- Fixed polygon stipple. Was still using SubmitPackets...ati-4-1-1-20001202-mergeati-4-1-1-20001201-freezeati-4-1-1-branch

- Implemented HW-assisted depth span functions. All access to the depth
    buffer is done via the engine, so we don't have to worry about the
    differences between linear and tiled memory.
- Implemented significantly more 3D-friendly static partitioning of
    offscreen memory. 2D still gets enough pixmap cache to be performant.
    For example, 800x600@16bpp texture memory goes from ~3mb to 11.5mb on a
    16mb card.

9 files changed, 530 insertions, 50 deletions

diff --git a/linux-core/r128_drv.c b/linux-core/r128_drv.c
index f73d4c0b..89b68696 100644
--- a/linux-core/r128_drv.c
+++ b/linux-core/r128_drv.c
@@ -37,10 +37,10 @@
 
 #define R128_NAME		"r128"
 #define R128_DESC		"ATI Rage 128"
-#define R128_DATE		"20001129"
+#define R128_DATE		"20001201"
 #define R128_MAJOR		2
-#define R128_MINOR		0
-#define R128_PATCHLEVEL		1
+#define R128_MINOR		1
+#define R128_PATCHLEVEL		0
 
 static drm_device_t	r128_device;
 drm_ctx_t		r128_res_ctx;
@@ -119,6 +119,8 @@ static drm_ioctl_desc_t	      r128_ioctls[] = {
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_VERTEX)]  = { r128_cce_vertex,   1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_INDICES)] = { r128_cce_indices,  1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_BLIT)]    = { r128_cce_blit,     1, 0 },
+	[DRM_IOCTL_NR(DRM_IOCTL_R128_DEPTH)]   = { r128_cce_depth,    1, 0 },
+	[DRM_IOCTL_NR(DRM_IOCTL_R128_STIPPLE)] = { r128_cce_stipple,  1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_PACKET)]  = { r128_cce_packet,   1, 0 },
 };
 #define R128_IOCTL_COUNT DRM_ARRAY_SIZE(r128_ioctls)
diff --git a/linux/drm.h b/linux/drm.h
index 5d383427..20691ccf 100644
--- a/linux/drm.h
+++ b/linux/drm.h
@@ -374,7 +374,9 @@ typedef struct drm_agp_info {
 #define DRM_IOCTL_R128_VERTEX	 DRM_IOW( 0x49, drm_r128_vertex_t)
 #define DRM_IOCTL_R128_INDICES	 DRM_IOW( 0x4a, drm_r128_indices_t)
 #define DRM_IOCTL_R128_BLIT	 DRM_IOW( 0x4b, drm_r128_blit_t)
-#define DRM_IOCTL_R128_PACKET	 DRM_IOWR(0x4c, drm_r128_packet_t)
+#define DRM_IOCTL_R128_DEPTH	 DRM_IOW( 0x4c, drm_r128_depth_t)
+#define DRM_IOCTL_R128_STIPPLE	 DRM_IOW( 0x4d, drm_r128_stipple_t)
+#define DRM_IOCTL_R128_PACKET	 DRM_IOWR(0x4e, drm_r128_packet_t)
 
 /* SiS specific ioctls */
 #define SIS_IOCTL_FB_ALLOC     DRM_IOWR( 0x44, drm_sis_mem_t)
diff --git a/linux/r128_cce.c b/linux/r128_cce.c
index cab9acd6..0b1576a9 100644
--- a/linux/r128_cce.c
+++ b/linux/r128_cce.c
@@ -437,18 +437,13 @@ static int r128_do_init_cce( drm_device_t *dev, drm_r128_init_t *init )
 	dev_priv->fb_bpp	= init->fb_bpp;
 	dev_priv->front_offset	= init->front_offset;
 	dev_priv->front_pitch	= init->front_pitch;
-	dev_priv->front_x	= init->front_x;
-	dev_priv->front_y	= init->front_y;
 	dev_priv->back_offset	= init->back_offset;
 	dev_priv->back_pitch	= init->back_pitch;
-	dev_priv->back_x	= init->back_x;
-	dev_priv->back_y	= init->back_y;
 
