diff --git a/src/acpi.c b/src/acpi.c
index 350021e00..81c62b103 100644
--- a/src/acpi.c
+++ b/src/acpi.c
@@ -39,6 +39,9 @@
 #include <86box/i2c.h>
 
 
+int acpi_rtc_status = 0;
+
+
 #ifdef ENABLE_ACPI_LOG
 int acpi_do_log = ENABLE_ACPI_LOG;
 
@@ -119,6 +122,8 @@ acpi_reg_read_common_regs(int size, uint16_t addr, void *p)
 	case 0x00: case 0x01:
 		/* PMSTS - Power Management Status Register (IO) */
 		ret = (dev->regs.pmsts >> shift16) & 0xff;
+		if (addr == 0x01)
+			ret |= (acpi_rtc_status << 2);
 		break;
 	case 0x02: case 0x03:
 		/* PMEN - Power Management Resume Enable Register (IO) */
@@ -127,6 +132,8 @@ acpi_reg_read_common_regs(int size, uint16_t addr, void *p)
 	case 0x04: case 0x05:
 		/* PMCNTRL - Power Management Control Register (IO) */
 		ret = (dev->regs.pmcntrl >> shift16) & 0xff;
+		if (addr == 0x05)
+			ret = (ret & 0xdf) | 0x02;	/* Bit 5 is write-only. */
 		break;
 	case 0x08: case 0x09: case 0x0a: case 0x0b:
 		/* PMTMR - Power Management Timer Register (IO) */
@@ -138,7 +145,10 @@ acpi_reg_read_common_regs(int size, uint16_t addr, void *p)
 		break;
     }
 
-    acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#endif
     return ret;
 }
 
@@ -211,7 +221,10 @@ acpi_reg_read_intel(int size, uint16_t addr, void *p)
 		break;
     }
 
-    acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+		acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#endif
     return ret;
 }
 
@@ -285,7 +298,10 @@ acpi_reg_read_via_common(int size, uint16_t addr, void *p)
 		break;
     }
 
-    acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#endif
     return ret;
 }
 
@@ -338,7 +354,10 @@ acpi_reg_read_via(int size, uint16_t addr, void *p)
 		break;
     }
 
-    acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#endif
     return ret;
 }
 
@@ -372,7 +391,10 @@ acpi_reg_read_via_596b(int size, uint16_t addr, void *p)
 		break;
     }
 
-    acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#endif
     return ret;
 }
 
@@ -386,7 +408,10 @@ acpi_reg_read_smc(int size, uint16_t addr, void *p)
 
     ret = acpi_reg_read_common_regs(size, addr, p);
 
-    acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
+#endif
     return ret;
 }
 
@@ -436,13 +461,18 @@ acpi_reg_write_common_regs(int size, uint16_t addr, uint8_t val, void *p)
     int shift16, sus_typ;
 
     addr &= 0x3f;
-    acpi_log("(%i) ACPI Write (%i) %02X: %02X\n", in_smm, size, addr, val);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Write (%i) %02X: %02X\n", in_smm, size, addr, val);
+#endif
     shift16 = (addr & 1) << 3;
 
     switch (addr) {
 	case 0x00: case 0x01:
 		/* PMSTS - Power Management Status Register (IO) */
 		dev->regs.pmsts &= ~((val << shift16) & 0x8d31);
+		if ((addr == 0x01) && (val & 0x04))
+			acpi_rtc_status = 0;
 		acpi_update_irq(dev);
 		break;
 	case 0x02: case 0x03:
@@ -452,8 +482,7 @@ acpi_reg_write_common_regs(int size, uint16_t addr, uint8_t val, void *p)
 		break;
 	case 0x04: case 0x05:
 		/* PMCNTRL - Power Management Control Register (IO) */
-		dev->regs.pmcntrl = ((dev->regs.pmcntrl & ~(0xff << shift16)) | (val << shift16)) & 0x3c07;
-		if (dev->regs.pmcntrl & 0x2000) {
+		if ((addr == 0x05) && (dev->regs.pmcntrl & 0x2000)) {
 			sus_typ = (dev->regs.pmcntrl >> 10) & 7;
 			switch (sus_typ) {
 				case 0:
@@ -479,8 +508,12 @@ acpi_reg_write_common_regs(int size, uint16_t addr, uint8_t val, void *p)
 
 					resetx86();
 					break;
+				default:
+					dev->regs.pmcntrl = ((dev->regs.pmcntrl & ~(0xff << shift16)) | (val << shift16)) & 0x3c07;
+					break;
 			}
-		}
+		} else
+			dev->regs.pmcntrl = ((dev->regs.pmcntrl & ~(0xff << shift16)) | (val << shift16)) & 0x3c07;
 		break;
     }
 }
@@ -493,7 +526,10 @@ acpi_reg_write_intel(int size, uint16_t addr, uint8_t val, void *p)
     int shift16, shift32;
 
     addr &= 0x3f;
-    acpi_log("(%i) ACPI Write (%i) %02X: %02X\n", in_smm, size, addr, val);
+#ifdef ENABLE_ACPI_LOG
+    if (size != 1)
+	acpi_log("(%i) ACPI Write (%i) %02X: %02X\n", in_smm, size, addr, val);
+#endif
     shift16 = (addr & 1) << 3;
     shift32 = (addr & 3) << 3;
 
@@ -876,6 +912,8 @@ acpi_reg_read(uint16_t addr, void *p)
 
     ret = acpi_reg_read_common(1, addr, p);
 
+    acpi_log("ACPI: Read B %02X from %04X\n", ret, addr);
+
     return ret;
 }
 
@@ -924,6 +962,8 @@ acpi_aux_read_read(uint16_t addr, void *p)
 static void
 acpi_reg_writel(uint16_t addr, uint32_t val, void *p)
 {
+    acpi_log("ACPI: Write L %08X to %04X\n", val, addr);
+
     acpi_reg_write_common(4, addr, val & 0xff, p);
     acpi_reg_write_common(4, addr + 1, (val >> 8) & 0xff, p);
     acpi_reg_write_common(4, addr + 2, (val >> 16) & 0xff, p);
@@ -934,6 +974,8 @@ acpi_reg_writel(uint16_t addr, uint32_t val, void *p)
 static void
 acpi_reg_writew(uint16_t addr, uint16_t val, void *p)
 {
+    acpi_log("ACPI: Write W %04X to %04X\n", val, addr);
+
     acpi_reg_write_common(2, addr, val & 0xff, p);
     acpi_reg_write_common(2, addr + 1, (val >> 8) & 0xff, p);
 }
@@ -942,6 +984,8 @@ acpi_reg_writew(uint16_t addr, uint16_t val, void *p)
 static void
 acpi_reg_write(uint16_t addr, uint8_t val, void *p)
 {
+    acpi_log("ACPI: Write B %02X to %04X\n", val, addr);
+
     acpi_reg_write_common(1, addr, val, p);
 }
 
@@ -1188,6 +1232,8 @@ acpi_reset(void *priv)
 	   - Bit  1: 80-conductor cable on primary IDE channel (active low) */
 	dev->regs.gpi_val = !strcmp(machines[machine].internal_name, "wcf681") ? 0xffffffe3 : 0xffffffe5;
     }
+
+    acpi_rtc_status = 0;
 }
 
 
diff --git a/src/chipset/intel_sio.c b/src/chipset/intel_sio.c
index cf587170a..d456c96df 100644
--- a/src/chipset/intel_sio.c
+++ b/src/chipset/intel_sio.c
@@ -105,7 +105,7 @@ sio_timer_read(uint16_t addr, void *priv)
     uint8_t ret = 0xff;
 
     if (!(addr & 0x0002)) {
-	sub_cycles((int)(PITCONST >> 32));
+	cycles -= ((int) (PITCONST >> 32));
 
 	sio_timer_latch = timer_get_remaining_us(&dev->timer);
 
@@ -126,7 +126,7 @@ sio_timer_readw(uint16_t addr, void *priv)
     uint16_t ret = 0xffff;
 
     if (!(addr & 0x0002)) {
-	sub_cycles((int)(PITCONST >> 32));
+	cycles -= ((int) (PITCONST >> 32));
 
 	ret = timer_get_remaining_us(&dev->timer);
     }
diff --git a/src/config.c b/src/config.c
index 85c9c8d65..a2aa1b19d 100644
--- a/src/config.c
+++ b/src/config.c
@@ -113,6 +113,7 @@ static list_t	config_head;
 
 /* TODO: Backwards compatibility, get rid of this when enough time has passed. */
 static int	backwards_compat = 0;
+static int	backwards_compat2 = 0;
 
 
 #ifdef ENABLE_CONFIG_LOG
@@ -873,14 +874,16 @@ load_ports(void)
 }
 
 
-/* Load "Other Peripherals" section. */
+/* Load "Storage Controllers" section. */
 static void
-load_other_peripherals(void)
+load_storage_controllers(void)
 {
-    char *cat = "Other peripherals";
+    char *cat = "Storage controllers";
     char *p;
-    char temp[512];
-    int c, free_p = 0;
+    int free_p = 0;
+
+    /* TODO: Backwards compatibility, get rid of this when enough time has passed. */
+    backwards_compat2 = (find_section(cat) == NULL);
 	
     p = config_get_string(cat, "scsicard", NULL);
     if (p != NULL)
@@ -921,19 +924,6 @@ load_other_peripherals(void)
 
     ide_ter_enabled = !!config_get_int(cat, "ide_ter", 0);
     ide_qua_enabled = !!config_get_int(cat, "ide_qua", 0);
-
-    bugger_enabled = !!config_get_int(cat, "bugger_enabled", 0);
-    postcard_enabled = !!config_get_int(cat, "postcard_enabled", 0);
-
-    for (c = 0; c < ISAMEM_MAX; c++) {
-	sprintf(temp, "isamem%d_type", c);
-
-	p = config_get_string(cat, temp, "none");
-	isamem_type[c] = isamem_get_from_internal_name(p);
-    }
-
-    p = config_get_string(cat, "isartc_type", "none");
-    isartc_type = isartc_get_from_internal_name(p);	
 }
 
 
@@ -1518,7 +1508,6 @@ load_other_removable_devices(void)
 	sprintf(temp, "zip_%02i_iso_path", c+1);
 	config_delete_var(cat, temp);
     }
-    
 
     memset(temp, 0x00, sizeof(temp));
     for (c=0; c<MO_NUM; c++) {
@@ -1584,6 +1573,78 @@ load_other_removable_devices(void)
 }
 
 
+/* Load "Other Peripherals" section. */
+static void
+load_other_peripherals(void)
+{
+    char *cat = "Other peripherals";
+    char *p;
+    char temp[512];
+    int c, free_p = 0;
+
+    if (backwards_compat2) {	
+	p = config_get_string(cat, "scsicard", NULL);
+	if (p != NULL)
+		scsi_card_current = scsi_card_get_from_internal_name(p);
+	else
+		scsi_card_current = 0;
+	config_delete_var(cat, "scsicard");
+
+	p = config_get_string(cat, "fdc", NULL);
+	if (p != NULL)
+		fdc_type = fdc_card_get_from_internal_name(p);
+	else
+		fdc_type = FDC_INTERNAL;
+	config_delete_var(cat, "fdc");
+
+	p = config_get_string(cat, "hdc", NULL);
+	if (p == NULL) {
+		if (machines[machine].flags & MACHINE_HDC) {
+			p = (char *)malloc((strlen("internal")+1)*sizeof(char));
+			strcpy(p, "internal");
+		} else {
+			p = (char *)malloc((strlen("none")+1)*sizeof(char));
+			strcpy(p, "none");
+		}
+		free_p = 1;
+	}
+	if (!strcmp(p, "mfm_xt"))
+		hdc_current = hdc_get_from_internal_name("st506_xt");
+	else if (!strcmp(p, "mfm_xt_dtc5150x"))
+		hdc_current = hdc_get_from_internal_name("st506_xt_dtc5150x");
+	else if (!strcmp(p, "mfm_at"))
+		hdc_current = hdc_get_from_internal_name("st506_at");
+	else
+		hdc_current = hdc_get_from_internal_name(p);
+	config_delete_var(cat, "hdc");
+
+	if (free_p) {
+		free(p);
+		p = NULL;
+	}
+
+	ide_ter_enabled = !!config_get_int(cat, "ide_ter", 0);
+	config_delete_var(cat, "ide_ter");
+	ide_qua_enabled = !!config_get_int(cat, "ide_qua", 0);
+	config_delete_var(cat, "ide_qua");
+    }
+    backwards_compat2 = 0;
+
+    bugger_enabled = !!config_get_int(cat, "bugger_enabled", 0);
+    postcard_enabled = !!config_get_int(cat, "postcard_enabled", 0);
+
+    for (c = 0; c < ISAMEM_MAX; c++) {
+	sprintf(temp, "isamem%d_type", c);
+
+	p = config_get_string(cat, temp, "none");
+	isamem_type[c] = isamem_get_from_internal_name(p);
+    }
+
+    p = config_get_string(cat, "isartc_type", "none");
+    isartc_type = isartc_get_from_internal_name(p);	
+}
+
+
 /* Load the specified or a default configuration file. */
 void
 config_load(void)
@@ -1648,12 +1709,13 @@ config_load(void)
     load_sound();			/* Sound */
     load_network();			/* Network */
     load_ports();			/* Ports (COM & LPT) */
-    load_other_peripherals();		/* Other peripherals */
+    load_storage_controllers();		/* Storage controllers */
     load_hard_disks();			/* Hard disks */
     load_floppy_and_cdrom_drives();	/* Floppy and CD-ROM drives */
     /* TODO: Backwards compatibility, get rid of this when enough time has passed. */
     load_floppy_drives();		/* Floppy drives */
     load_other_removable_devices();	/* Other removable devices */
+    load_other_peripherals();		/* Other peripherals */
 
