Greatly improve the correctness and performance of the MGA render

implementation. Includes cache flushing to prevent bad first reads of
    the framebuffer, fixing blending of many formats, falling back on many
    unsupported operations, and falling back early to prevent migration.
    Passes all of rendercheck except some of the blend (!) tests.

hw/kdrive/mga/g400_common.h

@@ -51,6 +51,7 @@
 #    define MGA_SRC_DST_ALPHA                	(0x00000006)
 #    define MGA_SRC_ONE_MINUS_DST_ALPHA      	(0x00000007)
 #    define MGA_SRC_SRC_ALPHA_SATURATE       	(0x00000008)
+#    define MGA_SRC_BLEND_MASK			(0x0000000f)
 
 #    define MGA_DST_ZERO                     	(0x00000000)
 #    define MGA_DST_ONE                      	(0x00000010)
@@ -60,6 +61,7 @@
 #    define MGA_DST_ONE_MINUS_SRC_ALPHA      	(0x00000050)
 #    define MGA_DST_DST_ALPHA                	(0x00000060)
 #    define MGA_DST_ONE_MINUS_DST_ALPHA      	(0x00000070)
+#    define MGA_DST_BLEND_MASK			(0x00000070)
 
 #    define MGA_ALPHACHANNEL                 	(0x00000100)
 #    define MGA_VIDEOALPHA                   	(0x00000200)
@@ -93,7 +95,7 @@
 #    define MGA_TW8AL         			(0x00000008)
 #    define MGA_TW422         			(0x0000000A)
 #    define MGA_TW422UYVY     			(0x0000000B)
-#    define MGA_PITCHEXT			(0x00000100)
+#    define MGA_PITCHLIN			(0x00000100)
 #    define MGA_NOPERSPECTIVE			(0x00200000)
 #    define MGA_TAKEY           	     	(0x02000000)
 #    define MGA_TAMASK				(0x04000000)
@@ -168,14 +170,8 @@
 
 /* Composition Prototypes. MMIO Access */
 Bool
-mgaPrepareBlend(int op, PicturePtr pSrcPicture, PicturePtr pDstPicture,
+mgaCheckComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
-		PixmapPtr pSrc, PixmapPtr pDst);
+		  PicturePtr pDstPicture);
-
-void
-mgaBlend(int srcX, int srcY, int dstX, int dstY, int width, int height);
-
-void
-mgaDoneBlend(void);
 
 Bool
 mgaPrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,

hw/kdrive/mga/g400_composite.c

@@ -34,37 +34,56 @@
 static PixmapPtr currentSrc;
 static PixmapPtr currentMask;
 
-static CARD32 mgaBlendOP[14] = {
+struct blendinfo {
-    /* Clear */
+	Bool dst_alpha;
-    MGA_SRC_ZERO	        | MGA_DST_ZERO,
+	Bool src_alpha;
-    /* Src */
+	CARD32 blend_cntl;
-    MGA_SRC_ONE			| MGA_DST_ZERO,
-    /* Dst */
-    MGA_SRC_ZERO	        | MGA_DST_ONE,
-    /* Over */
-    MGA_SRC_ALPHA               | MGA_DST_ONE_MINUS_SRC_ALPHA,
-    /* OverReverse */
-    MGA_SRC_ONE_MINUS_DST_ALPHA | MGA_DST_ONE,
-    /* In */
-    MGA_SRC_DST_ALPHA           | MGA_DST_ZERO,
-    /* InReverse */
-    MGA_SRC_ZERO	        | MGA_DST_SRC_ALPHA,
-    /* Out */
-    MGA_SRC_ONE_MINUS_DST_ALPHA | MGA_DST_ZERO,
-    /* OutReverse */
-    MGA_SRC_ZERO	        | MGA_DST_ONE_MINUS_SRC_ALPHA,
-    /* Atop */
-    MGA_SRC_DST_ALPHA	        | MGA_DST_ONE_MINUS_SRC_ALPHA,
-    /* AtopReverse */
-    MGA_SRC_ONE_MINUS_DST_ALPHA | MGA_DST_SRC_ALPHA,
-    /* Xor */
-    MGA_SRC_ONE_MINUS_DST_ALPHA | MGA_DST_ONE_MINUS_SRC_ALPHA,
-    /* Add */
-    MGA_SRC_ONE                 | MGA_DST_ONE,
-    /* Saturate */
-    MGA_SRC_SRC_ALPHA_SATURATE  | MGA_DST_ONE
 };
 
