- Implement acceleration of Composite on R128 when there is no mask, no

    transformation, no repeat, and only certain ops are used.
- Add debugging output for software fallbacks for Composite.
- Allow pixmaps in offscreen that don't match root depth.
- Clean up some mess in ati_reg.h.
Many thanks to andersca for a lot of this code.

7 files changed, 469 insertions, 48 deletions

diff --git a/hw/kdrive/ati/Makefile.am b/hw/kdrive/ati/Makefile.am
index b7cb8896a..69c88f019 100644
--- a/hw/kdrive/ati/Makefile.am
+++ b/hw/kdrive/ati/Makefile.am
@@ -44,6 +44,7 @@ libati_a_SOURCES = 	\
 	ati.c		\
 	ati.h		\
 	ati_reg.h	\
+	r128_blendtmp.h	\
 	$(DRI_SOURCES)
 
 Xati_SOURCES = \
diff --git a/hw/kdrive/ati/ati_draw.c b/hw/kdrive/ati/ati_draw.c
index c4835b958..6e8575ef7 100644
--- a/hw/kdrive/ati/ati_draw.c
+++ b/hw/kdrive/ati/ati_draw.c
@@ -73,9 +73,41 @@ CARD8 ATIBltRop[16] = {
     /* GXset        */      0xff,         /* 1 */
 };
 
+static CARD32 R128BlendOp[] = {
+	/* Clear */
+	R128_ALPHA_BLEND_SRC_ZERO	 | R128_ALPHA_BLEND_DST_ZERO,
+	/* Src */
+	R128_ALPHA_BLEND_SRC_ONE	 | R128_ALPHA_BLEND_DST_ZERO,
+	/* Dst */
+	R128_ALPHA_BLEND_SRC_ZERO	 | R128_ALPHA_BLEND_DST_ONE,
+	/* Over */
+	R128_ALPHA_BLEND_SRC_ONE	 | R128_ALPHA_BLEND_DST_INVSRCALPHA,
+	/* OverReverse */
+	R128_ALPHA_BLEND_SRC_INVDSTALPHA | R128_ALPHA_BLEND_DST_ONE,
+	/* In */
+	R128_ALPHA_BLEND_SRC_DSTALPHA	 | R128_ALPHA_BLEND_DST_ZERO,
+	/* InReverse */
+	R128_ALPHA_BLEND_SRC_ZERO	 | R128_ALPHA_BLEND_DST_SRCALPHA,
+	/* Out */
+	R128_ALPHA_BLEND_SRC_INVDSTALPHA | R128_ALPHA_BLEND_DST_ZERO,
+	/* OutReverse */
+	R128_ALPHA_BLEND_SRC_ZERO	 | R128_ALPHA_BLEND_DST_INVSRCALPHA,
+	/* Atop */
+	R128_ALPHA_BLEND_SRC_DSTALPHA	 | R128_ALPHA_BLEND_DST_INVSRCALPHA,
+	/* AtopReverse */
+	R128_ALPHA_BLEND_SRC_INVDSTALPHA | R128_ALPHA_BLEND_DST_SRCALPHA,
+	/* Xor */
+	R128_ALPHA_BLEND_SRC_INVDSTALPHA | R128_ALPHA_BLEND_DST_INVSRCALPHA,
+	/* Add */
+	R128_ALPHA_BLEND_SRC_ONE	 | R128_ALPHA_BLEND_DST_ONE,
+};
+
 int copydx, copydy;
 int fifo_size;
 ATIScreenInfo *accel_atis;
+int src_pitch;
+int src_offset;
+int src_bpp;
 
 static void
 ATIWaitAvailMMIO(int n)
@@ -313,14 +345,34 @@ drmBufPtr ATIDMAGetBuffer()
 	buf->used = 0;
 	return buf;
 }
+#endif /* USE_DRI */
+
+static Bool
+R128GetDatatype(CARD32 format, CARD32 *type)
+{
+	switch (format) {
+	case PICT_a8r8g8b8:
+		*type = R128_DATATYPE_ARGB_8888;
+		return TRUE;
+	case PICT_r5g6b5:
+		*type = R128_DATATYPE_RGB_565;
+		return TRUE;
+	}
+
+	ErrorF ("Unsupported format: %x\n", format);
 
+	return FALSE;
+}
+
+#ifdef USE_DRI
 #define USE_DMA
 #include "ati_drawtmp.h"
-
+#include "r128_blendtmp.h"
 #endif /* USE_DRI */
 