     /* Mark the configuration as changed. */
     config_changed = 1;
@@ -2080,13 +2142,11 @@ save_ports(void)
 }
 
 
-/* Save "Other Peripherals" section. */
+/* Save "Storage Controllers" section. */
 static void
-save_other_peripherals(void)
+save_storage_controllers(void)
 {
-    char *cat = "Other peripherals";
-    char temp[512];
-    int c;
+    char *cat = "Storage controllers";
 
     if (scsi_card_current == 0)
 	config_delete_var(cat, "scsicard");
@@ -2113,6 +2173,18 @@ save_other_peripherals(void)
       else
 	config_set_int(cat, "ide_qua", ide_qua_enabled);
 
+    delete_section_if_empty(cat);
+}
+
+
+/* Save "Other Peripherals" section. */
+static void
+save_other_peripherals(void)
+{
+    char *cat = "Other peripherals";
+    char temp[512];
+    int c;
+
     if (bugger_enabled == 0)
 	config_delete_var(cat, "bugger_enabled");
       else
@@ -2401,10 +2473,11 @@ config_save(void)
     save_sound();			/* Sound */
     save_network();			/* Network */
     save_ports();			/* Ports (COM & LPT) */
-    save_other_peripherals();		/* Other peripherals */
+    save_storage_controllers();		/* Storage controllers */
     save_hard_disks();			/* Hard disks */
     save_floppy_and_cdrom_drives();	/* Floppy and CD-ROM drives */
     save_other_removable_devices();	/* Other removable devices */
+    save_other_peripherals();		/* Other peripherals */
 
     config_write(cfg_path);
 }
diff --git a/src/cpu/808x.c b/src/cpu/808x.c
index 7e55a1b3a..4932b9494 100644
--- a/src/cpu/808x.c
+++ b/src/cpu/808x.c
@@ -271,6 +271,19 @@ sub_cycles(int c)
 }
 
 
+void
+resub_cycles(int old_cycles)
+{
+    int cyc_diff = 0;
+
+    if (old_cycles > cycles) {
+	cyc_diff = old_cycles - cycles;
+	cycles = old_cycles;
+	resub_cycles(cyc_diff);
+    }
+}
+
+
 #undef readmemb
 #undef readmemw
 #undef readmeml
@@ -280,6 +293,8 @@ sub_cycles(int c)
 static void
 cpu_io(int bits, int out, uint16_t port)
 {
+    int old_cycles = cycles;
+
     if (out) {
 	wait(4, 1);
 	if (bits == 16) {
@@ -305,6 +320,8 @@ cpu_io(int bits, int out, uint16_t port)
 	} else
 		AL = inb(port);
     }
+
+    resub_cycles(old_cycles);
 }
 
 
diff --git a/src/cpu/cpu_table.c b/src/cpu/cpu_table.c
index 585f85831..3381f400c 100644
--- a/src/cpu/cpu_table.c
+++ b/src/cpu/cpu_table.c
@@ -600,9 +600,9 @@ const cpu_family_t cpu_families[] = {
 	.name = "Pentium OverDrive",
 	.internal_name = "pentium_p54c_od3v",
 	.cpus = {
-		{"125",        CPU_PENTIUM,    fpus_internal, 125000000, 3.0, 3520,  0x52c,  0x52c, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC | CPU_FIXED_MULTIPLIER, 12,12,7,7, 16},
+		{"125",        CPU_PENTIUM,    fpus_internal, 125000000, 3.0, 3520,  0x52c,  0x52c, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC | CPU_FIXED_MULTIPLIER, 12,12,7,7, 15},
 		{"150",        CPU_PENTIUM,    fpus_internal, 150000000, 2.5, 3520,  0x52c,  0x52c, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC | CPU_FIXED_MULTIPLIER, 15,15,7,7, 35/2},
-		{"166",        CPU_PENTIUM,    fpus_internal, 166666666, 2.5, 3520,  0x52c,  0x52c, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC | CPU_FIXED_MULTIPLIER, 15,15,7,7, 40},
+		{"166",        CPU_PENTIUM,    fpus_internal, 166666666, 2.5, 3520,  0x52c,  0x52c, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC | CPU_FIXED_MULTIPLIER, 15,15,7,7, 20},
 		{"", 0}
 	}
     }, {
diff --git a/src/device/isartc.c b/src/device/isartc.c
index a14d7e82d..610bf0eae 100644
--- a/src/device/isartc.c
+++ b/src/device/isartc.c
@@ -388,7 +388,7 @@ mm67_read(uint16_t port, void *priv)
     uint8_t ret = 0xff;
 
     /* This chip is directly mapped on I/O. */
-    sub_cycles(ISA_CYCLES(4));
+    cycles -= ISA_CYCLES(4);
 
     switch(reg) {
 	case MM67_ISTAT:		/* IRQ status (RO) */
@@ -424,7 +424,7 @@ mm67_write(uint16_t port, uint8_t val, void *priv)
 #endif
 
     /* This chip is directly mapped on I/O. */
-    sub_cycles(ISA_CYCLES(4));
+    cycles -= ISA_CYCLES(4);
 
     switch(reg) {
 	case MM67_ISTAT:		/* intr status (RO) */
diff --git a/src/device/keyboard_at.c b/src/device/keyboard_at.c
index cd521da8b..1164c7007 100644
--- a/src/device/keyboard_at.c
+++ b/src/device/keyboard_at.c
@@ -2047,7 +2047,7 @@ kbd_read(uint16_t port, void *priv)
     uint8_t ret = 0xff;
 
     if ((dev->flags & KBC_TYPE_MASK) >= KBC_TYPE_PS2_NOREF)
-	sub_cycles(ISA_CYCLES(8));
+	cycles -= ISA_CYCLES(8);
 
     if (((dev->flags & KBC_VEN_MASK) == KBC_VEN_XI8088) && (port == 0x63))
 	port = 0x61;
diff --git a/src/device/mouse_serial.c b/src/device/mouse_serial.c
index 5ae1cd36b..8535e3466 100644
--- a/src/device/mouse_serial.c
+++ b/src/device/mouse_serial.c
@@ -173,7 +173,6 @@ sermouse_callback(struct serial_s *serial, void *priv)
 	dev->format = 7;
     dev->transmit_period = sermouse_transmit_period(dev, 1200, -1);
     timer_stop(&dev->command_timer);
-    sub_cycles(ISA_CYCLES(8));
 #ifdef USE_NEW_DYNAREC
     sermouse_timer_on(dev, 5000.0, 0);
 #else
diff --git a/src/device/serial.c b/src/device/serial.c
index b331d4d99..ed6915438 100644
--- a/src/device/serial.c
+++ b/src/device/serial.c
@@ -355,7 +355,7 @@ serial_write(uint16_t addr, uint8_t val, void *p)
 
     serial_log("UART: Write %02X to port %02X\n", val, addr);
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     switch (addr & 7) {
 	case 0:
@@ -513,7 +513,7 @@ serial_read(uint16_t addr, void *p)
     serial_t *dev = (serial_t *)p;
     uint8_t i, ret = 0;
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     switch (addr & 7) {
 	case 0:
diff --git a/src/disk/hdc_xtide.c b/src/disk/hdc_xtide.c
index 7dc6ec095..50403d9ae 100644
--- a/src/disk/hdc_xtide.c
+++ b/src/disk/hdc_xtide.c
@@ -276,7 +276,7 @@ const device_t xtide_acculogic_device = {
 
 const device_t xtide_at_ps2_device = {
     "PS/2 AT XTIDE (1.1.5)",
-    DEVICE_AT,
+    DEVICE_ISA | DEVICE_AT,
     0,
     xtide_at_ps2_init, xtide_at_close, NULL,
     { xtide_at_ps2_available }, NULL, NULL,
diff --git a/src/floppy/fdc.c b/src/floppy/fdc.c
index 82191881c..a48307b2f 100644
--- a/src/floppy/fdc.c
+++ b/src/floppy/fdc.c
@@ -735,7 +735,7 @@ fdc_write(uint16_t addr, uint8_t val, void *priv)
 
     fdc_log("Write FDC %04X %02X\n", addr, val);
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     switch (addr&7) {
 	case 0:
@@ -1261,7 +1261,7 @@ fdc_read(uint16_t addr, void *priv)
     uint8_t ret;
     int drive;
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     switch (addr & 7) {
 	case 0:		/* STA */
diff --git a/src/floppy/fdd.c b/src/floppy/fdd.c
index 7e1b21452..ece63c55e 100644
--- a/src/floppy/fdd.c
+++ b/src/floppy/fdd.c
@@ -552,7 +552,7 @@ fdd_hole(int drive)
 }
 
 
-uint64_t
+static __inline uint64_t
 fdd_byteperiod(int drive)
 {
     if (!fdd_get_turbo(drive) && drives[drive].byteperiod)
diff --git a/src/game/gameport.c b/src/game/gameport.c
index d1fc35492..883af2388 100644
--- a/src/game/gameport.c
+++ b/src/game/gameport.c
@@ -186,7 +186,7 @@ gameport_write(uint16_t addr, uint8_t val, void *priv)
 	
     p->joystick->write(p->joystick_dat);
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 }
 
 
@@ -198,7 +198,7 @@ gameport_read(uint16_t addr, void *priv)
 
     ret = p->state | p->joystick->read(p->joystick_dat);
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     return(ret);
 }
diff --git a/src/include/86box/86box.h b/src/include/86box/86box.h
index 2ddd11021..627a293c6 100644
--- a/src/include/86box/86box.h
+++ b/src/include/86box/86box.h
@@ -177,6 +177,7 @@ extern uint16_t	get_last_addr(void);
    should be in cpu.c but I put it here to avoid
    having to include cpu.c everywhere. */
 extern void	sub_cycles(int c);
+extern void	resub_cycles(int old_cycles);
 
 extern double	isa_timing;
 extern int	io_delay;
diff --git a/src/include/86box/acpi.h b/src/include/86box/acpi.h
index 0e5e688ef..6cc98cefb 100644
--- a/src/include/86box/acpi.h
+++ b/src/include/86box/acpi.h
@@ -89,6 +89,8 @@ typedef struct
 
 
 /* Global variables. */
+extern int		acpi_rtc_status;
+
 extern const device_t	acpi_intel_device;
 extern const device_t	acpi_smc_device;
 extern const device_t	acpi_via_device;
diff --git a/src/include/86box/resource.h b/src/include/86box/resource.h
index 5b0cbf0c8..24ede722a 100644
--- a/src/include/86box/resource.h
+++ b/src/include/86box/resource.h
@@ -295,8 +295,9 @@
 #define IDM_VID_REMEMBER	40041
 #define IDM_VID_SDL_SW		40050
 #define IDM_VID_SDL_HW		40051
+#define IDM_VID_SDL_OPENGL	40052
 #ifdef USE_VNC
-#define IDM_VID_VNC		40052
+#define IDM_VID_VNC		40053
 #endif
 #define IDM_VID_SCALE_1X	40055
 #define IDM_VID_SCALE_2X	40056
diff --git a/src/include/86box/vid_svga.h b/src/include/86box/vid_svga.h
index 9e4277789..ca7ae4c21 100644
--- a/src/include/86box/vid_svga.h
+++ b/src/include/86box/vid_svga.h
@@ -20,7 +20,8 @@
 
 #define FLAG_EXTRA_BANKS	1
 #define	FLAG_ADDR_BY8		2
-#define FLAG_LATCH8		4
+#define FLAG_EXT_WRITE		4
+#define FLAG_LATCH8		8
 
 
 typedef struct {
diff --git a/src/include/86box/video.h b/src/include/86box/video.h
index 8997d3857..6640827ce 100644
--- a/src/include/86box/video.h
+++ b/src/include/86box/video.h
@@ -214,10 +214,8 @@ extern const device_t gd5401_isa_device;
 extern const device_t gd5402_isa_device;
 extern const device_t gd5402_onboard_device;
 extern const device_t gd5420_isa_device;
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
 extern const device_t gd5422_isa_device;
 extern const device_t gd5424_vlb_device;
-#endif
 extern const device_t gd5426_vlb_device;
 extern const device_t gd5426_onboard_device;
 extern const device_t gd5428_isa_device;
diff --git a/src/include/86box/win.h b/src/include/86box/win.h
index 75c250ddb..4aceae0be 100644
--- a/src/include/86box/win.h
+++ b/src/include/86box/win.h
@@ -83,9 +83,9 @@ DECLARE_HANDLE(DPI_AWARENESS_CONTEXT);
 #define WM_HAS_SHUTDOWN		0x8897
 
