S3: various fixes

        * src/s3_video.c:

        Streams Processor initialization.

        XVideo support for 16, 24 and 32 bpp (only for TRIO64V2).

        * src/s3_driver.c:

        Add XVideo option (enabled by default).

        Replace "swcursor" by "hwcursor" option (HW cursor
        not implemented yet for all chipsets).

        Close bug #5527: 24-bit colour depth support (24bpp and
        32bpp framebuffer) for TRIO64V2. Acceleration doesn't work
        with packed colour mode (24 bpp FB) but works with 32bpp
        framebuffer mode (hardware limitation? -- need data).

        Fix system hang-up when switching between console and X session.

        Some minor changes of driver messages.

9 files changed, 604 insertions, 353 deletions

diff --git a/ChangeLog b/ChangeLog
index dc15e51..fab4ec7 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,27 @@
+2008-07-06  Evgeny M. Zubok <evgeny.zubok@tochka.ru>
+
+	* src/s3_video.c: 
+
+	Streams Processor initialization.
+
+	XVideo support for 16, 24 and 32 bpp (only for TRIO64V2).
+
+	* src/s3_driver.c: 
+
+	Add XVideo option (enabled by default).
+
+	Replace "swcursor" by "hwcursor" option (HW cursor 
+	not implemented yet for all chipsets).
+
+	Close bug #5527: 24-bit colour depth support (24bpp and 
+	32bpp framebuffer) for TRIO64V2. Acceleration doesn't work
+	with packed colour mode (24 bpp FB) but works with 32bpp 
+	framebuffer mode (hardware limitation? -- need data).
+
+	Fix system hang-up when switching between console and X session.
+
+	Some minor changes of driver messages.
+
 2008-02-23  Evgeny M. Zubok <evgeny.zubok@tochka.ru>
 
 	* src/s3_driver.c: Add DPMS support for S3 Trio64V2 and possibly
diff --git a/src/Makefile.am b/src/Makefile.am
index 0d9ae4b..743babd 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -39,13 +39,12 @@ s3_drv_la_SOURCES = \
 	 s3_pcirename.h \
          s3_reg.h \
          s3_Ti.c \
-         s3_Trio64DAC.c \
-         s3_video.c
+         s3_Trio64DAC.c
 
 noinst_LTLIBRARIES = libs3_accel_newmmio.la libs3_accel_pio.la
 s3_drv_la_LIBADD = libs3_accel_newmmio.la libs3_accel_pio.la
 
-libs3_accel_newmmio_la_SOURCES = s3_accel.c
+libs3_accel_newmmio_la_SOURCES = s3_accel.c s3_video.c
 libs3_accel_newmmio_la_CFLAGS = $(AM_CFLAGS) -DS3_NEWMMIO=1
 
 libs3_accel_pio_la_SOURCES = s3_accel.c
diff --git a/src/newmmio.h b/src/newmmio.h
index fdb0d34..e783647 100644
--- a/src/newmmio.h
+++ b/src/newmmio.h
@@ -248,10 +248,10 @@ typedef struct {
 
 	
 #define WaitQueue(v)					\
-	if(!(pS3->PCIRetry)) {				\
+        if(!(pS3->PCIRetry)) {				\
 	   mem_barrier();				\
-	   while(inb(GP_STAT) & (0x0100 >> (v)));	\
-	}
+	   while(INB_GP_STAT() & (0x0100 >> (v)));      \
+        }				
 
 #define CMD_REG_WIDTH  0x200  	/* select 32bit command register */
 
diff --git a/src/s3.h b/src/s3.h
index 7c59236..a12ff6e 100644
--- a/src/s3.h
+++ b/src/s3.h
@@ -64,7 +64,7 @@ typedef struct _S3RegRec {
 typedef struct {
         unsigned char brightness;
         unsigned char contrast;
-        FBAreaPtr     area;
+        FBLinearPtr   area;
         RegionRec     clip;
         CARD32        colorKey;
         CARD32        videoStatus;
@@ -98,12 +98,15 @@ typedef struct _S3Rec {
         OptionInfoPtr           Options;
         unsigned int            Flags;
         Bool                    NoAccel;
-	Bool			SWCursor;
+	Bool			HWCursor;
 	Bool			SlowDRAMRefresh;
 	Bool			SlowDRAM;
 	Bool			SlowEDODRAM;
 	Bool			SlowVRAM;
 	Bool			S3NewMMIO;
+	Bool                    hasStreams;
+	int                     Streams_FIFO;
+	Bool                    XVideo;
 	Bool			PCIRetry;
 	Bool			ColorExpandBug;
 
@@ -159,7 +162,7 @@ typedef struct _S3Rec {
 	unsigned char		*imageBuffer;
 	int			imageWidth;
 	int			imageHeight;
-    Bool			hwCursor;
+	Bool			hwCursor;
 } S3Rec, *S3Ptr;
 
 #define S3PTR(p)		((S3Ptr)((p)->driverPrivate))
@@ -234,9 +237,14 @@ Bool S3_CursorInit(ScreenPtr pScreen);
 
 #define S3_964_SERIES()		((pS3->Chipset == PCI_CHIP_964_0) ||	\
 			 	 (pS3->Chipset == PCI_CHIP_964_1))
+
 #define	S3_TRIO_SERIES()	((pS3->Chipset == PCI_CHIP_TRIO) ||	\
 			 	 (pS3->Chipset == PCI_CHIP_AURORA64VP) || \
 				 (pS3->Chipset == PCI_CHIP_TRIO64UVP) || \
 				 (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX))
 
+#define HAS_STREAMS_PROCESSOR() ((pS3->Chipset == PCI_CHIP_AURORA64VP) || \
+				 (pS3->Chipset == PCI_CHIP_TRIO64UVP) || \
+				 (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX))
+
 #endif /* _S3_H */
diff --git a/src/s3_Trio64DAC.c b/src/s3_Trio64DAC.c
index 5b32146..e2cf257 100644
--- a/src/s3_Trio64DAC.c
+++ b/src/s3_Trio64DAC.c
@@ -35,6 +35,7 @@
 #include "compiler.h"
 
 #include "s3.h"
+#include "s3_reg.h"
 
 /* this is really quite dumb */
 Bool S3Trio64DACProbe(ScrnInfoPtr pScrn)
@@ -232,33 +233,30 @@ static void S3TrioSetPLL(ScrnInfoPtr pScrn, int clk, unsigned char m,
 
 		outb(0x3c4, 0x08);
 		outb(0x3c5, 0x06);	/* unlock extended CR9-18 */
-
-		if (clk != 10) {
-			outb(0x3c4, 0x12);
-			outb(0x3c5, n);
-			outb(0x3c4, 0x13);
-			outb(0x3c5, m);
-
-			outb(0x3c4, 0x15);
-			tmp = inb(0x3c5) & ~0x21;
-			outb(0x3c5, tmp | 0x02);
-			outb(0x3c5, tmp | 0x22);
-			outb(0x3c5, tmp | 0x02);
-		} else {
-			outb(0x3c4, 0x10);
-			outb(0x3c5, n);
-			outb(0x3c4, 0x11);
-			outb(0x3c5, m);
-			outb(0x3c4, 0x1a);
-			outb(0x3c5, n);
-
-			outb(0x3c4, 0x15);
-			tmp = inb(0x3c5) & ~0x21;
-			outb(0x3c5, tmp | 0x01);
-			outb(0x3c5, tmp | 0x21);
-			outb(0x3c5, tmp | 0x01);
-			outb(0x3c5, tmp);
-		}
+		
+		outb(0x3c4, 0x12);      /* write N1 and N2 to DCLK PLL */
+		outb(0x3c5, n);
+		outb(0x3c4, 0x13);      /* write M to DCLK PLL */
+		outb(0x3c5, m);
+
+#if 0
+/* this code was in previous driver version but it was never called. 
+   So I decide to comment it. */
+		outb(0x3c4, 0x10);     
+		outb(0x3c5, n);
+		outb(0x3c4, 0x11);
+		outb(0x3c5, m); */
+			
+		outb(0x3c4, 0x1a);
+		outb(0x3c5, n);
+#endif
+		/* Toggle cr15_5 by sequence 0->1->0 to immediately apply 
+		   new PLL parameters */
+		outb(0x3c4, 0x15);
+		tmp = inb(0x3c5) & ~0x20;
+		outb(0x3c5, tmp);
+		outb(0x3c5, tmp | 0x20);
+		outb(0x3c5, tmp);
 
 		outb(0x3c4, 0x08);
 		outb(0x3c5, 0x00);	/* lock em */
@@ -331,18 +329,17 @@ void S3Trio64DAC_Init(ScrnInfoPtr pScrn, DisplayModePtr mode)
 		S3TrioSetClock(pScrn, mode->Clock, 2, 1, 1, 31, 0, 3, 2,
 			       135000, 270000);
 
