Add an OUT_RING_REG macro for use with DMA_PACKET0, which is like OUT_RING

    but includes debugging to ensure that the reg being submitted is the
    one that follows in the packet. Convert most uses of OUT_RING to it,
    and convert a couple of OUT_REG sets to DMA_PACKET0/OUT_RING_REG. Also,
    add checking to see if more registers are submitted to a DMA_PACKET0
    than should be, to avoid hangs during stupid mistakes (checking for
    less isn't done).

6 files changed, 193 insertions, 177 deletions

diff --git a/hw/kdrive/ati/ati_dma.h b/hw/kdrive/ati/ati_dma.h
index 62501fa8f..ad2f3e12a 100644
--- a/hw/kdrive/ati/ati_dma.h
+++ b/hw/kdrive/ati/ati_dma.h
@@ -24,15 +24,23 @@
 #ifndef _ATI_DMA_H_
 #define _ATI_DMA_H_
 
+#define CCE_DEBUG 1
+
+#if !CCE_DEBUG
 #define DMA_PACKET0(reg, count)						\
 	(ATI_CCE_PACKET0 | (((count) - 1) << 16) | ((reg) >> 2))
+#else
+#define DMA_PACKET0(reg, count)						\
+	(__packet0count = (count), __reg = (reg),			\
+	ATI_CCE_PACKET0 | (((count) - 1) << 16) | ((reg) >> 2))
+#endif
 #define DMA_PACKET1(reg1, reg2)						\
 	(ATI_CCE_PACKET1 |						\
 	(((reg2) >> 2) << ATI_CCE_PACKET1_REG_2_SHIFT) |  ((reg1) >> 2))
 #define DMA_PACKET3(type, count)					\
 	((type) | (((count) - 1) << 16))
 
-#if 0	/* CCE non-debug */
+#if !CCE_DEBUG
 
 #define RING_LOCALS	CARD32 *__head; int __count
 #define BEGIN_DMA(n)							\
@@ -51,7 +59,8 @@ do {									\
 
 #else
 
-#define RING_LOCALS	CARD32 *__head; int __count; int __total
+#define RING_LOCALS	\
+	CARD32 *__head; int __count, __total, __reg, __packet0count
 #define BEGIN_DMA(n)							\
 do {									\
 	if ((atis->indirectBuffer->used + 4*(n)) >			\
@@ -62,6 +71,8 @@ do {									\
 	    atis->indirectBuffer->used);				\
 	__count = 0;							\
 	__total = n;							\
+	__reg = 0;								\
+	__packet0count = 0;								\
 } while (0)
 #define END_DMA() do {							\
 	if (__count != __total)						\
@@ -72,8 +83,19 @@ do {									\
 
 #endif
 
-#define OUT_RING(x) do {						\
-	__head[__count++] = (x);					\
+#define OUT_RING(val) do {						\
+	__head[__count++] = (val);					\
+} while (0)
+
+#define OUT_RING_REG(reg, val) do {					\
+	if (__reg != reg)						\
+		FatalError("unexpected reg (0x%x vs 0x%x) at %s:%d\n",	\
+		    reg, __reg, __FILE__, __LINE__);			\
+	if (__packet0count-- <= 0)					\
+		FatalError("overrun of packet0 at %s:%d\n",		\
+		    __FILE__, __LINE__);				\
+	__head[__count++] = (val);					\
+	__reg += 4;							\
 } while (0)
 
 #define OUT_RING_F(x) OUT_RING(GET_FLOAT_BITS(x))
diff --git a/hw/kdrive/ati/ati_draw.c b/hw/kdrive/ati/ati_draw.c
index a6ea02521..12b85ed9b 100644
--- a/hw/kdrive/ati/ati_draw.c
+++ b/hw/kdrive/ati/ati_draw.c
@@ -423,8 +423,8 @@ ATISolid(int x1, int y1, int x2, int y2)
 #else
 	BEGIN_DMA(3);
 	OUT_RING(DMA_PACKET0(ATI_REG_DST_Y_X, 2));
-	OUT_RING((y1 << 16) | x1);
-	OUT_RING(((y2 - y1) << 16) | (x2 - x1));
+	OUT_RING_REG(ATI_REG_DST_Y_X, (y1 << 16) | x1);
+	OUT_RING_REG(ATI_REG_DST_HEIGHT_WIDTH, ((y2 - y1) << 16) | (x2 - x1));
 	END_DMA();
 #endif
 	LEAVE_DRAW(0);
@@ -542,9 +542,9 @@ ATICopy(int srcX, int srcY, int dstX, int dstY, int w, int h)
 #else
 	BEGIN_DMA(4);
 	OUT_RING(DMA_PACKET0(ATI_REG_SRC_Y_X, 3));
-	OUT_RING((srcY << 16) | srcX);
-	OUT_RING((dstY << 16) | dstX);
-	OUT_RING((h << 16) | w);
+	OUT_RING_REG(ATI_REG_SRC_Y_X, (srcY << 16) | srcX);
+	OUT_RING_REG(ATI_REG_DST_Y_X, (dstY << 16) | dstX);
+	OUT_RING_REG(ATI_REG_DST_HEIGHT_WIDTH, (h << 16) | w);
 	END_DMA();
 #endif
 }
diff --git a/hw/kdrive/ati/ati_reg.h b/hw/kdrive/ati/ati_reg.h
index 17b5db360..c0c8dd29c 100644
--- a/hw/kdrive/ati/ati_reg.h
+++ b/hw/kdrive/ati/ati_reg.h
@@ -1312,6 +1312,11 @@
 # define R200_BORDER_MODE_D3D			(1  << 31)
 
