- Allow acceleration between same-depth pixmaps, rather than between

    anything and a dst that matched the screen depth (fixes corruption for
    non-screen-depth src and makes more acceleration possible).
- Add ATI_FALLBACK macro and use it to allow verbose descriptions of why
    hardware acceleration fails.
- Check that src and dst alignment meet requirements of the card before
    accelerating. The BIOS may set up screens that don't meet the
    requirements.
- Fix the R128 offset alignment (32 bytes, not 8).
- Enable Blend operation even if screen is 24bpp (it will fail if the dest
    is 24bpp anyway).

6 files changed, 103 insertions, 71 deletions

diff --git a/hw/kdrive/ati/ati.h b/hw/kdrive/ati/ati.h
index 44e83a5ac..cacc264dd 100644
--- a/hw/kdrive/ati/ati.h
+++ b/hw/kdrive/ati/ati.h
@@ -146,13 +146,6 @@ typedef struct _ATIScreenInfo {
 
 	Bool using_dri;
 	Bool using_dma;
-	/* If is_24bpp is set, then we are using the accelerator in 8-bit mode due
-	 * to it being broken for 24bpp, so coordinates have to be multiplied by 3.
-	 */
-	Bool is_24bpp;
-
-	int datatype;
-	int dp_gui_master_cntl;
 
 #ifdef USE_DRI
 	drmSize         registerSize;
diff --git a/hw/kdrive/ati/ati_draw.c b/hw/kdrive/ati/ati_draw.c
index 6e8575ef7..5dab41ede 100644
--- a/hw/kdrive/ati/ati_draw.c
+++ b/hw/kdrive/ati/ati_draw.c
@@ -108,6 +108,10 @@ ATIScreenInfo *accel_atis;
 int src_pitch;
 int src_offset;
 int src_bpp;
+/* If is_24bpp is set, then we are using the accelerator in 8-bit mode due
+ * to it being broken for 24bpp, so coordinates have to be multiplied by 3.
+ */
+int is_24bpp;
 
 static void
 ATIWaitAvailMMIO(int n)
@@ -348,7 +352,7 @@ drmBufPtr ATIDMAGetBuffer()
 #endif /* USE_DRI */
 
 static Bool
-R128GetDatatype(CARD32 format, CARD32 *type)
+R128GetDatatypePict(CARD32 format, CARD32 *type)
 {
 	switch (format) {
 	case PICT_a8r8g8b8:
@@ -364,6 +368,34 @@ R128GetDatatype(CARD32 format, CARD32 *type)
 	return FALSE;
 }
 
+/* Assumes that depth 15 and 16 can be used as depth 16, which is okay since we
+ * require src and dest datatypes to be equal.
+ */
+static Bool
+ATIGetDatatypeBpp(int bpp, CARD32 *type)
+{
+	is_24bpp = FALSE;
+
+	switch (bpp) {
+	case 8:
+		*type = R128_DATATYPE_C8;
+		return TRUE;
+	case 16:
+		*type = R128_DATATYPE_RGB_565;
+		return TRUE;
+	case 24:
+		*type = R128_DATATYPE_C8;
+		is_24bpp = TRUE;
+		return TRUE;
+	case 32:
+		*type = R128_DATATYPE_ARGB_8888;
+		return TRUE;
+	default:
+		ErrorF("Unsupported bpp: %x\n", bpp);
+		return FALSE;
+	}
+}
+
 #ifdef USE_DRI
 #define USE_DMA
 #include "ati_drawtmp.h"
@@ -391,31 +423,6 @@ ATIDrawInit(ScreenPtr pScreen)
 	ATIScreenInfo(pScreenPriv);
 	ATICardInfo(pScreenPriv);
 
-	switch (pScreenPriv->screen->fb[0].depth)
-	{
-	case 8:
-		atis->datatype = 2;
-		break;
-	case 15:
-		atis->datatype = 3;
-		break;
-	case 16:
-		atis->datatype = 4;
-		break;
-	case 24:
-		if (pScreenPriv->screen->fb[0].bitsPerPixel == 24) {
-			atis->is_24bpp = TRUE;
-			atis->datatype = 2;
-		} else {
-			atis->datatype = 6;
-		}
-		break;
-	default:
-		FatalError("[ati]: depth %d unsupported\n",
-		    pScreenPriv->screen->fb[0].depth);
-		return FALSE;
-	}
-
 	ErrorF("Screen: %d/%d depth/bpp\n", pScreenPriv->screen->fb[0].depth,
 	    pScreenPriv->screen->fb[0].bitsPerPixel);
 #ifdef USE_DRI
@@ -427,9 +434,6 @@ ATIDrawInit(ScreenPtr pScreen)
 	}
 #endif /* USE_DRI */
 