 #ifdef USE_VNC
-#define RENDERERS_NUM		3
+#define RENDERERS_NUM		4
 #else
-#define RENDERERS_NUM		2
+#define RENDERERS_NUM		3
 #endif
 
 
diff --git a/src/include/86box/win_sdl.h b/src/include/86box/win_sdl.h
index 97ee42090..4f6bca231 100644
--- a/src/include/86box/win_sdl.h
+++ b/src/include/86box/win_sdl.h
@@ -53,8 +53,10 @@
 extern void	sdl_close(void);
 extern int	sdl_inits(HWND h);
 extern int	sdl_inith(HWND h);
+extern int	sdl_initho(HWND h);
 extern int	sdl_inits_fs(HWND h);
 extern int	sdl_inith_fs(HWND h);
+extern int	sdl_initho_fs(HWND h);
 extern int	sdl_pause(void);
 extern void	sdl_resize(int x, int y);
 extern void	sdl_enable(int enable);
diff --git a/src/io.c b/src/io.c
index daea584c4..0930da37d 100644
--- a/src/io.c
+++ b/src/io.c
@@ -307,7 +307,7 @@ inb(uint16_t port)
 	amstrad_latch = AMSTRAD_SW9;
 
     if (!found)
-	sub_cycles(io_delay);
+	cycles -= io_delay;
 
     io_log("[%04X:%08X] (%i, %i, %04i) in b(%04X) = %02X\n", CS, cpu_state.pc, in_smm, found, qfound, port, ret);
 
@@ -334,7 +334,7 @@ outb(uint16_t port, uint8_t val)
     }
 	
     if (!found) {
-	sub_cycles(io_delay);
+	cycles -= io_delay;
 #ifdef USE_DYNAREC
 	if (cpu_use_dynarec && ((port == 0xeb) || (port == 0xed)))
 		update_tsc();
@@ -392,7 +392,7 @@ inw(uint16_t port)
 	amstrad_latch = AMSTRAD_SW9;
 
     if (!found)
-	sub_cycles(io_delay);
+	cycles -= io_delay;
 
     io_log("[%04X:%08X] (%i, %i, %04i) in w(%04X) = %04X\n", CS, cpu_state.pc, in_smm, found, qfound, port, ret);
 
@@ -433,7 +433,7 @@ outw(uint16_t port, uint16_t val)
     }
 
     if (!found) {
-	sub_cycles(io_delay);
+	cycles -= io_delay;
 #ifdef USE_DYNAREC
 	if (cpu_use_dynarec && ((port == 0xeb) || (port == 0xed)))
 		update_tsc();
@@ -510,7 +510,7 @@ inl(uint16_t port)
 	amstrad_latch = AMSTRAD_SW9;
 
     if (!found)
-	sub_cycles(io_delay);
+	cycles -= io_delay;
 
     io_log("[%04X:%08X] (%i, %i, %04i) in l(%04X) = %08X\n", CS, cpu_state.pc, in_smm, found, qfound, port, ret);
 
@@ -566,7 +566,7 @@ outl(uint16_t port, uint32_t val)
     }
 
     if (!found) {
-	sub_cycles(io_delay);
+	cycles -= io_delay;
 #ifdef USE_DYNAREC
 	if (cpu_use_dynarec && ((port == 0xeb) || (port == 0xed)))
 		update_tsc();
diff --git a/src/machine/m_amstrad.c b/src/machine/m_amstrad.c
index bc22e11f9..87de506ef 100644
--- a/src/machine/m_amstrad.c
+++ b/src/machine/m_amstrad.c
@@ -301,7 +301,7 @@ vid_write_1512(uint32_t addr, uint8_t val, void *priv)
     amsvid_t *vid = (amsvid_t *)priv;
 
     egawrites++;
-    sub_cycles(12);
+    cycles -= 12;
     addr &= 0x3fff;
 
     if ((vid->cgamode & 0x12) == 0x12) {
@@ -320,7 +320,7 @@ vid_read_1512(uint32_t addr, void *priv)
     amsvid_t *vid = (amsvid_t *)priv;
 
     egareads++;
-    sub_cycles(12);
+    cycles -= 12;
     addr &= 0x3fff;
 
     if ((vid->cgamode & 0x12) == 0x12)
diff --git a/src/machine/m_at_t3100e_vid.c b/src/machine/m_at_t3100e_vid.c
index 1f27b7392..945b6b296 100644
--- a/src/machine/m_at_t3100e_vid.c
+++ b/src/machine/m_at_t3100e_vid.c
@@ -204,7 +204,7 @@ void t3100e_write(uint32_t addr, uint8_t val, void *p)
         egawrites++;
 
         t3100e->vram[addr & 0x7fff] = val;
-        sub_cycles(4);
+        cycles -= 4;
 }
 	
 
@@ -213,7 +213,7 @@ uint8_t t3100e_read(uint32_t addr, void *p)
 {
         t3100e_t *t3100e = (t3100e_t *)p;
         egareads++;
-	sub_cycles(4);
+	cycles -= 4;
 
         return t3100e->vram[addr & 0x7fff];
 }
diff --git a/src/machine/m_xt_t1000_vid.c b/src/machine/m_xt_t1000_vid.c
index 15a9f699c..2f291ab63 100644
--- a/src/machine/m_xt_t1000_vid.c
+++ b/src/machine/m_xt_t1000_vid.c
@@ -194,14 +194,14 @@ static void t1000_write(uint32_t addr, uint8_t val, void *p)
         egawrites++;
 
         t1000->vram[addr & 0x3fff] = val;
-        sub_cycles(4);
+        cycles -= 4;
 }
 	
 static uint8_t t1000_read(uint32_t addr, void *p)
 {
         t1000_t *t1000 = (t1000_t *)p;
         egareads++;
-	sub_cycles(4);
+	cycles -= 4;
 
         return t1000->vram[addr & 0x3fff];
 }
diff --git a/src/mem/mem.c b/src/mem/mem.c
index ee92d995c..52fae564d 100644
--- a/src/mem/mem.c
+++ b/src/mem/mem.c
@@ -273,7 +273,7 @@ mem_flush_write_page(uint32_t addr, uint32_t virt)
 #define rammap(x)	((uint32_t *)(_mem_exec[(x) >> MEM_GRANULARITY_BITS]))[((x) >> 2) & MEM_GRANULARITY_QMASK]
 #define rammap64(x)	((uint64_t *)(_mem_exec[(x) >> MEM_GRANULARITY_BITS]))[((x) >> 3) & MEM_GRANULARITY_PMASK]
 
-static uint64_t
+static __inline uint64_t
 mmutranslatereal_normal(uint32_t addr, int rw)
 {
     uint32_t temp,temp2,temp3;
@@ -335,7 +335,7 @@ mmutranslatereal_normal(uint32_t addr, int rw)
 }
 
 
-static uint64_t
+static __inline uint64_t
 mmutranslatereal_pae(uint32_t addr, int rw)
 {
     uint64_t temp,temp2,temp3,temp4;
@@ -428,7 +428,7 @@ mmutranslatereal32(uint32_t addr, int rw)
 }
 
 
-static uint64_t
+static __inline uint64_t
 mmutranslate_noabrt_normal(uint32_t addr, int rw)
 {
     uint32_t temp,temp2,temp3;
@@ -461,7 +461,7 @@ mmutranslate_noabrt_normal(uint32_t addr, int rw)
 }
 
 
-static uint64_t
+static __inline uint64_t
 mmutranslate_noabrt_pae(uint32_t addr, int rw)
 {
     uint64_t temp,temp2,temp3,temp4;
@@ -571,7 +571,7 @@ addreadlookup(uint32_t virt, uint32_t phys)
     readlookup[readlnext++] = virt >> 12;
     readlnext &= (cachesize-1);
 
-    sub_cycles(9);
+    cycles -= 9;
 }
 
 
@@ -624,7 +624,7 @@ addwritelookup(uint32_t virt, uint32_t phys)
     writelookup[writelnext++] = virt >> 12;
     writelnext &= (cachesize - 1);
 
-    sub_cycles(9);
+    cycles -= 9;
 }
 
 
@@ -664,15 +664,19 @@ uint8_t
 read_mem_b(uint32_t addr)
 {
     mem_mapping_t *map;
+    uint8_t ret = 0xff;
+    int old_cycles = cycles;
 
     mem_logical_addr = addr;
     addr &= rammask;
 
     map = read_mapping[addr >> MEM_GRANULARITY_BITS];
     if (map && map->read_b)
-	return map->read_b(addr, map->p);
+	ret = map->read_b(addr, map->p);
 
-    return 0xff;
+    resub_cycles(old_cycles);
+
+    return ret;
 }
 
 
@@ -680,22 +684,26 @@ uint16_t
 read_mem_w(uint32_t addr)
 {
     mem_mapping_t *map;
+    uint16_t ret = 0xffff;
+    int old_cycles = cycles;
 
     mem_logical_addr = addr;
     addr &= rammask;
 
     if (addr & 1)
-	return read_mem_b(addr) | (read_mem_b(addr + 1) << 8);
+	ret = read_mem_b(addr) | (read_mem_b(addr + 1) << 8);
+    else {
+	map = read_mapping[addr >> MEM_GRANULARITY_BITS];
 
-    map = read_mapping[addr >> MEM_GRANULARITY_BITS];
+	if (map && map->read_w)
+		ret = map->read_w(addr, map->p);
+	else if (map && map->read_b)
+		ret = map->read_b(addr, map->p) | (map->read_b(addr + 1, map->p) << 8);
+    }
 
-    if (map && map->read_w)
-	return map->read_w(addr, map->p);
+    resub_cycles(old_cycles);
 
-    if (map && map->read_b)
-	return map->read_b(addr, map->p) | (map->read_b(addr + 1, map->p) << 8);
-
-    return 0xffff;
+    return ret;
 }
 
 
@@ -703,6 +711,7 @@ void
 write_mem_b(uint32_t addr, uint8_t val)
 {
     mem_mapping_t *map;
+    int old_cycles = cycles;
 
     mem_logical_addr = addr;
     addr &= rammask;
@@ -710,6 +719,8 @@ write_mem_b(uint32_t addr, uint8_t val)
     map = write_mapping[addr >> MEM_GRANULARITY_BITS];
     if (map && map->write_b)
 	map->write_b(addr, val, map->p);
+
+    resub_cycles(old_cycles);
 }
 
 
@@ -717,6 +728,7 @@ void
 write_mem_w(uint32_t addr, uint16_t val)
 {
     mem_mapping_t *map;
+    int old_cycles = cycles;
 
     mem_logical_addr = addr;
     addr &= rammask;
@@ -724,18 +736,19 @@ write_mem_w(uint32_t addr, uint16_t val)
     if (addr & 1) {
 	write_mem_b(addr, val);
 	write_mem_b(addr + 1, val >> 8);
-	return;
-    }
-
-    map = write_mapping[addr >> MEM_GRANULARITY_BITS];
-    if (map) {
-	if (map->write_w)
-		map->write_w(addr, val, map->p);
-	else if (map->write_b) {
-		map->write_b(addr, val, map->p);
-		map->write_b(addr + 1, val >> 8, map->p);
+    } else {
+	map = write_mapping[addr >> MEM_GRANULARITY_BITS];
+	if (map) {
+		if (map->write_w)
+			map->write_w(addr, val, map->p);
+		else if (map->write_b) {
+			map->write_b(addr, val, map->p);
+			map->write_b(addr + 1, val >> 8, map->p);
+		}
 	}
     }
+
+    resub_cycles(old_cycles);
 }
 
 
@@ -802,7 +815,7 @@ readmemwl(uint32_t addr)
 
     if (addr64 & 1) {
 	if (!cpu_cyrix_alignment || (addr64 & 7) == 7)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr64 & 0xfff) > 0xffe) {
 		if (cr0 >> 31) {
 			if (mmutranslate_read(addr)   == 0xffffffffffffffffULL)
@@ -846,7 +859,7 @@ writememwl(uint32_t addr, uint16_t val)
 
     if (addr & 1) {
 	if (!cpu_cyrix_alignment || (addr & 7) == 7)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr & 0xFFF) > 0xFFE) {
 		if (cr0 >> 31) {
 			if (mmutranslate_write(addr)   == 0xffffffff)
@@ -899,7 +912,7 @@ readmemll(uint32_t addr)
 
     if (addr & 3) {
 	if (!cpu_cyrix_alignment || (addr & 7) > 4)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr & 0xfff) > 0xffc) {
 		if (cr0>>31) {
 			if (mmutranslate_read(addr)   == 0xffffffffffffffffULL)
@@ -949,7 +962,7 @@ writememll(uint32_t addr, uint32_t val)
 
     if (addr & 3) {
 	if (!cpu_cyrix_alignment || (addr & 7) > 4)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr & 0xFFF) > 0xFFC) {
 		if (cr0>>31) {
 			if (mmutranslate_write(addr)   == 0xffffffffffffffffULL)
@@ -1005,7 +1018,7 @@ readmemql(uint32_t addr)
     mem_logical_addr = addr;
 
     if (addr & 7) {
-	sub_cycles(timing_misaligned);
+	cycles -= timing_misaligned;
 	if ((addr & 0xFFF) > 0xFF8) {
 		if (cr0>>31) {
 			if (mmutranslate_read(addr)   == 0xffffffffffffffffULL)
@@ -1045,7 +1058,7 @@ writememql(uint32_t addr, uint64_t val)
     mem_logical_addr = addr;
 
     if (addr & 7) {
-	sub_cycles(timing_misaligned);
+	cycles -= timing_misaligned;
 	if ((addr & 0xFFF) > 0xFF8) {
 		if (cr0>>31) {
 			if (mmutranslate_write(addr)   == 0xffffffffffffffffULL)
@@ -1122,7 +1135,7 @@ readmemwl(uint32_t seg, uint32_t addr)
 