 #undef USE_DMA
 #include "ati_drawtmp.h"
+#include "r128_blendtmp.h"
 
 static void
 ATIDoneSolid(void)
@@ -385,6 +437,11 @@ ATIDrawInit(ScreenPtr pScreen)
 		atis->kaa.Solid = ATISolidDMA;
 		atis->kaa.PrepareCopy = ATIPrepareCopyDMA;
 		atis->kaa.Copy = ATICopyDMA;
+		if (!atic->is_radeon && !atis->is_24bpp) {
+			atis->kaa.PrepareBlend = R128PrepareBlendDMA;
+			atis->kaa.Blend = R128BlendDMA;
+			atis->kaa.DoneBlend = R128DoneBlendDMA;
+		}
 	} else {
 #else
 	{
@@ -393,6 +450,11 @@ ATIDrawInit(ScreenPtr pScreen)
 		atis->kaa.Solid = ATISolidMMIO;
 		atis->kaa.PrepareCopy = ATIPrepareCopyMMIO;
 		atis->kaa.Copy = ATICopyMMIO;
+		if (!atic->is_radeon && !atis->is_24bpp) {
+			atis->kaa.PrepareBlend = R128PrepareBlendMMIO;
+			atis->kaa.Blend = R128BlendMMIO;
+			atis->kaa.DoneBlend = R128DoneBlendMMIO;
+		}
 	}
 	atis->kaa.DoneSolid = ATIDoneSolid;
 	atis->kaa.DoneCopy = ATIDoneCopy;
diff --git a/hw/kdrive/ati/ati_drawtmp.h b/hw/kdrive/ati/ati_drawtmp.h
index 890800d2a..6573aef68 100644
--- a/hw/kdrive/ati/ati_drawtmp.h
+++ b/hw/kdrive/ati/ati_drawtmp.h
@@ -25,7 +25,8 @@
 
 #ifdef USE_DMA
 #define TAG(x)		x##DMA
-#define LOCALS		RING_LOCALS
+#define LOCALS		(void)atic; \
+			RING_LOCALS
 #define BEGIN(x)	BEGIN_RING(x * 2)
 #define OUT_REG(reg, val) OUT_RING_REG(reg, val)
 #define END()		ADVANCE_RING()
@@ -43,7 +44,7 @@ TAG(ATISetup)(PixmapPtr pDst, PixmapPtr pSrc)
 	KdScreenPriv(pDst->drawable.pScreen);
 	ATIScreenInfo(pScreenPriv);
 	ATICardInfo(pScreenPriv);
-	int dst_offset, dst_pitch, src_offset = 0, src_pitch = 0;
+	int dst_offset, dst_pitch;
 	int bpp = pScreenPriv->screen->fb[0].bitsPerPixel;
 	LOCALS;
 
diff --git a/hw/kdrive/ati/ati_reg.h b/hw/kdrive/ati/ati_reg.h
index 776c5ed33..4b0341c3b 100644
--- a/hw/kdrive/ati/ati_reg.h
+++ b/hw/kdrive/ati/ati_reg.h
@@ -48,20 +48,23 @@
 #define RADEON_REG_DST_Y_X			0x1438
 #define RADEON_REG_DST_HEIGHT_WIDTH		0x143c
 #define RADEON_REG_DP_GUI_MASTER_CNTL		0x146c
-#define RADEON_REG_DP_BRUSH_FRGD_CLR		0x147c
 # define RADEON_GMC_SRC_PITCH_OFFSET_CNTL	(1    <<  0)
 # define RADEON_GMC_DST_PITCH_OFFSET_CNTL	(1    <<  1)
 # define RADEON_GMC_BRUSH_SOLID_COLOR		(13   <<  4)
 # define RADEON_GMC_BRUSH_NONE			(15   <<  4)
 # define RADEON_GMC_SRC_DATATYPE_COLOR		(3    << 12)
 # define RADEON_DP_SRC_SOURCE_MEMORY		(2    << 24)
+# define RADEON_GMC_3D_FCN_EN			(1    << 27)
 # define RADEON_GMC_CLR_CMP_CNTL_DIS		(1    << 28)
 # define RADEON_GMC_AUX_CLIP_DIS		(1    << 29)
+#define RADEON_REG_DP_BRUSH_FRGD_CLR		0x147c
+#define RADEON_REG_DST_WIDTH_HEIGHT		0x1598
+#define RADEON_REG_CLR_CMP_CNTL			0x15c0
+#define RADEON_REG_AUX_SC_CNTL			0x1660
 #define RADEON_REG_DP_CNTL			0x16c0
 # define RADEON_DST_X_LEFT_TO_RIGHT		(1 <<  0)
 # define RADEON_DST_Y_TOP_TO_BOTTOM		(1 <<  1)
-#define RADEON_REG_DST_WIDTH_HEIGHT		0x1598
-#define RADEON_REG_AUX_SC_CNTL			0x1660
+#define RADEON_REG_DP_MIX			0x16c8
 #define RADEON_REG_DP_WRITE_MASK		0x16cc
 #define RADEON_REG_DEFAULT_OFFSET		0x16e0
 #define RADEON_REG_DEFAULT_PITCH		0x16e4
@@ -76,28 +79,112 @@
 # define RADEON_RB2D_DC_FLUSH_ALL		0xf
 # define RADEON_RB2D_DC_BUSY			(1 << 31)
 
