvr_t *nvr;
-	void *gameport;
-	serial_t *uart[2];
+    /* these regs are not affected by reset */
+    uint8_t	regs[15];		/* there are 16 indexes, but there is no need to store the last one which is: R = cri_addr / 4, W = exit config mode */
+    fdc_t *	fdc;
+    nvr_t *	nvr;
+    void *	gameport;
+    serial_t *	uart[2];
 } upc_t;
 
+
 static void 
-f82c710_update_ports(upc_t *upc)
+f82c710_update_ports(upc_t *dev, int set)
 {
-	uint16_t com_addr = 0;
-	uint16_t lpt_addr = 0;
-	
-        serial_remove(upc->uart[0]);
-        serial_remove(upc->uart[1]);
-        lpt1_remove();
-        lpt2_remove();
-        fdc_remove(upc->fdc);
-        ide_pri_disable();
+    uint16_t com_addr = 0;
+    uint16_t lpt_addr = 0;
 
-        if (upc->regs[0] & 4) {
-                com_addr = upc->regs[4] * 4;
-		if (com_addr == SERIAL1_ADDR) {
-			serial_setup(upc->uart[0], com_addr, 4);
-		} else if (com_addr == SERIAL2_ADDR) {
-			serial_setup(upc->uart[1], com_addr, 3);
-		}
-        }
-                
+    serial_remove(dev->uart[0]);
+    serial_remove(dev->uart[1]);
+    lpt1_remove();
+    lpt2_remove();
+    fdc_remove(dev->fdc);
+    ide_pri_disable();
 
-        if (upc->regs[0] & 8) {
-		lpt_addr = upc->regs[6] * 4;
-                lpt1_init(lpt_addr);
-		if ((lpt_addr == 0x378) || (lpt_addr == 0x3bc)) {
-			lpt1_irq(7);
-		} else if (lpt_addr == 0x278) {
-			lpt1_irq(5);
-		}
-        }
+    if (!set)
+	return;
 
-        if (upc->regs[12] & 0x80) {
-                ide_pri_enable();
-        }
+    if (dev->regs[0] & 4) {
+	com_addr = dev->regs[4] * 4;
+	if (com_addr == SERIAL1_ADDR)
+		serial_setup(dev->uart[0], com_addr, 4);
+	else if (com_addr == SERIAL2_ADDR)
+		serial_setup(dev->uart[1], com_addr, 3);
+    }
 
-        if (upc->regs[12] & 0x20) {
-                fdc_set_base(upc->fdc, 0x03f0);
-        }
+    if (dev->regs[0] & 8) {
+	lpt_addr = dev->regs[6] * 4;
+	lpt1_init(lpt_addr);
+	if ((lpt_addr == 0x378) || (lpt_addr == 0x3bc))
+		lpt1_irq(7);
+	else if (lpt_addr == 0x278)
+		lpt1_irq(5);
+    }
+
+    if (dev->regs[12] & 0x80)
+	ide_pri_enable();
+
+    if (dev->regs[12] & 0x20)
+	fdc_set_base(dev->fdc, 0x03f0);
 }
 
+
 static void
-f82c606_update_ports(upc_t *upc)
+f82c606_update_ports(upc_t *dev, int set)
 {
-	uint8_t uart1_int = 0xff;
-	uint8_t uart2_int = 0xff;
-	uint8_t lpt1_int = 0xff;
-	int nvr_int = -1;
+    uint8_t uart1_int = 0xff;
+    uint8_t uart2_int = 0xff;
+    uint8_t lpt1_int = 0xff;
+    int nvr_int = -1;
 
-	serial_remove(upc->uart[0]);
-	serial_remove(upc->uart[1]);
-	lpt1_remove();
-	lpt2_remove();
+    serial_remove(dev->uart[0]);
+    serial_remove(dev->uart[1]);
+    lpt1_remove();
+    lpt2_remove();
 
-	nvr_at_handler(0, upc->regs[3] * 4, upc->nvr);
-	nvr_at_handler(0, 0x70, upc->nvr);
+    nvr_at_handler(0, ((uint16_t) dev->regs[3]) << 2, dev->nvr);
+    nvr_at_handler(0, 0x70, dev->nvr);
 
-	switch (upc->regs[8] & 0xc0) {
-	case 0x40: nvr_int = 4; break;
-	case 0x80: uart1_int = 4; break;
-	case 0xc0: uart2_int = 4; break;
-	}
+    gameport_remap(dev->gameport, 0);
 
-	switch (upc->regs[8] & 0x30) {
-	case 0x10: nvr_int = 3; break;
-	case 0x20: uart1_int = 3; break;
-	case 0x30: uart2_int = 3; break;
-	}
+    if (!set)
+	return;
 
-	switch (upc->regs[8] & 0x0c) {
+    switch (dev->regs[8] & 0xc0) {
+	case 0x40: nvr_int = 3; break;
+	case 0x80: uart1_int = 3; break;
+	case 0xc0: uart2_int = 3; break;
+    }
+
+    switch (dev->regs[8] & 0x30) {
+	case 0x10: nvr_int = 4; break;
+	case 0x20: uart1_int = 4; break;
+	case 0x30: uart2_int = 4; break;
+    }
+
+    switch (dev->regs[8] & 0x0c) {
 	case 0x04: nvr_int = 5; break;
 	case 0x08: uart1_int = 5; break;
 	case 0x0c: lpt1_int = 5; break;
-	}
+    }
 
-	switch (upc->regs[8] & 0x03) {
+    switch (dev->regs[8] & 0x03) {
 	case 0x01: nvr_int = 7; break;
 	case 0x02: uart2_int = 7; break;
 	case 0x03: lpt1_int = 7; break;
-	}
+    }
 
-	if (upc->regs[0] & 1)
-		gameport_remap(upc->gameport, upc->regs[7] * 4);
-	else
-		gameport_remap(upc->gameport, 0);
+    if (dev->regs[0] & 1) {
+	gameport_remap(dev->gameport, ((uint16_t) dev->regs[7]) << 2);
+	pclog("Game port at %04X\n", ((uint16_t) dev->regs[7]) << 2);
+    }
 