     if (addr2 & 1) {
 	if (!cpu_cyrix_alignment || (addr2 & 7) == 7)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr2 & 0xfff) > 0xffe) {
 		if (cr0 >> 31) {
 			if (mmutranslate_read(addr2)   == 0xffffffffffffffffULL)
@@ -1174,7 +1187,7 @@ writememwl(uint32_t seg, uint32_t addr, uint16_t val)
 
     if (addr2 & 1) {
 	if (!cpu_cyrix_alignment || (addr2 & 7) == 7)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr2 & 0xFFF) > 0xffe) {
 		if (cr0 >> 31) {
 			if (mmutranslate_write(addr2)   == 0xffffffffffffffffULL) return;
@@ -1234,7 +1247,7 @@ readmemll(uint32_t seg, uint32_t addr)
 
     if (addr2 & 3) {
 	if (!cpu_cyrix_alignment || (addr2 & 7) > 4)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr2 & 0xfff) > 0xffc) {
 		if (cr0 >> 31) {
 			if (mmutranslate_read(addr2)   == 0xffffffffffffffffULL) return 0xffffffff;
@@ -1284,7 +1297,7 @@ writememll(uint32_t seg, uint32_t addr, uint32_t val)
 
     if (addr2 & 3) {
 	if (!cpu_cyrix_alignment || (addr2 & 7) > 4)
-		sub_cycles(timing_misaligned);
+		cycles -= timing_misaligned;
 	if ((addr2 & 0xfff) > 0xffc) {
 		if (cr0 >> 31) {
 			if (mmutranslate_write(addr2)   == 0xffffffffffffffffULL) return;
@@ -1344,7 +1357,7 @@ readmemql(uint32_t seg, uint32_t addr)
     uint32_t addr2 = mem_logical_addr = seg + addr;
 
     if (addr2 & 7) {
-	sub_cycles(timing_misaligned);
+	cycles -= timing_misaligned;
 	if ((addr2 & 0xfff) > 0xff8) {
 		if (cr0 >> 31) {
 			if (mmutranslate_read(addr2)   == 0xffffffffffffffffULL) return 0xffffffffffffffffULL;
@@ -1382,7 +1395,7 @@ writememql(uint32_t seg, uint32_t addr, uint64_t val)
     uint32_t addr2 = mem_logical_addr = seg + addr;
 
     if (addr2 & 7) {
-	sub_cycles(timing_misaligned);
+	cycles -= timing_misaligned;
 	if ((addr2 & 0xfff) > 0xff8) {
 		if (cr0 >> 31) {
 			if (mmutranslate_write(addr2)   == 0xffffffffffffffffULL) return;
diff --git a/src/nvr_at.c b/src/nvr_at.c
index 1b5612593..e7a3f7f4e 100644
--- a/src/nvr_at.c
+++ b/src/nvr_at.c
@@ -638,7 +638,7 @@ nvr_write(uint16_t addr, uint8_t val, void *priv)
     local_t *local = (local_t *)nvr->data;
     uint8_t addr_id = (addr & 0x0e) >> 1;
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     if (local->bank[addr_id] == 0xff)
 	return;
@@ -673,7 +673,7 @@ nvr_read(uint16_t addr, void *priv)
     uint8_t addr_id = (addr & 0x0e) >> 1;
     uint16_t i, checksum = 0x0000;
 
-    sub_cycles(ISA_CYCLES(8));
+    cycles -= ISA_CYCLES(8);
 
     if (/* (addr & 1) && */(local->bank[addr_id] == 0xff))
 	return 0xff;
diff --git a/src/pic.c b/src/pic.c
index ad2efc67e..418f4e03a 100644
--- a/src/pic.c
+++ b/src/pic.c
@@ -31,6 +31,10 @@
 #include <86box/pic.h>
 #include <86box/timer.h>
 #include <86box/pit.h>
+#include <86box/device.h>
+#include <86box/apm.h>
+#include <86box/nvr.h>
+#include <86box/acpi.h>
 
 
 enum
@@ -142,7 +146,7 @@ pic_cascade_mode(pic_t *dev)
 }
 
 
-static uint8_t
+static __inline uint8_t
 pic_slave_on(pic_t *dev, int channel)
 {
     pic_log("pic_slave_on(%i): %i, %02X, %02X\n", channel, pic_cascade_mode(dev), dev->icw4 & 0x0c, dev->icw3 & (1 << channel));
@@ -152,7 +156,7 @@ pic_slave_on(pic_t *dev, int channel)
 }
 
 
-static int
+static __inline int
 find_best_interrupt(pic_t *dev)
 {
     uint8_t b, s;
@@ -184,7 +188,7 @@ find_best_interrupt(pic_t *dev)
 }
 
 
-void
+static __inline void
 pic_update_pending(void)
 {
     int is_at = IS_ARCH(machine, (MACHINE_BUS_ISA16 | MACHINE_BUS_MCA | MACHINE_BUS_PCMCIA));
@@ -321,7 +325,7 @@ pic_action(pic_t *dev, uint8_t irq, uint8_t eoi, uint8_t rotate)
 
 
 /* Automatic non-specific EOI. */
-static void
+static __inline void
 pic_auto_non_specific_eoi(pic_t *dev)
 {
     uint8_t irq;
@@ -520,6 +524,9 @@ picint_common(uint16_t num, int level, int set)
 	return;
     }
 
+    if (num & 0x0100)
+	acpi_rtc_status = !!set;
+
     if (set) {
 	if (num & 0xff00) {
 		if (level)
@@ -650,8 +657,13 @@ picinterrupt()
 		pit_ctr_set_gate(&pit2->counters[0], 0);
 
 	/* Two ACK's - do them in a loop to avoid potential compiler misoptimizations. */
-	for (i = 0; i < 2; i++)
-		ret = pic_irq_ack();
+	for (i = 0; i < 2; i++) {
+		ret = pic_irq_ack_read(&pic, pic.ack_bytes);
+		pic.ack_bytes = (pic.ack_bytes + 1) % (pic_i86_mode(&pic) ? 2 : 3);
+
+		if (pic.ack_bytes == 0)
+			pic_update_pending();
+	}
     }
 
     return ret;
diff --git a/src/pit.c b/src/pit.c
index 03fd80c93..2e10a7b38 100644
--- a/src/pit.c
+++ b/src/pit.c
@@ -358,7 +358,7 @@ ctr_load(ctr_t *ctr)
 }
 
 
-static void
+static __inline void
 ctr_latch_status(ctr_t *ctr)
 {
     ctr->read_status = (ctr->ctrl & 0x3f) | (ctr->out ? 0x80 : 0) | (ctr->null_count ? 0x40 : 0);
@@ -366,7 +366,7 @@ ctr_latch_status(ctr_t *ctr)
 }
 
 
-static void
+static __inline void
 ctr_latch_count(ctr_t *ctr)
 {
     int count = (ctr->latch || (ctr->state == 1)) ? ctr->l : ctr->count;
@@ -459,8 +459,8 @@ pit_ctr_set_gate(ctr_t *ctr, int gate)
 }
 
 
-void
-pit_ctr_set_clock(ctr_t *ctr, int clock)
+static __inline void
+pit_ctr_set_clock_common(ctr_t *ctr, int clock)
 {
     int old = ctr->clock;
 
@@ -491,6 +491,13 @@ pit_ctr_set_clock(ctr_t *ctr, int clock)
 }
 
 
+void
+pit_ctr_set_clock(ctr_t *ctr, int clock)
+{
+    pit_ctr_set_clock_common(ctr, clock);
+}
+
+
 void
 pit_ctr_set_using_timer(ctr_t *ctr, int using_timer)
 {
@@ -509,7 +516,7 @@ pit_timer_over(void *p)
     dev->clock ^= 1;
 
     for (i = 0; i < 3; i++)
-	pit_ctr_set_clock(&dev->counters[i], dev->clock);
+	pit_ctr_set_clock_common(&dev->counters[i], dev->clock);
 
     timer_advance_u64(&dev->callback_timer, PITCONST >> 1ULL);
 }
diff --git a/src/sound/snd_opl.c b/src/sound/snd_opl.c
index 0547c19c2..f8326b053 100644
--- a/src/sound/snd_opl.c
+++ b/src/sound/snd_opl.c
@@ -228,7 +228,7 @@ opl2_read(uint16_t port, void *priv)
     opl_t *dev = (opl_t *)priv;
 
     if (dev->flags & FLAG_CYCLES)
-	sub_cycles((int) (isa_timing * 8));
+	cycles -= ((int) (isa_timing * 8));
 
     opl2_update(dev);
 
@@ -277,7 +277,7 @@ opl3_read(uint16_t port, void *priv)
     opl_t *dev = (opl_t *)priv;
 
     if (dev->flags & FLAG_CYCLES)
-	sub_cycles((int)(isa_timing * 8));
+	cycles -= ((int)(isa_timing * 8));
 
     opl3_update(dev);
 
diff --git a/src/sound/snd_sb.c b/src/sound/snd_sb.c
index 1565bb8f2..353d25987 100644
--- a/src/sound/snd_sb.c
+++ b/src/sound/snd_sb.c
@@ -1207,7 +1207,7 @@ sb_pro_v1_opl_read(uint16_t port, void *priv)
 {
     sb_t *sb = (sb_t *)priv;
 
-    sub_cycles((int)(isa_timing * 8));
+    cycles -= ((int) (isa_timing * 8));
 
     (void)opl2_read(port, &sb->opl2);	// read, but ignore
     return(opl2_read(port, &sb->opl));
diff --git a/src/usb.c b/src/usb.c
index fec6797ae..9692068b2 100644
--- a/src/usb.c
+++ b/src/usb.c
@@ -102,8 +102,10 @@ uhci_reg_writew(uint16_t addr, uint16_t val, void *p)
 
     switch (addr) {
 	case 0x00:
-		if ((regs[0x00] & 0x0001) && !(val & 0x0001))
-			regs[0x02] |= 0x20;
+		if ((val & 0x0001) && !(regs[0x00] & 0x0001))
+			regs[0x01] &= ~0x20;
+		else if (!(val & 0x0001))
+			regs[0x01] |= 0x20;
 		regs[0x00] = (val & 0x00ff);
 		break;
 	case 0x06:
diff --git a/src/video/vid_cga.c b/src/video/vid_cga.c
index e7a5ffd61..60a662490 100644
--- a/src/video/vid_cga.c
+++ b/src/video/vid_cga.c
@@ -123,7 +123,7 @@ cga_waitstates(void *p)
     int ws;
 
     ws = ws_array[cycles & 0xf];
-    sub_cycles(ws);
+    cycles -= ws;
 }
 
 
diff --git a/src/video/vid_cl54xx.c b/src/video/vid_cl54xx.c
index bb808e1dc..ca76ad23b 100644
--- a/src/video/vid_cl54xx.c
+++ b/src/video/vid_cl54xx.c
@@ -42,11 +42,7 @@
 #define BIOS_GD5402_PATH		L"roms/video/cirruslogic/avga2.rom"
 #define BIOS_GD5402_ONBOARD_PATH	L"roms/machines/cbm_sl386sx25/Commodore386SX-25_AVGA2.bin"
 #define BIOS_GD5420_PATH		L"roms/video/cirruslogic/5420.vbi"
-
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
 #define BIOS_GD5422_PATH		L"roms/video/cirruslogic/cl5422.bin"
-#endif
-
 #define BIOS_GD5426_PATH		L"roms/video/cirruslogic/Diamond SpeedStar PRO VLB v3.04.bin"
 #define BIOS_GD5428_ISA_PATH		L"roms/video/cirruslogic/5428.bin"
 #define BIOS_GD5428_PATH		L"roms/video/cirruslogic/vlbusjapan.BIN"
@@ -535,6 +531,8 @@ gd54xx_out(uint16_t addr, uint8_t val, void *p)
 						svga->adv_flags = FLAG_EXTRA_BANKS;
 					if (svga->gdcreg[0xb] & 0x02)
 						svga->adv_flags |= FLAG_ADDR_BY8;
+					if (svga->gdcreg[0xb] & 0x04)
+						svga->adv_flags |= FLAG_EXT_WRITE;
 					if (svga->gdcreg[0xb] & 0x08)
 						svga->adv_flags |= FLAG_LATCH8;
 					gd54xx_recalc_banking(gd54xx);
@@ -1447,6 +1445,7 @@ gd54xx_write_modes45(svga_t *svga, uint8_t val, uint32_t addr)
 	case 4:
 		if (svga->gdcreg[0xb] & 0x10) {
 			addr <<= 2;
+			addr &= svga->decode_mask;
 
 			for (i = 0; i < 8; i++) {
 				if (val & svga->seqregs[2] & (0x80 >> i)) {
@@ -1456,6 +1455,7 @@ gd54xx_write_modes45(svga_t *svga, uint8_t val, uint32_t addr)
 			}
 		} else {
 			addr <<= 1;
+			addr &= svga->decode_mask;
 
 			for (i = 0; i < 8; i++) {
 				if (val & svga->seqregs[2] & (0x80 >> i))
@@ -1467,6 +1467,7 @@ gd54xx_write_modes45(svga_t *svga, uint8_t val, uint32_t addr)
 	case 5:
 		if (svga->gdcreg[0xb] & 0x10) {
 			addr <<= 2;
+			addr &= svga->decode_mask;
 
 			for (i = 0; i < 8; i++) {
 				j = (0x80 >> i);
@@ -1479,6 +1480,7 @@ gd54xx_write_modes45(svga_t *svga, uint8_t val, uint32_t addr)
 			}
 		} else {
 			addr <<= 1;
+			addr &= svga->decode_mask;
 
 			for (i = 0; i < 8; i++) {
 				j = (0x80 >> i);
@@ -1602,7 +1604,7 @@ gd54xx_readw_linear(uint32_t addr, void *p)
 		temp |= (svga_readb_linear(addr, svga) << 8);
 
 		if (svga->fast)
-		        sub_cycles(video_timing_read_w);
+		        cycles -= video_timing_read_w;
 
 		return temp;
 	case 3:
@@ -1653,7 +1655,7 @@ gd54xx_readl_linear(uint32_t addr, void *p)
 		temp |= (svga_readb_linear(addr + 2, svga) << 24);
 
 		if (svga->fast)
-		        sub_cycles(video_timing_read_l);
+		        cycles -= video_timing_read_l;
 
 		return temp;
 	case 2:
@@ -1663,7 +1665,7 @@ gd54xx_readl_linear(uint32_t addr, void *p)
 		temp |= (svga_readb_linear(addr, svga) << 24);
 
 		if (svga->fast)
-		        sub_cycles(video_timing_read_l);
+		        cycles -= video_timing_read_l;
 
 		return temp;
 	case 3:
@@ -1852,7 +1854,7 @@ gd54xx_writew_linear(uint32_t addr, uint16_t val, void *p)
 			svga_writeb_linear(addr, val >> 8, svga);
 
 			if (svga->fast)
-		        	sub_cycles(video_timing_write_w);
+				cycles -= video_timing_write_w;
 		case 3:
 			return;
 	}
@@ -3072,12 +3074,10 @@ static void
 		romfn = BIOS_GD5420_PATH;
 		break;
 