-
-	outb(0x3c4, 1);
+	outb(0x3c4, 0x01);
 	blank = inb(0x3c5);
 	outb(0x3c5, blank | 0x20);	/* blank the screen */
 
 	outb(0x3c4, 0x08);
 	sr8 = inb(0x3c5);
-	outb(0x3c5, 0x06);
+	outb(0x3c5, 0x06);       /* unlock extended sequenser register */
 
-	outb(0x3c4, 0x0d0);
-	tmp = inb(0x3c5) & ~1;
-	outb(0x3c5, tmp);
+	outb(0x3c4, 0x0d);
+	tmp = inb(0x3c5) & ~0x01;
+	outb(0x3c5, tmp);       /* VCLK, HSYNC, VSYNC are outputs */
 
 	outb(0x3c4, 0x15);
 	sr15 = inb(0x3c5) & ~0x10;
@@ -352,11 +349,6 @@ void S3Trio64DAC_Init(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	outb(pS3->vgaCRIndex, 0x33);
 	cr33 = inb(pS3->vgaCRReg) & ~0x28;
 
-	if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)
-	{
-	  cr33 |= 0x20;
-	}
-
 	/* ! pixmux */
 	switch (pScrn->depth) {
 	case 8:
@@ -369,6 +361,7 @@ void S3Trio64DAC_Init(ScrnInfoPtr pScrn, DisplayModePtr mode)
 		cr33 |= 0x08;
 		pixmux = 0x50;
 		break;
+	case 24:
 	case 32:
 		pixmux = 0xd0;
 		break;
@@ -377,6 +370,7 @@ void S3Trio64DAC_Init(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	outb(pS3->vgaCRReg, cr33);
 
 	outb(pS3->vgaCRIndex, 0x67);
+	WaitVSync();
 	outb(pS3->vgaCRReg, pixmux | invert_vclk);
 
 	outb(0x3c4, 0x15);
@@ -392,6 +386,6 @@ void S3Trio64DAC_Init(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	outb(0x3c4, 0x08);
 	outb(0x3c5, sr8);
 
-	outb(0x3c4, 1);
+	outb(0x3c4, 0x01);
 	outb(0x3c5, blank);	/* unblank the screen */
 }
diff --git a/src/s3_accel.c b/src/s3_accel.c
index 675b823..78db3b1 100644
--- a/src/s3_accel.c
+++ b/src/s3_accel.c
@@ -140,7 +140,7 @@ static void S3SubsequentScreenToScreenCopy(ScrnInfoPtr pScrn,
 	}
 }
 
-
+#if 0
 static void S3SetupForColor8x8PatternFill(ScrnInfoPtr pScrn,
 					  int patx, int paty,
 					  int rop, unsigned int planemask,
@@ -182,6 +182,7 @@ static void S3SubsequentColor8x8PatternFillRect(ScrnInfoPtr pScrn,
 		SET_MULT_MISC(CMD_REG_WIDTH);
 	}
 }
+#endif
 
 #ifdef S3_NEWMMIO
 static void S3SetupForCPUToScreenColorExpandFill(ScrnInfoPtr pScrn,
@@ -569,9 +570,22 @@ Bool S3AccelInitPIO(ScreenPtr pScreen)
 
 	pXAA->SetupForScreenToScreenCopy = S3SetupForScreenToScreenCopy;
 	pXAA->SubsequentScreenToScreenCopy = S3SubsequentScreenToScreenCopy;
+	pXAA->ScreenToScreenCopyFlags = NO_TRANSPARENCY;
 
+#if 0
+/*
+  8x8 color pattern filling doesn't work properly after introducing 
+  framebuffer manager initialization before XAA initialization. There 
+  are problems with addressing a colour patterns from offscreen area.
+*/
 	pXAA->SetupForColor8x8PatternFill = S3SetupForColor8x8PatternFill;
 	pXAA->SubsequentColor8x8PatternFillRect = S3SubsequentColor8x8PatternFillRect;
+	pXAA->Color8x8PatternFillFlags = NO_TRANSPARENCY | 
+		HARDWARE_PATTERN_SCREEN_ORIGIN |
+		BIT_ORDER_IN_BYTE_MSBFIRST;
+
+	pXAA->CachePixelGranularity = 0;
+#endif
 
 #ifdef S3_NEWMMIO
 	pXAA->SetupForCPUToScreenColorExpandFill =
diff --git a/src/s3_driver.c b/src/s3_driver.c
index 1a68202..797e0b3 100644
--- a/src/s3_driver.c
+++ b/src/s3_driver.c
@@ -152,20 +152,22 @@ static PciChipsets S3PciChipsets[] = {
 
 typedef enum {
 	OPTION_NOACCEL,
-	OPTION_SWCURS,
+	OPTION_HWCURS,
 	OPTION_SLOW_DRAM_REFRESH,
 	OPTION_SLOW_DRAM,
 	OPTION_SLOW_EDODRAM,
-	OPTION_SLOW_VRAM
+	OPTION_SLOW_VRAM,
+	OPTION_XVIDEO
 } S3Opts;
 
 static OptionInfoRec S3Options[] = {
 	{ OPTION_NOACCEL, "noaccel", OPTV_BOOLEAN, {0}, FALSE },
-	{ OPTION_SWCURS, "swcursor", OPTV_BOOLEAN, {0}, FALSE },
+	{ OPTION_HWCURS, "hwcursor", OPTV_BOOLEAN, {0}, FALSE },
 	{ OPTION_SLOW_DRAM_REFRESH, "slow_dram_refresh", OPTV_BOOLEAN, {0}, FALSE },
 	{ OPTION_SLOW_DRAM, "slow_dram", OPTV_BOOLEAN, {0}, FALSE },
 	{ OPTION_SLOW_EDODRAM, "slow_edodram", OPTV_BOOLEAN, {0}, FALSE },
 	{ OPTION_SLOW_VRAM, "slow_vram", OPTV_BOOLEAN, {0}, FALSE },
+	{ OPTION_XVIDEO, "XVideo", OPTV_BOOLEAN, {0}, FALSE },
 	{ -1, NULL, OPTV_NONE, {0}, FALSE }
 };
 
@@ -388,21 +390,20 @@ static Bool S3PreInit(ScrnInfoPtr pScrn, int flags)
         pScrn->monitor = pScrn->confScreen->monitor;
         
         if (!xf86SetDepthBpp(pScrn, 0, 0, 0, Support24bppFb | Support32bppFb))
-                return FALSE;
+	    return FALSE;
 
         switch (pScrn->depth) {
-            	case 8:
-                case 15:
-                case 16:
-                case 24:
-                case 32:
-      	               /* OK */
-                       break;
-                default:
-                       xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
-                                  "Given depth (%d) is not  supported by this driver\n",
-                                   pScrn->depth);
-	                       return FALSE;
+	case 8:
+	case 15:
+	case 16:
+	case 24:
+		/* OK */
+		break;
+	default:
+		xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
+			   "Given depth (%d) is not  supported by this driver\n",
+			   pScrn->depth);
+		return FALSE;
         }
 
         xf86PrintDepthBpp(pScrn);
@@ -427,37 +428,13 @@ static Bool S3PreInit(ScrnInfoPtr pScrn, int flags)
         xf86CollectOptions(pScrn, NULL);
         xf86ProcessOptions(pScrn->scrnIndex, pScrn->options, S3Options);
          