 #define R200_REG_PP_TXFORMAT_0			0x2c04
+#define R200_REG_PP_TXFORMAT_1			0x2c24
+#define R200_REG_PP_TXFORMAT_2			0x2c44
+#define R200_REG_PP_TXFORMAT_3			0x2c64
+#define R200_REG_PP_TXFORMAT_4			0x2c84
+#define R200_REG_PP_TXFORMAT_5			0x2ca4
 # define R200_TXFORMAT_I8			(0 << 0)
 # define R200_TXFORMAT_AI88			(1 << 0)
 # define R200_TXFORMAT_RGB332			(2 << 0)
@@ -1352,6 +1357,11 @@
 # define R200_TXFORMAT_CUBIC_MAP_ENABLE		(1 << 30)
 
 #define R200_REG_PP_TXFORMAT_X_0		0x2c08
+#define R200_REG_PP_TXFORMAT_X_1		0x2c28
+#define R200_REG_PP_TXFORMAT_X_2		0x2c48
+#define R200_REG_PP_TXFORMAT_X_3		0x2c68
+#define R200_REG_PP_TXFORMAT_X_4		0x2c88
+#define R200_REG_PP_TXFORMAT_X_5		0x2ca8
 # define R200_DEPTH_LOG2_MASK			(0xf << 0)
 # define R200_DEPTH_LOG2_SHIFT			0
 # define R200_VOLUME_FILTER_SHIFT		4
@@ -1383,7 +1393,19 @@
 # define R200_LOD_BIAS_SHIFT			19
 
 #define R200_REG_PP_TXSIZE_0			0x2c0c /* NPOT only */
+#define R200_REG_PP_TXSIZE_1			0x2c2c /* NPOT only */
+#define R200_REG_PP_TXSIZE_2			0x2c4c /* NPOT only */
+#define R200_REG_PP_TXSIZE_3			0x2c6c /* NPOT only */
+#define R200_REG_PP_TXSIZE_4			0x2c8c /* NPOT only */
+#define R200_REG_PP_TXSIZE_5			0x2cac /* NPOT only */
+
 #define R200_REG_PP_TXPITCH_0			0x2c10 /* NPOT only */
+#define R200_REG_PP_TXPITCH_1			0x2c30 /* NPOT only */
+#define R200_REG_PP_TXPITCH_2			0x2c50 /* NPOT only */
+#define R200_REG_PP_TXPITCH_3			0x2c70 /* NPOT only */
+#define R200_REG_PP_TXPITCH_4			0x2c90 /* NPOT only */
+#define R200_REG_PP_TXPITCH_5			0x2cb0 /* NPOT only */
+
 #define R200_REG_PP_BORDER_COLOR_0		0x2c14
 #define R200_REG_PP_TXMULTI_CTL_0		0x2c1c
 
diff --git a/hw/kdrive/ati/ati_video.c b/hw/kdrive/ati/ati_video.c
index 3960e6e8e..8e47d9ed0 100644
--- a/hw/kdrive/ati/ati_video.c
+++ b/hw/kdrive/ati/ati_video.c
@@ -217,11 +217,13 @@ R128DisplayVideo(KdScreenInfo *screen, ATIPortPrivPtr pPortPriv)
 	OUT_REG(R128_REG_SCALE_3D_DATATYPE, srcDatatype);
 