+static struct blendinfo mgaBlendOP[] = {
+	/* Clear */
+	{0, 0, MGA_SRC_ZERO			| MGA_DST_ZERO},
+	/* Src */
+	{0, 0, MGA_SRC_ONE			| MGA_DST_ZERO},
+	/* Dst */
+	{0, 0, MGA_SRC_ZERO			| MGA_DST_ONE},
+	/* Over */
+	{0, 1, MGA_SRC_ONE			| MGA_DST_ONE_MINUS_SRC_ALPHA},
+	/* OverReverse */
+	{1, 0, MGA_SRC_ONE_MINUS_DST_ALPHA	| MGA_DST_ONE},
+	/* In */
+	{1, 0, MGA_SRC_DST_ALPHA		| MGA_DST_ZERO},
+	/* InReverse */
+	{0, 1, MGA_SRC_ZERO			| MGA_DST_SRC_ALPHA},
+	/* Out */
+	{1, 0, MGA_SRC_ONE_MINUS_DST_ALPHA	| MGA_DST_ZERO},
+	/* OutReverse */
+	{0, 1, MGA_SRC_ZERO			| MGA_DST_ONE_MINUS_SRC_ALPHA},
+	/* Atop */
+	{1, 1, MGA_SRC_DST_ALPHA		| MGA_DST_ONE_MINUS_SRC_ALPHA},
+	/* AtopReverse */
+	{1, 1, MGA_SRC_ONE_MINUS_DST_ALPHA	| MGA_DST_SRC_ALPHA},
+	/* Xor */
+	{1, 1, MGA_SRC_ONE_MINUS_DST_ALPHA	| MGA_DST_ONE_MINUS_SRC_ALPHA},
+	/* Add */
+	{0, 0, MGA_SRC_ONE			| MGA_DST_ONE},
+};
+
+struct formatinfo {
+	int fmt;
+	CARD32 card_fmt;
+};
+
+static struct formatinfo texformats[] = {
+    {PICT_a8r8g8b8, MGA_TW32},
+    {PICT_x8r8g8b8, MGA_TW32},
+    {PICT_r5g6b5, MGA_TW16},
+    {PICT_a1r5g5b5, MGA_TW15},
+    {PICT_x1r5g5b5, MGA_TW15},
+    {PICT_a4r4g4b4, MGA_TW12},
+    {PICT_x4r4g4b4, MGA_TW12},
+    {PICT_a8, MGA_TW8A},
+};
 
 static int MGA_LOG2( int val )
 {
@@ -76,37 +95,32 @@ static int MGA_LOG2( int val )
     return ((1 << (ret-1)) == val) ? (ret-1) : ret;
 }
 
-
+static Bool
-Bool
+mgaCheckSourceTexture (int		tmu,
-mgaPrepareBlend (int		op,
+		       PicturePtr 	pPict)
-		 PicturePtr	pSrcPicture,
-		 PicturePtr	pDstPicture,
-		 PixmapPtr	pSrc,
-		 PixmapPtr	pDst)
 {
-    return mgaPrepareComposite (op, pSrcPicture, NULL, pDstPicture,
+    int w = pPict->pDrawable->width;
-				pSrc, NULL, pDst);
+    int h = pPict->pDrawable->height;
+    int i;
+    CARD32 texctl = 0;
+
+    if ((w > 2047) || (h > 2047))
+	MGA_FALLBACK(("Picture too large for composition (%dx%d)\n", w, h));
+
+    for (i = 0; i < sizeof(texformats) / sizeof(texformats[0]); i++) {
+	if (texformats[i].fmt == pPict->format) {
+	    texctl = texformats[i].card_fmt;
+	    break;
+	}
+    }
+    if (texctl == 0) {
+	    MGA_FALLBACK(("Picture is in unsupported format 0x%x\n"
+			  pPict->format));
     }
 
-void
+    return TRUE;
-mgaBlend (int	srcX,
-	  int	srcY,
-	  int	dstX,
-	  int	dstY,
-	  int	width,
-	  int	height)
-{
-    mgaComposite (srcX, srcY, 0, 0, dstX, dstY, width, height);
 }
 