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
 	case CIRRUS_ID_CLGD5422:
 	case CIRRUS_ID_CLGD5424:
 		romfn = BIOS_GD5422_PATH;
 		break;		
-#endif
 
 	case CIRRUS_ID_CLGD5426:
 		if (info->local & 0x200)
@@ -3295,13 +3295,11 @@ gd5420_available(void)
     return rom_present(BIOS_GD5420_PATH);
 }
 
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
 static int
 gd5422_available(void)
 {
     return rom_present(BIOS_GD5422_PATH);
 }
-#endif
 
 static int
 gd5426_available(void)
@@ -3589,7 +3587,6 @@ const device_t gd5420_isa_device =
     gd5422_config,
 };
 
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
 const device_t gd5422_isa_device = {
     "Cirrus Logic GD-5422",
     DEVICE_AT | DEVICE_ISA,
@@ -3613,7 +3610,6 @@ const device_t gd5424_vlb_device = {
     gd54xx_force_redraw,
     gd5422_config,
 };
-#endif
 
 const device_t gd5426_vlb_device =
 {
diff --git a/src/video/vid_colorplus.c b/src/video/vid_colorplus.c
index 371d23e5e..223c4d1c6 100644
--- a/src/video/vid_colorplus.c
+++ b/src/video/vid_colorplus.c
@@ -100,7 +100,7 @@ void colorplus_write(uint32_t addr, uint8_t val, void *p)
 				colorplus->cga.charbuffer[offset | 1] = colorplus->cga.vram[addr & 0x7fff];
         }
         egawrites++;
-        sub_cycles(4);
+        cycles -= 4;
 }
 
 uint8_t colorplus_read(uint32_t addr, void *p)
@@ -117,7 +117,7 @@ uint8_t colorplus_read(uint32_t addr, void *p)
 	{
 		addr &= 0x3FFF;
 	}
-        sub_cycles(4);        
+        cycles -= 4;        
         if (colorplus->cga.snow_enabled)
         {
 				int offset = ((timer_get_remaining_u64(&colorplus->cga.timer) / CGACONST) * 2) & 0xfc;
diff --git a/src/video/vid_ega.c b/src/video/vid_ega.c
index e47aed584..bbd8e7da9 100644
--- a/src/video/vid_ega.c
+++ b/src/video/vid_ega.c
@@ -732,7 +732,7 @@ ega_write(uint32_t addr, uint8_t val, void *p)
     int writemask2 = ega->writemask;
 
     egawrites++;
-    sub_cycles(video_timing_write_b);
+    cycles -= video_timing_write_b;
 
     if (addr >= 0xB0000)	addr &= 0x7fff;
     else			addr &= 0xffff;
@@ -859,7 +859,7 @@ ega_read(uint32_t addr, void *p)
     int readplane = ega->readplane;
 
     egareads++;
-    sub_cycles(video_timing_read_b);
+    cycles -= video_timing_read_b;
     if (addr >= 0xb0000)	addr &= 0x7fff;
     else			addr &= 0xffff;
 
diff --git a/src/video/vid_genius.c b/src/video/vid_genius.c
index 6fbe02b94..d6afc1863 100644
--- a/src/video/vid_genius.c
+++ b/src/video/vid_genius.c
@@ -253,7 +253,7 @@ genius_waitstates(void)
     int ws;
 
     ws = ws_array[cycles & 0xf];
-    sub_cycles(ws);
+    cycles -= ws;
 }
 
 
diff --git a/src/video/vid_hercules.c b/src/video/vid_hercules.c
index 760774fe3..94faf3fea 100644
--- a/src/video/vid_hercules.c
+++ b/src/video/vid_hercules.c
@@ -212,7 +212,7 @@ hercules_waitstates(void *p)
     int ws;
 
     ws = ws_array[cycles & 0xf];
-    sub_cycles(ws);
+    cycles -= ws;
 }
 
 
diff --git a/src/video/vid_ht216.c b/src/video/vid_ht216.c
index 79cfb6bdf..083de1f26 100644
--- a/src/video/vid_ht216.c
+++ b/src/video/vid_ht216.c
@@ -729,7 +729,7 @@ ht216_write_common(ht216_t *ht216, uint32_t addr, uint8_t val)
     svga_t *svga = &ht216->svga;
     uint8_t bit_mask = 0, rop_select = 0;
 
-    sub_cycles(video_timing_write_b);
+    cycles -= video_timing_write_b;
 
     egawrites++;
 
@@ -911,7 +911,7 @@ ht216_read_common(ht216_t *ht216, uint32_t addr)
 
     addr &= 0xfffff;
 
-    sub_cycles(video_timing_read_b);
+    cycles -= video_timing_read_b;
 
     egareads++;
 
diff --git a/src/video/vid_mga.c b/src/video/vid_mga.c
index 136531457..3433b7539 100644
--- a/src/video/vid_mga.c
+++ b/src/video/vid_mga.c
@@ -2132,7 +2132,7 @@ mystique_readb_linear(uint32_t addr, void *p)
 
         egareads++;
 
-        sub_cycles(video_timing_read_b);
+        cycles -= video_timing_read_b;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -2149,7 +2149,7 @@ mystique_readw_linear(uint32_t addr, void *p)
 
         egareads += 2;
 
-        sub_cycles(video_timing_read_w);
+        cycles -= video_timing_read_w;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -2166,7 +2166,7 @@ mystique_readl_linear(uint32_t addr, void *p)
 
         egareads += 4;
 
-        sub_cycles(video_timing_read_l);
+        cycles -= video_timing_read_l;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -2183,7 +2183,7 @@ mystique_writeb_linear(uint32_t addr, uint8_t val, void *p)
 
         egawrites++;
 
-        sub_cycles(video_timing_write_b);
+        cycles -= video_timing_write_b;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -2201,7 +2201,7 @@ mystique_writew_linear(uint32_t addr, uint16_t val, void *p)
 
         egawrites += 2;
 
-        sub_cycles(video_timing_write_w);
+        cycles -= video_timing_write_w;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -2219,7 +2219,7 @@ mystique_writel_linear(uint32_t addr, uint32_t val, void *p)
 
         egawrites += 4;
 
-        sub_cycles(video_timing_write_l);
+        cycles -= video_timing_write_l;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
diff --git a/src/video/vid_sigma.c b/src/video/vid_sigma.c
index e8862a68b..78812d27d 100644
--- a/src/video/vid_sigma.c
+++ b/src/video/vid_sigma.c
@@ -344,7 +344,7 @@ sigma_write(uint32_t addr, uint8_t val, void *p)
 
     sigma->vram[sigma->plane * 0x8000 + (addr & 0x7fff)] = val;
     egawrites++;
-    sub_cycles(4);
+    cycles -= 4;
 }
 
 
@@ -353,7 +353,7 @@ sigma_read(uint32_t addr, void *p)
 {
     sigma_t *sigma = (sigma_t *)p;
 
-    sub_cycles(4);
+    cycles -= 4;
     egareads++;
     return sigma->vram[sigma->plane * 0x8000 + (addr & 0x7fff)];
 }
diff --git a/src/video/vid_svga.c b/src/video/vid_svga.c
index 7a8fc0981..df81a9cb2 100644
--- a/src/video/vid_svga.c
+++ b/src/video/vid_svga.c
@@ -1015,7 +1015,7 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
 
     egawrites++;
 
-    sub_cycles(video_timing_write_b);
+    cycles -= video_timing_write_b;
 
     if (!linear) {
 	addr = svga_decode_addr(svga, addr, 1);
@@ -1057,14 +1057,23 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
     if (svga->adv_flags & FLAG_LATCH8)
 	count = 8;
 
+    /* Undocumented Cirrus Logic behavior: The datasheet says that, with EXT_WRITE and FLAG_ADDR_BY8, the write mask only
+       changes meaning in write modes 4 and 5, as well as write mode 1. In reality, however, all other write modes are also
+       affected, as proven by the Windows 3.1 CL-GD 5422/4 drivers in 8bpp modes. */
     switch (svga->writemode) {
 	case 0:
-		if (svga->gdcreg[3] & 7) 
-			val = svga_rotate[svga->gdcreg[3] & 7][val];
+		val = ((val >> (svga->gdcreg[3] & 7)) | (val << (8 - (svga->gdcreg[3] & 7))));
+		// if (svga->gdcreg[3] & 7) 
+			// val = svga_rotate[svga->gdcreg[3] & 7][val];
 		if ((svga->gdcreg[8] == 0xff) && !(svga->gdcreg[3] & 0x18) && (!svga->gdcreg[1] || svga->set_reset_disabled)) {
 			for (i = 0; i < count; i++) {
-				if (writemask2 & (1 << i))
-					svga->vram[addr | i] = val;
+				if ((svga->adv_flags & FLAG_EXT_WRITE) && (svga->adv_flags & FLAG_ADDR_BY8)) {
+					if (writemask2 & (0x80 >> i))
+						svga->vram[addr | i] = val;
+				} else {
+					if (writemask2 & (1 << i))
+						svga->vram[addr | i] = val;
+				}
 			}
 			return;
 		} else {
@@ -1078,8 +1087,13 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
 		break;
 	case 1:
 		for (i = 0; i < count; i++) {
-			if (writemask2 & (1 << i))
-				svga->vram[addr | i] = svga->latch.b[i];
+			if ((svga->adv_flags & FLAG_EXT_WRITE) && (svga->adv_flags & FLAG_ADDR_BY8)) {
+				if (writemask2 & (0x80 >> i))
+					svga->vram[addr | i] = svga->latch.b[i];
+			} else {
+				if (writemask2 & (1 << i))
+					svga->vram[addr | i] = svga->latch.b[i];
+			}
 		}
 		return;
 	case 2:
@@ -1095,8 +1109,9 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
 		}
 		break;
 	case 3:
-                if (svga->gdcreg[3] & 7) 
-                        val = svga_rotate[svga->gdcreg[3] & 7][val];
+		val = ((val >> (svga->gdcreg[3] & 7)) | (val << (8 - (svga->gdcreg[3] & 7))));
+                // if (svga->gdcreg[3] & 7) 
+                        // val = svga_rotate[svga->gdcreg[3] & 7][val];
                 wm = svga->gdcreg[8];
                 svga->gdcreg[8] &= val;
 