 	dev_priv->depth_bpp	= init->depth_bpp;
 	dev_priv->depth_offset	= init->depth_offset;
 	dev_priv->depth_pitch	= init->depth_pitch;
-	dev_priv->depth_x	= init->depth_x;
-	dev_priv->depth_y	= init->depth_y;
+	dev_priv->span_offset	= init->span_offset;
 
 	dev_priv->front_pitch_offset_c = (((dev_priv->front_pitch/8) << 21) |
 					  (dev_priv->front_offset >> 5));
@@ -457,6 +452,8 @@ static int r128_do_init_cce( drm_device_t *dev, drm_r128_init_t *init )
 	dev_priv->depth_pitch_offset_c = (((dev_priv->depth_pitch/8) << 21) |
 					  (dev_priv->depth_offset >> 5) |
 					  R128_DST_TILE);
+	dev_priv->span_pitch_offset_c = (((dev_priv->depth_pitch/8) << 21) |
+					 (dev_priv->span_offset >> 5));
 
 	/* FIXME: We want multiple shared areas, including one shared
 	 * only by the X Server and kernel module.
diff --git a/linux/r128_drm.h b/linux/r128_drm.h
index 51e31cd3..b81d2fca 100644
--- a/linux/r128_drm.h
+++ b/linux/r128_drm.h
@@ -94,7 +94,7 @@
 #define R128_LOCAL_TEX_HEAP		0
 #define R128_AGP_TEX_HEAP		1
 #define R128_NR_TEX_HEAPS		2
-#define R128_NR_TEX_REGIONS		16
+#define R128_NR_TEX_REGIONS		64
 #define R128_LOG_TEX_GRANULARITY	16
 
 #define R128_NR_CONTEXT_REGS		12
@@ -196,12 +196,10 @@ typedef struct drm_r128_init {
 
 	unsigned int fb_bpp;
 	unsigned int front_offset, front_pitch;
-	unsigned int front_x, front_y;
 	unsigned int back_offset, back_pitch;
-	unsigned int back_x, back_y;
 	unsigned int depth_bpp;
 	unsigned int depth_offset, depth_pitch;
-	unsigned int depth_x, depth_y;
+	unsigned int span_offset;
 
 	unsigned int fb_offset;
 	unsigned int mmio_offset;
@@ -249,6 +247,24 @@ typedef struct drm_r128_blit {
 	unsigned short width, height;
 } drm_r128_blit_t;
 
+typedef struct drm_r128_depth {
+	enum {
+		R128_WRITE_SPAN		= 0x01,
+		R128_WRITE_PIXELS	= 0x02,
+		R128_READ_SPAN		= 0x03,
+		R128_READ_PIXELS	= 0x04
+	} func;
+	int n;
+	int *x;
+	int *y;
+	unsigned int *buffer;
+	unsigned char *mask;
+} drm_r128_depth_t;
+
+typedef struct drm_r128_stipple {
+	unsigned int *mask;
+} drm_r128_stipple_t;
+
 typedef struct drm_r128_packet {
 	unsigned int *buffer;
 	int count;
diff --git a/linux/r128_drv.c b/linux/r128_drv.c
index f73d4c0b..89b68696 100644
--- a/linux/r128_drv.c
+++ b/linux/r128_drv.c
@@ -37,10 +37,10 @@
 
 #define R128_NAME		"r128"
 #define R128_DESC		"ATI Rage 128"
-#define R128_DATE		"20001129"
+#define R128_DATE		"20001201"
 #define R128_MAJOR		2
-#define R128_MINOR		0
-#define R128_PATCHLEVEL		1
+#define R128_MINOR		1
+#define R128_PATCHLEVEL		0
 