 	OUT_RING(DMA_PACKET0(R128_REG_SCALE_PITCH, 5));
-	OUT_RING(pPortPriv->src_pitch / 16);
-	OUT_RING((pPortPriv->src_w << 16) / pPortPriv->dst_w);
-	OUT_RING((pPortPriv->src_h << 16) / pPortPriv->dst_h);
-	OUT_RING(0x0);
-	OUT_RING(0x0);
+	OUT_RING_REG(R128_REG_SCALE_PITCH, pPortPriv->src_pitch / 16);
+	OUT_RING_REG(R128_REG_SCALE_X_INC,
+	    (pPortPriv->src_w << 16) / pPortPriv->dst_w);
+	OUT_RING_REG(R128_REG_SCALE_Y_INC,
+	    (pPortPriv->src_h << 16) / pPortPriv->dst_h);
+	OUT_RING_REG(R128_REG_SCALE_HACC, 0x0);
+	OUT_RING_REG(R128_REG_SCALE_VACC, 0x0);
 
 	END_DMA();
 
@@ -240,19 +242,16 @@ R128DisplayVideo(KdScreenInfo *screen, ATIPortPrivPtr pPortPriv)
 		srch = pPortPriv->src_h - srcY;
 
 		BEGIN_DMA(6);
-		/* R128_REG_SCALE_SRC_HEIGHT_WIDTH,
-		 * R128_REG_SCALE_OFFSET_0
-		 */
 		OUT_RING(DMA_PACKET0(R128_REG_SCALE_SRC_HEIGHT_WIDTH, 2));
-		OUT_RING((srch << 16) | srcw);
-		OUT_RING(pPortPriv->src_offset + srcY * pPortPriv->src_pitch +
-		    srcX * 2);
-		/* R128_REG_SCALE_DST_X_Y
-		 * R128_REG_SCALE_DST_HEIGHT_WIDTH
-		 */
+		OUT_RING_REG(R128_REG_SCALE_SRC_HEIGHT_WIDTH,
+		    (srch << 16) | srcw);
+		OUT_RING_REG(R128_REG_SCALE_OFFSET_0, pPortPriv->src_offset +
+		    srcY * pPortPriv->src_pitch + srcX * 2);
+
 		OUT_RING(DMA_PACKET0(R128_REG_SCALE_DST_X_Y, 2));
-		OUT_RING((dstX << 16) | dstY);
-		OUT_RING((dsth << 16) | dstw);
+		OUT_RING_REG(R128_REG_SCALE_DST_X_Y, (dstX << 16) | dstY);
+		OUT_RING_REG(R128_REG_SCALE_DST_HEIGHT_WIDTH,
+		    (dsth << 16) | dstw);
 		END_DMA();
 		pBox++;
 	}
@@ -340,14 +339,12 @@ RadeonDisplayVideo(KdScreenInfo *screen, ATIPortPrivPtr pPortPriv)
 
 	BEGIN_DMA(8);
 
-	/* RADEON_REG_PP_CNTL,
-	 * RADEON_REG_RB3D_CNTL, 
-	 * RADEON_REG_RB3D_COLOROFFSET
-	 */
 	OUT_RING(DMA_PACKET0(RADEON_REG_PP_CNTL, 3));
-	OUT_RING(RADEON_TEX_0_ENABLE | RADEON_TEX_BLEND_0_ENABLE);
-	OUT_RING(dst_format | RADEON_ALPHA_BLEND_ENABLE);
-	OUT_RING(dst_offset);
+	OUT_RING_REG(RADEON_REG_PP_CNTL,
+	    RADEON_TEX_0_ENABLE | RADEON_TEX_BLEND_0_ENABLE);
+	OUT_RING_REG(RADEON_REG_RB3D_CNTL,
+	    dst_format | RADEON_ALPHA_BLEND_ENABLE);
+	OUT_RING_REG(RADEON_REG_RB3D_COLOROFFSET, dst_offset);
 
 	OUT_REG(RADEON_REG_RB3D_COLORPITCH, dst_pitch >> pixel_shift);
 
@@ -363,79 +360,66 @@ RadeonDisplayVideo(KdScreenInfo *screen, ATIPortPrivPtr pPortPriv)
 		OUT_REG(R200_REG_SE_VTX_FMT_1,
 		    (2 << R200_VTX_TEX0_COMP_CNT_SHIFT));
 
-		/* R200_REG_PP_TXFILTER_0,
-		 * R200_REG_PP_TXFORMAT_0,
-		 * R200_REG_PP_TXFORMAT_X_0,
-		 * R200_REG_PP_TXSIZE_0,
-		 * R200_REG_PP_TXPITCH_0
-		 */
 		OUT_RING(DMA_PACKET0(R200_REG_PP_TXFILTER_0, 5));
-		OUT_RING(R200_MAG_FILTER_LINEAR |
+		OUT_RING_REG(R200_REG_PP_TXFILTER_0,
+		    R200_MAG_FILTER_LINEAR |
 		    R200_MIN_FILTER_LINEAR |
 		    R200_YUV_TO_RGB);
-		OUT_RING(txformat);
-		OUT_RING(0);
-		OUT_RING((pPixmap->drawable.width - 1) |
+		OUT_RING_REG(R200_REG_PP_TXFORMAT_0, txformat);
+		OUT_RING_REG(R200_REG_PP_TXFORMAT_X_0, 0);
+		OUT_RING_REG(R200_REG_PP_TXSIZE_0,
+		    (pPixmap->drawable.width - 1) |
 		    ((pPixmap->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
-		OUT_RING(pPortPriv->src_pitch - 32);
+		OUT_RING_REG(R200_REG_PP_TXPITCH_0, pPortPriv->src_pitch - 32);
 