-#define RADEON_CP_PACKET0                           0x00000000
-#define RADEON_CP_PACKET1                           0x40000000
-#define RADEON_CP_PACKET2                           0x80000000
+#define RADEON_CP_PACKET0			0x00000000
+#define RADEON_CP_PACKET1			0x40000000
+#define RADEON_CP_PACKET2			0x80000000
 
 #define R128_REG_PC_NGUI_CTLSTAT		0x0184
 # define R128_PC_BUSY				(1 << 31)
+#define R128_REG_PCI_GART_PAGE			0x017c
+#define R128_REG_PC_NGUI_CTLSTAT		0x0184
+#define R128_REG_BM_CHUNK_0_VAL			0x0a18
+# define R128_BM_PTR_FORCE_TO_PCI		(1 << 21)
+# define R128_BM_PM4_RD_FORCE_TO_PCI		(1 << 22)
+# define R128_BM_GLOBAL_FORCE_TO_PCI		(1 << 23)
 #define R128_REG_GUI_STAT			0x1740
 # define R128_GUI_ACTIVE			(1 << 31)
-#define R128_REG_PCI_GART_PAGE          0x017c
-#define R128_REG_PC_NGUI_CTLSTAT	0x0184
-#define R128_REG_BM_CHUNK_0_VAL         0x0a18
-# define R128_BM_PTR_FORCE_TO_PCI    (1 << 21)
-# define R128_BM_PM4_RD_FORCE_TO_PCI (1 << 22)
-# define R128_BM_GLOBAL_FORCE_TO_PCI (1 << 23)
+#define R128_REG_TEX_CNTL			0x1800
+
+#define R128_REG_SCALE_SRC_HEIGHT_WIDTH		0x1994
+#define R128_REG_SCALE_OFFSET_0			0x1998
+#define R128_REG_SCALE_PITCH			0x199c
+#define R128_REG_SCALE_X_INC			0x19a0
+#define R128_REG_SCALE_Y_INC			0x19a4
+#define R128_REG_SCALE_HACC			0x19a8
+#define R128_REG_SCALE_VACC			0x19ac
+#define R128_REG_SCALE_DST_X_Y			0x19b0
+#define R128_REG_SCALE_DST_HEIGHT_WIDTH		0x19b4
+#define R128_REG_SCALE_3D_CNTL			0x1a00
+# define R128_SCALE_DITHER_ERR_DIFF		(0  <<  1)
+# define R128_SCALE_DITHER_TABLE		(1  <<  1)
+# define R128_TEX_CACHE_SIZE_FULL		(0  <<  2)
+# define R128_TEX_CACHE_SIZE_HALF		(1  <<  2)
+# define R128_DITHER_INIT_CURR			(0  <<  3)
+# define R128_DITHER_INIT_RESET			(1  <<  3)
+# define R128_ROUND_24BIT			(1  <<  4)
+# define R128_TEX_CACHE_DISABLE			(1  <<  5)
+# define R128_SCALE_3D_NOOP			(0  <<  6)
+# define R128_SCALE_3D_SCALE			(1  <<  6)
+# define R128_SCALE_3D_TEXMAP_SHADE		(2  <<  6)
+# define R128_SCALE_PIX_BLEND			(0  <<  8)
+# define R128_SCALE_PIX_REPLICATE		(1  <<  8)
+# define R128_TEX_CACHE_SPLIT			(1  <<  9)
+# define R128_APPLE_YUV_MODE			(1  << 10)
+# define R128_TEX_CACHE_PALLETE_MODE		(1  << 11)
+# define R128_ALPHA_COMB_ADD_CLAMP		(0  << 12)
+# define R128_ALPHA_COMB_ADD_NCLAMP		(1  << 12)
+# define R128_ALPHA_COMB_SUB_DST_SRC_CLAMP	(2  << 12)
+# define R128_ALPHA_COMB_SUB_DST_SRC_NCLAMP	(3  << 12)