-	if (upc->regs[0] & 2)
-		serial_setup(upc->uart[0], upc->regs[4] * 4, uart1_int);
+    if (dev->regs[0] & 2) {
+	serial_setup(dev->uart[0], ((uint16_t) dev->regs[4]) << 2, uart1_int);
+	pclog("UART 1 at %04X, IRQ %i\n", ((uint16_t) dev->regs[4]) << 2, uart1_int);
+    }
 
-	if (upc->regs[0] & 4)
-		serial_setup(upc->uart[1], upc->regs[5] * 4, uart2_int);
+    if (dev->regs[0] & 4) {
+	serial_setup(dev->uart[1], ((uint16_t) dev->regs[5]) << 2, uart2_int);
+	pclog("UART 2 at %04X, IRQ %i\n", ((uint16_t) dev->regs[5]) << 2, uart2_int);
+    }
 
-	if (upc->regs[0] & 8) {
-		lpt1_init(upc->regs[6] * 4);
-		lpt1_irq(lpt1_int);
-	}
+    if (dev->regs[0] & 8) {
+	lpt1_init(((uint16_t) dev->regs[6]) << 2);
+	lpt1_irq(lpt1_int);
+	pclog("LPT1 at %04X, IRQ %i\n", ((uint16_t) dev->regs[6]) << 2, lpt1_int);
+    }
 
-	nvr_at_handler(1, upc->regs[3] * 4, upc->nvr);
-	nvr_irq_set(nvr_int, upc->nvr);
+    nvr_at_handler(1, ((uint16_t) dev->regs[3]) << 2, dev->nvr);
+    nvr_irq_set(nvr_int, dev->nvr);
+    pclog("RTC at %04X, IRQ %i\n", ((uint16_t) dev->regs[3]) << 2, nvr_int);
 }
 
+
 static uint8_t 
 f82c710_config_read(uint16_t port, void *priv)
 {
-	upc_t *upc = (upc_t *)priv;
-	uint8_t temp = 0xff;
+    upc_t *dev = (upc_t *) priv;
+    uint8_t temp = 0xff;
 
-        if (upc->configuration_mode) {
-		if (port == upc->cri_addr) {
-			temp = upc->cri;
-                } else if (port == upc->cap_addr) {
-			if (upc->cri == 0xf)
-				temp = upc->cri_addr / 4;
-			else
-				temp = upc->regs[upc->cri];
-                }
-        }
+    if (dev->configuration_mode) {
+	if (port == dev->cri_addr) {
+		temp = dev->cri;
+	} else if (port == dev->cap_addr) {
+		if (dev->cri == 0xf)
+			temp = dev->cri_addr / 4;
+		else
+			temp = dev->regs[dev->cri];
+	}
+    }
 
-        return temp;
+    return temp;
 }
 
+
 static void 
 f82c710_config_write(uint16_t port, uint8_t val, void *priv)
 {
-	upc_t *upc = (upc_t *)priv;
-	int configuration_state_event = 0;
+    upc_t *dev = (upc_t *) priv;
+    int configuration_state_event = 0;
 
-        switch(port) {
-                case 0x2fa:
-			if (upc->configuration_state == 0 && val == 0x55)
-				configuration_state_event = 1;
-                        else if (upc->configuration_state == 4) {
-                                uint8_t addr_verify = upc->cri_addr / 4;
-				addr_verify += val;
-				if (addr_verify == 0xff) {
-                                        upc->configuration_mode = 1;
-                                        /* TODO: is the value of cri reset here or when exiting configuration mode? */
-                                        io_sethandler(upc->cri_addr, 0x0002, f82c710_config_read, NULL, NULL, f82c710_config_write, NULL, NULL, upc);
-                                } else {
-					upc->configuration_mode = 0;
-				}
-			}
-			break;
-                case 0x3fa:
-			if (upc->configuration_state == 1 && val == 0xaa)
-				configuration_state_event = 1;
-                        else if (upc->configuration_state == 2 && val == 0x36)
-				configuration_state_event = 1;
-                        else if (upc->configuration_state == 3) {
-				upc->cri_addr = val * 4;
-				upc->cap_addr = upc->cri_addr + 1;
-				configuration_state_event = 1;
-			}
-			break;
-                default:
-                        break;
-        }
-        
-	if (upc->configuration_mode) {
-                if (port == upc->cri_addr) {
-                        upc->cri = val & 0xf;
-                } else if (port == upc->cap_addr) {
-                        if (upc->cri == 0xf) {
-				upc->configuration_mode = 0;
-				io_removehandler(upc->cri_addr, 0x0002, f82c710_config_read, NULL, NULL, f82c710_config_write, NULL, NULL, upc);
-				/* TODO: any benefit in updating at each register write instead of when exiting config mode? */
-				if (upc->local == 710)
-					f82c710_update_ports(upc);
-				if (upc->local == 606)
-					f82c606_update_ports(upc);
-			} else {
-                                upc->regs[upc->cri] = val;
-                        }
-                }
-        }
+    switch (port) {
+	case 0x2fa:
+		if ((dev->configuration_state == 0) && (val != 0x00) && (val != 0xff) && (dev->local == 606)) {
+			configuration_state_event = 1;
+			dev->last_write = val;
+		} else if ((dev->configuration_state == 0) && (val == 0x55) && (dev->local == 710))
+			configuration_state_event = 1;
+		else if (dev->configuration_state == 4) {
+			if ((val | dev->last_write) == 0xff) {
+				dev->cri_addr = ((uint16_t) dev->last_write) << 2;
+				dev->cap_addr = dev->cri_addr + 1;
+				dev->configuration_mode = 1;
+				if (dev->local == 606)
+					f82c606_update_ports(dev, 0);
+				else if (dev->local == 710)
+					f82c710_update_ports(dev, 0);
+				/* TODO: is the value of cri reset here or when exiting configuration mode? */
+				io_sethandler(dev->cri_addr, 0x0002, f82c710_config_read, NULL, NULL, f82c710_config_write, NULL, NULL, dev);
+			} else
+				dev->configuration_mode = 0;
+		}
+		break;
+	case 0x3fa:
+		if ((dev->configuration_state == 1) && ((val | dev->last_write) == 0xff) && (dev->local == 606))
+			configuration_state_event = 1;
+		else if ((dev->configuration_state == 1) && (val == 0xaa) && (dev->local == 710))
+			configuration_state_event = 1;
+		else if ((dev->configuration_state == 2) && (val == 0x36))
+			configuration_state_event = 1;
+		else if (dev->configuration_state == 3) {
+			dev->last_write = val;
+			configuration_state_event = 1;
+		}
+		break;
+	default:
+		break;
+    }
 