-        if (xf86ReturnOptValBool(S3Options, OPTION_NOACCEL, FALSE)) {
-                pS3->NoAccel = TRUE;
-                xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Option: NoAccel - acceleration disabled\n");
-        } else          
-                pS3->NoAccel = FALSE;
-	if (xf86ReturnOptValBool(S3Options, OPTION_SWCURS, FALSE)) {
-		pS3->SWCursor = TRUE;
-		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Option: SWCursor - using software cursor\n");
-        } else
-		pS3->SWCursor = FALSE;
-	if (xf86ReturnOptValBool(S3Options, OPTION_SLOW_DRAM_REFRESH, FALSE)) {
-		pS3->SlowDRAMRefresh = TRUE;
-		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Option: Slow DRAM Refresh enabled\n");
-	} else
-		pS3->SlowDRAMRefresh = FALSE;
-	if (xf86ReturnOptValBool(S3Options, OPTION_SLOW_DRAM, FALSE)) {
-		pS3->SlowDRAM = TRUE;
-		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Option: Slow DRAM enabled\n");
-	} else
-		pS3->SlowDRAM = FALSE;
-	if (xf86ReturnOptValBool(S3Options, OPTION_SLOW_EDODRAM, FALSE)) {
-		pS3->SlowEDODRAM = TRUE;
-		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Option: Slow EDO DRAM enabled\n");
-	} else
-		pS3->SlowEDODRAM = FALSE;
-	if (xf86ReturnOptValBool(S3Options, OPTION_SLOW_DRAM, FALSE)) {
-		pS3->SlowVRAM = TRUE;
-		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Option: Slow VRAM enabled\n");
-	} else
-		pS3->SlowVRAM = FALSE;
-
+	pS3->XVideo = xf86ReturnOptValBool(S3Options, OPTION_XVIDEO, TRUE);
+	pS3->NoAccel = xf86ReturnOptValBool(S3Options, OPTION_NOACCEL, FALSE);
+	pS3->HWCursor = xf86ReturnOptValBool(S3Options, OPTION_HWCURS, FALSE);
+	pS3->SlowDRAMRefresh = xf86ReturnOptValBool(S3Options, OPTION_SLOW_DRAM_REFRESH, FALSE);
+	pS3->SlowDRAM = xf86ReturnOptValBool(S3Options, OPTION_SLOW_DRAM, FALSE);
+	pS3->SlowEDODRAM = xf86ReturnOptValBool(S3Options, OPTION_SLOW_EDODRAM, FALSE);
+	pS3->SlowVRAM = xf86ReturnOptValBool(S3Options, OPTION_SLOW_VRAM, FALSE);
         if (pScrn->numEntities > 1) {      
                 S3FreeRec(pScrn);
                 return FALSE;
@@ -532,6 +509,11 @@ static Bool S3PreInit(ScrnInfoPtr pScrn, int flags)
 		break;
 	}
 
+	if (HAS_STREAMS_PROCESSOR() && pS3->S3NewMMIO)
+        	pS3->hasStreams = TRUE;
+        else
+        	pS3->hasStreams = FALSE;
+ 
 	pS3->FBAddress = PCI_REGION_BASE(pS3->PciInfo, 0, REGION_MEM);
 	pScrn->memPhysBase = pS3->FBAddress;
 	pScrn->fbOffset = 0;
@@ -581,9 +563,10 @@ static Bool S3PreInit(ScrnInfoPtr pScrn, int flags)
 
 	if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)
 	{
+          /* disable DAC power saving to avoid bright left edge */
 	  outb (0x3d4, 0x86);
 	  outb (0x3d5, 0x80);
-	  
+	  /* disable the stream display fetch length control */
 	  outb (0x3d4, 0x90);
 	  outb (0x3d5, 0x00);
 	}
@@ -664,17 +647,28 @@ static Bool S3PreInit(ScrnInfoPtr pScrn, int flags)
 #endif
 		switch(pScrn->bitsPerPixel) {
 		case 8:
-			pS3->MaxClock = 135000;
+			if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)
+				pS3->MaxClock = 170000;
+			else
+				pS3->MaxClock = 135000;
+
+			pScrn->rgbBits = 6;
 			break;
 		case 16:
 			pS3->MaxClock = 80000;
+			pScrn->rgbBits = 6;
 			break;
 		case 24:
 		case 32:
-			pS3->MaxClock = 50000;
+			if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)
+				pS3->MaxClock = 56700;
+			else
+				pS3->MaxClock = 50000;
+
+			pScrn->rgbBits = 8;
 			break;
 		}
-	        pScrn->rgbBits = 6;
+
 		pS3->LoadPalette = S3GenericLoadPalette;
 	}
 
@@ -684,7 +678,7 @@ static Bool S3PreInit(ScrnInfoPtr pScrn, int flags)
 		return FALSE;
 	}
 
-	if (pS3->SWCursor)
+	if (!pS3->HWCursor)
 		pS3->CursorInit = NULL;
 
 	pS3->RefClock = S3GetRefClock(pScrn);
@@ -749,6 +743,7 @@ static Bool S3ScreenInit(int scrnIndex, ScreenPtr pScreen, int argc,
 {
 	ScrnInfoPtr pScrn = xf86Screens[pScreen->myNum];
 	S3Ptr pS3 = S3PTR(pScrn);
+	BoxRec ScreenArea;
 
 	pScrn->fbOffset = 0;
 
@@ -811,31 +806,71 @@ static Bool S3ScreenInit(int scrnIndex, ScreenPtr pScreen, int argc,
         miInitializeBackingStore(pScreen);
         xf86SetBackingStore(pScreen);
 
+	/* framebuffer manager setup */
+	ScreenArea.x1 = 0;
+	ScreenArea.y1 = 0;
+	ScreenArea.x2 = pScrn->displayWidth;
+	ScreenArea.y2 = (pScrn->videoRam * 1024) / pS3->s3BppDisplayWidth;
+
+	if (!xf86InitFBManager(pScreen, &ScreenArea)) {
+		xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
+			   "Memory manager initialization to (%d,%d) (%d,%d) failed\n",
+			   ScreenArea.x1, ScreenArea.y1,
+			   ScreenArea.x2, ScreenArea.y2);
+		return FALSE;
+	} else
+		xf86DrvMsg(pScrn->scrnIndex, X_INFO,
+			   "Memory manager initialized to (%d,%d) (%d,%d)\n",
+			   ScreenArea.x1, ScreenArea.y1, 
+			   ScreenArea.x2, ScreenArea.y2);
+	
+
+	/* 2D acceleration setup */
+
+	if (pS3->NoAccel)
+		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, 
+			   "Acceleration disabled (by option)\n");
+
+	/* It seems that acceleration isn't supported for 24-bit packed
+	   colour. Disable it for S3 Trio64V2 */
+	if (!pS3->NoAccel && (pScrn->bitsPerPixel == 24) &&
+	    (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)) {
+		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Acceleration isn't supported for 24 bpp. Disabled.\n");
+		pS3->NoAccel = TRUE;
+	}
+
 	if (!pS3->NoAccel) {
-		if (pS3->S3NewMMIO) {
+		if (pS3->S3NewMMIO)
 			if (S3AccelInitNewMMIO(pScreen)) {
-				xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Acceleration enabled\n");
-				xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Using NewMMIO\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+					   "Acceleration enabled\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+					   "Using NewMMIO\n");
 			} else {
-				xf86DrvMsg(pScrn->scrnIndex, X_ERROR, "Acceleration initialization failed\n");
-				xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Acceleration disabled\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_ERROR, 
+					   "Acceleration initialization failed\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+					   "Acceleration disabled\n");
 			}
-		} else {
+		else {
 			if (S3AccelInitPIO(pScreen)) {
-				xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Acceleration enabled\n");
-				xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Using PIO\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+					   "Acceleration enabled\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+					   "Using PIO\n");
 			} else {
-				xf86DrvMsg(pScrn->scrnIndex, X_ERROR, "Acceleration initialization failed\n");
-				xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Acceleration disabled\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_ERROR, 
+					   "Acceleration initialization failed\n");
+				xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+					   "Acceleration disabled\n");
 			}
 		}
-	} else {
-		xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Acceleration disabled by option\n");
 	}
-
+	
         miDCInitialize(pScreen, xf86GetPointerScreenFuncs());
 
-	/* hw cursor setup */
+	/* HW cursor setup */
+
 	if (pS3->CursorInit) {
 		if (pS3->CursorInit(pScreen))
 			xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Using HW cursor\n");
@@ -869,10 +904,54 @@ static Bool S3ScreenInit(int scrnIndex, ScreenPtr pScreen, int argc,
 	pScrn->racIoFlags = pScrn->racMemFlags = RAC_COLORMAP
 	    | RAC_FB | RAC_VIEWPORT | RAC_CURSOR;
 