@@ -1114,26 +1129,46 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
     switch (svga->gdcreg[3] & 0x18) {
 	case 0x00:	/* Set */
 		for (i = 0; i < count; i++) {
-			if (writemask2 & (1 << i))
-				svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) | (svga->latch.b[i] & ~svga->gdcreg[8]);
+			if ((svga->adv_flags & FLAG_EXT_WRITE) && (svga->adv_flags & FLAG_ADDR_BY8)) {
+				if (writemask2 & (0x80 >> i))
+					svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) | (svga->latch.b[i] & ~svga->gdcreg[8]);
+			} else {
+				if (writemask2 & (1 << i))
+					svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) | (svga->latch.b[i] & ~svga->gdcreg[8]);
+			}
 		}
 		break;
 	case 0x08:	/* AND */
 		for (i = 0; i < count; i++) {
-			if (writemask2 & (1 << i))
-				svga->vram[addr | i] = (vall.b[i] | ~svga->gdcreg[8]) & svga->latch.b[i];
+			if ((svga->adv_flags & FLAG_EXT_WRITE) && (svga->adv_flags & FLAG_ADDR_BY8)) {
+				if (writemask2 & (0x80 >> i))
+					svga->vram[addr | i] = (vall.b[i] | ~svga->gdcreg[8]) & svga->latch.b[i];
+			} else {
+				if (writemask2 & (1 << i))
+					svga->vram[addr | i] = (vall.b[i] | ~svga->gdcreg[8]) & svga->latch.b[i];
+			}
 		}
 		break;
 	case 0x10:	/* OR */
 		for (i = 0; i < count; i++) {
-			if (writemask2 & (1 << i))
-				svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) | svga->latch.b[i];
+			if ((svga->adv_flags & FLAG_EXT_WRITE) && (svga->adv_flags & FLAG_ADDR_BY8)) {
+				if (writemask2 & (0x80 >> i))
+					svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) | svga->latch.b[i];
+			} else {
+				if (writemask2 & (1 << i))
+					svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) | svga->latch.b[i];
+			}
 		}
 		break;
 	case 0x18:	/* XOR */
 		for (i = 0; i < count; i++) {
-			if (writemask2 & (1 << i))
-				svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) ^ svga->latch.b[i];
+			if ((svga->adv_flags & FLAG_EXT_WRITE) && (svga->adv_flags & FLAG_ADDR_BY8)) {
+				if (writemask2 & (0x80 >> i))
+					svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) ^ svga->latch.b[i];
+			} else {
+				if (writemask2 & (1 << i))
+					svga->vram[addr | i] = (vall.b[i] & svga->gdcreg[8]) ^ svga->latch.b[i];
+			}
 		}
 		break;
     }
@@ -1156,7 +1191,7 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
     if (svga->adv_flags & FLAG_ADDR_BY8)
 	readplane = svga->gdcreg[4] & 7;
 
-    sub_cycles(video_timing_read_b);
+    cycles -= video_timing_read_b;
 
     egareads++;
 
@@ -1182,6 +1217,9 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
 		addr = svga->translate_address(addr, p);
 	if (addr >= svga->vram_max)
 		return 0xff;
+	latch_addr = (addr & svga->vram_mask) & ~3;
+	for (i = 0; i < count; i++)
+		svga->latch.b[i] = svga->vram[latch_addr | i];
 	return svga->vram[addr & svga->vram_mask];
     } else if (svga->chain2_read) {
 	readplane = (readplane & 2) | (addr & 1);
@@ -1381,7 +1419,7 @@ svga_writew_common(uint32_t addr, uint16_t val, uint8_t linear, void *p)
 
     egawrites += 2;
 
-    sub_cycles(video_timing_write_w);
+    cycles -= video_timing_write_w;
 
     if (!linear) {
 	addr = svga_decode_addr(svga, addr, 1);
@@ -1442,7 +1480,7 @@ svga_writel_common(uint32_t addr, uint32_t val, uint8_t linear, void *p)
 
     egawrites += 4;
 
-    sub_cycles(video_timing_write_l);
+    cycles -= video_timing_write_l;
 
     if (!linear) {
 	addr = svga_decode_addr(svga, addr, 1);
@@ -1526,7 +1564,7 @@ svga_readw_common(uint32_t addr, uint8_t linear, void *p)
 
     egareads += 2;
 
-    sub_cycles(video_timing_read_w);
+    cycles -= video_timing_read_w;
 
     if (!linear) {
 	addr = svga_decode_addr(svga, addr, 0);
@@ -1579,7 +1617,7 @@ svga_readl_common(uint32_t addr, uint8_t linear, void *p)
 
     egareads += 4;
 
-    sub_cycles(video_timing_read_l);
+    cycles -= video_timing_read_l;
 
     if (!linear) {
 	addr = svga_decode_addr(svga, addr, 0);
diff --git a/src/video/vid_table.c b/src/video/vid_table.c
index 611be2524..e8605e16d 100644
--- a/src/video/vid_table.c
+++ b/src/video/vid_table.c
@@ -71,9 +71,7 @@ video_cards[] = {
     { "cl_gd5401_isa",		&gd5401_isa_device			},
     { "cl_gd5402_isa",		&gd5402_isa_device			},
     { "cl_gd5420_isa",		&gd5420_isa_device			},
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
     { "cl_gd5422_isa",		&gd5422_isa_device			},
-#endif
     { "cl_gd5428_isa",		&gd5428_isa_device			},
     { "cl_gd5429_isa",		&gd5429_isa_device			},
     { "cl_gd5434_isa",		&gd5434_isa_device			},
@@ -154,9 +152,7 @@ video_cards[] = {
     { "voodoo3_3k_pci",		&voodoo_3_3000_device 			},
     { "mach64gx_vlb",		&mach64gx_vlb_device			},
     { "et4000w32p_vlb",		&et4000w32p_cardex_vlb_device		},
-#if defined(DEV_BRANCH) && defined(USE_CL5422)
     { "cl_gd5424_vlb",		&gd5424_vlb_device			},
-#endif
     { "cl_gd5428_vlb",		&gd5428_vlb_device			},
     { "cl_gd5429_vlb",		&gd5429_vlb_device			},
     { "cl_gd5434_vlb",		&gd5434_vlb_device			},
diff --git a/src/video/vid_tgui9440.c b/src/video/vid_tgui9440.c
index d5f28fefc..17415c10a 100644
--- a/src/video/vid_tgui9440.c
+++ b/src/video/vid_tgui9440.c
@@ -798,7 +798,7 @@ static uint8_t tgui_ext_linear_read(uint32_t addr, void *p)
         tgui_t *tgui = (tgui_t *)svga->p;
         int c;
 
-        sub_cycles(video_timing_read_b);
+        cycles -= video_timing_read_b;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -829,7 +829,7 @@ static void tgui_ext_linear_write(uint32_t addr, uint8_t val, void *p)
         uint8_t bg[2] = {tgui->ext_gdc_regs[1], tgui->ext_gdc_regs[2]};
         uint8_t mask = tgui->ext_gdc_regs[7];
 
-        sub_cycles(video_timing_write_b);
+        cycles -= video_timing_write_b;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
@@ -898,7 +898,7 @@ static void tgui_ext_linear_writew(uint32_t addr, uint16_t val, void *p)
         uint8_t bg[2] = {tgui->ext_gdc_regs[1], tgui->ext_gdc_regs[2]};
         uint16_t mask = (tgui->ext_gdc_regs[7] << 8) | tgui->ext_gdc_regs[8];
         
-        sub_cycles(video_timing_write_w);
+        cycles -= video_timing_write_w;
 
         addr &= svga->decode_mask;
         if (addr >= svga->vram_max)
diff --git a/src/video/vid_voodoo.c b/src/video/vid_voodoo.c
index e8d2ad7d3..c806bf243 100644
--- a/src/video/vid_voodoo.c
+++ b/src/video/vid_voodoo.c
@@ -151,7 +151,7 @@ static uint16_t voodoo_readw(uint32_t addr, void *p)
         
         addr &= 0xffffff;
 
-        sub_cycles(voodoo->read_time);
+        cycles -= voodoo->read_time;
         
         if ((addr & 0xc00000) == 0x400000) /*Framebuffer*/
         {
@@ -190,7 +190,7 @@ static uint32_t voodoo_readl(uint32_t addr, void *p)
         voodoo->rd_count++;
         addr &= 0xffffff;
         
-        sub_cycles(voodoo->read_time);
+        cycles -= voodoo->read_time;
 
         if (addr & 0x800000) /*Texture*/
         {
@@ -403,7 +403,7 @@ static void voodoo_writew(uint32_t addr, uint16_t val, void *p)
         voodoo->wr_count++;
         addr &= 0xffffff;
         
-        sub_cycles(voodoo->write_time);
+        cycles -= voodoo->write_time;
 
         if ((addr & 0xc00000) == 0x400000) /*Framebuffer*/
                 voodoo_queue_command(voodoo, addr | FIFO_WRITEW_FB, val);
@@ -418,9 +418,9 @@ static void voodoo_writel(uint32_t addr, uint32_t val, void *p)
         addr &= 0xffffff;
         
         if (addr == voodoo->last_write_addr+4)
-                sub_cycles(voodoo->burst_time);
+                cycles -= voodoo->burst_time;
         else
-                sub_cycles(voodoo->write_time);
+                cycles -= voodoo->write_time;
         voodoo->last_write_addr = addr;
 
         if (addr & 0x800000) /*Texture*/
diff --git a/src/video/vid_voodoo_banshee.c b/src/video/vid_voodoo_banshee.c
index 932fb9ee6..0b664d91e 100644
--- a/src/video/vid_voodoo_banshee.c
+++ b/src/video/vid_voodoo_banshee.c
@@ -835,7 +835,7 @@ static uint32_t banshee_ext_inl(uint16_t addr, void *p)
         svga_t *svga = &banshee->svga;
         uint32_t ret = 0xffffffff;
 
-        sub_cycles(voodoo->read_time);
+        cycles -= voodoo->read_time;
         
         switch (addr & 0xff)
         {
@@ -1026,7 +1026,7 @@ static uint32_t banshee_reg_readl(uint32_t addr, void *p)
         voodoo_t *voodoo = banshee->voodoo;
         uint32_t ret = 0xffffffff;
         
-        sub_cycles(voodoo->read_time);
+        cycles -= voodoo->read_time;
 
         switch (addr & 0x1f00000)
         {
@@ -1178,7 +1178,7 @@ static void banshee_reg_writew(uint32_t addr, uint16_t val, void *p)
         banshee_t *banshee = (banshee_t *)p;
         voodoo_t *voodoo = banshee->voodoo;
 
-        sub_cycles(voodoo->write_time);
+        cycles -= voodoo->write_time;
         
 //        banshee_log("banshee_reg_writew: addr=%08x val=%04x\n", addr, val);
         switch (addr & 0x1f00000)
@@ -1253,9 +1253,9 @@ static void banshee_reg_writel(uint32_t addr, uint32_t val, void *p)
         voodoo_t *voodoo = banshee->voodoo;
         
         if (addr == voodoo->last_write_addr+4)
-                sub_cycles(voodoo->burst_time);
+                cycles -= voodoo->burst_time;
         else
-                sub_cycles(voodoo->write_time);
+                cycles -= voodoo->write_time;
         voodoo->last_write_addr = addr;
 
 //        banshee_log("banshee_reg_writel: addr=%08x val=%08x\n", addr, val);
@@ -1351,7 +1351,7 @@ static uint8_t banshee_read_linear(uint32_t addr, void *p)
         voodoo_t *voodoo = banshee->voodoo;
         svga_t *svga = &banshee->svga;
         
-	sub_cycles(voodoo->read_time);
+	cycles -= voodoo->read_time;
 
         addr &= svga->decode_mask;
         if (addr >= voodoo->tile_base)
@@ -1369,7 +1369,7 @@ static uint8_t banshee_read_linear(uint32_t addr, void *p)
                 return 0xff;
 
         egareads++;
-        sub_cycles(video_timing_read_b);
+        cycles -= video_timing_read_b;
         
 //        banshee_log("read_linear: addr=%08x val=%02x\n", addr, svga->vram[addr & svga->vram_mask]);
 
@@ -1385,8 +1385,7 @@ static uint16_t banshee_read_linear_w(uint32_t addr, void *p)
         if (addr & 1)
                 return banshee_read_linear(addr, p) | (banshee_read_linear(addr+1, p) << 8);	
 	
-        sub_cycles(voodoo->read_time);
-
+        cycles -= voodoo->read_time;
         addr &= svga->decode_mask;
         if (addr >= voodoo->tile_base)
         {
@@ -1403,7 +1402,7 @@ static uint16_t banshee_read_linear_w(uint32_t addr, void *p)
                 return 0xff;
 
         egareads++;
-        sub_cycles(video_timing_read_w);
+        cycles -= video_timing_read_w;
 	
 //        banshee_log("read_linear: addr=%08x val=%02x\n", addr, svga->vram[addr & svga->vram_mask]);
 
@@ -1419,7 +1418,7 @@ static uint32_t banshee_read_linear_l(uint32_t addr, void *p)
         if (addr & 3)
                 return banshee_read_linear_w(addr, p) | (banshee_read_linear_w(addr+2, p) << 16);	
 	
-        sub_cycles(voodoo->read_time);
+        cycles -= voodoo->read_time;
 
         addr &= svga->decode_mask;
         if (addr >= voodoo->tile_base)
@@ -1437,7 +1436,7 @@ static uint32_t banshee_read_linear_l(uint32_t addr, void *p)
                 return 0xff;
 
         egareads++;
-	sub_cycles(video_timing_read_l);
+	cycles -= video_timing_read_l;
 
 //        banshee_log("read_linear: addr=%08x val=%02x\n", addr, svga->vram[addr & svga->vram_mask]);
 
@@ -1450,7 +1449,7 @@ static void banshee_write_linear(uint32_t addr, uint8_t val, void *p)
         voodoo_t *voodoo = banshee->voodoo;
         svga_t *svga = &banshee->svga;
         
-	sub_cycles(voodoo->write_time);
+	cycles -= voodoo->write_time;
 
 //        banshee_log("write_linear: addr=%08x val=%02x\n", addr, val);
         addr &= svga->decode_mask;
@@ -1470,7 +1469,7 @@ static void banshee_write_linear(uint32_t addr, uint8_t val, void *p)
 
         egawrites++;
 	
-	sub_cycles(video_timing_write_b);
+	cycles -= video_timing_write_b;
 
         svga->changedvram[addr >> 12] = changeframecount;
         svga->vram[addr & svga->vram_mask] = val;
@@ -1489,8 +1488,7 @@ static void banshee_write_linear_w(uint32_t addr, uint16_t val, void *p)
                 return;
         }	
 	