+# define R128_FOG_TABLE				(1  << 14)
+# define R128_SIGNED_DST_CLAMP			(1  << 15)
+# define R128_ALPHA_BLEND_SRC_ZERO		(0  << 16)
+# define R128_ALPHA_BLEND_SRC_ONE		(1  << 16)
+# define R128_ALPHA_BLEND_SRC_SRCCOLOR		(2  << 16)
+# define R128_ALPHA_BLEND_SRC_INVSRCCOLOR	(3  << 16)
+# define R128_ALPHA_BLEND_SRC_SRCALPHA		(4  << 16)
+# define R128_ALPHA_BLEND_SRC_INVSRCALPHA	(5  << 16)
+# define R128_ALPHA_BLEND_SRC_DSTALPHA		(6  << 16)
+# define R128_ALPHA_BLEND_SRC_INVDSTALPHA	(7  << 16)
+# define R128_ALPHA_BLEND_SRC_DSTCOLOR		(8  << 16)
+# define R128_ALPHA_BLEND_SRC_INVDSTCOLOR	(9  << 16)
+# define R128_ALPHA_BLEND_SRC_SAT		(10 << 16)
+# define R128_ALPHA_BLEND_SRC_BLEND		(11 << 16)
+# define R128_ALPHA_BLEND_SRC_INVBLEND		(12 << 16)
+# define R128_ALPHA_BLEND_DST_ZERO		(0  << 20)
+# define R128_ALPHA_BLEND_DST_ONE		(1  << 20)
+# define R128_ALPHA_BLEND_DST_SRCCOLOR		(2  << 20)
+# define R128_ALPHA_BLEND_DST_INVSRCCOLOR	(3  << 20)
+# define R128_ALPHA_BLEND_DST_SRCALPHA		(4  << 20)
+# define R128_ALPHA_BLEND_DST_INVSRCALPHA	(5  << 20)
+# define R128_ALPHA_BLEND_DST_DSTALPHA		(6  << 20)
+# define R128_ALPHA_BLEND_DST_INVDSTALPHA	(7  << 20)
+# define R128_ALPHA_BLEND_DST_DSTCOLOR		(8  << 20)
+# define R128_ALPHA_BLEND_DST_INVDSTCOLOR	(9  << 20)
+# define R128_ALPHA_TEST_NEVER			(0  << 24)
+# define R128_ALPHA_TEST_LESS			(1  << 24)
+# define R128_ALPHA_TEST_LESSEQUAL		(2  << 24)
+# define R128_ALPHA_TEST_EQUAL			(3  << 24)
+# define R128_ALPHA_TEST_GREATEREQUAL		(4  << 24)
+# define R128_ALPHA_TEST_GREATER		(5  << 24)
+# define R128_ALPHA_TEST_NEQUAL			(6  << 24)
+# define R128_ALPHA_TEST_ALWAYS			(7  << 24)
+# define R128_COMPOSITE_SHADOW_CMP_EQUAL	(0  << 28)
+# define R128_COMPOSITE_SHADOW_CMP_NEQUAL	(1  << 28)
+# define R128_COMPOSITE_SHADOW			(1  << 29)
+# define R128_TEX_MAP_ALPHA_IN_TEXTURE		(1  << 30)
+# define R128_TEX_CACHE_LINE_SIZE_8QW		(0  << 31)
+# define R128_TEX_CACHE_LINE_SIZE_4QW		(1  << 31)
+#define R128_REG_SCALE_3D_DATATYPE		0x1a20
+
+#define R128_DATATYPE_C8			2
+#define R128_DATATYPE_ARGB_1555			3
+#define R128_DATATYPE_RGB_565			4
+#define R128_DATATYPE_ARGB_8888			6
+#define R128_DATATYPE_RGB_332			7
+#define R128_DATATYPE_Y8			8
+#define R128_DATATYPE_RGB_8			9
+#define R128_DATATYPE_VYUY_422			11
+#define R128_DATATYPE_YVYU_422			12
+#define R128_DATATYPE_AYUV_444			14
+#define R128_DATATYPE_ARGB_4444			15
 