-#if 0
-	S3InitVideo(pScreen);
-#endif
+	if (pS3->SlowEDODRAM)
+		xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+			   "SlowEDODRAM: Setting 2-cycle EDO\n");
+
+	if (pS3->SlowVRAM)
+		xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+			   "SlowVRAM: -RAS low time is 4.5 MCLKs\n");
+
+	if (pS3->SlowDRAM)
+		xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+			   "SlowDRAM: -RAS precharge time is 3.5 MCLKs\n");
+
+	if (pS3->SlowDRAMRefresh)
+		xf86DrvMsg(pScrn->scrnIndex, X_INFO, 
+			   "SlowDRAMRefresh: three refresh cycles per scanline\n");
+
+	/* XVideo setup */
+
+	if (pS3->XVideo) {
+		if (!pS3->hasStreams) {
+			xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Overlay video isn't supported by video hardware. Disabled.\n");
+			pS3->XVideo = FALSE;
+		} else if (pScrn->bitsPerPixel < 16) {
+			xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "Overlay video isn't supported for %d bpp. Disabled.\n", pScrn->bitsPerPixel);
+			pS3->XVideo = FALSE;
+		}
+	} else
+		xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, 
+			   "Overlay video disabled by option\n");
 	
+	/* At present time we support XV only for chips with New MMIO */
+	if ((pS3->XVideo) && (pS3->S3NewMMIO))
+		S3InitVideo(pScreen);
+	
+	switch (pScrn->bitsPerPixel) {
+	case 8:
+		pS3->Streams_FIFO = FIFO_PS16_SS8;
+		break;
+	case 15:
+	case 16:
+		pS3->Streams_FIFO = FIFO_PS12_SS12;
+		break;
+	case 24:
+	case 32: 
+		pS3->Streams_FIFO = FIFO_PS8_SS16;
+		break;
+	}
+		
         return TRUE;
 }                 
 
@@ -1108,8 +1187,11 @@ static int S3GetPixMuxShift(ScrnInfoPtr pScrn)
 	else if (pS3->Chipset == PCI_CHIP_TRIO || 
 	         pS3->Chipset == PCI_CHIP_TRIO64UVP || 
 	         pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)
-		shift = -(pS3->s3Bpp >> 1);
-
+		if (pS3->s3Bpp == 2)
+			shift = -1; 
+		else 
+			shift = 0;
+				
 	return shift;
 }
 
@@ -1134,13 +1216,22 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	pS3->s3ScissB = ((pScrn->videoRam * 1024) / pS3->s3BppDisplayWidth) - 1;
 	pS3->s3ScissR = pScrn->displayWidth - 1;
 
-	if (mode->HTotal == mode->CrtcHTotal) {
+	/* 
+	   Set correct blanking for S3 Trio64V2. It's also needed
+	   to clear cr33_5.
+	*/
+	if (pS3->Chipset = PCI_CHIP_TRIO64V2_DXGX)
+		mode->CrtcHBlankStart = mode->CrtcHDisplay + 8;
+
+	if ((mode->HTotal == mode->CrtcHTotal) && (pS3->pixMuxShift != 0)) {
 		if (pS3->pixMuxShift > 0) {
 			/* XXX hack */
 /* 			mode->Flags |= V_PIXMUX; */
 
 			mode->CrtcHTotal >>= pS3->pixMuxShift;
 			mode->CrtcHDisplay >>= pS3->pixMuxShift;
+			mode->CrtcHBlankStart >>= pS3->pixMuxShift;
+			mode->CrtcHBlankEnd >>= pS3->pixMuxShift;
 			mode->CrtcHSyncStart >>= pS3->pixMuxShift;
 			mode->CrtcHSyncEnd >>= pS3->pixMuxShift;
 			mode->CrtcHSkew >>= pS3->pixMuxShift;
@@ -1149,6 +1240,8 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 
 			mode->CrtcHTotal <<= -pS3->pixMuxShift;
 			mode->CrtcHDisplay <<= -pS3->pixMuxShift;
+			mode->CrtcHBlankStart <<= -pS3->pixMuxShift;
+			mode->CrtcHBlankEnd <<= -pS3->pixMuxShift;
 			mode->CrtcHSyncStart <<= -pS3->pixMuxShift;
 			mode->CrtcHSyncEnd <<= -pS3->pixMuxShift;
 			mode->CrtcHSkew <<= -pS3->pixMuxShift;
@@ -1202,13 +1295,12 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	}
 
 	/* We need to set this first - S3 *is* broken */
-	outw(vgaCRIndex, (pVga->CRTC[17] << 8) | 17);
+	outw(vgaCRIndex, (pVga->CRTC[17] << 8) | 17); 
 	for(r=0; r<25; r++)
 		outw(vgaCRIndex, (pVga->CRTC[r] << 8) | r);
 
-	for(r=0; r<9; r++) {
+	for(r=0; r<9; r++)
 		outw(0x3ce, (pVga->Graphics[r] << 8) | r);
-	}
 
 	inb(vgaIOBase + 0x0a);
 
@@ -1235,19 +1327,14 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	if ((pS3->Chipset == PCI_CHIP_964_0) ||
 	    (pS3->Chipset == PCI_CHIP_964_1))
 		new->cr33 = 0x20;
+	else if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX)
+		new->cr33 &= ~0x20;
 	outb(vgaCRReg, new->cr33);
 	
 	new->cr34 = 0x10;
 	outb(vgaCRIndex, 0x34);
 	outb(vgaCRReg, new->cr34);
 
-	if (pS3->SlowDRAMRefresh)
-		new->cr3a = 0xb7;
-	else
-		new->cr3a = 0xb5;
-	outb(vgaCRIndex, 0x3a);
-	outb(vgaCRReg, new->cr3a);
-
 	if (pS3->Chipset != PCI_CHIP_AURORA64VP) {
 		new->cr3b = (pVga->CRTC[0] + pVga->CRTC[4] + 1) / 2;
 		outb(vgaCRIndex, 0x3b);
@@ -1302,7 +1389,7 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 		tmp |= 0x10;
 		break;
 	case 24:
-		tmp |= 0x20;
+		tmp |= 0x20; /* there is no such value in spec s3.txt */
 		break;
 	case 32:
 		tmp |= 0x30;
@@ -1373,6 +1460,16 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	new->cr60 = n;
 	outb(vgaCRReg, new->cr60);
 
+	if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX) {
+		new->cr60 = 255;		
+		outb(vgaCRIndex, 0x60);
+		outb(vgaCRReg, new->cr60);		
+
+		new->cr54 = 31 << 3;
+		outb(vgaCRIndex, 0x54);
+		outb(vgaCRReg, new->cr54);		
+	}
+
 	outb(vgaCRIndex, 0x55);
 	new->cr55 = (inb(vgaCRReg) & 0x08) | 0x40;
 	outb(vgaCRReg, new->cr55);
@@ -1387,10 +1484,10 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	{
 		int i;
 		unsigned int j;
-
-		i = ((((mode->CrtcHTotal >> 3) - 5) & 0x100) >> 8) 	|
-		    ((((mode->CrtcHDisplay >> 3) - 1) & 0x100) >> 7) 	|
-		    ((((mode->CrtcHSyncStart >> 3) - 1) & 0x100) >> 6)	|
+		
+		i = ((((mode->CrtcHTotal >> 3) - 5) & 0x100) >> 8)      |
+		    ((((mode->CrtcHDisplay >> 3) - 1) & 0x100) >> 7)    |
+		    ((((mode->CrtcHSyncStart >> 3) - 1) & 0x100) >> 6)  |
 		    ((mode->CrtcHSyncStart & 0x800) >> 7);
 		if ((mode->CrtcHSyncEnd >> 3) - (mode->CrtcHSyncStart >> 3) > 64)
 			i |= 0x08;
@@ -1452,6 +1549,33 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 		outb(vgaCRReg, new->cr42);
 	}
 