 static drm_device_t	r128_device;
 drm_ctx_t		r128_res_ctx;
@@ -119,6 +119,8 @@ static drm_ioctl_desc_t	      r128_ioctls[] = {
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_VERTEX)]  = { r128_cce_vertex,   1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_INDICES)] = { r128_cce_indices,  1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_BLIT)]    = { r128_cce_blit,     1, 0 },
+	[DRM_IOCTL_NR(DRM_IOCTL_R128_DEPTH)]   = { r128_cce_depth,    1, 0 },
+	[DRM_IOCTL_NR(DRM_IOCTL_R128_STIPPLE)] = { r128_cce_stipple,  1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_R128_PACKET)]  = { r128_cce_packet,   1, 0 },
 };
 #define R128_IOCTL_COUNT DRM_ARRAY_SIZE(r128_ioctls)
diff --git a/linux/r128_drv.h b/linux/r128_drv.h
index 1aa6bde5..cffd0800 100644
--- a/linux/r128_drv.h
+++ b/linux/r128_drv.h
@@ -73,22 +73,18 @@ typedef struct drm_r128_private {
 	unsigned int fb_bpp;
 	unsigned int front_offset;
 	unsigned int front_pitch;
-	unsigned int front_x;
-	unsigned int front_y;
 	unsigned int back_offset;
 	unsigned int back_pitch;
-	unsigned int back_x;
-	unsigned int back_y;
 
 	unsigned int depth_bpp;
 	unsigned int depth_offset;
 	unsigned int depth_pitch;
-	unsigned int depth_x;
-	unsigned int depth_y;
+	unsigned int span_offset;
 
 	u32 front_pitch_offset_c;
 	u32 back_pitch_offset_c;
 	u32 depth_pitch_offset_c;
+	u32 span_pitch_offset_c;
 
 	drm_map_t *sarea;
 	drm_map_t *fb;
@@ -154,6 +150,10 @@ extern int r128_cce_indices( struct inode *inode, struct file *filp,
 			     unsigned int cmd, unsigned long arg );
 extern int r128_cce_blit( struct inode *inode, struct file *filp,
 			  unsigned int cmd, unsigned long arg );
+extern int r128_cce_depth( struct inode *inode, struct file *filp,
+			   unsigned int cmd, unsigned long arg );
+extern int r128_cce_stipple( struct inode *inode, struct file *filp,
+			     unsigned int cmd, unsigned long arg );
 
 				/* r128_bufs.c */
 extern int r128_addbufs(struct inode *inode, struct file *filp,
@@ -208,6 +208,7 @@ extern int  r128_context_switch_complete(drm_device_t *dev, int new);
 #define R128_AUX3_SC_TOP		0x168c
 #define R128_AUX3_SC_BOTTOM		0x1690
 
+#define R128_BRUSH_DATA0		0x1480
 #define R128_BUS_CNTL			0x0030
 #	define R128_BUS_MASTER_DIS		(1 << 6)
 
@@ -425,7 +426,7 @@ extern int R128_READ_PLL(drm_device_t *dev, int addr);
 
 #define R128_VERBOSE	0
 
-#define RING_LOCALS	int write; unsigned int mask; volatile u32 *ring;
+#define RING_LOCALS	int write; unsigned int tail_mask; volatile u32 *ring;
 
 #define BEGIN_RING( n ) do {						\
 	if ( R128_VERBOSE ) {						\
@@ -438,7 +439,7 @@ extern int R128_READ_PLL(drm_device_t *dev, int addr);
 	dev_priv->ring.space -= n * sizeof(u32);			\
 	ring = dev_priv->ring.start;					\
 	write = dev_priv->ring.tail;					\
-	mask = dev_priv->ring.tail_mask;				\
+	tail_mask = dev_priv->ring.tail_mask;				\
 } while (0)
 
 #define ADVANCE_RING() do {						\
@@ -457,7 +458,7 @@ extern int R128_READ_PLL(drm_device_t *dev, int addr);
 			   (unsigned int)(x), write );			\
 	}								\
 	ring[write++] = x;						\
-	write &= mask;							\
+	write &= tail_mask;						\
 } while (0)
 