-        /* TODO: is the state only reset when accessing 0x2fa and 0x3fa wrongly? */
-	if ((port == 0x2fa || port == 0x3fa) && configuration_state_event)
-		upc->configuration_state++;
-	else
-		upc->configuration_state = 0;
+    if (dev->configuration_mode) {
+	if (port == dev->cri_addr) {
+		dev->cri = val & 0xf;
+	} else if (port == dev->cap_addr) {
+		if (dev->cri == 0xf) {
+			dev->configuration_mode = 0;
+			io_removehandler(dev->cri_addr, 0x0002, f82c710_config_read, NULL, NULL, f82c710_config_write, NULL, NULL, dev);
+			/* TODO: any benefit in updating at each register write instead of when exiting config mode? */
+			if (dev->local == 606)
+				f82c606_update_ports(dev, 1);
+			else if (dev->local == 710)
+				f82c710_update_ports(dev, 1);
+		} else
+			dev->regs[dev->cri] = val;
+	}
+    }
+
+    /* TODO: is the state only reset when accessing 0x2fa and 0x3fa wrongly? */
+    if ((port == 0x2fa || port == 0x3fa) && configuration_state_event)
+	dev->configuration_state++;
+    else
+	dev->configuration_state = 0;
 }
 
 
 static void
-f82c710_reset(upc_t *upc)
+f82c710_reset(void *priv)
 {
-	serial_remove(upc->uart[0]);
-	serial_setup(upc->uart[0], SERIAL1_ADDR, SERIAL1_IRQ);
+    upc_t *dev = (upc_t *) priv;
 
-	serial_remove(upc->uart[1]);
-	serial_setup(upc->uart[1], SERIAL2_ADDR, SERIAL2_IRQ);
-	
-	lpt1_remove();
-	lpt1_init(0x378);
-	lpt1_irq(7);
+    /* Set power-on defaults. */
+    if (dev->local == 606) {
+	dev->regs[0] = 0x00;	/* Enable */
+	dev->regs[1] = 0x00;	/* Configuration Register */
+	dev->regs[2] = 0x00;	/* Ext Baud Rate Select */
+	dev->regs[3] = 0xb0;	/* RTC Base */
+	dev->regs[4] = 0xfe;	/* UART1 Base */
+	dev->regs[5] = 0xbe;	/* UART2 Base */
+	dev->regs[6] = 0x9e;	/* Parallel Base */
+	dev->regs[7] = 0x80;	/* Game Base */
+	dev->regs[8] = 0xec;	/* Interrupt Select */
+    } else if (dev->local == 710) {
+	dev->regs[0] = 0x0c;
+	dev->regs[1] = 0x00;
+	dev->regs[2] = 0x00;
+	dev->regs[3] = 0x00;
+	dev->regs[4] = 0xfe;
+	dev->regs[5] = 0x00;
+	dev->regs[6] = 0x9e;
+	dev->regs[7] = 0x00;
+	dev->regs[8] = 0x00;
+	dev->regs[9] = 0xb0;
+	dev->regs[10] = 0x00;
+	dev->regs[11] = 0x00;
+	dev->regs[12] = 0xa0;
+	dev->regs[13] = 0x00;
+	dev->regs[14] = 0x00;
+    }
 
-	if (upc->local == 710)
-		fdc_reset(upc->fdc);
+    if (dev->local == 606)
+	f82c606_update_ports(dev, 1);
+    else if (dev->local == 710)
+	f82c710_update_ports(dev, 1);
 }
 
-static void *
-f82c710_init(const device_t *info)
-{
-	upc_t *upc = (upc_t *) malloc(sizeof(upc_t));
-	memset(upc, 0, sizeof(upc_t));
-	upc->local = info->local;
-
-	if (upc->local == 710) {
-		upc->regs[0] = 0x0c;
-		upc->regs[1] = 0x00;
-		upc->regs[2] = 0x00;
-		upc->regs[3] = 0x00;
-		upc->regs[4] = 0xfe;
-		upc->regs[5] = 0x00;
-		upc->regs[6] = 0x9e;
-		upc->regs[7] = 0x00;
-		upc->regs[8] = 0x00;
-		upc->regs[9] = 0xb0;
-		upc->regs[10] = 0x00;
-		upc->regs[11] = 0x00;
-		upc->regs[12] = 0xa0;
-		upc->regs[13] = 0x00;
-		upc->regs[14] = 0x00;
-
-		upc->fdc = device_add(&fdc_at_device);
-	}
-
-	if (upc->local == 606) {
-		/* Set power-on defaults. */
-		upc->regs[0] = 0x00; /* Enable */
-		upc->regs[1] = 0x00; /* Configuration Register */
-		upc->regs[2] = 0x00; /* Ext Baud Rate Select */
-		upc->regs[3] = 0xb0; /* RTC Base */
-		upc->regs[4] = 0xfe; /* UART1 Base */
-		upc->regs[5] = 0xbe; /* UART2 Base */
-		upc->regs[6] = 0x9e; /* Parallel Base */
-		upc->regs[7] = 0x80; /* Game Base */
-		upc->regs[8] = 0xec; /* Interrupt Select */
-
-		upc->nvr = device_add(&at_nvr_old_device);
-		upc->gameport = gameport_add(&gameport_sio_device);
-	}
-
-	upc->uart[0] = device_add_inst(&ns16450_device, 1);
-	upc->uart[1] = device_add_inst(&ns16450_device, 2);	
-
-        io_sethandler(0x02fa, 0x0001, NULL, NULL, NULL, f82c710_config_write, NULL, NULL, upc);
-        io_sethandler(0x03fa, 0x0001, NULL, NULL, NULL, f82c710_config_write, NULL, NULL, upc);
-
-	f82c710_reset(upc);
-
-	if (upc->local == 710)
-		f82c710_update_ports(upc);
-	if (upc->local == 606)
-		f82c606_update_ports(upc);
-
-	return upc;
-}
 