-	atis->dp_gui_master_cntl = (atis->datatype << 8) |
-	    RADEON_GMC_CLR_CMP_CNTL_DIS | RADEON_GMC_AUX_CLIP_DIS;
-
 	memset(&atis->kaa, 0, sizeof(KaaScreenInfoRec));
 #ifdef USE_DRI
 	if (atis->using_dma) {
@@ -437,7 +441,7 @@ ATIDrawInit(ScreenPtr pScreen)
 		atis->kaa.Solid = ATISolidDMA;
 		atis->kaa.PrepareCopy = ATIPrepareCopyDMA;
 		atis->kaa.Copy = ATICopyDMA;
-		if (!atic->is_radeon && !atis->is_24bpp) {
+		if (!atic->is_radeon) {
 			atis->kaa.PrepareBlend = R128PrepareBlendDMA;
 			atis->kaa.Blend = R128BlendDMA;
 			atis->kaa.DoneBlend = R128DoneBlendDMA;
@@ -450,7 +454,7 @@ ATIDrawInit(ScreenPtr pScreen)
 		atis->kaa.Solid = ATISolidMMIO;
 		atis->kaa.PrepareCopy = ATIPrepareCopyMMIO;
 		atis->kaa.Copy = ATICopyMMIO;
-		if (!atic->is_radeon && !atis->is_24bpp) {
+		if (!atic->is_radeon) {
 			atis->kaa.PrepareBlend = R128PrepareBlendMMIO;
 			atis->kaa.Blend = R128BlendMMIO;
 			atis->kaa.DoneBlend = R128DoneBlendMMIO;
@@ -463,13 +467,11 @@ ATIDrawInit(ScreenPtr pScreen)
 		atis->kaa.offscreenByteAlign = 1024;
 		atis->kaa.offscreenPitch = 1024;
 	} else {
-		atis->kaa.offscreenByteAlign = 8;
-		/* Workaround for corrupation at 8 and 24bpp. Why? */
-		if (atis->datatype == 2)
-			atis->kaa.offscreenPitch = 16;
-		else
-			atis->kaa.offscreenPitch =
-			    pScreenPriv->screen->fb[0].bitsPerPixel;
+		atis->kaa.offscreenByteAlign = 32;
+		/* Pitch alignment is in sets of 8 pixels, and we need to cover
+		 * 32bpp, so 32 bytes.
+		 */
+		atis->kaa.offscreenPitch = 32;
 	}
 	if (!kaaDrawInit(pScreen, &atis->kaa))
 		return FALSE;
diff --git a/hw/kdrive/ati/ati_draw.h b/hw/kdrive/ati/ati_draw.h
index 95b3b875c..eddc8721e 100644
--- a/hw/kdrive/ati/ati_draw.h
+++ b/hw/kdrive/ati/ati_draw.h
@@ -69,4 +69,14 @@ void ATIDMAStop(ScreenPtr pScreen);
 
 #endif /* USE_DRI */
 
+#if 0
+#define ATI_FALLBACK(x)		\
+do {				\
+	ErrorF x;		\
+	return FALSE;		\
+} while (0)
+#else
+#define ATI_FALLBACK(x) return FALSE
+#endif
+
 #endif /* _ATI_DRAW_H_ */
diff --git a/hw/kdrive/ati/ati_drawtmp.h b/hw/kdrive/ati/ati_drawtmp.h
index 6573aef68..a8a0a7dac 100644
--- a/hw/kdrive/ati/ati_drawtmp.h
+++ b/hw/kdrive/ati/ati_drawtmp.h
@@ -32,7 +32,8 @@
 #define END()		ADVANCE_RING()
 #else
 #define TAG(x)		x##MMIO
-#define LOCALS		char *mmio = atic->reg_base
+#define LOCALS		(void)atis; \
+			char *mmio = atic->reg_base
 #define BEGIN(x)	ATIWaitAvailMMIO(x)
 #define OUT_REG(reg, val) MMIO_OUT32((mmio), (reg), (val))
 #define END()
@@ -45,22 +46,27 @@ TAG(ATISetup)(PixmapPtr pDst, PixmapPtr pSrc)
 	ATIScreenInfo(pScreenPriv);
 	ATICardInfo(pScreenPriv);
 	int dst_offset, dst_pitch;
-	int bpp = pScreenPriv->screen->fb[0].bitsPerPixel;
+	int bpp = pDst->drawable.bitsPerPixel;
 	LOCALS;
 
 	accel_atis = atis;
 