 #define R128_PERFORMANCE_BOXES	0
diff --git a/linux/r128_state.c b/linux/r128_state.c
index 6671518f..faad25a2 100644
--- a/linux/r128_state.c
+++ b/linux/r128_state.c
@@ -534,10 +534,10 @@ static void r128_cce_dispatch_swap( drm_device_t *dev )
 		OUT_RING( R128_GMC_SRC_PITCH_OFFSET_CNTL
 			  | R128_GMC_DST_PITCH_OFFSET_CNTL
 			  | R128_GMC_BRUSH_NONE
-			  | R128_GMC_SRC_DATATYPE_COLOR
-			  | R128_DP_SRC_SOURCE_MEMORY
 			  | fb_bpp
+			  | R128_GMC_SRC_DATATYPE_COLOR
 			  | R128_ROP3_S
+			  | R128_DP_SRC_SOURCE_MEMORY
 			  | R128_GMC_CLR_CMP_CNTL_DIS
 			  | R128_GMC_AUX_CLIP_DIS
 			  | R128_GMC_WR_MSK_DIS );
@@ -550,12 +550,12 @@ static void r128_cce_dispatch_swap( drm_device_t *dev )
 		OUT_RING( (x << 16) | y );
 		OUT_RING( (w << 16) | h );
 #else
-		OUT_RING( dev_priv->depth_pitch_offset_c & ~R128_DST_TILE );
+		OUT_RING( dev_priv->depth_pitch_offset_c /*& ~R128_DST_TILE*/ );
 		OUT_RING( dev_priv->front_pitch_offset_c );
 
 		OUT_RING( (0 << 16) | 0 );
 		OUT_RING( (0 << 16) | 0 );
-		OUT_RING( (1024 << 16) | 768 );
+		OUT_RING( (800 << 16) | 600 );
 #endif
 
 		ADVANCE_RING();
@@ -581,9 +581,7 @@ static void r128_cce_dispatch_vertex( drm_device_t *dev,
 	drm_r128_private_t *dev_priv = dev->dev_private;
 	drm_r128_buf_priv_t *buf_priv = buf->dev_private;
 	drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
-	int vertsize = sarea_priv->vertsize;
 	int format = sarea_priv->vc_format;
-	int index = buf->idx;
 	int offset = dev_priv->buffers->offset + buf->offset - dev->agp->base;
 	int size = buf->used;
 	int prim = buf_priv->prim;
@@ -691,21 +689,6 @@ static void r128_cce_dispatch_indirect( drm_device_t *dev,
 			data[dwords++] = R128_CCE_PACKET2;
 		}
 
-		DRM_DEBUG( "indirect: offset=0x%x dwords=%d\n",
-			   offset, dwords );
-
-		if ( 0 ) {
-			u32 *data = (u32 *)
-				((char *)dev_priv->buffers->handle
-				 + buf->offset + start);
-			int i;
-			DRM_INFO( "data = %p\n", data );
-			for ( i = 0 ; i < dwords ; i++ ) {
-				DRM_INFO( "data[0x%x] = 0x%08x\n",
-					  i, data[i] );
-			}
-		}
-
 		buf_priv->dispatched = 1;
 
 		/* Fire off the indirect buffer */
@@ -941,7 +924,422 @@ static int r128_cce_dispatch_blit( drm_device_t *dev,
 