-        sub_cycles(voodoo->write_time);
-
+        cycles -= voodoo->write_time;
 //        banshee_log("write_linear: addr=%08x val=%02x\n", addr, val);
         addr &= svga->decode_mask;
         if (addr >= voodoo->tile_base)
@@ -1509,7 +1507,7 @@ static void banshee_write_linear_w(uint32_t addr, uint16_t val, void *p)
 
         egawrites++;
 
-        sub_cycles(video_timing_write_w);
+        cycles -= video_timing_write_w;
 
         svga->changedvram[addr >> 12] = changeframecount;
         *(uint16_t *)&svga->vram[addr & svga->vram_mask] = val;
@@ -1533,7 +1531,7 @@ static void banshee_write_linear_l(uint32_t addr, uint32_t val, void *p)
                 timing = voodoo->burst_time;
         else
                 timing = voodoo->write_time;
-	sub_cycles(timing);
+	cycles -= timing;
         voodoo->last_write_addr = addr;
 
 //        /*if (val) */banshee_log("write_linear_l: addr=%08x val=%08x  %08x\n", addr, val, voodoo->tile_base);
@@ -1555,7 +1553,7 @@ static void banshee_write_linear_l(uint32_t addr, uint32_t val, void *p)
 
         egawrites += 4;
 
-        sub_cycles(video_timing_write_l);
+        cycles -= video_timing_write_l;
 
         svga->changedvram[addr >> 12] = changeframecount;
         *(uint32_t *)&svga->vram[addr & svga->vram_mask] = val;
diff --git a/src/video/video.c b/src/video/video.c
index ce77e9abf..2b1a9bcd5 100644
--- a/src/video/video.c
+++ b/src/video/video.c
@@ -336,7 +336,7 @@ static png_infop	info_ptr;
 
 
 static void
-video_take_screenshot(const wchar_t *fn, int startx, int starty, int w, int h)
+video_take_screenshot(const wchar_t *fn, int startx, int starty, int y1, int y2, int w, int h)
 {
     int i, x, y;
     png_bytep *b_rgb = NULL;
@@ -382,7 +382,7 @@ video_take_screenshot(const wchar_t *fn, int startx, int starty, int w, int h)
     for (y = 0; y < h; ++y) {
 	b_rgb[y] = (png_byte *) malloc(png_get_rowbytes(png_ptr, info_ptr));
     	for (x = 0; x < w; ++x) {
-		temp = render_buffer->line[y + starty][x + startx];
+		temp = render_buffer->dat[(y * w) + x];
 
 		b_rgb[y][(x) * 3 + 0] = (temp >> 16) & 0xff;
 		b_rgb[y][(x) * 3 + 1] = (temp >> 8) & 0xff;
@@ -407,7 +407,7 @@ video_take_screenshot(const wchar_t *fn, int startx, int starty, int w, int h)
 
 
 static void
-video_screenshot(int x, int y, int w, int h)
+video_screenshot(int x, int y, int y1, int y2, int w, int h)
 {
     wchar_t path[1024], fn[128];
 
@@ -426,7 +426,7 @@ video_screenshot(int x, int y, int w, int h)
 
     video_log("taking screenshot to: %S\n", path);
 
-    video_take_screenshot((const wchar_t *) path, x, y, w, h);
+    video_take_screenshot((const wchar_t *) path, x, y, y1, y2, w, h);
     png_destroy_write_struct(&png_ptr, &info_ptr);
 }
 
@@ -449,20 +449,20 @@ video_blit_memtoscreen(int x, int y, int y1, int y2, int w, int h)
 {
     int yy;
 
-    if ((w > 0) && (h > 0)) {
-	for (yy = 0; yy < h; yy++) {
+    if (y2 > 0) {
+	for (yy = y1; yy < y2; yy++) {
 		if (((y + yy) >= 0) && ((y + yy) < buffer32->h)) {
 			if (video_grayscale || invert_display)
-				video_transform_copy(&(render_buffer->line[y + yy][x]), &(buffer32->line[y + yy][x]), w);
+				video_transform_copy(&(render_buffer->dat)[yy * w], &(buffer32->line[y + yy][x]), w);
 			else
-				memcpy(&(render_buffer->line[y + yy][x]), &(buffer32->line[y + yy][x]), w << 2);
+				memcpy(&(render_buffer->dat)[yy * w], &(buffer32->line[y + yy][x]), w << 2);
 		}
 	}
     }
 
     if (screenshots) {
 	if (render_buffer != NULL)
-		video_screenshot(x, y, w, h);
+		video_screenshot(x, y, y1, y2, w, h);
 	screenshots--;
 	video_log("screenshot taken, %i left\n", screenshots);
     }
diff --git a/src/vnc.c b/src/vnc.c
index 5813767fb..09c00831f 100644
--- a/src/vnc.c
+++ b/src/vnc.c
@@ -176,7 +176,7 @@ vnc_blit(int x, int y, int y1, int y2, int w, int h)
 	p = (uint32_t *)&(((uint32_t *)rfb->frameBuffer)[yy*VNC_MAX_X]);
 
 	if ((y+yy) >= 0 && (y+yy) < VNC_MAX_Y)
-		memcpy(p, &(render_buffer->line[y+yy][x]), w*4);
+		memcpy(p, &(render_buffer->dat[yy * w]), w*4);
     }
  
     video_blit_complete();
diff --git a/src/win/86Box.rc b/src/win/86Box.rc
index 237e7f783..ae7f632e6 100644
--- a/src/win/86Box.rc
+++ b/src/win/86Box.rc
@@ -66,7 +66,8 @@ BEGIN
         POPUP "Re&nderer"
         BEGIN
             MENUITEM "&SDL (Software)",         IDM_VID_SDL_SW
-            MENUITEM "&SDL (Hardware)",         IDM_VID_SDL_HW
+            MENUITEM "SDL (&Hardware)",         IDM_VID_SDL_HW
+            MENUITEM "SDL (&OpenGL)",           IDM_VID_SDL_OPENGL
 #ifdef USE_VNC
             MENUITEM "&VNC",                    IDM_VID_VNC
 #endif
diff --git a/src/win/Makefile.mingw b/src/win/Makefile.mingw
index f35641d58..6fd9b04e8 100644
--- a/src/win/Makefile.mingw
+++ b/src/win/Makefile.mingw
@@ -509,10 +509,6 @@ ifeq ($(AMD_K5), y)
 OPTS		+= -DUSE_AMD_K5
 endif
 
-ifeq ($(CL5422), y)
-OPTS		+= -DUSE_CL5422
-endif
-
 ifeq ($(CYRIX_6X86), y)
 OPTS		+= -DUSE_CYRIX_6X86
 endif
diff --git a/src/win/win.c b/src/win/win.c
index 541a05d1a..0d465f2f3 100644
--- a/src/win/win.c
+++ b/src/win/win.c
@@ -89,14 +89,16 @@ static const struct {
 } vid_apis[2][RENDERERS_NUM] = {
   {
     {	"SDL_Software", 1, (int(*)(void*))sdl_inits, sdl_close, NULL, sdl_pause, sdl_enable		},
-    {	"SDL_Hardware", 1, (int(*)(void*))sdl_inith, sdl_close, NULL, sdl_pause, sdl_enable		}
+    {	"SDL_Hardware", 1, (int(*)(void*))sdl_inith, sdl_close, NULL, sdl_pause, sdl_enable		},
+    {	"SDL_OpenGL", 1, (int(*)(void*))sdl_initho, sdl_close, NULL, sdl_pause, sdl_enable		}
 #ifdef USE_VNC
     ,{	"VNC", 0, vnc_init, vnc_close, vnc_resize, vnc_pause, NULL					}
 #endif
   },
   {
     {	"SDL_Software", 1, (int(*)(void*))sdl_inits_fs, sdl_close, sdl_resize, sdl_pause, sdl_enable	},
-    {	"SDL_Hardware", 1, (int(*)(void*))sdl_inith_fs, sdl_close, sdl_resize, sdl_pause, sdl_enable	}
+    {	"SDL_Hardware", 1, (int(*)(void*))sdl_inith_fs, sdl_close, sdl_resize, sdl_pause, sdl_enable	},
+    {	"SDL_OpenGL", 1, (int(*)(void*))sdl_initho_fs, sdl_close, sdl_resize, sdl_pause, sdl_enable	}
 #ifdef USE_VNC
     ,{	"VNC", 0, vnc_init, vnc_close, vnc_resize, vnc_pause, NULL					}
 #endif
@@ -700,9 +702,12 @@ plat_vidapi_name(int api)
 		break;
 	case 1:
 		break;
+	case 2:
+		name = "sdl_opengl";
+		break;
 
 #ifdef USE_VNC
-	case 2:
+	case 3:
 		name = "vnc";
 		break;
 #endif
@@ -762,12 +767,21 @@ plat_vidsize(int x, int y)
 void
 plat_vidapi_enable(int enable)
 {
+    int i = 1;
+
     if (!vid_api_inited || !vid_apis[video_fullscreen][vid_api].enable) return;
 
     startblit();
     video_wait_for_blit();
-    vid_apis[video_fullscreen][vid_api].enable(enable);
+
+    vid_apis[video_fullscreen][vid_api].enable(enable & 1);
+
     endblit();
+
+    if (! i) return;
+
+   if (enable)
+	device_force_redraw();
 }
 
 int
diff --git a/src/win/win_sdl.c b/src/win/win_sdl.c
index 9d3a9b4cd..db12f2cc8 100644
--- a/src/win/win_sdl.c
+++ b/src/win/win_sdl.c
@@ -75,6 +75,7 @@
 
 #define RENDERER_FULL_SCREEN	1
 #define RENDERER_HARDWARE	2
+#define RENDERER_OPENGL		4
 
 
 static SDL_Window	*sdl_win = NULL;
@@ -83,12 +84,55 @@ static SDL_Texture	*sdl_tex = NULL;
 static HWND		sdl_parent_hwnd = NULL;
 static HWND		sdl_hwnd = NULL;
 static int		sdl_w, sdl_h;
-static int		sdl_fs;
+static int		sdl_fs, sdl_flags = -1;
 static int		cur_w, cur_h;
+static int		cur_wx = 0, cur_wy = 0, cur_ww =0, cur_wh = 0;
 static volatile int	sdl_enabled = 0;
 static SDL_mutex*	sdl_mutex = NULL;
 
 
+typedef struct
+{
+    const void *magic;
+    Uint32 id;
+    char *title;
+    SDL_Surface *icon;
+    int x, y;
+    int w, h;
+    int min_w, min_h;
+    int max_w, max_h;
+    Uint32 flags;
+    Uint32 last_fullscreen_flags;
+
+    /* Stored position and size for windowed mode */
+    SDL_Rect windowed;
+
+    SDL_DisplayMode fullscreen_mode;
+
+    float brightness;
+    Uint16 *gamma;
+    Uint16 *saved_gamma;        /* (just offset into gamma) */
+
+    SDL_Surface *surface;
+    SDL_bool surface_valid;
+
+    SDL_bool is_hiding;
+    SDL_bool is_destroying;
+
+    void *shaper;
+
+    SDL_HitTest hit_test;
+    void *hit_test_data;
+
+    void *data;
+
+    void *driverdata;
+
+    SDL_Window *prev;
+    SDL_Window *next;
+} SDL_Window_Ex;
+
+
 #ifdef ENABLE_SDL_LOG
 int sdl_do_log = ENABLE_SDL_LOG;
 
@@ -190,9 +234,7 @@ static void
 sdl_blit(int x, int y, int y1, int y2, int w, int h)
 {
     SDL_Rect r_src;
-    void *pixeldata;
-    int pitch;
-    int yy, ret;
+    int ret;
 
     if (!sdl_enabled) {
 	video_blit_complete();
@@ -211,22 +253,13 @@ sdl_blit(int x, int y, int y1, int y2, int w, int h)
 
     SDL_LockMutex(sdl_mutex);
 
-    /*
-     * TODO:
-     * SDL_UpdateTexture() might be better here, as it is
-     * (reportedly) slightly faster.
-     */
-    SDL_LockTexture(sdl_tex, 0, &pixeldata, &pitch);
-
-    for (yy = y1; yy < y2; yy++) {
-       	if ((y + yy) >= 0 && (y + yy) < render_buffer->h)
-		memcpy((uint32_t *) &(((uint8_t *)pixeldata)[yy * pitch]), &(render_buffer->line[y + yy][x]), w * 4);
-    }
-
+    r_src.x = 0;
+    r_src.y = y1;
+    r_src.w = w;
+    r_src.h = y2 - y1;
+    SDL_UpdateTexture(sdl_tex, &r_src, &(render_buffer->dat)[y1 * w], w * 4);
     video_blit_complete();
 
-    SDL_UnlockTexture(sdl_tex);
-
     if (sdl_fs) {
 	sdl_log("sdl_blit(%i, %i, %i, %i, %i, %i) (%i, %i)\n", x, y, y1, y2, w, h, unscaled_size_x, efscrnsz_y);
 	if (w == unscaled_size_x)
@@ -234,6 +267,8 @@ sdl_blit(int x, int y, int y1, int y2, int w, int h)
 	sdl_log("(%08X, %08X, %08X)\n", sdl_win, sdl_render, sdl_tex);
     }
 