 		OUT_REG(R200_PP_TXOFFSET_0, pPortPriv->src_offset);
 
-		/* R200_REG_PP_TXCBLEND_0,
-		 * R200_REG_PP_TXCBLEND2_0
-		 * R200_REG_PP_TXABLEND_0
-		 * R200_REG_PP_TXABLEND2_0
-		 */
 		OUT_RING(DMA_PACKET0(R200_REG_PP_TXCBLEND_0, 4));
-		OUT_RING(
+		OUT_RING_REG(R200_REG_PP_TXCBLEND_0,
 		    R200_TXC_ARG_A_ZERO |
 		    R200_TXC_ARG_B_ZERO |
 		    R200_TXC_ARG_C_R0_COLOR |
 		    R200_TXC_OP_MADD);
-		OUT_RING(R200_TXC_CLAMP_0_1 | R200_TXC_OUTPUT_REG_R0);
-		OUT_RING(
+		OUT_RING_REG(R200_REG_PP_TXCBLEND2_0,
+		    R200_TXC_CLAMP_0_1 | R200_TXC_OUTPUT_REG_R0);
+		OUT_RING_REG(R200_REG_PP_TXABLEND_0,
 		    R200_TXA_ARG_A_ZERO |
 		    R200_TXA_ARG_B_ZERO |
 		    R200_TXA_ARG_C_R0_ALPHA |
 		    R200_TXA_OP_MADD);
-		OUT_RING(R200_TXA_CLAMP_0_1 | R200_TXA_OUTPUT_REG_R0);
+		OUT_RING_REG(R200_REG_PP_TXABLEND2_0,
+		    R200_TXA_CLAMP_0_1 | R200_TXA_OUTPUT_REG_R0);
 
 		END_DMA();
 	} else {
-		BEGIN_DMA(11);
-
-		/* RADEON_REG_PP_TXFILTER_0,
-		 * RADEON_REG_PP_TXFORMAT_0,
-		 * RADEON_REG_PP_TXOFFSET_0
-		 */
-		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TXFILTER_0, 3));
-		OUT_RING(RADEON_MAG_FILTER_LINEAR |
+		BEGIN_DMA(9);
+
+		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TXFILTER_0, 5));
+		OUT_RING_REG(RADEON_REG_PP_TXFILTER_0, RADEON_MAG_FILTER_LINEAR |
 		    RADEON_MIN_FILTER_LINEAR |
 		    RADEON_YUV_TO_RGB);
-		OUT_RING(txformat);
-		OUT_RING(pPortPriv->src_offset);
-
-		/* RADEON_REG_PP_TEX_SIZE_0,
-		 * RADEON_REG_PP_TEX_PITCH_0
-		 */
-		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TEX_SIZE_0, 2));
-		OUT_RING((pPixmap->drawable.width - 1) |
-		    ((pPixmap->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
-		OUT_RING(pPortPriv->src_pitch - 32);
-
-		OUT_REG(RADEON_REG_PP_TXCBLEND_0,
+		OUT_RING_REG(RADEON_REG_PP_TXFORMAT_0, txformat);
+		OUT_RING_REG(RADEON_REG_PP_TXOFFSET_0, pPortPriv->src_offset);
+		OUT_RING_REG(RADEON_REG_PP_TXCBLEND_0,
 		    RADEON_COLOR_ARG_A_ZERO |
 		    RADEON_COLOR_ARG_B_ZERO |
 		    RADEON_COLOR_ARG_C_T0_COLOR |
 		    RADEON_BLEND_CTL_ADD |
 		    RADEON_CLAMP_TX);
-		OUT_REG(RADEON_REG_PP_TXABLEND_0,
+		OUT_RING_REG(RADEON_REG_PP_TXABLEND_0,
 		    RADEON_ALPHA_ARG_A_ZERO |
 		    RADEON_ALPHA_ARG_B_ZERO |
 		    RADEON_ALPHA_ARG_C_T0_ALPHA |
 		    RADEON_BLEND_CTL_ADD |
 		    RADEON_CLAMP_TX);
 