+	if (pS3->Chipset == PCI_CHIP_TRIO64V2_DXGX) {
+		unsigned char a;
+
+		outb(vgaCRIndex, 0x67);
+		a = inb(vgaCRReg) & 0xfe;
+
+		switch (pScrn->depth) {
+		case 8:
+			break;
+		case 15:
+			a |= (3 << 4);
+			break;
+		case 16:
+			a |= (5 << 4);
+			break;
+		case 24:
+			a |= (13 << 4);
+			break;		 
+		}
+
+		if (pS3->hasStreams)
+			a |= (3 << 2);
+
+		WaitVSync(); /* Wait for VSync before setting mode */
+		outb(vgaCRReg, a);
+	}	    
+
 	if (pS3->Chipset == PCI_CHIP_968) {
 		unsigned char a;
 
@@ -1513,6 +1637,35 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 		new->cr66 |= 0x80;
 	outb(vgaCRReg, new->cr66);
 
+	if (pS3->SlowDRAMRefresh)
+		new->cr3a = 0xb7;
+	else
+		new->cr3a = 0xb5;
+	outb(vgaCRIndex, 0x3a);
+	outb(vgaCRReg, new->cr3a);
+
+	/* 
+	   Set 3.5 MCLKs for -RAS low, 2.5 MCLKs for -RAS precharge, 
+	   disable -CAS/-OE adjustment. It seems that cr68 has different 
+	   format for 96x and TRIOs
+	*/
+	if (!((pS3->Chipset == PCI_CHIP_968) || 
+	      (pS3->Chipset == PCI_CHIP_964_0) ||
+	      (pS3->Chipset == PCI_CHIP_964_1))) {
+
+		outb(vgaCRIndex, 0x39);
+		outb(vgaCRReg, 0xa5);
+
+		outb(vgaCRIndex, 0x68);
+		tmp = inb(vgaCRReg) & ~0x0f;
+		outb(vgaCRReg, tmp | 0x0f);
+
+		/* Enable 1-cycle EDO access */
+		outb(vgaCRIndex, 0x36);
+		tmp = inb(vgaCRReg);
+		outb(vgaCRReg, tmp & 0xf3);
+	}
+	
 	if (pS3->SlowVRAM) {
 		/*
 		 * some Diamond Stealth 64 VRAM cards have a problem with
@@ -1530,7 +1683,7 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	if (pS3->SlowDRAM) {
 		/*
 		 * fixes some pixel errors for a SPEA Trio64V+ card
-		 * increas -RAS precharge timing from 2.5 MCLKs
+		 * increase -RAS precharge timing from 2.5 MCLKs
 		 * to 3.5 MCLKs
 		 */
 		outb(vgaCRIndex, 0x39);
@@ -1543,7 +1696,7 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	if (pS3->SlowEDODRAM) {
 		/*
 		 * fixes some pixel errors for a SPEA Trio64V+ card
-		 * increas from 1-cycle to 2-cycle EDO mode
+		 * increase from 1-cycle to 2-cycle EDO mode
 		 */
 		outb(vgaCRIndex, 0x39);
 
@@ -1626,13 +1779,8 @@ static Bool S3ModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 	WaitQueue(5);
 	SET_SCISSORS(0, 0, pS3->s3ScissR, pS3->s3ScissB);
 
-	outb(vgaCRIndex, 0x6f);
-
-#if 0
-	if (((pScrn->bitsPerPixel == 16) || 
-	     (pScrn->bitsPerPixel == 24)) && (pS3->S3NewMMIO))
+	if (pS3->hasStreams)
 		S3InitStreams(pScrn, mode);
-#endif
 
 	return TRUE;
 }
@@ -1784,7 +1932,7 @@ void S3Regdump(ScrnInfoPtr pScrn)
         S3Ptr pS3 = S3PTR(pScrn);
         int vgaCRIndex = pS3->vgaCRIndex, vgaCRReg = pS3->vgaCRReg;
 
-#if 1
+#if 0
 	outb(vgaCRIndex, 0x31);
 	ErrorF("cr31 = 0x%x\n", inb(vgaCRReg));
 	outb(vgaCRIndex, 0x32);
diff --git a/src/s3_reg.h b/src/s3_reg.h
index 64e3633..8ef9fb5 100644
--- a/src/s3_reg.h
+++ b/src/s3_reg.h
@@ -139,6 +139,16 @@ the
 #define VECDIR_315      0x00e0
 #define SSVDRAW         0x0010
 
+/* 
+ * Some values for Streams FIFO.
+ * primary stream threshold | secondary stream threshold | 
+ * secondary stream slots (can be 0, 8, 12, 16, 24) from 24 total
+ */
+#define FIFO_PS0_SS24  (0 << 10) | (12 << 5) | 24
+#define FIFO_PS8_SS16  (8 << 10) | (12 << 5) | 16
+#define FIFO_PS12_SS12 (6 << 10) | (8 << 5) | 12
+#define FIFO_PS16_SS8  (8 << 10) | (4 << 5) | 8
+#define FIFO_PS24_SS0  (12 << 10) | (0 << 5) | 0
 
 #define S3_OUTW(p,n)		outw(p, n)
 #define S3_OUTL(p,n)		outl(p, n)
@@ -153,6 +163,10 @@ the
 		while(inw(GP_STAT) & GPBUSY);	\
 	} while(0)
 
+#define WaitVSync() {					\
+	while (inb(0x3da) & 8);                         \
+	while (!(inb(0x3da) & 8));			\
+}
 
 #ifdef S3_NEWMMIO
 #include "newmmio.h"
@@ -162,13 +176,15 @@ the
  */
 #define SET_BLEND_CNTL(val)     ((mmtr)s3MmioMem)->streams_regs.regs.blend_cntl = (val)
 #define SET_PSTREAM_CNTL(val)   ((mmtr)s3MmioMem)->streams_regs.regs.prim_stream_cntl = (val)
-#define SET_PSTREAM_FBADDR(val) ((mmtr)s3MmioMem)->streams_regs.regs.prim_fbaddr0 = (val)
+#define SET_PSTREAM_FBADDR0(val) ((mmtr)s3MmioMem)->streams_regs.regs.prim_fbaddr0 = (val)
+#define SET_PSTREAM_FBADDR1(val) ((mmtr)s3MmioMem)->streams_regs.regs.prim_fbaddr1 = (val)
 #define SET_PSTREAM_STRIDE(val) ((mmtr)s3MmioMem)->streams_regs.regs.prim_stream_stride = (val)
 #define SET_PSTREAM_START(val)  ((mmtr)s3MmioMem)->streams_regs.regs.prim_start_coord = (val)
 #define SET_PSTREAM_WIND(val)   ((mmtr)s3MmioMem)->streams_regs.regs.prim_window_size = (val)
 #define SET_SSTREAM_CNTL(val)   ((mmtr)s3MmioMem)->streams_regs.regs.second_stream_cntl = (val)
 #define SET_SSTRETCH(val)       ((mmtr)s3MmioMem)->streams_regs.regs.second_stream_stretch = (val)
-#define SET_SSTREAM_FBADDR(val) ((mmtr)s3MmioMem)->streams_regs.regs.second_fbaddr0 = (val)
+#define SET_SSTREAM_FBADDR0(val) ((mmtr)s3MmioMem)->streams_regs.regs.second_fbaddr0 = (val)
+#define SET_SSTREAM_FBADDR1(val) ((mmtr)s3MmioMem)->streams_regs.regs.second_fbaddr1 = (val)
 #define SET_SSTREAM_STRIDE(val) ((mmtr)s3MmioMem)->streams_regs.regs.second_stream_stride = (val)
 #define SET_SSTREAM_START(val)  ((mmtr)s3MmioMem)->streams_regs.regs.second_start_coord = (val)
 #define SET_SSTREAM_WIND(val)   ((mmtr)s3MmioMem)->streams_regs.regs.second_window_size = (val)
@@ -178,6 +194,7 @@ the
 #define SET_CHROMA_KEY(val)     ((mmtr)s3MmioMem)->streams_regs.regs.col_chroma_key_cntl = (val)
 #define SET_DOUBLE_BUFFER(val)  ((mmtr)s3MmioMem)->streams_regs.regs.double_buffer = (val)
 #define SET_OPAQUE_OVERLAY(val) ((mmtr)s3MmioMem)->streams_regs.regs.opaq_overlay_cntl = (val)
+#define SET_FIFO_CNTL(val)     ((mmtr)s3MmioMem)->streams_regs.regs.streams_fifo = (val)
 