-void
-mgaDoneBlend (void)
-{
-    mgaDoneComposite ();
-}
-
-
-
 static Bool
 PrepareSourceTexture (int		tmu,
 		      PicturePtr 	pSrcPicture,
@@ -115,39 +129,31 @@ PrepareSourceTexture (int		tmu,
     KdScreenPriv (pSrc->drawable.pScreen);  
     int mem_base=(int)pScreenPriv->screen->memory_base;
     int pitch = pSrc->devKind / (pSrc->drawable.bitsPerPixel >> 3); 
+    int i;
 
     int w = pSrc->drawable.width;
     int h = pSrc->drawable.height;
     int w_log2 = MGA_LOG2(w);
     int h_log2 = MGA_LOG2(h);
 
-    int texctl=MGA_PITCHEXT | ((pitch&0x7ff)<<9) | 
+    int texctl = MGA_PITCHLIN | ((pitch & (2048 - 1)) << 9) | 
-	MGA_TAKEY | MGA_CLAMPUV | MGA_NOPERSPECTIVE;  
+	MGA_CLAMPUV | MGA_NOPERSPECTIVE;  
     int texctl2=MGA_G400_TC2_MAGIC;
  
-    if ((w > 2047) || (h > 2047))
+    for (i = 0; i < sizeof(texformats) / sizeof(texformats[0]); i++) {
-	MGA_FALLBACK(("Picture too large for composition (%dx%d)\n", w, h));
+	if (texformats[i].fmt == pSrcPicture->format) {
-
+	    texctl |= texformats[i].card_fmt;
-    switch (pSrcPicture->format) {
-    case PICT_a8r8g8b8:
-    case PICT_x8r8g8b8:
-    case PICT_a8b8g8r8:
-    case PICT_x8b8g8r8:
-	texctl |= MGA_TW32;
 	    break;
-    case PICT_r5g6b5:
+	}
-    case PICT_b5g6r5:
+    }
-	texctl |= MGA_TW16;
+    if (PICT_FORMAT_A(pSrcPicture->format) != 0) {
-	break;
+	texctl |= MGA_TAKEY;
-    case PICT_a8:
+    } else {
-	texctl |= MGA_TW8A;
+	texctl |= MGA_TAMASK | MGA_TAKEY;
-	break;
-    default:
-	MGA_FALLBACK(("unsupported Picture format for composition (%x)\n",
-		      pSrcPicture->format));
     }
 
-    if (tmu == 1) texctl2 |= MGA_TC2_DUALTEX | MGA_TC2_SELECT_TMU1;
+    if (tmu == 1)
+	texctl2 |= MGA_TC2_DUALTEX | MGA_TC2_SELECT_TMU1;
   
     mgaWaitAvail (6);
     MGA_OUT32 (mmio, MGA_REG_TEXCTL2, texctl2);
@@ -167,6 +173,50 @@ PrepareSourceTexture (int		tmu,
     return TRUE;
 }
 