 static void
 f82c710_close(void *priv)
 {
-	upc_t *upc = (upc_t *)priv;
+    upc_t *dev = (upc_t *) priv;
 
-	free(upc);
+    free(dev);
 }
 
+
+static void *
+f82c710_init(const device_t *info)
+{
+    upc_t *dev = (upc_t *) malloc(sizeof(upc_t));
+    memset(dev, 0, sizeof(upc_t));
+    dev->local = info->local;
+
+    if (dev->local == 606) {
+	dev->nvr = device_add(&at_nvr_old_device);
+	dev->gameport = gameport_add(&gameport_sio_device);
+    } else if (dev->local == 710)
+	dev->fdc = device_add(&fdc_at_device);
+
+    dev->uart[0] = device_add_inst(&ns16450_device, 1);
+    dev->uart[1] = device_add_inst(&ns16450_device, 2);	
+
+    io_sethandler(0x02fa, 0x0001, NULL, NULL, NULL, f82c710_config_write, NULL, NULL, dev);
+    io_sethandler(0x03fa, 0x0001, NULL, NULL, NULL, f82c710_config_write, NULL, NULL, dev);
+
+    f82c710_reset(dev);
+
+    return dev;
+}
+
+
 const device_t f82c606_device = {
-	"82C606 CHIPSpak Multifunction Controller",
-	0,
-	606,
-	f82c710_init, f82c710_close, NULL,
-	{ NULL }, NULL, NULL,
-	NULL
+    "82C606 CHIPSpak Multifunction Controller",
+    0,
+    606,
+    f82c710_init, f82c710_close, f82c710_reset,
+    { NULL }, NULL, NULL,
+    NULL
 };
 