-# define R128_PM4_NONPM4                 (0  << 28)
-# define R128_PM4_192PIO                 (1  << 28)
-# define R128_PM4_192BM                  (2  << 28)
-# define R128_PM4_128PIO_64INDBM         (3  << 28)
-# define R128_PM4_128BM_64INDBM          (4  << 28)
-# define R128_PM4_64PIO_128INDBM         (5  << 28)
-# define R128_PM4_64BM_128INDBM          (6  << 28)
-# define R128_PM4_64PIO_64VCBM_64INDBM   (7  << 28)
-# define R128_PM4_64BM_64VCBM_64INDBM    (8  << 28)
-# define R128_PM4_64PIO_64VCPIO_64INDPIO (15 << 28)
+#define R128_PM4_NONPM4				(0  << 28)
+#define R128_PM4_192PIO				(1  << 28)
+#define R128_PM4_192BM				(2  << 28)
+#define R128_PM4_128PIO_64INDBM			(3  << 28)
+#define R128_PM4_128BM_64INDBM			(4  << 28)
+#define R128_PM4_64PIO_128INDBM			(5  << 28)
+#define R128_PM4_64BM_128INDBM			(6  << 28)
+#define R128_PM4_64PIO_64VCBM_64INDBM		(7  << 28)
+#define R128_PM4_64BM_64VCBM_64INDBM		(8  << 28)
+#define R128_PM4_64PIO_64VCPIO_64INDPIO		(15 << 28)
diff --git a/hw/kdrive/ati/r128_blendtmp.h b/hw/kdrive/ati/r128_blendtmp.h
new file mode 100644
index 000000000..af2fc5cdd
--- /dev/null
+++ b/hw/kdrive/ati/r128_blendtmp.h
@@ -0,0 +1,119 @@
+/*
+ * $Id$
+ *
+ * Copyright © 2003 Eric Anholt, Anders Carlsson
+ *
+ * Permission to use, copy, modify, distribute, and sell this software and its
+ * documentation for any purpose is hereby granted without fee, provided that
+ * the above copyright notice appear in all copies and that both that
+ * copyright notice and this permission notice appear in supporting
+ * documentation, and that the name of Eric Anholt not be used in
+ * advertising or publicity pertaining to distribution of the software without
+ * specific, written prior permission.  Eric Anholt makes no
+ * representations about the suitability of this software for any purpose.  It
+ * is provided "as is" without express or implied warranty.
+ *
+ * ERIC ANHOLT DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
+ * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO
+ * EVENT SHALL ERIC ANHOLT BE LIABLE FOR ANY SPECIAL, INDIRECT OR
+ * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE,
+ * DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER
+ * TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+/* $Header$ */
+
+#ifdef USE_DMA
+#define TAG(x)		x##DMA
+#define LOCALS		(void)atic; \
+			RING_LOCALS
+#define BEGIN(x)	BEGIN_RING(x * 2)
+#define OUT_REG(reg, val) OUT_RING_REG(reg, val)
+#define END()		ADVANCE_RING()
+#else
+#define TAG(x)		x##MMIO