+		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TEX_SIZE_0, 2));
+		OUT_RING_REG(RADEON_REG_PP_TEX_SIZE_0,
+		    (pPixmap->drawable.width - 1) |
+		    ((pPixmap->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
+		OUT_RING_REG(RADEON_REG_PP_TEX_PITCH_0,
+		    pPortPriv->src_pitch - 32);
+
 		END_DMA();
 	}
 
diff --git a/hw/kdrive/ati/r128_composite.c b/hw/kdrive/ati/r128_composite.c
index 39ec33da2..46d4edb72 100644
--- a/hw/kdrive/ati/r128_composite.c
+++ b/hw/kdrive/ati/r128_composite.c
@@ -308,13 +308,8 @@ R128PrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
 	 * in the mask, depending on componentAlpha.
 	 */
 	BEGIN_DMA(15);
-	/* R128_REG_PRIM_TEX_CNTL_C,
-	 * R128_REG_PRIM_TEXTURE_COMBINE_CNTL_C,
-	 * R128_REG_TEX_SIZE_PITCH_C,
-	 * R128_REG_PRIM_TEX_0_OFFSET_C - R128_REG_PRIM_TEX_10_OFFSET_C
-	 */
 	OUT_RING(DMA_PACKET0(R128_REG_PRIM_TEX_CNTL_C, 14));
-	OUT_RING(prim_tex_cntl_c);
+	OUT_RING_REG(R128_REG_PRIM_TEX_CNTL_C, prim_tex_cntl_c);
 
 	/* If this is the only stage and the dest is a8, route the alpha result 
 	 * to the color (red channel, in particular), too.  Otherwise, be sure
@@ -332,29 +327,28 @@ R128PrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
 	else
 		alpha_comb = R128_COMB_ALPHA_DIS;
 
-	OUT_RING(R128_COMB_COPY |
+	OUT_RING_REG(R128_REG_PRIM_TEXTURE_COMBINE_CNTL_C,
+	    R128_COMB_COPY |
 	    color_factor |
 	    R128_INPUT_FACTOR_INT_COLOR |
 	    alpha_comb |
 	    R128_ALPHA_FACTOR_TEX_ALPHA |
 	    R128_INP_FACTOR_A_CONST_ALPHA);
-	OUT_RING(txsize);
+	OUT_RING_REG(R128_REG_TEX_SIZE_PITCH_C, txsize);
 	/* We could save some output by only writing the offset register that
 	 * will actually be used.  On the other hand, this is easy.
 	 */
-	for (i = 0; i <= 10; i++)
-		OUT_RING(((CARD8 *)pSrc->devPrivate.ptr -
+	for (i = 0; i <= 10; i++) {
+		OUT_RING_REG(R128_REG_PRIM_TEX_0_OFFSET_C + 4 * i,
+		    ((CARD8 *)pSrc->devPrivate.ptr -
 		    pScreenPriv->screen->memory_base));
+	}
 	END_DMA();
 
 	if (pMask != NULL) {
 		BEGIN_DMA(14);
-		/* R128_REG_SEC_TEX_CNTL_C,
-		 * R128_REG_SEC_TEXTURE_COMBINE_CNTL_C,
-		 * R128_REG_SEC_TEX_0_OFFSET_C - R128_REG_SEC_TEX_10_OFFSET_C
-		 */
 		OUT_RING(DMA_PACKET0(R128_REG_SEC_TEX_CNTL_C, 13));
-		OUT_RING(sec_tex_cntl_c);
+		OUT_RING_REG(R128_REG_SEC_TEX_CNTL_C, sec_tex_cntl_c);
 
 		if (pDstPicture->format == PICT_a8) {
 			color_factor = R128_COLOR_FACTOR_ALPHA;
@@ -367,15 +361,18 @@ R128PrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
 			in_color_factor = R128_INPUT_FACTOR_PREV_COLOR;
 		}
 
-		OUT_RING(R128_COMB_MODULATE |
+		OUT_RING_REG(R128_REG_SEC_TEXTURE_COMBINE_CNTL_C,
+		    R128_COMB_MODULATE |
 		    color_factor |
 		    in_color_factor |
 		    R128_COMB_ALPHA_MODULATE |
 		    R128_ALPHA_FACTOR_TEX_ALPHA |
 		    R128_INP_FACTOR_A_PREV_ALPHA);
-		for (i = 0; i <= 10; i++)
-			OUT_RING(((CARD8 *)pMask->devPrivate.ptr -
+		for (i = 0; i <= 10; i++) {
+			OUT_RING_REG(R128_REG_SEC_TEX_0_OFFSET_C + 4 * i,
+			    ((CARD8 *)pMask->devPrivate.ptr -
 			    pScreenPriv->screen->memory_base));
+		}
 		END_DMA();
 	}
 