 const device_t f82c710_device = {
-	"F82C710 UPC Super I/O",
-	0,
-	710,
-	f82c710_init, f82c710_close, NULL,
-	{ NULL }, NULL, NULL,
-	NULL
+    "F82C710 UPC Super I/O",
+    0,
+    710,
+    f82c710_init, f82c710_close, f82c710_reset,
+    { NULL }, NULL, NULL,
+    NULL
 };
diff --git a/src/thread.c b/src/thread.c
new file mode 100644
index 000000000..916132fb2
--- /dev/null
+++ b/src/thread.c
@@ -0,0 +1,195 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include <time.h>
+#include <pthread.h>
+#include <sys/param.h>
+#include <unistd.h>
+#include <inttypes.h>
+#include <86box/86box.h>
+#include <86box/plat.h>
+
+
+typedef struct event_pthread_t
+{
+    pthread_cond_t	cond;
+    pthread_mutex_t	mutex;
+    int state;
+} event_pthread_t;
+
+
+typedef struct thread_param
+{
+    void		(*thread_rout)(void*);
+    void *		param;
+} thread_param;
+
+
+typedef struct pt_mutex_t
+{
+    pthread_mutex_t	mutex;
+} pt_mutex_t;
+
+
+void *
+thread_run_wrapper(thread_param* arg)
+{
+    thread_param localparam = *arg;
+    free(arg);
+    localparam.thread_rout(localparam.param);
+    return NULL;
+}
+
+
+thread_t *
+thread_create(void (*thread_rout)(void *param), void *param)
+{
+    pthread_t *thread = malloc(sizeof(pthread_t));
+    thread_param *thrparam = malloc(sizeof(thread_param));
+    thrparam->thread_rout = thread_rout;
+    thrparam->param = param;
+
+    pthread_create(thread, NULL, (void* (*)(void*))thread_run_wrapper, thrparam);
+
+    return thread;
+}
+
+
+int
+thread_wait(thread_t *arg, int timeout)
+{
+    return pthread_join(*(pthread_t*)(arg), NULL) != 0;
+}
+
+
+event_t *
+thread_create_event()
+{
+    event_pthread_t *event = malloc(sizeof(event_pthread_t));
+
+    pthread_cond_init(&event->cond, NULL);
+    pthread_mutex_init(&event->mutex, NULL);
+    event->state = 0;
+
+    return (event_t *)event;
+}
+
+
+void
+thread_set_event(event_t *handle)
+{
+    event_pthread_t *event = (event_pthread_t *)handle;
+
+    pthread_mutex_lock(&event->mutex);
+    event->state = 1;
+    pthread_cond_broadcast(&event->cond);
+    pthread_mutex_unlock(&event->mutex);
+}
+
+
+void
+thread_reset_event(event_t *handle)
+{
+    event_pthread_t *event = (event_pthread_t *)handle;
+
+    pthread_mutex_lock(&event->mutex);
+    event->state = 0;
+    pthread_mutex_unlock(&event->mutex);
+}
+
+
+int
+thread_wait_event(event_t *handle, int timeout)
+{
+    event_pthread_t *event = (event_pthread_t *)handle;
+    struct timespec abstime;
+
+    clock_gettime(CLOCK_REALTIME, &abstime);
+    abstime.tv_nsec += (timeout % 1000) * 1000000;
+    abstime.tv_sec += (timeout / 1000);
+    if (abstime.tv_nsec > 1000000000) {
+	abstime.tv_nsec -= 1000000000;
+	abstime.tv_sec++;
+    }
+
+    pthread_mutex_lock(&event->mutex);
+    if (timeout == -1) {
+	while (!event->state)
+		pthread_cond_wait(&event->cond, &event->mutex);
+    } else if (!event->state)
+	pthread_cond_timedwait(&event->cond, &event->mutex, &abstime);
+    pthread_mutex_unlock(&event->mutex);
+
+    return 0;
+}
+
+
+void
+thread_destroy_event(event_t *handle)
+{
+    event_pthread_t *event = (event_pthread_t *)handle;
+
+    pthread_cond_destroy(&event->cond);
+    pthread_mutex_destroy(&event->mutex);
+
+    free(event);
+}
+
+
+void
+thread_sleep(int t)
+{
+    usleep(t * 1000);
+}
+
+
+mutex_t *
+thread_create_mutex(void)
+{
+    pt_mutex_t *mutex = malloc(sizeof(pt_mutex_t));
+
+    pthread_mutex_init(&mutex->mutex, NULL);
+
+    return mutex;
+}
+
+
+mutex_t *
+thread_create_mutex_with_spin_count(unsigned int spin_count)
+{
+    /* Setting spin count of a mutex is not possible with pthreads. */
+    return thread_create_mutex();
+}
+
+
+int
+thread_wait_mutex(mutex_t *_mutex)
+{
+    if (_mutex == NULL)
+	return(0);
+    pt_mutex_t *mutex = (pt_mutex_t *)_mutex;
+
+    return
+	pthread_mutex_lock(&mutex->mutex) != 0;
+}
+
+
+int
+thread_release_mutex(mutex_t *_mutex)
+{
+    if (_mutex == NULL)
+	return(0);
+    pt_mutex_t *mutex = (pt_mutex_t *)_mutex;
+
+    return pthread_mutex_unlock(&mutex->mutex) != 0;
+}
+
+
+void
+thread_close_mutex(mutex_t *_mutex)
+{
+    pt_mutex_t *mutex = (pt_mutex_t *)_mutex;
+
+    pthread_mutex_destroy(&mutex->mutex);
+
+    free(mutex);
+}
diff --git a/src/unix/CMakeLists.txt b/src/unix/CMakeLists.txt
index 2371124df..10cc8c53e 100644
--- a/src/unix/CMakeLists.txt
+++ b/src/unix/CMakeLists.txt
@@ -1,13 +1,12 @@
-set(PLAT_SOURCES unix_thread.c)
 if (${CMAKE_SYSTEM_NAME} MATCHES "Linux")
     find_package(ALSA)
     if (ALSA_FOUND)
-        set(PLAT_SOURCES ${PLAT_SOURCES} linux_midi_alsa.c)
+        set(PLAT_SOURCES linux_midi_alsa.c)
     else()
-        set(PLAT_SOURCES ${PLAT_SOURCES} unix_midi.c)
+        set(PLAT_SOURCES unix_midi.c)
     endif()
 else()
-    set(PLAT_SOURCES ${PLAT_SOURCES} unix_midi.c)
+    set(PLAT_SOURCES unix_midi.c)
 endif()
 add_library(plat STATIC ${PLAT_SOURCES})
 add_library(ui STATIC unix.c unix_sdl.c unix_cdrom.c)
diff --git a/src/unix/unix_thread.c b/src/unix/unix_thread.c
deleted file mode 100644
index 071178418..000000000
--- a/src/unix/unix_thread.c
+++ /dev/null
@@ -1,174 +0,0 @@
-#include <stdlib.h>
-#include <stdio.h>
-#include <time.h>
-#include <pthread.h>
-#include <sys/param.h>
-#include <unistd.h>
-#include <inttypes.h>
-#include <86box/86box.h>
-#include <86box/plat.h>
-
-typedef struct event_pthread_t
-{
-	pthread_cond_t cond;
-	pthread_mutex_t mutex;
-	int state;
-} event_pthread_t;
-
-typedef struct thread_param
-{
-	void (*thread_rout)(void*);
-	void* param;
-} thread_param;
-
-void* thread_run_wrapper(thread_param* arg)
-{
-	thread_param localparam = *arg;
-	free(arg);
-	localparam.thread_rout(localparam.param);
-	return NULL;
-}
-
-thread_t *thread_create(void (*thread_rout)(void *param), void *param)
-{
-	pthread_t *thread = malloc(sizeof(pthread_t));
-	thread_param *thrparam = malloc(sizeof(thread_param));
-	thrparam->thread_rout = thread_rout;
-	thrparam->param = param;
-        
-	pthread_create(thread, NULL, (void* (*)(void*))thread_run_wrapper, thrparam);
-
-	return thread;
-}
-
-int
-thread_wait(thread_t *arg, int timeout)
-{
-	return pthread_join(*(pthread_t*)(arg), NULL) != 0;
-}
-
-event_t *thread_create_event()
-{
-	event_pthread_t *event = malloc(sizeof(event_pthread_t));
-
-	pthread_cond_init(&event->cond, NULL);
-	pthread_mutex_init(&event->mutex, NULL);
-	event->state = 0;
-
-        return (event_t *)event;
-}
-
-void thread_set_event(event_t *handle)
-{
-	event_pthread_t *event = (event_pthread_t *)handle;
-
-	pthread_mutex_lock(&event->mutex);
-	event->state = 1;
-	pthread_cond_broadcast(&event->cond);
-	pthread_mutex_unlock(&event->mutex);
-}
-
-void thread_reset_event(event_t *handle)
-{
-	event_pthread_t *event = (event_pthread_t *)handle;
-
-	pthread_mutex_lock(&event->mutex);
-	event->state = 0;
-	pthread_mutex_unlock(&event->mutex);
-}
-
-int thread_wait_event(event_t *handle, int timeout)
-{
-	event_pthread_t *event = (event_pthread_t *)handle;
-	struct timespec abstime;
-
-#if defined __linux__ || defined BSD
-	clock_gettime(CLOCK_REALTIME, &abstime);
-#else
-        struct timeval now;
-        gettimeofday(&now, 0);
-        abstime.tv_sec = now.tv_sec;
-        abstime.tv_nsec = now.tv_usec*1000UL;
-#endif
-	abstime.tv_nsec += (timeout % 1000) * 1000000;
-	abstime.tv_sec += (timeout / 1000);
-	if (abstime.tv_nsec > 1000000000)
-	{
-		abstime.tv_nsec -= 1000000000;
-		abstime.tv_sec++;
-	}
-
-	pthread_mutex_lock(&event->mutex);
-	if (timeout == -1)
-	{
-		while (!event->state)
-			pthread_cond_wait(&event->cond, &event->mutex);
-	}
-	else if (!event->state)
-		pthread_cond_timedwait(&event->cond, &event->mutex, &abstime);
-	pthread_mutex_unlock(&event->mutex);
-
-        return 0;
-}
-
-void thread_destroy_event(event_t *handle)
-{
-	event_pthread_t *event = (event_pthread_t *)handle;
-
-	pthread_cond_destroy(&event->cond);
-	pthread_mutex_destroy(&event->mutex);
-
-	free(event);
-}
-
-void thread_sleep(int t)
-{
-	usleep(t * 1000);
-}
-
-
-typedef struct pt_mutex_t
-{
-        pthread_mutex_t mutex;
-} pt_mutex_t;
-
-mutex_t *thread_create_mutex(void)
-{
-        pt_mutex_t *mutex = malloc(sizeof(pt_mutex_t));
-        
-        pthread_mutex_init(&mutex->mutex, NULL);
-        
-        return mutex;
-}
-
-mutex_t *
-thread_create_mutex_with_spin_count(unsigned int spin_count)
-{
-	/* Setting spin count of a mutex is not possible with pthreads. */
-	return thread_create_mutex();
-}
-
-int thread_wait_mutex(mutex_t *_mutex)
-{
-	if (_mutex == NULL) return(0);
-    pt_mutex_t *mutex = (pt_mutex_t *)_mutex;
-        
-    return pthread_mutex_lock(&mutex->mutex) != 0;
-}
-
-int thread_release_mutex(mutex_t *_mutex)
-{
-	if (_mutex == NULL) return(0);
-    pt_mutex_t *mutex = (pt_mutex_t *)_mutex;
-        
-	return pthread_mutex_unlock(&mutex->mutex) != 0;
-}
-
-void thread_close_mutex(mutex_t *_mutex)
-{
-        pt_mutex_t *mutex = (pt_mutex_t *)_mutex;
-
-        pthread_mutex_destroy(&mutex->mutex);
-
-        free(mutex);
-}
\ No newline at end of file
diff --git a/src/video/vid_f82c425.c b/src/video/vid_f82c425.c
index a9d2f14ba..e2f2f0677 100644
--- a/src/video/vid_f82c425.c
+++ b/src/video/vid_f82c425.c
@@ -144,7 +144,11 @@ static inline uint32_t f82c425_makecol(uint8_t rgbi, int gs4, int inv)
 	}
 	c = 0x10 * gs4 * ((rgbi >> gs4) + 2);
 