 #else
 
diff --git a/src/s3_video.c b/src/s3_video.c
index e320329..ad68de0 100644
--- a/src/s3_video.c
+++ b/src/s3_video.c
@@ -29,8 +29,6 @@
 #include "config.h"
 #endif
 
-#define S3_NEWMMIO	/* previously defined in Imakefile in monolith */
-
 #include "xf86.h"
 #include "xf86_OSproc.h"
 
@@ -59,7 +57,7 @@ static int  S3PutImage(ScrnInfoPtr, short, short, short, short, short,
 static int  S3QueryImageAttributes(ScrnInfoPtr, int, unsigned short *,
                         	   unsigned short *,  int *, int *);
 static void S3ResetVideoOverlay(ScrnInfoPtr);
-
+static FBLinearPtr S3XVMemAlloc(ScrnInfoPtr pScrn, pointer pVideo, int size);
 
 
 void S3InitVideo(ScreenPtr pScreen)
@@ -70,12 +68,8 @@ void S3InitVideo(ScreenPtr pScreen)
 	XF86VideoAdaptorPtr newAdaptor = NULL;
 	int num_adaptors;
 
-	if (((pScrn->bitsPerPixel == 16) ||
-	     (pScrn->bitsPerPixel == 24)) && (pS3->S3NewMMIO)) {
-		newAdaptor = S3SetupImageVideoOverlay(pScreen);
-		xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Using overlay video\n");
-	} else
-		return;
+	newAdaptor = S3SetupImageVideoOverlay(pScreen);
+	xf86DrvMsg(pScrn->scrnIndex, X_INFO, "Using overlay video\n");
 
 	num_adaptors = xf86XVListGenericAdaptors(pScrn, &adaptors);
 
@@ -92,7 +86,7 @@ void S3InitVideo(ScreenPtr pScreen)
                 		newAdaptors[num_adaptors] = newAdaptor;
                 		adaptors = newAdaptors;
                 		num_adaptors++;
-            		} 
+           		} 
         	}
     	}
 
@@ -122,12 +116,42 @@ static XF86VideoEncodingRec DummyEncoding[2] =
 };
 
 
-
+static FBLinearPtr S3XVMemAlloc(ScrnInfoPtr pScrn, pointer pVideo, int size)
+{
+	FBLinearPtr pLinear = (FBLinearPtr)pVideo;
+	ScreenPtr pScreen = pScrn->pScreen;
+
+	if (pLinear) {
+		if ((pLinear->size >= size) ||
+		    xf86ResizeOffscreenLinear(pLinear, size)) {
+			pLinear->MoveLinearCallback = NULL;
+			pLinear->RemoveLinearCallback = NULL;
+			return pLinear;
+		}
+		xf86FreeOffscreenLinear(pLinear);
+	}
+	pLinear = xf86AllocateOffscreenLinear(pScreen, size, 16, 
+					      NULL, NULL, NULL);
+	
+	if (!pLinear) {
+		int maxSize;
+		
+		xf86QueryLargestOffscreenLinear(pScreen, &maxSize, 16,
+						PRIORITY_EXTREME);
+		if (maxSize < size)
+			return NULL;
+		
+		xf86PurgeUnlockedOffscreenAreas(pScreen);
+		pLinear = xf86AllocateOffscreenLinear(pScreen, size, 16, 
+						      NULL, NULL, NULL);
+	}
+	return pLinear;
+}
 
 static XF86VideoFormatRec Formats[NUM_FORMATS_TEXTURE] =
 {
-  	/*{15, TrueColor},*/ {16, TrueColor}, {24, TrueColor} /* ,
-    	{15, DirectColor}*/, {16, DirectColor}, {24, DirectColor}
+  	{16, TrueColor}, {24, TrueColor},
+	{16, DirectColor}, {24, DirectColor}
 };
 
 
@@ -188,7 +212,7 @@ static XF86VideoAdaptorPtr S3AllocAdaptor(ScrnInfoPtr pScrn)
         	return NULL;
    
     	if(!(pPriv = xcalloc(1, sizeof(S3PortPrivRec)  +
-        	                (sizeof(DevUnion) * S3_MAX_PORTS))))
+			     (sizeof(DevUnion) * S3_MAX_PORTS))))
     	{
         	xfree(adapt);
         	return NULL;
@@ -199,15 +223,16 @@ static XF86VideoAdaptorPtr S3AllocAdaptor(ScrnInfoPtr pScrn)
     	for(i = 0; i < S3_MAX_PORTS; i++)
         	adapt->pPortPrivates[i].val = i;
 
-	pPriv->colorKey = (1 << pScrn->offset.red) | (1 << pScrn->offset.green) |
-			  (((pScrn->mask.blue >> pScrn->offset.blue) - 1) << pScrn->offset.blue);
+	pPriv->colorKey = (1 << pScrn->offset.red) | 
+		(1 << pScrn->offset.green) |
+		(((pScrn->mask.blue >> pScrn->offset.blue) - 1) << pScrn->offset.blue);
 
 	pPriv->videoStatus = 0;
 	pPriv->lastPort = -1;
 
 	pS3->adaptor = adapt;
 	pS3->portPrivate = pPriv;
-
+	
 	return adapt;
 }
 
@@ -262,11 +287,13 @@ static void S3StopVideo(ScrnInfoPtr pScrn, pointer data, Bool exit)
 	REGION_EMPTY(pScrn->pScreen, &pPriv->clip);
 
 	if (exit) {
+		SET_FIFO_CNTL(0x00080000 | FIFO_PS24_SS0);
+
 		if (pPriv->videoStatus & CLIENT_VIDEO_ON)
 			SET_BLEND_CNTL(0x01000000);
-
+		
 		if (pPriv->area) {
-			xf86FreeOffscreenArea(pPriv->area);
+			xf86FreeOffscreenLinear(pPriv->area);
 	        	pPriv->area = NULL;
 		}
 
@@ -274,47 +301,6 @@ static void S3StopVideo(ScrnInfoPtr pScrn, pointer data, Bool exit)
 	}
 }
 
-
-static FBAreaPtr S3AllocateMemory(ScrnInfoPtr pScrn, FBAreaPtr area,
-				  int numlines)
-{
-	ScreenPtr pScreen;
-	FBAreaPtr new_area;
-
-   if(area) {
-        if((area->box.y2 - area->box.y1) >= numlines)
-           return area;
-            
-        if(xf86ResizeOffscreenArea(area, pScrn->displayWidth, numlines))
-           return area;
-     
-        xf86FreeOffscreenArea(area);
-   }
-      
-   pScreen = screenInfo.screens[pScrn->scrnIndex];
-  
-   new_area = xf86AllocateOffscreenArea(pScreen, pScrn->displayWidth,
-                                numlines, 0, NULL, NULL, NULL);
-
-   if(!new_area) {
-        int max_w, max_h;
-
-        xf86QueryLargestOffscreenArea(pScreen, &max_w, &max_h, 0,
-                        FAVOR_WIDTH_THEN_AREA, PRIORITY_EXTREME);
-           
-        if((max_w < pScrn->displayWidth) || (max_h < numlines))
-           return NULL;
-
-        xf86PurgeUnlockedOffscreenAreas(pScreen);
-        new_area = xf86AllocateOffscreenArea(pScreen, pScrn->displayWidth,
-                                numlines, 0, NULL, NULL, NULL);
-   }
-  
-   return new_area;  
-}
-
-
-
 static void S3DisplayVideoOverlay(ScrnInfoPtr pScrn, int id, int offset,
 				  short width, short height, int pitch,
 				  int x1, int y1, int x2, int y2,
@@ -330,139 +316,162 @@ static void S3DisplayVideoOverlay(ScrnInfoPtr pScrn, int id, int offset,
 	else
 		tmp = 2;
 