 
 /* ================================================================
- *
+ * Tiled depth buffer management
+ */
+
+static int r128_cce_dispatch_write_span( drm_device_t *dev,
+					 drm_r128_depth_t *depth )
+{
+	drm_r128_private_t *dev_priv = dev->dev_private;
+	int count, x, y;
+	u32 *buffer;
+	u8 *mask;
+	u32 depth_bpp;
+	int i;
+	RING_LOCALS;
+	DRM_DEBUG( "%s\n", __FUNCTION__ );
+
+	r128_update_ring_snapshot( dev_priv );
+
+	switch ( dev_priv->depth_bpp ) {
+	case 16:
+		depth_bpp = R128_GMC_DST_16BPP;
+		break;
+	case 24:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	case 32:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	count = depth->n;
+	if ( copy_from_user( &x, depth->x, sizeof(x) ) ) {
+		return -EFAULT;
+	}
+	if ( copy_from_user( &y, depth->y, sizeof(y) ) ) {
+		return -EFAULT;
+	}
+
+	buffer = kmalloc( depth->n * sizeof(u32), 0 );
+	if ( buffer == NULL )
+		return -ENOMEM;
+	if ( copy_from_user( buffer, depth->buffer,
+			     depth->n * sizeof(u32) ) ) {
+		kfree( buffer );
+		return -EFAULT;
+	}
+
+	if ( depth->mask ) {
+		mask = kmalloc( depth->n * sizeof(u8), 0 );
+		if ( mask == NULL ) {
+			kfree( buffer );
+			return -ENOMEM;
+		}
+		if ( copy_from_user( mask, depth->mask,
+				     depth->n * sizeof(u8) ) ) {
+			kfree( buffer );
+			kfree( mask );
+			return -EFAULT;
+		}
+
+		for ( i = 0 ; i < count ; i++, x++ ) {
+			if ( mask[i] ) {
+				BEGIN_RING( 6 );
+
+				OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI,
+						       4 ) );
+				OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL
+					  | R128_GMC_BRUSH_SOLID_COLOR
+					  | depth_bpp
+					  | R128_GMC_SRC_DATATYPE_COLOR
+					  | R128_ROP3_P
+					  | R128_GMC_CLR_CMP_CNTL_DIS
+					  | R128_GMC_WR_MSK_DIS );
+
+				OUT_RING( dev_priv->depth_pitch_offset_c );
+				OUT_RING( buffer[i] );
+
+				OUT_RING( (x << 16) | y );
+				OUT_RING( (1 << 16) | 1 );
+
+				ADVANCE_RING();
+			}
+		}
+
+		kfree( mask );
+	} else {
+		for ( i = 0 ; i < count ; i++, x++ ) {
+			BEGIN_RING( 6 );
+
+			OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
+			OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL
+				  | R128_GMC_BRUSH_SOLID_COLOR
+				  | depth_bpp
+				  | R128_GMC_SRC_DATATYPE_COLOR
+				  | R128_ROP3_P
+				  | R128_GMC_CLR_CMP_CNTL_DIS
+				  | R128_GMC_WR_MSK_DIS );
+
+			OUT_RING( dev_priv->depth_pitch_offset_c );
+			OUT_RING( buffer[i] );
+
+			OUT_RING( (x << 16) | y );
+			OUT_RING( (1 << 16) | 1 );
+
+			ADVANCE_RING();
+		}
+	}
+
+	kfree( buffer );
+
+	return 0;
+}
+
+static int r128_cce_dispatch_write_pixels( drm_device_t *dev,
+					   drm_r128_depth_t *depth )
+{
+	drm_r128_private_t *dev_priv = dev->dev_private;
+	int count, *x, *y;
+	u32 *buffer;
+	u8 *mask;
+	u32 depth_bpp;
+	int i;
+	RING_LOCALS;
+	DRM_DEBUG( "%s\n", __FUNCTION__ );
+
+	r128_update_ring_snapshot( dev_priv );
+
+	switch ( dev_priv->depth_bpp ) {
+	case 16:
+		depth_bpp = R128_GMC_DST_16BPP;
+		break;
+	case 24:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	case 32:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	count = depth->n;
+
+	x = kmalloc( count * sizeof(*x), 0 );
+	if ( x == NULL ) {
+		return -ENOMEM;
+	}
+	y = kmalloc( count * sizeof(*y), 0 );
+	if ( y == NULL ) {
+		kfree( x );
+		return -ENOMEM;
+	}
+	if ( copy_from_user( x, depth->x, count * sizeof(int) ) ) {
+		kfree( x );
+		kfree( y );
+		return -EFAULT;
+	}
+	if ( copy_from_user( y, depth->y, count * sizeof(int) ) ) {
+		kfree( x );
+		kfree( y );
+		return -EFAULT;
+	}
+
+	buffer = kmalloc( depth->n * sizeof(u32), 0 );
+	if ( buffer == NULL ) {
+		kfree( x );
+		kfree( y );
+		return -ENOMEM;
+	}
+	if ( copy_from_user( buffer, depth->buffer,
+			     depth->n * sizeof(u32) ) ) {
+		kfree( x );
+		kfree( y );
+		kfree( buffer );
+		return -EFAULT;
+	}
+
+	if ( depth->mask ) {
+		mask = kmalloc( depth->n * sizeof(u8), 0 );
+		if ( mask == NULL ) {
+			kfree( x );
+			kfree( y );
+			kfree( buffer );
+			return -ENOMEM;
+		}
+		if ( copy_from_user( mask, depth->mask,
+				     depth->n * sizeof(u8) ) ) {
+			kfree( x );
+			kfree( y );
+			kfree( buffer );
+			kfree( mask );
+			return -EFAULT;
+		}
+
+		for ( i = 0 ; i < count ; i++ ) {
+			if ( mask[i] ) {
+				BEGIN_RING( 6 );
+
+				OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI,
+						       4 ) );
+				OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL
+					  | R128_GMC_BRUSH_SOLID_COLOR
+					  | depth_bpp
+					  | R128_GMC_SRC_DATATYPE_COLOR