+Bool
+mgaCheckComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
+		  PicturePtr pDstPicture)
+{
+    if (op >= sizeof(mgaBlendOP) / sizeof(mgaBlendOP[0]))
+	MGA_FALLBACK(("unsupported op %x", op));
+    if (!mgaCheckSourceTexture (0, pSrcPicture))
+	return FALSE;
+    if (pSrcPicture->transform != NULL)
+	    MGA_FALLBACK(("Transformed src unsupported"));
+    if (pMaskPicture != NULL) {
+	if (PICT_FORMAT_A(pMaskPicture->format) == 0)
+	    MGA_FALLBACK(("Mask without alpha unsupported"));
+	if (!mgaCheckSourceTexture (1, pMaskPicture))
+	    return FALSE;
+	if (pMaskPicture->transform != NULL)
+	    MGA_FALLBACK(("Transformed mask unsupported"));
+    }
+
+    if (pMaskPicture->componentAlpha)
+	MGA_FALLBACK(("Component alpha unsupported"));
+
+    switch (pDstPicture->format) {
+    case PICT_a8:
+	MGA_FALLBACK(("render to A8 unsupported"));
+    }
+
+    return TRUE;
+}
+
+#define C_ARG1_CUR		0x0	
+#define C_ARG2_DIFFUSE		MGA_TDS_COLOR_ARG2_DIFFUSE
+#define C_ARG2_PREV		MGA_TDS_COLOR_ARG2_PREVSTAGE
+#define COLOR_MUL		MGA_TDS_COLOR_SEL_MUL
+#define COLOR_ARG1		MGA_TDS_COLOR_SEL_ARG1
+#define COLOR_ARG2		MGA_TDS_COLOR_SEL_ARG2
+#define A_ARG1_CUR		0x0
+#define A_ARG2_IGN		A_ARG2_DIFFUSE
+#define A_ARG2_DIFFUSE		MGA_TDS_ALPHA_ARG2_DIFFUSE
+#define A_ARG2_PREV		MGA_TDS_ALPHA_ARG2_PREVSTAGE
+#define ALPHA_MUL		MGA_TDS_ALPHA_SEL_MUL
+#define ALPHA_ARG1		MGA_TDS_ALPHA_SEL_ARG1
+#define ALPHA_ARG2		MGA_TDS_ALPHA_SEL_ARG2
+
 Bool
 mgaPrepareComposite (int		op,
 		     PicturePtr		pSrcPicture,
@@ -178,17 +228,7 @@ mgaPrepareComposite (int		op,
 {
     KdScreenPriv (pSrc->drawable.pScreen);  
     int mem_base=(int)pScreenPriv->screen->memory_base;
-    int cmd;
+    int cmd, blendcntl;
-
-    /* sometimes mgaPrepareComposite is given the same pointer for Src
-     * and Mask. PSrcPicture is an unsupported format (x8b8g8r8) whereas
-     * pMaskPicture is supported (a8r8g8b8). We Choose to render a
-     * simple blend and using pMaskPicture as the Source.
-     */
-    if (pMask == pSrc) {
-	pMask=NULL;
-	pMaskPicture=NULL;
-    }
 
     mgaWaitIdle();
     /* Init MGA (clipping) */
@@ -202,6 +242,18 @@ mgaPrepareComposite (int		op,
     MGA_OUT32 (mmio, MGA_REG_TMR7, 0);
     MGA_OUT32 (mmio, MGA_REG_TMR8, 0x10000); 
 
+    /* Initialize colors to 0, used in the src = A8 case */
+    mgaWaitAvail(9);
+    MGA_OUT32 (mmio, MGA_REG_DR4, 0); 
+    MGA_OUT32 (mmio, MGA_REG_DR6, 0);
+    MGA_OUT32 (mmio, MGA_REG_DR7, 0);
+    MGA_OUT32 (mmio, MGA_REG_DR8, 0); 
+    MGA_OUT32 (mmio, MGA_REG_DR10, 0); 
+    MGA_OUT32 (mmio, MGA_REG_DR11, 0);
+    MGA_OUT32 (mmio, MGA_REG_DR12, 0);
+    MGA_OUT32 (mmio, MGA_REG_DR14, 0); 
+    MGA_OUT32 (mmio, MGA_REG_DR15, 0);
+
     /* Destination flags */
     mgaWaitAvail (2);
     MGA_OUT32 (mmio, MGA_REG_DSTORG, ((int)pDst->devPrivate.ptr - mem_base));
@@ -215,16 +267,28 @@ mgaPrepareComposite (int		op,
 