-	SET_SSTREAM_CNTL(tmp << 28 | 0x01000000 |
-			 ((((src_w-1)<<1)-(drw_w-1)) & 0xfff));
-	SET_SSTRETCH(((src_w - 1) & 0x7ff) | (((src_w-drw_w) & 0x7ff) << 16));
+	SET_SSTREAM_CNTL((tmp << 28) | 0x01000000 |
+			 ((((src_w - 1) << 1) - (drw_w - 1)) & 0xfff));
+	SET_SSTRETCH(((src_w - 1) & 0x7ff) | 
+		     (((src_w - drw_w) & 0x7ff) << 16));
 	SET_BLEND_CNTL(0x05000000);
-	SET_SSTREAM_FBADDR(offset & 0x3fffff);
+	SET_SSTREAM_FBADDR0(offset & 0x3fffff);
+	SET_SSTREAM_FBADDR1(offset & 0x3fffff);
 	SET_SSTREAM_STRIDE(pitch & 0xfff);
+	SET_SSTREAM_START(((dstBox->x1 + 1) << 16) | (dstBox->y1 + 1));
+	SET_SSTREAM_WIND((((drw_w - 1) << 16) | drw_h) & 0x7ff07ff);
 
 	SET_K1_VSCALE(src_h - 1);
 	SET_K2_VSCALE((src_h - drw_h) & 0x7ff);
-
-	SET_DDA_VERT((((~drw_h)-1)) & 0xfff);
-
-	SET_SSTREAM_START(((dstBox->x1 +1) << 16) | (dstBox->y1 +1));
-	SET_SSTREAM_WIND(( ((drw_w-1) << 16) | (drw_h ) ) & 0x7ff07ff);
+	SET_DDA_VERT(((~drw_h - 1)) & 0xfff);
 
 	SET_CHROMA_KEY(0x10000000 |
 		       ((pScrn->weight.red-1) << 24) |
 		       ((pPriv->colorKey & pScrn->mask.red) >> pScrn->offset.red) <<
-			(16 + 8-pScrn->weight.red) |
+		       (16 + 8-pScrn->weight.red) |
 		       ((pPriv->colorKey & pScrn->mask.green) >> pScrn->offset.green) <<
-			(8 + 8-pScrn->weight.green) |
+		       (8 + 8-pScrn->weight.green) |
 		       ((pPriv->colorKey & pScrn->mask.blue) >> pScrn->offset.blue) <<
-			(8-pScrn->weight.blue));
-}
+		       (8-pScrn->weight.blue));
 
+	SET_FIFO_CNTL(0x00080000 | pS3->Streams_FIFO);
+}
 
 static int S3PutImage(ScrnInfoPtr pScrn, short src_x, short src_y,
-		  short drw_x, short drw_y, short src_w, short src_h,
-		  short drw_w, short drw_h, int id, unsigned char *buf,
-		  short width, short height, Bool sync, RegionPtr clipBoxes,
-		  pointer data, DrawablePtr pDraw)
+		      short drw_x, short drw_y, short src_w, short src_h,
+		      short drw_w, short drw_h, int id, unsigned char *buf,
+		      short width, short height, Bool sync, 
+		      RegionPtr clipBoxes, pointer data, DrawablePtr pDraw)
 {
 	S3Ptr pS3 = S3PTR(pScrn);
 	S3PortPrivPtr pPriv = pS3->portPrivate;
    	INT32 x1, x2, y1, y2;
-   	unsigned char *dst_start; 
-   	int pitch, new_h, offset, offset2=0, offset3=0;
-   	int srcPitch, srcPitch2=0, dstPitch;
-   	int top, left, npixels, nlines;
+   	CARD8 *dst_start; 
+   	int pitch, new_h, offset, offsetV = 0, offsetU = 0;
+   	int srcPitch, srcPitchUV = 0, dstPitch, dstSize;
+   	int top, bottom, right, left, npixels, nlines;
    	BoxRec dstBox;
-   	CARD32 tmp;
-
-   /* Clip */
-   x1 = src_x;
-   x2 = src_x + src_w;
-   y1 = src_y;
-   y2 = src_y + src_h;
+	CARD32 tmp;
+	int cpp = (pScrn->bitsPerPixel + 7) >> 3;
+
+	/* Clip */
+	x1 = src_x;
+	x2 = src_x + src_w;
+	y1 = src_y;
+	y2 = src_y + src_h;
    
-   dstBox.x1 = drw_x;
-   dstBox.x2 = drw_x + drw_w;
-   dstBox.y1 = drw_y;
-   dstBox.y2 = drw_y + drw_h;
+	dstBox.x1 = drw_x;
+	dstBox.x2 = drw_x + drw_w;
+	dstBox.y1 = drw_y;
+	dstBox.y2 = drw_y + drw_h;
    
-   if(!xf86XVClipVideoHelper(&dstBox, &x1, &x2, &y1, &y2,
-			     clipBoxes, width, height))
-        return Success;
-        
-   /*if(!pMga->TexturedVideo) {*/
-        dstBox.x1 -= pScrn->frameX0;
-        dstBox.x2 -= pScrn->frameX0;
-        dstBox.y1 -= pScrn->frameY0;
-        dstBox.y2 -= pScrn->frameY0;
-        /*}*/
-        
-   pitch = pScrn->bitsPerPixel * pScrn->displayWidth >> 3;
-   dstPitch = ((width << 1) + 15) & ~15;
-   new_h = ((dstPitch * height) + pitch - 1) / pitch;
+	if(!xf86XVClipVideoHelper(&dstBox, &x1, &x2, &y1, &y2, clipBoxes, 
+				  width, height))
+		return Success;
    
-   switch(id) {
-   case FOURCC_YV12:
-   case FOURCC_I420:
-        srcPitch = (width + 3) & ~3;
-        offset2 = srcPitch * height;
-        srcPitch2 = ((width >> 1) + 3) & ~3;
-        offset3 = (srcPitch2 * (height >> 1)) + offset2;
-        break;
-   case FOURCC_UYVY:  
-   case FOURCC_YUY2:
-   default:
-        srcPitch = (width << 1);
-        break;
-   }
+	dstBox.x1 -= pScrn->frameX0;
+	dstBox.x2 -= pScrn->frameX0;
+	dstBox.y1 -= pScrn->frameY0;
+	dstBox.y2 -= pScrn->frameY0;
+        
+	/* requested size in bytes */
+	dstPitch = ((width << 1) + 15) & ~15;
+	dstSize = dstPitch * height;
+
+	pPriv->area = S3XVMemAlloc(pScrn, pPriv->area,
+				   (dstSize + cpp - 1) / cpp);
+	if (!pPriv->area) 
+		return BadAlloc;
+
+	offset = pPriv->area->offset * cpp;
+	dst_start = pS3->FBBase + offset;
+
+	switch (id) {
+        case FOURCC_YV12:
+        case FOURCC_I420:
+		left = (x1 >> 16) & ~1;
+		right = ((x2 + 0x1ffff) >> 16) & ~1;
+		top = (y1 >> 16) & ~1;
+		bottom = ((y2 + 0x1ffff) >> 16) & ~1;
+
+		if ((right < width) && ((x1 & 0x1ffff) <= (x2 & 0x1ffff)))
+			right += 2;
+		if ((bottom < height) && ((y1 & 0x1ffff) <= (y2 & 0x1ffff)))
+			bottom += 2;
+
+		npixels = right - left;
+		nlines = bottom - top;
+
+		srcPitch = (width + 3) & ~3;
+		offsetV = srcPitch * height;
+		srcPitchUV = ((width >> 1) + 3) & ~3;
+		offsetU = ((height >> 1) * srcPitchUV) + offsetV;
+
+		tmp = ((top >> 1) * srcPitchUV) + (left >> 1);
+		offsetV += tmp;
+		offsetU += tmp;
+
+		if (id == FOURCC_I420)
+		{
+			tmp = offsetV;
+			offsetV = offsetU;
+			offsetU = tmp;
+		}
+
+		dst_start += top * dstPitch + (left << 1);
+
+		xf86XVCopyYUV12ToPacked(buf + (top * srcPitch) + left,
+					buf + offsetV, buf + offsetU, 
+					dst_start, srcPitch, srcPitchUV,
+					dstPitch, nlines, npixels);
+		break;
+
+        case FOURCC_UYVY:
+        case FOURCC_YUY2:
+        default:
+		left = (x1 >> 16) & ~1;
+		right = ((x2 + 0x1ffff) >> 16) & ~1;
+		top = y1 >> 16;
+		bottom = (y2 + 0x0ffff) >> 16;
+
+		if ((right < width) && ((x1 & 0x1ffff) <= (x2 & 0x1ffff)))
+			right += 2;
+		if ((bottom < height) && ((y1 & 0x0ffff) <= (y2 & 0x0ffff)))
+			bottom++;
+
+		npixels = right - left;
+		nlines = bottom - top;
+
+		srcPitch = width << 1;
+		buf += (top * srcPitch) + (left << 1);
+		dst_start += top * dstPitch + (left << 1);
+
+		xf86XVCopyPacked(buf, dst_start, srcPitch, dstPitch, 
+				 nlines, npixels);
+		break;
+	}
+
+	/* update cliplist */
+	if(!REGION_EQUAL(pScrn->pScreen, &pPriv->clip, clipBoxes)) {
+		REGION_COPY(pScrn->pScreen, &pPriv->clip, clipBoxes);
+		/* draw these */
+		xf86XVFillKeyHelper(pScrn->pScreen, pPriv->colorKey, 
+				    clipBoxes);
+	}
    