@@ -502,18 +499,12 @@ R128PrepareTrapezoids(PicturePtr pDstPicture, PixmapPtr pDst)
 	    R128_ALPHA_ENABLE);
 	OUT_REG(R128_REG_PC_GUI_CTLSTAT, R128_PC_FLUSH_GUI);
 
-	/* R128_REG_AUX_SC_CNTL,
-	 * R128_REG_AUX1_SC_LEFT
-	 * R128_REG_AUX1_SC_RIGHT
-	 * R128_REG_AUX1_SC_TOP
-	 * R128_REG_AUX1_SC_BOTTOM
-	 */
 	OUT_RING(DMA_PACKET0(R128_REG_AUX_SC_CNTL, 5));
-	OUT_RING(R128_AUX1_SC_ENB);
-	OUT_RING(0);
-	OUT_RING(pDst->drawable.width);
-	OUT_RING(0);
-	OUT_RING(pDst->drawable.height);
+	OUT_RING_REG(R128_REG_AUX_SC_CNTL, R128_AUX1_SC_ENB);
+	OUT_RING_REG(R128_REG_AUX1_SC_LEFT, 0);
+	OUT_RING_REG(R128_REG_AUX1_SC_RIGHT, pDst->drawable.width);
+	OUT_RING_REG(R128_REG_AUX1_SC_TOP, 0);
+	OUT_RING_REG(R128_REG_AUX1_SC_BOTTOM, pDst->drawable.height);
 	END_DMA();
 
 	return TRUE;
diff --git a/hw/kdrive/ati/radeon_composite.c b/hw/kdrive/ati/radeon_composite.c
index e2fdbe831..19b2d25ff 100644
--- a/hw/kdrive/ati/radeon_composite.c
+++ b/hw/kdrive/ati/radeon_composite.c
@@ -196,25 +196,28 @@ R100TextureSetup(PicturePtr pPict, PixmapPtr pPix, int unit)
 	if ((txpitch & 0x1f) != 0)
 		ATI_FALLBACK(("Bad texture pitch 0x%x\n", txpitch));
 
-	/* RADEON_REG_PP_TXFILTER_0,
-	 * RADEON_REG_PP_TXFORMAT_0,
-	 * RADEON_REG_PP_TXOFFSET_0
-	 */
-	BEGIN_DMA(4);
-	OUT_RING(DMA_PACKET0(RADEON_REG_PP_TXFILTER_0 + 0x18 * unit, 3));
-	OUT_RING(0);
-	OUT_RING(txformat);
-	OUT_RING(txoffset);
-	END_DMA();
-
-	/* RADEON_REG_PP_TEX_SIZE_0,
-	 * RADEON_REG_PP_TEX_PITCH_0
-	 */
-	BEGIN_DMA(3);
-	OUT_RING(DMA_PACKET0(RADEON_REG_PP_TEX_SIZE_0 + 0x8 * unit, 2));
-	OUT_RING((pPix->drawable.width - 1) |
-	    ((pPix->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
-	OUT_RING(txpitch - 32);
+	BEGIN_DMA(7);
+	if (unit == 0) {
+		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TXFILTER_0, 3));
+		OUT_RING_REG(RADEON_REG_PP_TXFILTER_0, 0);
+		OUT_RING_REG(RADEON_REG_PP_TXFORMAT_0, txformat);
+		OUT_RING_REG(RADEON_REG_PP_TXOFFSET_0, txoffset);
+		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TEX_SIZE_0, 2));
+		OUT_RING_REG(RADEON_REG_PP_TEX_SIZE_0,
+		    (pPix->drawable.width - 1) |
+		    ((pPix->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
+		OUT_RING_REG(RADEON_REG_PP_TEX_PITCH_0, txpitch - 32);
+	} else {
+		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TXFILTER_1, 3));
+		OUT_RING_REG(RADEON_REG_PP_TXFILTER_1, 0);
+		OUT_RING_REG(RADEON_REG_PP_TXFORMAT_1, txformat);
+		OUT_RING_REG(RADEON_REG_PP_TXOFFSET_1, txoffset);
+		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TEX_SIZE_1, 2));
+		OUT_RING_REG(RADEON_REG_PP_TEX_SIZE_1,
+		    (pPix->drawable.width - 1) |
+		    ((pPix->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
+		OUT_RING_REG(RADEON_REG_PP_TEX_PITCH_1, txpitch - 32);
+	}
 	END_DMA();
 
 	if (pPict->transform != 0) {
@@ -302,14 +305,11 @@ R100PrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
 
 	BEGIN_DMA(12);
 