+#ifdef NO_BLUE
 	return makecol(c, c + 0x08, c - 0x20);
+#else
+	return makecol(c, c + 0x08, 0x70);
+#endif
 }
 
 /* Saturating/non-saturating addition for SMARTMAP(see below). */
diff --git a/src/video/vid_s3.c b/src/video/vid_s3.c
index 24a322d25..3f4987bab 100644
--- a/src/video/vid_s3.c
+++ b/src/video/vid_s3.c
@@ -1877,6 +1877,22 @@ s3_queue(s3_t *s3, uint32_t addr, uint32_t val, uint32_t type)
 		wake_fifo_thread(s3);
 }
 
+
+static uint32_t
+s3_hwcursor_convert_addr(svga_t *svga)
+{
+    if ((svga->bpp == 8) && ((svga->gdcreg[5] & 0x60) >= 0x20) && (svga->crtc[0x45] & 0x10)) {
+	if ((svga->gdcreg[5] & 0x60) >= 0x40)
+		return ((svga->hwcursor_latch.addr & 0xfffff1ff) | ((svga->hwcursor_latch.addr & 0x200) << 2)) | 0x600;
+	else if ((svga->gdcreg[5] & 0x60) == 0x20)
+		return ((svga->hwcursor_latch.addr & 0xfffff0ff) | ((svga->hwcursor_latch.addr & 0x300) << 2)) | 0x300;
+	else
+		return svga->hwcursor_latch.addr;
+    } else
+	return svga->hwcursor_latch.addr;
+}
+
+
 static void
 s3_hwcursor_draw(svga_t *svga, int displine)
 {
@@ -1887,6 +1903,7 @@ s3_hwcursor_draw(svga_t *svga, int displine)
 	int xx;
 	int offset = svga->hwcursor_latch.x - svga->hwcursor_latch.xoff;
 	uint32_t fg, bg;
+	uint32_t real_addr;
 
 	switch (svga->bpp)
 	{	       
@@ -1939,10 +1956,12 @@ s3_hwcursor_draw(svga_t *svga, int displine)
 	if (svga->interlace && svga->hwcursor_oddeven)
 		svga->hwcursor_latch.addr += 16;
 