-   if(!(pPriv->area = S3AllocateMemory(pScrn, pPriv->area, new_h)))
-        return BadAlloc;
+	offset += (left << 1) + (top * dstPitch);
+	S3DisplayVideoOverlay(pScrn, id, offset, width, height, dstPitch,
+			      x1, y1, x2, y2, &dstBox, 
+			      src_w, src_h, drw_w, drw_h);
    
-    /* copy data */
-    top = y1 >> 16;
-    left = (x1 >> 16) & ~1;
-    npixels = ((((x2 + 0xffff) >> 16) + 1) & ~1) - left;
-    left <<= 1;
-        
-    offset = pPriv->area->box.y1 * pitch;
-    dst_start = pS3->FBBase + offset + left + (top * dstPitch);
-    switch(id) {
-    case FOURCC_YV12:
-    case FOURCC_I420:
-        top &= ~1;    
-        tmp = ((top >> 1) * srcPitch2) + (left >> 2);
-        offset2 += tmp;
-        offset3 += tmp;
-        if(id == FOURCC_I420) {
-           tmp = offset2;
-           offset2 = offset3;
-           offset3 = tmp;
-        }
-        nlines = ((((y2 + 0xffff) >> 16) + 1) & ~1) - top;
-        xf86XVCopyYUV12ToPacked(buf + (top * srcPitch) + (left >> 1),
-                                buf + offset2, buf + offset3, dst_start,
-                                srcPitch, srcPitch2, dstPitch, nlines, npixels);
-        break; 
-    case FOURCC_UYVY:
-    case FOURCC_YUY2:
-    default:
-        buf += (top * srcPitch) + left;
-        nlines = ((y2 + 0xffff) >> 16) - top;
-        xf86XVCopyPacked(buf, dst_start, srcPitch, dstPitch, nlines, npixels);
-        break;
-    }
-        
-    /* update cliplist */
-        if(!REGION_EQUAL(pScrn->pScreen, &pPriv->clip, clipBoxes)) {
-            REGION_COPY(pScrn->pScreen, &pPriv->clip, clipBoxes);
-            /* draw these */
-	    xf86XVFillKeyHelper(pScrn->pScreen, pPriv->colorKey, clipBoxes);
-        }
-                          
-        offset += left + (top * dstPitch);
-        S3DisplayVideoOverlay(pScrn, id, offset, width, height, dstPitch,
-             x1, y1, x2, y2, &dstBox, src_w, src_h, drw_w, drw_h);
-    
-        pPriv->videoStatus = CLIENT_VIDEO_ON;
-    return Success;
+	pPriv->videoStatus = CLIENT_VIDEO_ON;
+	return Success;
 }       
 
 
@@ -472,54 +481,92 @@ static int S3QueryImageAttributes(ScrnInfoPtr pScrn, int id,
 				  int *pitches, int *offsets)
 {
 	int size, tmp;
-
-    *w = (*w + 1) & ~1;
-    if(offsets) offsets[0] = 0;
-    
-    switch(id) {
-    case FOURCC_YV12:
-    case FOURCC_I420:
-        *h = (*h + 1) & ~1;
-        size = (*w + 3) & ~3;
-        if(pitches) pitches[0] = size;
-        size *= *h;
-        if(offsets) offsets[1] = size;
-        tmp = ((*w >> 1) + 3) & ~3;
-        if(pitches) pitches[1] = pitches[2] = tmp;
-        tmp *= (*h >> 1);
-        size += tmp;
-        if(offsets) offsets[2] = size;
-        size += tmp;
-        break;
-    case FOURCC_UYVY:
-    case FOURCC_YUY2:
-    default:
-        size = *w << 1;
-        if(pitches) pitches[0] = size;
-        size *= *h;
-        break;
-    } 
+	
+	*w = (*w + 1) & ~1;
+	if(offsets) offsets[0] = 0;
+	
+	switch(id) {
+	case FOURCC_YV12:
+	case FOURCC_I420:
+		*h = (*h + 1) & ~1;
+		size = (*w + 3) & ~3;
+		if(pitches) pitches[0] = size;
+		size *= *h;
+		if(offsets) offsets[1] = size;
+		tmp = ((*w >> 1) + 3) & ~3;
+		if(pitches) pitches[1] = pitches[2] = tmp;
+		tmp *= (*h >> 1);
+		size += tmp;
+		if(offsets) offsets[2] = size;
+		size += tmp;
+		break;
+	case FOURCC_UYVY:
+	case FOURCC_YUY2:
+	default:
+		size = *w << 1;
+		if(pitches) pitches[0] = size;
+		size *= *h;
+		break;
+	} 
         
-    return size;
+	return size;
 }
 
 
-
 void S3InitStreams(ScrnInfoPtr pScrn, DisplayModePtr mode)
 {       
         S3Ptr pS3 = S3PTR(pScrn);
         unsigned int pst_wind = (mode->HDisplay-1) << 16 | (mode->VDisplay);
         
-        SET_PSTREAM_CNTL(0x05000000 & 0x77000000);
-        SET_CHROMA_KEY(0x00);
-        SET_SSTREAM_CNTL(0x03000000);
-        SET_BLEND_CNTL(0x01000000);
-        SET_PSTREAM_STRIDE((pScrn->displayWidth * 2) & 0x0fff);
-        SET_SSTREAM_STRIDE(0x01);
-        SET_OPAQUE_OVERLAY(0x40000000);
-        SET_PSTREAM_START(0x00010001);
+	WaitVSync();
+
+	switch (pScrn->bitsPerPixel) {
+	case 8: 
+		SET_PSTREAM_CNTL(0x00000000);
+		break;
+	case 15:
+		SET_PSTREAM_CNTL(0x03000000);
+		break;
+	case 16: 
+		SET_PSTREAM_CNTL(0x05000000);
+		break;
+	case 24:
+		SET_PSTREAM_CNTL(0x06000000);
+		break;
+	case 32:
+		SET_PSTREAM_CNTL(0x07000000);
+		break;
+	}
+
+	SET_PSTREAM_FBADDR0(0x00000000);
+	SET_PSTREAM_FBADDR1(0x00000000);
+	
+	SET_PSTREAM_STRIDE(pS3->s3BppDisplayWidth & 0x0fff);
+
         SET_PSTREAM_WIND(pst_wind & 0x07ff07ff);
+        SET_PSTREAM_START(0x00010001);
+
+        SET_CHROMA_KEY(0x00000000);
+	SET_SSTRETCH(0x00000000);
+        SET_BLEND_CNTL(0x01000000);
+	SET_DOUBLE_BUFFER(0x00000000);
+
+        SET_SSTREAM_CNTL(0x03000000);
+	SET_SSTREAM_FBADDR0(0x00000000);
+	SET_SSTREAM_FBADDR1(0x00000000);
+        SET_SSTREAM_STRIDE(0x00000001);
         SET_SSTREAM_START(0x07ff07ff);
         SET_SSTREAM_WIND(0x00010001);
+
+        SET_OPAQUE_OVERLAY(0x40000000);
+	SET_K1_VSCALE(0x00000000);
+	SET_K2_VSCALE(0x00000000);
+	SET_DDA_VERT(0x00000000);
+
+	/*
+	  ps thr | ss thr | ss fifo slots
+	  set primary stream FIFO to 24 slots and 12 slots for threshold 
+	*/
+	SET_FIFO_CNTL(0x00080000 | FIFO_PS24_SS0);
 }