+					  | R128_ROP3_P
+					  | R128_GMC_CLR_CMP_CNTL_DIS
+					  | R128_GMC_WR_MSK_DIS );
+
+				OUT_RING( dev_priv->depth_pitch_offset_c );
+				OUT_RING( buffer[i] );
+
+				OUT_RING( (x[i] << 16) | y[i] );
+				OUT_RING( (1 << 16) | 1 );
+
+				ADVANCE_RING();
+			}
+		}
+
+		kfree( mask );
+	} else {
+		for ( i = 0 ; i < count ; i++ ) {
+			BEGIN_RING( 6 );
+
+			OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
+			OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL
+				  | R128_GMC_BRUSH_SOLID_COLOR
+				  | depth_bpp
+				  | R128_GMC_SRC_DATATYPE_COLOR
+				  | R128_ROP3_P
+				  | R128_GMC_CLR_CMP_CNTL_DIS
+				  | R128_GMC_WR_MSK_DIS );
+
+			OUT_RING( dev_priv->depth_pitch_offset_c );
+			OUT_RING( buffer[i] );
+
+			OUT_RING( (x[i] << 16) | y[i] );
+			OUT_RING( (1 << 16) | 1 );
+
+			ADVANCE_RING();
+		}
+	}
+
+	kfree( x );
+	kfree( y );
+	kfree( buffer );
+
+	return 0;
+}
+
+static int r128_cce_dispatch_read_span( drm_device_t *dev,
+					drm_r128_depth_t *depth )
+{
+	drm_r128_private_t *dev_priv = dev->dev_private;
+	int count, x, y;
+	u32 depth_bpp;
+	RING_LOCALS;
+	DRM_DEBUG( "%s\n", __FUNCTION__ );
+
+	r128_update_ring_snapshot( dev_priv );
+
+	switch ( dev_priv->depth_bpp ) {
+	case 16:
+		depth_bpp = R128_GMC_DST_16BPP;
+		break;
+	case 24:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	case 32:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	count = depth->n;
+	if ( copy_from_user( &x, depth->x, sizeof(x) ) ) {
+		return -EFAULT;
+	}
+	if ( copy_from_user( &y, depth->y, sizeof(y) ) ) {
+		return -EFAULT;
+	}
+
+	BEGIN_RING( 7 );
+
+	OUT_RING( CCE_PACKET3( R128_CNTL_BITBLT_MULTI, 5 ) );
+	OUT_RING( R128_GMC_SRC_PITCH_OFFSET_CNTL
+		  | R128_GMC_DST_PITCH_OFFSET_CNTL
+		  | R128_GMC_BRUSH_NONE
+		  | depth_bpp
+		  | R128_GMC_SRC_DATATYPE_COLOR
+		  | R128_ROP3_S
+		  | R128_DP_SRC_SOURCE_MEMORY
+		  | R128_GMC_CLR_CMP_CNTL_DIS
+		  | R128_GMC_WR_MSK_DIS );
+
+	OUT_RING( dev_priv->depth_pitch_offset_c );
+	OUT_RING( dev_priv->span_pitch_offset_c );
+
+	OUT_RING( (x << 16) | y );
+	OUT_RING( (0 << 16) | 0 );
+	OUT_RING( (count << 16) | 1 );
+
+	ADVANCE_RING();
+
+	return 0;
+}
+
+static int r128_cce_dispatch_read_pixels( drm_device_t *dev,
+					  drm_r128_depth_t *depth )
+{
+	drm_r128_private_t *dev_priv = dev->dev_private;
+	int count, *x, *y;
+	u32 depth_bpp;
+	int i;
+	RING_LOCALS;
+	DRM_DEBUG( "%s\n", __FUNCTION__ );
+
+	r128_update_ring_snapshot( dev_priv );
+
+	switch ( dev_priv->depth_bpp ) {
+	case 16:
+		depth_bpp = R128_GMC_DST_16BPP;
+		break;
+	case 24:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	case 32:
+		depth_bpp = R128_GMC_DST_32BPP;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	count = depth->n;
+	if ( count > dev_priv->depth_pitch ) {
+		count = dev_priv->depth_pitch;
+	}
+
+	x = kmalloc( count * sizeof(*x), 0 );
+	if ( x == NULL ) {
+		return -ENOMEM;
+	}
+	y = kmalloc( count * sizeof(*y), 0 );
+	if ( y == NULL ) {
+		kfree( x );
+		return -ENOMEM;
+	}
+	if ( copy_from_user( x, depth->x, count * sizeof(int) ) ) {
+		kfree( x );
+		kfree( y );
+		return -EFAULT;
+	}
+	if ( copy_from_user( y, depth->y, count * sizeof(int) ) ) {
+		kfree( x );
+		kfree( y );
+		return -EFAULT;
+	}
+
+	for ( i = 0 ; i < count ; i++ ) {
+		BEGIN_RING( 7 );
+
+		OUT_RING( CCE_PACKET3( R128_CNTL_BITBLT_MULTI, 5 ) );
+		OUT_RING( R128_GMC_SRC_PITCH_OFFSET_CNTL
+			  | R128_GMC_DST_PITCH_OFFSET_CNTL
+			  | R128_GMC_BRUSH_NONE
+			  | depth_bpp
+			  | R128_GMC_SRC_DATATYPE_COLOR
+			  | R128_ROP3_S
+			  | R128_DP_SRC_SOURCE_MEMORY
+			  | R128_GMC_CLR_CMP_CNTL_DIS
+			  | R128_GMC_WR_MSK_DIS );
+
+		OUT_RING( dev_priv->depth_pitch_offset_c );
+		OUT_RING( dev_priv->span_pitch_offset_c );
+
+		OUT_RING( (x[i] << 16) | y[i] );
+		OUT_RING( (i << 16) | 0 );
+		OUT_RING( (1 << 16) | 1 );
+
+		ADVANCE_RING();
+	}
+
+	kfree( x );
+	kfree( y );
+
+	return 0;
+}
+
+
+/* ================================================================
+ * Polygon stipple
+ */
+
+static void r128_cce_dispatch_stipple( drm_device_t *dev, u32 *stipple )
+{
+	drm_r128_private_t *dev_priv = dev->dev_private;
+	int i;
+	RING_LOCALS;
+	DRM_DEBUG( "%s\n", __FUNCTION__ );
+
+	r128_update_ring_snapshot( dev_priv );
+
+	BEGIN_RING( 33 );
+
+	OUT_RING(  CCE_PACKET0( R128_BRUSH_DATA0, 31 ) );
+	for ( i = 0 ; i < 32 ; i++ ) {
+		OUT_RING( stipple[i] );
+	}
+
+	ADVANCE_RING();
+}
+
+
+/* ================================================================
+ * IOCTL functions
  */
 