+#define LOCALS		char *mmio = atic->reg_base
+#define BEGIN(x)	ATIWaitAvailMMIO(x)
+#define OUT_REG(reg, val) MMIO_OUT32(mmio, (reg), (val))
+#define END()
+#endif
+
+static Bool
+TAG(R128PrepareBlend)(int op, PicturePtr pSrcPicture, PicturePtr pDstPicture,
+    PixmapPtr pSrc, PixmapPtr pDst)
+{
+	KdScreenPriv(pDst->drawable.pScreen);
+	ATIScreenInfo(pScreenPriv);
+	ATICardInfo(pScreenPriv);
+	CARD32 dstDatatype, srcDatatype;
+	LOCALS;
+
+	accel_atis = atis;
+
+	if (!TAG(ATISetup)(pDst, pSrc))
+		return FALSE;
+
+	src_bpp = pSrc->drawable.bitsPerPixel;
+
+	if (op >= sizeof(R128BlendOp)/sizeof(R128BlendOp[0]))
+		return FALSE;
+	if (pSrcPicture->repeat || pSrcPicture->transform != NULL)
+		return FALSE;
+
+	if (!R128GetDatatype(pDstPicture->format, &dstDatatype) ||
+	    !R128GetDatatype(pSrcPicture->format, &srcDatatype))
+		return FALSE;
+
+	BEGIN(9);
+	OUT_REG(RADEON_REG_DP_GUI_MASTER_CNTL,
+	    (dstDatatype << 8) |
+	    RADEON_GMC_SRC_DATATYPE_COLOR |
+	    RADEON_GMC_DST_PITCH_OFFSET_CNTL |
+	    RADEON_GMC_CLR_CMP_CNTL_DIS |
+	    RADEON_GMC_AUX_CLIP_DIS |
+	    (ATIBltRop[3] << 16) |
+	    RADEON_GMC_3D_FCN_EN);
+	OUT_REG(R128_REG_SCALE_3D_CNTL,
+	    R128_SCALE_3D_SCALE |
+	    R128BlendOp[op] |
+	    R128_TEX_MAP_ALPHA_IN_TEXTURE);
+	OUT_REG(R128_REG_SCALE_3D_DATATYPE, srcDatatype);
+	OUT_REG(R128_REG_SCALE_PITCH, src_pitch / src_bpp);
+	/* 4.16 fixed point scaling factor? */
+	OUT_REG(R128_REG_SCALE_X_INC, 65536);
+	OUT_REG(R128_REG_SCALE_Y_INC, 65536);
+	OUT_REG(R128_REG_SCALE_HACC, 0x00000000);
+	OUT_REG(R128_REG_SCALE_VACC, 0x00000000);
+	OUT_REG(RADEON_REG_DP_CNTL, 
+	    RADEON_DST_X_LEFT_TO_RIGHT | RADEON_DST_Y_TOP_TO_BOTTOM );
+	END();
+
+	return TRUE;
+}
+
+static void
+TAG(R128Blend)(int srcX, int srcY, int dstX, int dstY, int width, int height)
+{
+	ATIScreenInfo *atis = accel_atis;
+	ATICardInfo *atic = atis->atic;
+	LOCALS;
+
+	BEGIN(4);
+	OUT_REG(R128_REG_SCALE_OFFSET_0, src_offset + srcY * src_pitch + srcX *
+	    (src_bpp >> 3));
+	OUT_REG(R128_REG_SCALE_SRC_HEIGHT_WIDTH, (height << 16) | width);
+	OUT_REG(R128_REG_SCALE_DST_X_Y, (dstX << 16) | dstY);
+	OUT_REG(R128_REG_SCALE_DST_HEIGHT_WIDTH, (height << 16) | width);
+	END();
+}
+
+static void
+TAG(R128DoneBlend)(void)
+{
+}
+
+#undef TAG
+#undef LOCALS
+#undef BEGIN
+#undef OUT_REG
+#undef END
diff --git a/hw/kdrive/src/kaa.c b/hw/kdrive/src/kaa.c
index c9a72a95d..1a39c25e9 100644
--- a/hw/kdrive/src/kaa.c
+++ b/hw/kdrive/src/kaa.c
@@ -153,7 +153,6 @@ kaaMoveInPixmap (PixmapPtr pPixmap)
     unsigned char *dst, *src;
     int i;
 