-	/* No acceleration for other formats (yet) */
-	if (pDst->drawable.bitsPerPixel != bpp)
-		return FALSE;
-
 	dst_pitch = pDst->devKind;
 	dst_offset = ((CARD8 *)pDst->devPrivate.ptr -
 	    pScreenPriv->screen->memory_base);
+	if ((dst_pitch & (atis->kaa.offscreenPitch - 1)) != 0)
+		ATI_FALLBACK(("Bad dst pitch 0x%x\n", dst_pitch));
+	if ((dst_offset & (atis->kaa.offscreenByteAlign - 1)) != 0)
+		ATI_FALLBACK(("Bad dst offset 0x%x\n", dst_offset));
+
 	if (pSrc != NULL) {
 		src_pitch = pSrc->devKind;
 		src_offset = ((CARD8 *)pSrc->devPrivate.ptr -
 		    pScreenPriv->screen->memory_base);
+		if ((src_pitch & (atis->kaa.offscreenPitch - 1)) != 0)
+			ATI_FALLBACK(("Bad src pitch 0x%x\n", src_pitch));
+		if ((src_offset & (atis->kaa.offscreenByteAlign - 1)) != 0)
+			ATI_FALLBACK(("Bad src offset 0x%x\n", src_offset));
 	}
 
 	BEGIN((pSrc != NULL) ? 3 : 2);
@@ -72,7 +78,7 @@ TAG(ATISetup)(PixmapPtr pDst, PixmapPtr pSrc)
 			    ((src_pitch >> 6) << 22) | (src_offset >> 10));
 		}
 	} else {
-		if (atis->is_24bpp) {
+		if (is_24bpp) {
 			dst_pitch *= 3;
 			src_pitch *= 3;
 		}
@@ -97,25 +103,31 @@ TAG(ATIPrepareSolid)(PixmapPtr pPixmap, int alu, Pixel pm, Pixel fg)
 	KdScreenPriv(pPixmap->drawable.pScreen);
 	ATIScreenInfo(pScreenPriv);
 	ATICardInfo(pScreenPriv);
+	CARD32 datatype;
 	LOCALS;
 
-	if (atis->is_24bpp) {
-		if (pm != 0xffffffff)
-			return FALSE;
+	if (is_24bpp) {
 		/* Solid fills in fake-24bpp mode only work if the pixel color
-		 * is all the same byte.
+		 * and planemask are all the same byte.
 		 */
 		if ((fg & 0xffffff) != (((fg & 0xff) << 16) | ((fg >> 8) &
 		    0xffff)))
-			return FALSE;
+			ATI_FALLBACK(("Can't do solid color %d in 24bpp\n"));
+		if ((pm & 0xffffff) != (((pm & 0xff) << 16) | ((pm >> 8) &
+		    0xffff)))
+			ATI_FALLBACK(("Can't do planemask %d in 24bpp\n"));
 	}
 
+	if (!ATIGetDatatypeBpp(pPixmap->drawable.bitsPerPixel, &datatype))
+		return FALSE;
 	if (!TAG(ATISetup)(pPixmap, NULL))
 		return FALSE;
 
 	BEGIN(4);
 	OUT_REG(RADEON_REG_DP_GUI_MASTER_CNTL,
-	    atis->dp_gui_master_cntl |
+	    (datatype << 8) |
+	    RADEON_GMC_CLR_CMP_CNTL_DIS |
+	    RADEON_GMC_AUX_CLIP_DIS |
 	    RADEON_GMC_BRUSH_SOLID_COLOR |
 	    RADEON_GMC_DST_PITCH_OFFSET_CNTL |
 	    RADEON_GMC_SRC_DATATYPE_COLOR |
@@ -136,7 +148,7 @@ TAG(ATISolid)(int x1, int y1, int x2, int y2)
 	ATICardInfo *atic = atis->atic;
 	LOCALS;
 	