-	/* RADEON_REG_PP_CNTL,
-	 * RADEON_REG_RB3D_CNTL, 
-	 * RADEON_REG_RB3D_COLOROFFSET
-	 */
 	OUT_RING(DMA_PACKET0(RADEON_REG_PP_CNTL, 3));
-	OUT_RING(pp_cntl);
-	OUT_RING(dst_format | RADEON_ALPHA_BLEND_ENABLE);
-	OUT_RING(dst_offset);
+	OUT_RING_REG(RADEON_REG_PP_CNTL, pp_cntl);
+	OUT_RING_REG(RADEON_REG_RB3D_CNTL,
+	    dst_format | RADEON_ALPHA_BLEND_ENABLE);
+	OUT_RING_REG(RADEON_REG_RB3D_COLOROFFSET, dst_offset);
 
 	OUT_REG(RADEON_REG_RB3D_COLORPITCH, dst_pitch >> pixel_shift);
 
@@ -428,21 +428,29 @@ R200TextureSetup(PicturePtr pPict, PixmapPtr pPix, int unit)
 	if ((txpitch & 0x1f) != 0)
 		ATI_FALLBACK(("Bad texture pitch 0x%x\n", txpitch));
 
-	/* R200_REG_PP_TXFILTER_0,
-	 * R200_REG_PP_TXFORMAT_0,
-	 * R200_REG_PP_TXFORMAT_X_0,
-	 * R200_REG_PP_TXSIZE_0,
-	 * R200_REG_PP_TXPITCH_0
-	 */
-	BEGIN_DMA(6);
-	OUT_RING(DMA_PACKET0(R200_REG_PP_TXFILTER_0 + 0x20 * unit, 5));
-	OUT_RING(0);
-	OUT_RING(txformat);
-	OUT_RING(0);
-	OUT_RING((pPix->drawable.width - 1) |
-	    ((pPix->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
-	OUT_RING(txpitch - 32);
-	END_DMA();
+	if (unit == 0) {
+		BEGIN_DMA(6);
+		OUT_RING(DMA_PACKET0(R200_REG_PP_TXFILTER_0 + 0x20 * unit, 5));
+		OUT_RING_REG(R200_REG_PP_TXFILTER_0, 0);
+		OUT_RING_REG(R200_REG_PP_TXFORMAT_0, txformat);
+		OUT_RING_REG(R200_REG_PP_TXFORMAT_X_0, 0);
+		OUT_RING_REG(R200_REG_PP_TXSIZE_0,
+		    (pPix->drawable.width - 1) |
+		    ((pPix->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
+		OUT_RING_REG(R200_REG_PP_TXPITCH_0, txpitch - 32);
+		END_DMA();
+	} else {
+		BEGIN_DMA(6);
+		OUT_RING(DMA_PACKET0(R200_REG_PP_TXFILTER_1, 5));
+		OUT_RING_REG(R200_REG_PP_TXFILTER_1, 0);
+		OUT_RING_REG(R200_REG_PP_TXFORMAT_1, txformat);
+		OUT_RING_REG(R200_REG_PP_TXFORMAT_X_1, 0);
+		OUT_RING_REG(R200_REG_PP_TXSIZE_1,
+		    (pPix->drawable.width - 1) |
+		    ((pPix->drawable.height - 1) << RADEON_TEX_VSIZE_SHIFT));
+		OUT_RING_REG(R200_REG_PP_TXPITCH_1, txpitch - 32);
+		END_DMA();
+	}
 
 	BEGIN_DMA(2);
 	OUT_REG(R200_PP_TXOFFSET_0 + 0x18 * unit, txoffset);
@@ -521,16 +529,12 @@ R200PrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
 
 	RadeonSwitchTo3D(atis);
 
-	BEGIN_DMA(20);
+	BEGIN_DMA(17);
 
-	/* RADEON_REG_PP_CNTL,
-	 * RADEON_REG_RB3D_CNTL, 
-	 * RADEON_REG_RB3D_COLOROFFSET
-	 */
 	OUT_RING(DMA_PACKET0(RADEON_REG_PP_CNTL, 3));
-	OUT_RING(pp_cntl);
-	OUT_RING(dst_format | RADEON_ALPHA_BLEND_ENABLE);
-	OUT_RING(dst_offset);
+	OUT_RING_REG(RADEON_REG_PP_CNTL, pp_cntl);
+	OUT_RING_REG(RADEON_REG_RB3D_CNTL, dst_format | RADEON_ALPHA_BLEND_ENABLE);
+	OUT_RING_REG(RADEON_REG_RB3D_COLOROFFSET, dst_offset);
 
 	OUT_REG(R200_REG_SE_VTX_FMT_0, R200_VTX_XY);
 	OUT_REG(R200_REG_SE_VTX_FMT_1,
@@ -567,14 +571,13 @@ R200PrepareComposite(int op, PicturePtr pSrcPicture, PicturePtr pMaskPicture,
 		ablend |= R200_TXA_ARG_B_ZERO | R200_TXA_COMP_ARG_B;
 	}
 