+    SDL_RenderClear(sdl_render);
+
     r_src.x = 0;
     r_src.y = 0;
     r_src.w = w;
@@ -252,6 +287,8 @@ sdl_blit(int x, int y, int y1, int y2, int w, int h)
 void
 sdl_close(void)
 {
+    SDL_LockMutex(sdl_mutex);
+
     /* Unregister our renderer! */
     video_setblit(NULL);
 
@@ -297,6 +334,8 @@ sdl_close(void)
 
     /* Quit. */
     SDL_Quit();
+
+    sdl_flags = -1;
 }
 
 
@@ -340,8 +379,12 @@ sdl_init_common(int flags)
 	return(0);
     }
 
-    if (flags & RENDERER_HARDWARE)
-	sdl_select_best_hw_driver();
+    if (flags & RENDERER_HARDWARE) {
+	if (flags & RENDERER_OPENGL)
+		SDL_SetHint(SDL_HINT_RENDER_DRIVER, "OpenGL");
+	else
+		sdl_select_best_hw_driver();
+    }
 
     if (flags & RENDERER_FULL_SCREEN) {
 	/* Get the size of the (current) desktop. */
@@ -461,6 +504,8 @@ sdl_init_common(int flags)
 
     sdl_mutex = SDL_CreateMutex();
 
+    sdl_flags = flags;
+
     return(1);
 }
 
@@ -479,6 +524,13 @@ sdl_inith(HWND h)
 }
 
 
+int
+sdl_initho(HWND h)
+{
+    return sdl_init_common(RENDERER_HARDWARE | RENDERER_OPENGL);
+}
+
+
 int
 sdl_inits_fs(HWND h)
 {
@@ -493,15 +545,30 @@ sdl_inith_fs(HWND h)
 }
 
 
+int
+sdl_initho_fs(HWND h)
+{
+    return sdl_init_common(RENDERER_FULL_SCREEN | RENDERER_HARDWARE | RENDERER_OPENGL);
+}
+
+
 void
 sdl_reinit_texture()
 {
-    if (sdl_render == NULL)
+    if ((sdl_render == NULL) || (sdl_flags == -1))
         return;
 
     SDL_DestroyTexture(sdl_tex);
+    SDL_DestroyRenderer(sdl_render);
+    if (sdl_flags & RENDERER_HARDWARE) {
+	sdl_render = SDL_CreateRenderer(sdl_win, -1, SDL_RENDERER_ACCELERATED);
+	SDL_SetHint(SDL_HINT_RENDER_SCALE_QUALITY, "2");
+    } else
+	sdl_render = SDL_CreateRenderer(sdl_win, -1, SDL_RENDERER_SOFTWARE);
     sdl_tex = SDL_CreateTexture(sdl_render, SDL_PIXELFORMAT_ARGB8888,
 				SDL_TEXTUREACCESS_STREAMING, 2048, 2048);
+    SDL_SetWindowSize(sdl_win, cur_ww, cur_wh);
+    SDL_SetWindowPosition(sdl_win, cur_wx, cur_wy);
 }
 
 
@@ -520,6 +587,8 @@ sdl_resize(int x, int y)
     if ((x == cur_w) && (y == cur_h))
 	return;
 
+    SDL_LockMutex(sdl_mutex);
+
     ww = x;
     wh = y;
 
@@ -531,12 +600,26 @@ sdl_resize(int x, int y)
     cur_w = x;
     cur_h = y;
 
+    cur_wx = wx;
+    cur_wy = wy;
+    cur_ww = ww;
+    cur_wh = wh;
+
     sdl_reinit_texture();
+
+    SDL_UnlockMutex(sdl_mutex);
 }
 
 
 void
 sdl_enable(int enable)
 {
+    if (sdl_flags == -1)
+	return;
+
+    SDL_LockMutex(sdl_mutex);
     sdl_enabled = enable;
+    if (enable)
+	sdl_reinit_texture();
+    SDL_UnlockMutex(sdl_mutex);
 }
diff --git a/src/win/win_settings.c b/src/win/win_settings.c
index c18006d9a..ea102e36b 100644
--- a/src/win/win_settings.c
+++ b/src/win/win_settings.c
@@ -2232,10 +2232,18 @@ win_settings_hard_disks_resize_columns(HWND hdlg)
 {
     /* Bus, File, Cylinders, Heads, Sectors, Size */
     int iCol, width[C_COLUMNS_HARD_DISKS] = {104, 177, 50, 26, 32, 50};
+    int total = 0;
     HWND hwndList = GetDlgItem(hdlg, IDC_LIST_HARD_DISKS);
+    RECT r;
 
-    for (iCol = 0; iCol < C_COLUMNS_HARD_DISKS; iCol++)
+    GetWindowRect(hwndList, &r);
+    for (iCol = 0; iCol < (C_COLUMNS_HARD_DISKS - 1); iCol++) {
+	width[iCol] = MulDiv(width[iCol], dpi, 96);
+	total += width[iCol];
 	ListView_SetColumnWidth(hwndList, iCol, MulDiv(width[iCol], dpi, 96));
+    }
+    width[C_COLUMNS_HARD_DISKS - 1] = (r.right - r.left) - 4 - total;
+    ListView_SetColumnWidth(hwndList, C_COLUMNS_HARD_DISKS - 1, width[C_COLUMNS_HARD_DISKS - 1]);
 }
 
 
@@ -3678,11 +3686,19 @@ win_settings_zip_drives_recalc_list(HWND hdlg)
 static void
 win_settings_floppy_drives_resize_columns(HWND hdlg)
 {
+    int iCol, width[3] = {292, 58, 89};
+    int total = 0;
     HWND hwndList = GetDlgItem(hdlg, IDC_LIST_FLOPPY_DRIVES);
+    RECT r;
 
-    ListView_SetColumnWidth(hwndList, 0, MulDiv(292, dpi, 96));
-    ListView_SetColumnWidth(hwndList, 1, MulDiv(58, dpi, 96));
-    ListView_SetColumnWidth(hwndList, 2, MulDiv(89, dpi, 96));
+    GetWindowRect(hwndList, &r);
+    for (iCol = 0; iCol < 2; iCol++) {
+	width[iCol] = MulDiv(width[iCol], dpi, 96);
+	total += width[iCol];
+	ListView_SetColumnWidth(hwndList, iCol, MulDiv(width[iCol], dpi, 96));
+    }
+    width[2] = (r.right - r.left) - 4 - total;
+    ListView_SetColumnWidth(hwndList, 2, width[2]);
 }
 
 
@@ -3729,10 +3745,15 @@ win_settings_floppy_drives_init_columns(HWND hdlg)
 static void
 win_settings_cdrom_drives_resize_columns(HWND hdlg)
 {
+    int width[2] = {292, 147};
     HWND hwndList = GetDlgItem(hdlg, IDC_LIST_CDROM_DRIVES);
+    RECT r;
 
-    ListView_SetColumnWidth(hwndList, 0, MulDiv(292, dpi, 96));
-    ListView_SetColumnWidth(hwndList, 1, MulDiv(147, dpi, 96));
+    GetWindowRect(hwndList, &r);
+    width[0] = MulDiv(width[0], dpi, 96);
+    ListView_SetColumnWidth(hwndList, 0, MulDiv(width[0], dpi, 96));
+    width[1] = (r.right - r.left) - 4 - width[0];
+    ListView_SetColumnWidth(hwndList, 1, width[1]);
 }
 
 
@@ -3770,10 +3791,15 @@ win_settings_cdrom_drives_init_columns(HWND hdlg)
 static void
 win_settings_mo_drives_resize_columns(HWND hdlg)
 {
+    int width[2] = {292, 147};
     HWND hwndList = GetDlgItem(hdlg, IDC_LIST_MO_DRIVES);
+    RECT r;
 
-    ListView_SetColumnWidth(hwndList, 0, MulDiv(292, dpi, 96));
-    ListView_SetColumnWidth(hwndList, 1, MulDiv(147, dpi, 96));
+    GetWindowRect(hwndList, &r);
+    width[0] = MulDiv(width[0], dpi, 96);
+    ListView_SetColumnWidth(hwndList, 0, MulDiv(width[0], dpi, 96));
+    width[1] = (r.right - r.left) - 4 - width[0];
+    ListView_SetColumnWidth(hwndList, 1, width[1]);
 }
 
 
@@ -3811,10 +3837,15 @@ win_settings_mo_drives_init_columns(HWND hdlg)
 static void
 win_settings_zip_drives_resize_columns(HWND hdlg)
 {
+    int width[2] = {292, 147};
     HWND hwndList = GetDlgItem(hdlg, IDC_LIST_ZIP_DRIVES);
+    RECT r;
 
-    ListView_SetColumnWidth(hwndList, 0, MulDiv(292, dpi, 96));
-    ListView_SetColumnWidth(hwndList, 1, MulDiv(147, dpi, 96));
+    GetWindowRect(hwndList, &r);
+    width[0] = MulDiv(width[0], dpi, 96);
+    ListView_SetColumnWidth(hwndList, 0, MulDiv(width[0], dpi, 96));
+    width[1] = (r.right - r.left) - 4 - width[0];
+    ListView_SetColumnWidth(hwndList, 1, width[1]);
 }
 
 
@@ -4981,6 +5012,17 @@ win_settings_confirm(HWND hdlg)
 }
 
 
+static void
+win_settings_categories_resize_columns(HWND hdlg)
+{
+    HWND hwndList = GetDlgItem(hdlg, IDC_SETTINGSCATLIST);
+    RECT r;
+
+    GetWindowRect(hwndList, &r);
+    ListView_SetColumnWidth(hwndList, 0, (r.right - r.left) + 1 - 5);
+}
+
+
 static BOOL
 win_settings_categories_init_columns(HWND hdlg)
 {
@@ -4999,20 +5041,11 @@ win_settings_categories_init_columns(HWND hdlg)
     if (ListView_InsertColumn(hwndList, iCol, &lvc) == -1)
 	return FALSE;
 
-    win_settings_hard_disks_resize_columns(hdlg);
+    win_settings_categories_resize_columns(hdlg);
     return TRUE;
 }
 
 
-static void
-win_settings_categories_resize_columns(HWND hdlg)
-{
-    HWND hwndList = GetDlgItem(hdlg, IDC_SETTINGSCATLIST);
-
-    ListView_SetColumnWidth(hwndList, 0, MulDiv(171, dpi, 96));
-}
-
-
 #if defined(__amd64__) || defined(__aarch64__)
 static LRESULT CALLBACK
 #else
diff --git a/src/win/win_ui.c b/src/win/win_ui.c
index 51dbb9e46..5dd4026c5 100644
--- a/src/win/win_ui.c
+++ b/src/win/win_ui.c
@@ -216,6 +216,7 @@ ResetAllMenus(void)
     CheckMenuItem(menuMain, IDM_VID_RESIZE, MF_UNCHECKED);
     CheckMenuItem(menuMain, IDM_VID_SDL_SW, MF_UNCHECKED);
     CheckMenuItem(menuMain, IDM_VID_SDL_HW, MF_UNCHECKED);
+    CheckMenuItem(menuMain, IDM_VID_SDL_OPENGL, MF_UNCHECKED);
 #ifdef USE_VNC
     CheckMenuItem(menuMain, IDM_VID_VNC, MF_UNCHECKED);
 #endif
@@ -418,7 +419,6 @@ MainWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam)
 				break;
 
 			case IDM_ACTION_RESET_CAD:
-				pclog("-\n");
 				pc_send_cad();
 				break;
 
@@ -538,6 +538,7 @@ MainWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam)
 
 			case IDM_VID_SDL_SW:
 			case IDM_VID_SDL_HW:
+			case IDM_VID_SDL_OPENGL:
 #ifdef USE_VNC
 			case IDM_VID_VNC:
 #endif
@@ -730,6 +731,7 @@ MainWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam)
 		temp_y = (lParam >> 16);
 
 		if ((temp_x <= 0) || (temp_y <= 0)) {
+			plat_vidapi_enable(0);
 			minimized = 1;
 			break;
 		} else if (minimized == 1) {
@@ -936,6 +938,14 @@ MainWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam)
 #endif
 		break;
 
+	case WM_ACTIVATE:
+		if (wParam != WA_INACTIVE) {
+			video_force_resize_set(1);
+			plat_vidapi_enable(0);
+			plat_vidapi_enable(1);
+		}
+		break;
+
 	case WM_ENTERSIZEMOVE:
 		user_resize = 1;
 		break;
@@ -1071,18 +1081,18 @@ ui_init(int nCmdShow)
 
     ui_window_title(title);
 
-	/* Get the current DPI */
-	dpi = win_get_dpi(hwndMain);
+    /* Get the current DPI */
+    dpi = win_get_dpi(hwndMain);
 
-	/* Check if we have a padded window frame */
-	padded_frame = (GetSystemMetrics(SM_CXPADDEDBORDER) != 0);
+    /* Check if we have a padded window frame */
+    padded_frame = (GetSystemMetrics(SM_CXPADDEDBORDER) != 0);
 
     /* Create the status bar window. */
     StatusBarCreate(hwndMain, IDC_STATUS, hinstance);
 
-	/* Get the actual height of the status bar */
-	GetWindowRect(hwndSBAR, &sbar_rect);
-	sbar_height = sbar_rect.bottom - sbar_rect.top;
+    /* Get the actual height of the status bar */
+    GetWindowRect(hwndSBAR, &sbar_rect);
+    sbar_height = sbar_rect.bottom - sbar_rect.top;
 
     /* Set up main window for resizing if configured. */
     if (vid_resize)