+	real_addr = s3_hwcursor_convert_addr(svga);
+
 	for (x = 0; x < 64; x += 16)
 	{
-		dat[0] = (svga->vram[svga->hwcursor_latch.addr & svga->vram_display_mask] << 8) | svga->vram[(svga->hwcursor_latch.addr + 1) & svga->vram_display_mask];
-		dat[1] = (svga->vram[(svga->hwcursor_latch.addr + 2) & svga->vram_display_mask] << 8) | svga->vram[(svga->hwcursor_latch.addr + 3) & svga->vram_display_mask];
+		dat[0] = (svga->vram[real_addr & svga->vram_display_mask] << 8) | svga->vram[(real_addr + 1) & svga->vram_display_mask];
+		dat[1] = (svga->vram[(real_addr + 2) & svga->vram_display_mask] << 8) | svga->vram[(real_addr + 3) & svga->vram_display_mask];
                 if (svga->crtc[0x55] & 0x10) {
                         /*X11*/
                         for (xx = 0; xx < 16; xx++) {
@@ -1971,6 +1990,7 @@ s3_hwcursor_draw(svga_t *svga, int displine)
                         }
                 }
 		svga->hwcursor_latch.addr += 4;
+		real_addr = s3_hwcursor_convert_addr(svga);
 	}
 	if (svga->interlace && !svga->hwcursor_oddeven)
 		svga->hwcursor_latch.addr += 16;
diff --git a/src/video/video.c b/src/video/video.c
index a98f125cb..1258b1824 100644
--- a/src/video/video.c
+++ b/src/video/video.c
@@ -367,7 +367,9 @@ video_take_screenshot(const char *fn)
 			memset(&(b_rgb[y][x * 3]), 0x00, 3);
 		else {
 			temp = buffer32->line[blit_data.y + y][blit_data.x + x];
-			memcpy(&(b_rgb[y][x * 3]), &temp, 3);
+			b_rgb[y][x * 3] = (temp >> 16) & 0xff;
+			b_rgb[y][(x * 3) + 1] = (temp >> 8) & 0xff;
+			b_rgb[y][(x * 3) + 2] = temp & 0xff;
 		}
 	}
     }
diff --git a/src/win/CMakeLists.txt b/src/win/CMakeLists.txt
index 6102cde68..75d3daf4a 100644
--- a/src/win/CMakeLists.txt
+++ b/src/win/CMakeLists.txt
@@ -15,7 +15,7 @@
 
 enable_language(RC)
 
-add_library(plat OBJECT win.c win_dynld.c win_thread.c win_cdrom.c
+add_library(plat OBJECT win.c win_dynld.c win_cdrom.c
 	win_keyboard.c win_crashdump.c win_midi.c win_mouse.c)
 
 add_library(ui OBJECT win_ui.c win_stbar.c win_sdl.c win_dialog.c win_about.c
diff --git a/src/win/Makefile.mingw b/src/win/Makefile.mingw
index 29d0c3456..d168796ec 100644
--- a/src/win/Makefile.mingw
+++ b/src/win/Makefile.mingw
@@ -26,6 +26,10 @@ ifndef DEV_BUILD
 DEV_BUILD	:= n
 endif
 
+ifneq ($(PTHREAD), n)
+  PTHREAD	:= y
+endif
+
 ifeq ($(DEV_BUILD), y)
  ifndef DEBUG
   DEBUG		:= y
@@ -595,6 +599,7 @@ MAINOBJ		:= 86box.o config.o random.o timer.o io.o acpi.o apm.o dma.o ddma.o \
 		   nmi.o pic.o pit.o port_92.o ppi.o pci.o mca.o \
 		   usb.o device.o nvr.o nvr_at.o nvr_ps2.o \
 		   $(VNCOBJ)
+endif
 
 MEMOBJ		:= catalyst_flash.o i2c_eeprom.o intel_flash.o mem.o rom.o smram.o spd.o sst_flash.o
 
@@ -796,11 +801,19 @@ VOODOOOBJ	:= vid_voodoo.o vid_voodoo_banshee.o \
 		    vid_voodoo_render.o vid_voodoo_setup.o \
 		    vid_voodoo_texture.o
 
+ifeq ($(PTHREAD), y)
+PLATOBJ		:= win.o \
+		    win_dynld.o \
+		    win_cdrom.o win_keyboard.o \
+		    win_crashdump.o win_midi.o \
+		    win_mouse.o
+else
 PLATOBJ		:= win.o \
 		    win_dynld.o win_thread.o \
 		    win_cdrom.o win_keyboard.o \
 		    win_crashdump.o win_midi.o \
 		    win_mouse.o
+endif
 
 ifeq ($(DINPUT), y)
  PLATOBJ	+= win_joystick.o
@@ -826,7 +839,11 @@ endif
 ifneq ($(WX), n)
 LIBS		+= $(WX_LIBS) -lm
 endif
+ifeq ($(PTHREAD), y)
+LIBS		+= -lpng -lz -lwsock32 -lshell32 -liphlpapi -lpsapi -lSDL2 -limm32 -lhid -lsetupapi -loleaut32 -luxtheme -lversion -lwinmm -static -lstdc++ -lpthread
+else
 LIBS		+= -lpng -lz -lwsock32 -lshell32 -liphlpapi -lpsapi -lSDL2 -limm32 -lhid -lsetupapi -loleaut32 -luxtheme -lversion -lwinmm -static -lstdc++
+endif
 ifneq ($(X64), y)
 ifneq ($(ARM64), y)
 LIBS		+= -Wl,--large-address-aware
diff --git a/src/win/win.c b/src/win/win.c
index 11b833860..b922c0da3 100644
--- a/src/win/win.c
+++ b/src/win/win.c
@@ -1094,6 +1094,8 @@ plat_setfullscreen(int on)
 				ResizeWindowByClientArea(hwndMain, temp_x, temp_y);
 			else
 				ResizeWindowByClientArea(hwndMain, temp_x, temp_y + sbar_height);
+
+			SetWindowPos(hwndMain, HWND_TOP, window_x, window_y, 0, 0, SWP_NOSIZE);
 		}
 