     /* MultiTexture modulation */
     mgaWaitAvail (2);
-    MGA_OUT32 (mmio, MGA_REG_TDUALSTAGE0, MGA_TDS_COLOR_SEL_ARG1);
+    if (pSrcPicture->format == PICT_a8) {
+	/* C = 0	A = As */
+	MGA_OUT32 (mmio, MGA_REG_TDUALSTAGE0,
+		   C_ARG2_DIFFUSE | COLOR_ARG2 |
+		   A_ARG1_CUR | ALPHA_ARG1);
+    } else {
+	/* C = Cs	A = As */
+	MGA_OUT32 (mmio, MGA_REG_TDUALSTAGE0,
+		   C_ARG1_CUR | COLOR_ARG1 |
+		   A_ARG1_CUR | ALPHA_ARG1);
+    }
     if (pMask != NULL) {
 	if (PICT_FORMAT_A (pSrcPicture->format) == 0) {
 	    MGA_OUT32 (mmio, MGA_REG_TDUALSTAGE1, 
-		       MGA_TDS_COLOR_ARG2_PREVSTAGE | MGA_TDS_COLOR_SEL_ARG2 |
+		       C_ARG1_CUR | C_ARG2_PREV | COLOR_MUL |
-		       MGA_TDS_ALPHA_SEL_ARG1);
+		       A_ARG1_CUR | A_ARG2_IGN | ALPHA_ARG1 |
+		       MGA_TDS_COLOR_ARG1_REPLICATEALPHA);
 	} else {
 	    MGA_OUT32 (mmio, MGA_REG_TDUALSTAGE1, 
-		       MGA_TDS_COLOR_ARG2_PREVSTAGE | MGA_TDS_COLOR_SEL_ARG2 |
+		       C_ARG1_CUR | C_ARG2_PREV | COLOR_MUL |
-		       MGA_TDS_ALPHA_ARG2_PREVSTAGE | MGA_TDS_ALPHA_SEL_MUL);
+		       A_ARG1_CUR | A_ARG2_PREV | ALPHA_MUL |
+		       MGA_TDS_COLOR_ARG1_REPLICATEALPHA);
 	}
     } else {
 	MGA_OUT32 (mmio, MGA_REG_TDUALSTAGE1, 0);    
@@ -234,9 +298,17 @@ mgaPrepareComposite (int		op,
 	MGA_DWGCTL_SHIFTZERO | MGA_DWGCTL_SGNZERO | MGA_DWGCTL_ARZERO |
 	MGA_ATYPE_I;
     
+    blendcntl = mgaBlendOP[op].blend_cntl;
+    if (PICT_FORMAT_A(pDstPicture->format) == 0 && mgaBlendOP[op].dst_alpha) {
+	if ((blendcntl & MGA_SRC_BLEND_MASK) == MGA_SRC_DST_ALPHA)
+	    blendcntl = (blendcntl & ~MGA_SRC_BLEND_MASK) | MGA_SRC_ONE;
+	else if ((blendcntl & MGA_SRC_BLEND_MASK) == MGA_SRC_ONE_MINUS_DST_ALPHA)
+	    blendcntl = (blendcntl & ~MGA_SRC_BLEND_MASK) | MGA_SRC_ZERO;
+    }
+
     mgaWaitAvail (2);
     MGA_OUT32 (mmio, MGA_REG_DWGCTL, cmd);
-    MGA_OUT32 (mmio, MGA_REG_ALPHACTRL, MGA_ALPHACHANNEL | mgaBlendOP[op]);
+    MGA_OUT32 (mmio, MGA_REG_ALPHACTRL, MGA_ALPHACHANNEL | blendcntl);
     
     currentSrc=pSrc;
     currentMask=pMask;

hw/kdrive/mga/mga.h

@@ -56,6 +56,7 @@
 #define MGA_REG_YBOT		(0x1c9c)
 #define MGA_REG_FIFOSTATUS 	(0x1e10)
 #define MGA_REG_STATUS	 	(0x1e14)
+#define MGA_REG_CACHEFLUSH 	(0x1fff)
 #define MGA_REG_SRCORG 		(0x2cb4)
 #define MGA_REG_DSTORG 		(0x2cb8)
 

hw/kdrive/mga/mgadraw.c

@@ -68,7 +68,11 @@ mgaWaitAvail (int n)
 void
 mgaWaitIdle (void)
 {
-    while (MGA_IN32 (mmio, MGA_REG_STATUS) & 0x10000);
+    while (MGA_IN32 (mmio, MGA_REG_STATUS) & 0x10000)
+	;
+
+    mgaWaitAvail (1);
+    MGA_OUT32(mmio, MGA_REG_CACHEFLUSH, 0);
 }
 
 static void
@@ -268,9 +272,7 @@ mgaDrawInit (ScreenPtr pScreen)
     mgas->kaa.flags		= KAA_OFFSCREEN_PIXMAPS;
 
     if (card->attr.deviceID == MGA_G4XX_DEVICE_ID) {
-        mgas->kaa.PrepareBlend	= mgaPrepareBlend;
+	mgas->kaa.CheckComposite = mgaCheckComposite;
-	mgas->kaa.Blend		= mgaBlend;
-	mgas->kaa.DoneBlend	= mgaDoneBlend;
 	mgas->kaa.PrepareComposite = mgaPrepareComposite;
 	mgas->kaa.Composite	= mgaComposite;
 	mgas->kaa.DoneComposite	= mgaDoneComposite;