-    return;
     KdCheckSync (pPixmap->drawable.pScreen);
     
     DBG_MIGRATE (("-> 0x%08x (0x%x) (%dx%d)\n",
@@ -271,6 +270,7 @@ kaaCreatePixmap(ScreenPtr pScreen, int w, int h, int depth)
 		break;
 	    }
     }
+
     pPixmap = fbCreatePixmapBpp (pScreen, w, h, depth, bpp);
     if (!pPixmap)
 	return NULL;
@@ -278,8 +278,7 @@ kaaCreatePixmap(ScreenPtr pScreen, int w, int h, int depth)
     pKaaPixmap->score = 0;
     pKaaPixmap->area = NULL;
     
-    if (depth == pScreen->rootDepth &&
-	(pPixmap->devKind * h) >= MIN_OFFPIX_SIZE)
+    if ((pPixmap->devKind * h) >= MIN_OFFPIX_SIZE)
 	kaaPixmapAllocArea (pPixmap);
     return pPixmap;
 }
@@ -939,6 +938,76 @@ kaaPaintWindow(WindowPtr pWin, RegionPtr pRegion, int what)
 #ifdef RENDER
 #include "mipict.h"
 
+#define KAA_DEBUG_FALLBACKS 0
+
+#if KAA_DEBUG_FALLBACKS
+static void
+kaaPrintCompositeFallback(CARD8 op,
+			  PicturePtr pSrc,
+			  PicturePtr pMask,
+			  PicturePtr pDst)
+{
+    char sop[20], ssfmt[20], smfmt[20], sdfmt[20];
+    char sloc, mloc, dloc;
+    int temp;
+
+    switch(op)
+    {
+    case PictOpSrc:
+	sprintf(sop, "Src");
+	break;
+    case PictOpOver:
+	sprintf(sop, "Over");
+	break;
+    default:
+	sprintf(sop, "0x%x", (int)op);
+	break;
+    }
+    switch(pSrc->format)
+    {
+    case PICT_a8r8g8b8:
+	sprintf(ssfmt, "ARGB8888");
+	break;
+    default:
+	sprintf(ssfmt, "0x%x", (int)pSrc->format);
+	break;
+    }
+    sprintf(smfmt, "(none)");
+    if (pMask) {
+	switch(pMask->format)
+	{
+	case PICT_a8:
+	    sprintf(smfmt, "A8");
+	    break;
+	default:
+	    sprintf(smfmt, "0x%x", (int)pMask->format);
+	    break;
+	}
+    }
+    switch(pDst->format)
+    {
+    case PICT_r5g6b5:
+	sprintf(sdfmt, "RGB565");
+	break;
+    default:
+	sprintf(sdfmt, "0x%x", (int)pDst->format);
+	break;
+    }
+    strcpy (smfmt, ("None"));
+    pMask = 0x0;
+
+    sloc = kaaGetOffscreenPixmap(pSrc->pDrawable, &temp, &temp) ? 's' : 'm';
+    mloc = (pMask && kaaGetOffscreenPixmap(pMask->pDrawable, &temp, &temp)) ?
+	    's' : 'm';
+    dloc = kaaGetOffscreenPixmap(pDst->pDrawable, &temp, &temp) ? 's' : 'm';
+
+    ErrorF("Composite fallback: op %s, src 0x%x (%c), mask 0x%x (%c), "
+	   "dst 0x%x (%c)\n"
+	   "                    srcfmt %s, maskfmt %s, dstfmt %s\n",
+	   sop, pSrc, sloc, pMask, mloc, pDst, dloc, ssfmt, smfmt, sdfmt);
+}
+#endif
+
 static void
 kaaComposite(CARD8	op,
 	     PicturePtr pSrc,
@@ -953,42 +1022,111 @@ kaaComposite(CARD8	op,
 	     CARD16	width,
 	     CARD16	height)
 {
-    if (op == PictOpSrc && !pMask)
+    KdScreenPriv (pDst->pDrawable->pScreen);
+    KaaScreenPriv (pDst->pDrawable->pScreen);
+    
+    if (!pMask)
     {
-	/*
-	 * Check for two special cases -- solid fill and copy area
-	 */
-	if (pSrc->pDrawable->width == 1 && pSrc->pDrawable->height == 1 &&
-	    pSrc->repeat)
+	if (op == PictOpSrc)
 	{
-	    ;
+	    /*
+	     * Check for two special cases -- solid fill and copy area
+	     */
+	    if (pSrc->pDrawable->width == 1 && pSrc->pDrawable->height == 1 &&
+		pSrc->repeat)
+	    {
+		;
+	    }
+	    else if (!pSrc->repeat && pSrc->format == pDst->format)
+	    {
+		RegionRec	region;
+		
+		xDst += pDst->pDrawable->x;
+		yDst += pDst->pDrawable->y;
+		xSrc += pSrc->pDrawable->x;
+		ySrc += pSrc->pDrawable->y;
+
+		if (!miComputeCompositeRegion (&region, pSrc, pMask, pDst,
+					       xSrc, ySrc, xMask, yMask, xDst, yDst,
+					       width, height))
+		    return;
+		
+		
+		kaaCopyNtoN (pSrc->pDrawable, pDst->pDrawable, 0,
+			     REGION_RECTS(&region), REGION_NUM_RECTS(&region),
+			     xSrc - xDst, ySrc - yDst,
+			     FALSE, FALSE, 0, 0);
+		return;
+	    }
 	}
-	else if (!pSrc->repeat && pSrc->format == pDst->format)
+
+	if (pScreenPriv->enabled && pKaaScr->info->PrepareBlend)
 	{
-	    RegionRec	region;
-	    
+	    RegionRec region;
+	    BoxPtr pbox;
+	    int nbox;
+	    int src_off_x, src_off_y;
+	    int dst_off_x, dst_off_y;
+	    PixmapPtr pSrcPix, pDstPix;
+
 	    xDst += pDst->pDrawable->x;
 	    yDst += pDst->pDrawable->y;
+
 	    xSrc += pSrc->pDrawable->x;
 	    ySrc += pSrc->pDrawable->y;
-	    if (pMask)
-	    {
-		xMask += pMask->pDrawable->x;
-		yMask += pMask->pDrawable->y;
-	    }
+
 	    if (!miComputeCompositeRegion (&region, pSrc, pMask, pDst,
 					   xSrc, ySrc, xMask, yMask, xDst, yDst,
 					   width, height))
 		return;
 