 		/* Render window. */
diff --git a/src/win/win_opengl.c b/src/win/win_opengl.c
index 7e92e40b8..a73ffd981 100644
--- a/src/win/win_opengl.c
+++ b/src/win/win_opengl.c
@@ -64,6 +64,7 @@ static const int INIT_HEIGHT = 400;
 static const int BUFFERPIXELS = 4460544;	/* Same size as render_buffer, pow(2048+64,2). */
 static const int BUFFERBYTES = 17842176;	/* Pixel is 4 bytes. */
 static const int BUFFERCOUNT = 3;		/* How many buffers to use for pixel transfer (2-3 is commonly recommended). */
+static const int ROW_LENGTH = 2112;		/* Source buffer row lenght (including padding) */
 
 /**
  * @brief A dedicated OpenGL thread.
@@ -106,7 +107,7 @@ static union
 */
 typedef struct
 {
-	int y1, y2, w, h;
+	int w, h;
 	void* buffer;			/* Buffer for pixel transfer, allocated by gpu driver. */
 	volatile atomic_flag in_use;	/* Is buffer currently in use. */
 	GLsync sync;			/* Fence sync object used by opengl thread to track pixel transfer completion. */
@@ -641,18 +642,16 @@ static void opengl_main(void* param)
 					SetEvent(sync_objects.resize);
 			}
 
-			/* Clip to height. y2 can be out-of-bounds. */
-			int sub_height = MIN(info->y2, info->h) - info->y1;
-
 			if (!GLAD_GL_ARB_buffer_storage)
 			{
 				/* Fallback method, copy data to pixel buffer. */
-				glBufferSubData(GL_PIXEL_UNPACK_BUFFER, BUFFERBYTES * read_pos, info->w * sub_height * sizeof(uint32_t), info->buffer);
+				glBufferSubData(GL_PIXEL_UNPACK_BUFFER, BUFFERBYTES * read_pos, info->h * ROW_LENGTH * sizeof(uint32_t), info->buffer);
 			}
 
 			/* Update texture from pixel buffer. */
 			glPixelStorei(GL_UNPACK_SKIP_PIXELS, BUFFERPIXELS * read_pos);
-			glTexSubImage2D(GL_TEXTURE_2D, 0, 0, info->y1, info->w, sub_height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, NULL);
+			glPixelStorei(GL_UNPACK_ROW_LENGTH, ROW_LENGTH);
+			glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, info->w, info->h, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, NULL);
 
 			/* Add fence to track when above gl commands are complete. */
 			info->sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
@@ -819,16 +818,10 @@ static void opengl_blit(int x, int y, int w, int h)
 		return;
 	}
 
-	for (yy = 0; yy < h; yy++) {
-		if ((y + yy) >= 0 && (y + yy) < buffer32->h)
-			memcpy(blit_info[write_pos].buffer + (yy * w * sizeof(uint32_t)),
-			       &(((uint32_t *) buffer32->line[y + yy])[x]), w * sizeof(uint32_t));
-	}
+	memcpy(blit_info[write_pos].buffer, &(buffer32->line[y][x]), h * ROW_LENGTH * sizeof(uint32_t));
 
 	video_blit_complete();
 
-	blit_info[write_pos].y1 = 0;
-	blit_info[write_pos].y2 = h - 1;
 	blit_info[write_pos].w = w;
 	blit_info[write_pos].h = h;
 
diff --git a/src/win/win_specify_dim.c b/src/win/win_specify_dim.c
index 52336ad5e..d4727c68b 100644
--- a/src/win/win_specify_dim.c
+++ b/src/win/win_specify_dim.c
@@ -140,11 +140,13 @@ SpecifyDimensionsDialogProcedure(HWND hdlg, UINT message, WPARAM wParam, LPARAM
 				if (mouse_capture)
 					ClipCursor(&r);
 
-				if (!(vid_resize & 2) && window_remember) {
+				if (window_remember || (vid_resize & 2)) {
 					window_x = r.left;
 					window_y = r.top;
-					window_w = r.right - r.left;
-					window_h = r.bottom - r.top;
+					if (!(vid_resize & 2)) {
+						window_w = r.right - r.left;
+						window_h = r.bottom - r.top;
+					}
 				}
 
 				config_save();
diff --git a/src/win/win_ui.c b/src/win/win_ui.c
index a7cf7be8f..75b0aad92 100644
--- a/src/win/win_ui.c
+++ b/src/win/win_ui.c
@@ -697,11 +697,13 @@ MainWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam)
 				window_remember = !window_remember;
 				CheckMenuItem(hmenu, IDM_VID_REMEMBER, window_remember ? MF_CHECKED : MF_UNCHECKED);
 				GetWindowRect(hwnd, &rect);
-				if (!(vid_resize & 2) && window_remember) {
+				if (window_remember || (vid_resize & 2)) {
 					window_x = rect.left;
 					window_y = rect.top;
-					window_w = rect.right - rect.left;
-					window_h = rect.bottom - rect.top;
+					if (!(vid_resize & 2)) {
+						window_w = rect.right - rect.left;
+						window_h = rect.bottom - rect.top;
+					}
 				}
 				config_save();
 				break;
@@ -978,11 +980,13 @@ MainWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam)
 			video_force_resize_set(1);
 		}
 
-		if (window_remember) {
+		if (window_remember || (vid_resize & 2)) {
 			window_x = pos->x;
 			window_y = pos->y;
-			window_w = pos->cx;
-			window_h = pos->cy;
+			if (!(vid_resize & 2)) {
+				window_w = pos->cx;
+				window_h = pos->cy;
+			}
 			save_window_pos = 1;
 			config_save();
 		}
@@ -1382,6 +1386,8 @@ ui_init(int nCmdShow)
 		ResizeWindowByClientArea(hwndMain, scrnsz_x, scrnsz_y);
 	else
 		ResizeWindowByClientArea(hwndMain, scrnsz_x, scrnsz_y + sbar_height);
+
+	SetWindowPos(hwndMain, HWND_TOP, window_x, window_y, 0, 0, SWP_NOSIZE);
     }
 
     /* Reset all menus to their defaults. */