 int r128_cce_clear( struct inode *inode, struct file *filp,
@@ -1172,3 +1570,61 @@ int r128_cce_blit( struct inode *inode, struct file *filp,
 
 	return r128_cce_dispatch_blit( dev, &blit );
 }
+
+int r128_cce_depth( struct inode *inode, struct file *filp,
+		    unsigned int cmd, unsigned long arg )
+{
+	drm_file_t *priv = filp->private_data;
+	drm_device_t *dev = priv->dev;
+	drm_r128_depth_t depth;
+
+	if ( !_DRM_LOCK_IS_HELD( dev->lock.hw_lock->lock ) ||
+	     dev->lock.pid != current->pid ) {
+		DRM_ERROR( "%s called without lock held\n", __FUNCTION__ );
+		return -EINVAL;
+	}
+
+	if ( copy_from_user( &depth, (drm_r128_depth_t *)arg,
+			     sizeof(depth) ) )
+		return -EFAULT;
+
+	switch ( depth.func ) {
+	case R128_WRITE_SPAN:
+		return r128_cce_dispatch_write_span( dev, &depth );
+	case R128_WRITE_PIXELS:
+		return r128_cce_dispatch_write_pixels( dev, &depth );
+	case R128_READ_SPAN:
+		return r128_cce_dispatch_read_span( dev, &depth );
+	case R128_READ_PIXELS:
+		return r128_cce_dispatch_read_pixels( dev, &depth );
+	}
+
+	return -EINVAL;
+}
+
+int r128_cce_stipple( struct inode *inode, struct file *filp,
+		      unsigned int cmd, unsigned long arg )
+{
+	drm_file_t *priv = filp->private_data;
+	drm_device_t *dev = priv->dev;
+	drm_r128_stipple_t stipple;
+	u32 mask[32];
+
+	if ( !_DRM_LOCK_IS_HELD( dev->lock.hw_lock->lock ) ||
+	     dev->lock.pid != current->pid ) {
+		DRM_ERROR( "%s called without lock held\n", __FUNCTION__ );
+		return -EINVAL;
+	}
+
+	if ( copy_from_user( &stipple, (drm_r128_stipple_t *)arg,
+			     sizeof(stipple) ) )
+		return -EFAULT;
+
+	if ( copy_from_user( &mask, stipple.mask,
+			     32 * sizeof(u32) ) )
+		return -EFAULT;
+
+	r128_cce_dispatch_stipple( dev, mask );
+
+	return 0;
+}
diff --git a/shared-core/drm.h b/shared-core/drm.h
index 5d383427..20691ccf 100644
--- a/shared-core/drm.h
+++ b/shared-core/drm.h
@@ -374,7 +374,9 @@ typedef struct drm_agp_info {
 #define DRM_IOCTL_R128_VERTEX	 DRM_IOW( 0x49, drm_r128_vertex_t)
 #define DRM_IOCTL_R128_INDICES	 DRM_IOW( 0x4a, drm_r128_indices_t)
 #define DRM_IOCTL_R128_BLIT	 DRM_IOW( 0x4b, drm_r128_blit_t)
-#define DRM_IOCTL_R128_PACKET	 DRM_IOWR(0x4c, drm_r128_packet_t)
+#define DRM_IOCTL_R128_DEPTH	 DRM_IOW( 0x4c, drm_r128_depth_t)
+#define DRM_IOCTL_R128_STIPPLE	 DRM_IOW( 0x4d, drm_r128_stipple_t)
+#define DRM_IOCTL_R128_PACKET	 DRM_IOWR(0x4e, drm_r128_packet_t)
 