-				      
-	    kaaCopyNtoN (pSrc->pDrawable, pDst->pDrawable, 0,
-			 REGION_RECTS(&region), REGION_NUM_RECTS(&region),
-			 xSrc - xDst, ySrc - yDst,
-			 FALSE, FALSE, 0, 0);
+
+	    /* Migrate pixmaps to same place as destination */
+	    if (pSrc->pDrawable->type == DRAWABLE_PIXMAP)
+		kaaPixmapUseScreen ((PixmapPtr) pSrc->pDrawable);
+	    if (pDst->pDrawable->type == DRAWABLE_PIXMAP)
+		kaaPixmapUseScreen ((PixmapPtr) pDst->pDrawable);
+	    
+	    pSrcPix = kaaGetOffscreenPixmap (pSrc->pDrawable, &src_off_x,
+					     &src_off_y);
+	    pDstPix = kaaGetOffscreenPixmap (pDst->pDrawable, &dst_off_x,
+					     &dst_off_y);
+	    if (!pSrcPix || !pDstPix ||
+		!(*pKaaScr->info->PrepareBlend) (op, pSrc, pDst, pSrcPix,
+						 pDstPix))
+	    {
+		goto software;
+	    }
+	    
+	    nbox = REGION_NUM_RECTS(&region);
+	    pbox = REGION_RECTS(&region);
+
+	    xSrc -= xDst;
+	    ySrc -= yDst;
+	
+	    while (nbox--)
+	    {
+		(*pKaaScr->info->Blend) (pbox->x1 + xSrc + src_off_x,
+					 pbox->y1 + ySrc + src_off_y,
+					 pbox->x1 + dst_off_x,
+					 pbox->y1 + dst_off_y,
+					 pbox->x2 - pbox->x1,
+					 pbox->y2 - pbox->y1);
+		pbox++;
+	    }
+	    
+	    (*pKaaScr->info->DoneBlend) ();
+	    KdMarkSync(pDst->pDrawable->pScreen);
+
 	    return;
 	}
     }
+
+software:
+#if KAA_DEBUG_FALLBACKS
+    kaaPrintCompositeFallback (op, pSrc, pMask, pDst);
+#endif
+
     if (pSrc->pDrawable->type == DRAWABLE_PIXMAP)
 	kaaPixmapUseMemory ((PixmapPtr) pSrc->pDrawable);
     if (pMask && pMask->pDrawable->type == DRAWABLE_PIXMAP)
diff --git a/hw/kdrive/src/kdrive.h b/hw/kdrive/src/kdrive.h
index df2ead9f4..9bde73d21 100644
--- a/hw/kdrive/src/kdrive.h
+++ b/hw/kdrive/src/kdrive.h
@@ -332,6 +332,19 @@ typedef struct _KaaScreenInfo {
     int	        offscreenByteAlign;
     int         offscreenPitch;
     int		flags;
+
+    Bool        (*PrepareBlend) (int		op,
+				 PicturePtr	pSrcPicture,
+				 PicturePtr	pDstPicture,
+				 PixmapPtr	pSrc,
+				 PixmapPtr	pDst);
+    void        (*Blend) (int	srcX,
+			  int	srcY,
+			  int	dstX,
+			  int	dstY,
+			  int	width,
+			  int	height);
+    void	(*DoneBlend) (void);
 } KaaScreenInfoRec, *KaaScreenInfoPtr;
 
 #define KAA_OFFSCREEN_PIXMAPS (1 << 0)