-	OUT_REG(R200_REG_PP_TXCBLEND_0, cblend);
-	OUT_REG(R200_REG_PP_TXABLEND_0, ablend);
-	OUT_REG(R200_REG_PP_TXCBLEND2_0,
-	    R200_TXC_CLAMP_0_1 |
-	    R200_TXC_OUTPUT_REG_R0);
-	OUT_REG(R200_REG_PP_TXABLEND2_0,
-	    R200_TXA_CLAMP_0_1 |
-	    R200_TXA_OUTPUT_REG_R0);
+	OUT_RING(DMA_PACKET0(R200_REG_PP_TXCBLEND_0, 4));
+	OUT_RING_REG(R200_REG_PP_TXCBLEND_0, cblend);
+	OUT_RING_REG(R200_REG_PP_TXCBLEND2_0,
+	    R200_TXC_CLAMP_0_1 | R200_TXC_OUTPUT_REG_R0);
+	OUT_RING_REG(R200_REG_PP_TXABLEND_0, ablend);
+	OUT_RING_REG(R200_REG_PP_TXABLEND2_0,
+	    R200_TXA_CLAMP_0_1 | R200_TXA_OUTPUT_REG_R0);
 
 	/* Op operator. */
 	blendcntl = RadeonBlendOp[op].blend_cntl;
@@ -730,34 +733,28 @@ RadeonPrepareTrapezoids(PicturePtr pDstPicture, PixmapPtr pDst)
 
 	BEGIN_DMA(8);
 
-	/* RADEON_REG_PP_CNTL,
-	 * RADEON_REG_RB3D_CNTL, 
-	 * RADEON_REG_RB3D_COLOROFFSET,
-	 * RADEON_REG_RE_WIDTH_HEIGHT,
-	 * RADEON_REG_RB3D_COLORPITCH
-	 */
 	OUT_RING(DMA_PACKET0(RADEON_REG_PP_CNTL, 5));
-	OUT_RING(RADEON_TEX_BLEND_0_ENABLE);
-	OUT_RING(RADEON_COLOR_FORMAT_RGB8 | RADEON_ALPHA_BLEND_ENABLE);
-	OUT_RING(dst_offset);
-	OUT_RING(((pDst->drawable.height - 1) << 16) |
+	OUT_RING_REG(RADEON_REG_PP_CNTL, RADEON_TEX_BLEND_0_ENABLE);
+	OUT_RING_REG(RADEON_REG_RB3D_CNTL,
+	    RADEON_COLOR_FORMAT_RGB8 | RADEON_ALPHA_BLEND_ENABLE);
+	OUT_RING_REG(RADEON_REG_RB3D_COLOROFFSET, dst_offset);
+	OUT_RING_REG(RADEON_REG_RE_WIDTH_HEIGHT,
+	    ((pDst->drawable.height - 1) << 16) |
 	    (pDst->drawable.width - 1));
-	OUT_RING(dst_pitch >> pixel_shift);
+	OUT_RING_REG(RADEON_REG_RB3D_COLORPITCH, dst_pitch >> pixel_shift);
 	OUT_REG(RADEON_REG_RB3D_BLENDCNTL, RadeonBlendOp[PictOpAdd].blend_cntl);
 	END_DMA();
 
 	if (atic->is_r100) {
 		BEGIN_DMA(4);
-		/* RADEON_REG_PP_TXCBLEND_0,
-		 * RADEON_REG_PP_TXABLEND_0,
-		 * RADEON_REG_PP_TFACTOR_0
-		 */
 		OUT_RING(DMA_PACKET0(RADEON_REG_PP_TXCBLEND_0, 3));
-		OUT_RING(RADEON_BLEND_CTL_ADD | RADEON_CLAMP_TX |
+		OUT_RING_REG(RADEON_REG_PP_TXCBLEND_0,
+		    RADEON_BLEND_CTL_ADD | RADEON_CLAMP_TX |
 		    RADEON_COLOR_ARG_C_TFACTOR_ALPHA);
-		OUT_RING(RADEON_BLEND_CTL_ADD | RADEON_CLAMP_TX |
+		OUT_RING_REG(RADEON_REG_PP_TXABLEND_0,
+		    RADEON_BLEND_CTL_ADD | RADEON_CLAMP_TX |
 		    RADEON_ALPHA_ARG_C_TFACTOR_ALPHA);
-		OUT_RING(0x01000000);
+		OUT_RING_REG(RADEON_REG_PP_TFACTOR_0, 0x01000000);
 		END_DMA();
 	} else if (atic->is_r200) {
 		BEGIN_DMA(14);