 /* SiS specific ioctls */
 #define SIS_IOCTL_FB_ALLOC     DRM_IOWR( 0x44, drm_sis_mem_t)
diff --git a/shared/drm.h b/shared/drm.h
index 5d383427..20691ccf 100644
--- a/shared/drm.h
+++ b/shared/drm.h
@@ -374,7 +374,9 @@ typedef struct drm_agp_info {
 #define DRM_IOCTL_R128_VERTEX	 DRM_IOW( 0x49, drm_r128_vertex_t)
 #define DRM_IOCTL_R128_INDICES	 DRM_IOW( 0x4a, drm_r128_indices_t)
 #define DRM_IOCTL_R128_BLIT	 DRM_IOW( 0x4b, drm_r128_blit_t)
-#define DRM_IOCTL_R128_PACKET	 DRM_IOWR(0x4c, drm_r128_packet_t)
+#define DRM_IOCTL_R128_DEPTH	 DRM_IOW( 0x4c, drm_r128_depth_t)
+#define DRM_IOCTL_R128_STIPPLE	 DRM_IOW( 0x4d, drm_r128_stipple_t)
+#define DRM_IOCTL_R128_PACKET	 DRM_IOWR(0x4e, drm_r128_packet_t)
 
 /* SiS specific ioctls */
 #define SIS_IOCTL_FB_ALLOC     DRM_IOWR( 0x44, drm_sis_mem_t)