-	if (atis->is_24bpp) {
+	if (is_24bpp) {
 		x1 *= 3;
 		x2 *= 3;
 	}
@@ -152,20 +164,31 @@ TAG(ATIPrepareCopy)(PixmapPtr pSrc, PixmapPtr pDst, int dx, int dy, int alu, Pix
 	KdScreenPriv(pDst->drawable.pScreen);
 	ATIScreenInfo(pScreenPriv);
 	ATICardInfo(pScreenPriv);
+	CARD32 datatype;
 	LOCALS;
 
+	/* No acceleration between different formats */
+	if (pSrc->drawable.bitsPerPixel != pDst->drawable.bitsPerPixel)
+		ATI_FALLBACK(("src bpp != dst bpp (%d vs %d)\n",
+		    pSrc->drawable.bitsPerPixel, pDst->drawable.bitsPerPixel));
+
 	copydx = dx;
 	copydy = dy;
 
-	if (atis->is_24bpp && pm != 0xffffffff)
-		return FALSE;
+	if (is_24bpp && ((pm & 0xffffff) != (((pm & 0xff) << 16) | ((pm >> 8) &
+	    0xffff))))
+		ATI_FALLBACK(("Can't do planemask %d in 24bpp\n"));
 
+	if (!ATIGetDatatypeBpp(pDst->drawable.bitsPerPixel, &datatype))
+		return FALSE;
 	if (!TAG(ATISetup)(pDst, pSrc))
 		return FALSE;
 
 	BEGIN(3);
 	OUT_REG(RADEON_REG_DP_GUI_MASTER_CNTL,
-	    atis->dp_gui_master_cntl |
+	    (datatype << 8) |
+	    RADEON_GMC_CLR_CMP_CNTL_DIS |
+	    RADEON_GMC_AUX_CLIP_DIS |
 	    RADEON_GMC_BRUSH_SOLID_COLOR |
 	    RADEON_GMC_SRC_DATATYPE_COLOR |
 	    (ATIBltRop[alu] << 16) |
@@ -188,7 +211,7 @@ TAG(ATICopy)(int srcX, int srcY, int dstX, int dstY, int w, int h)
 	ATICardInfo *atic = atis->atic;
 	LOCALS;
 
-	if (atis->is_24bpp) {
+	if (is_24bpp) {
 		srcX *= 3;
 		dstX *= 3;
 		w *= 3;
diff --git a/hw/kdrive/ati/ati_dri.c b/hw/kdrive/ati/ati_dri.c
index e289573f8..65ca0f72e 100644
--- a/hw/kdrive/ati/ati_dri.c
+++ b/hw/kdrive/ati/ati_dri.c
@@ -449,7 +449,7 @@ static void ATIDRISwapContext(ScreenPtr pScreen, DRISyncType syncType,
 {
 	KdScreenPriv(pScreen);
 	ATIScreenInfo(pScreenPriv);
-	ATICardInfo(pScreenPriv);
+
 	if ((syncType==DRI_3D_SYNC) && (oldContextType==DRI_2D_CONTEXT) &&
 	    (newContextType==DRI_2D_CONTEXT)) {
 		/* Entering from Wakeup */
diff --git a/hw/kdrive/ati/r128_blendtmp.h b/hw/kdrive/ati/r128_blendtmp.h
index e40aaf4c0..ede48fe4c 100644
--- a/hw/kdrive/ati/r128_blendtmp.h
+++ b/hw/kdrive/ati/r128_blendtmp.h
@@ -56,13 +56,17 @@ TAG(R128PrepareBlend)(int op, PicturePtr pSrcPicture, PicturePtr pDstPicture,
 	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;
+		ATI_FALLBACK(("Unsupported op 0x%x\n", op));
+	if (pSrcPicture->repeat)
+		ATI_FALLBACK(("repeat unsupported\n"));
+	if (pSrcPicture->transform != NULL)
+		ATI_FALLBACK(("transform unsupported\n"));
+	if (!R128GetDatatypePict(pDstPicture->format, &dstDatatype))
+		ATI_FALLBACK(("Unsupported dest format 0x%x\n",
+		    pDstPicture->format));
+	if (!R128GetDatatypePict(pSrcPicture->format, &srcDatatype))
+		ATI_FALLBACK(("Unsupported src format 0x%x\n",
+		    pSrcPicture->format));
 
 	BEGIN(11);
 	OUT_REG(RADEON_REG_DP_GUI_MASTER_CNTL,