i915: Convert to using VBs instead of inline prims.

9 files changed, 299 insertions, 214 deletions

diff --git a/src/mesa/drivers/dri/i915/i830_reg.h b/src/mesa/drivers/dri/i915/i830_reg.h
index d1084a84c07..d210c2d08e4 100644
--- a/src/mesa/drivers/dri/i915/i830_reg.h
+++ b/src/mesa/drivers/dri/i915/i830_reg.h
@@ -494,10 +494,6 @@
 #define VFT1_TEX0_FMT(x)	(x)
 #define VFT1_TEX0_MASK          3
 #define VFT1_TEX1_SHIFT         2
-#define TEXCOORDFMT_2D		0
-#define TEXCOORDFMT_3D		1
-#define TEXCOORDFMT_4D		2
-#define TEXCOORDFMT_1D		3
 
 /*New stuff picked up along the way */
 
diff --git a/src/mesa/drivers/dri/i915/i830_vtbl.c b/src/mesa/drivers/dri/i915/i830_vtbl.c
index cff051b16bd..4d3ad0083a1 100644
--- a/src/mesa/drivers/dri/i915/i830_vtbl.c
+++ b/src/mesa/drivers/dri/i915/i830_vtbl.c
@@ -31,6 +31,7 @@
 #include "i830_reg.h"
 #include "intel_batchbuffer.h"
 #include "intel_regions.h"
+#include "intel_tris.h"
 #include "tnl/t_context.h"
 #include "tnl/t_vertex.h"
 
@@ -435,7 +436,8 @@ i830_emit_state(struct intel_context *intel)
     * Set the space as LOOP_CLIPRECTS now, since that's what our primitives
     * will be emitted under.
     */
-   intel_batchbuffer_require_space(intel->batch, get_state_size(state) + 8,
+   intel_batchbuffer_require_space(intel->batch,
+				   get_state_size(state) + INTEL_PRIM_EMIT_SIZE,
 				   LOOP_CLIPRECTS);
    count = 0;
  again:
@@ -675,6 +677,9 @@ i830_new_batch(struct intel_context *intel)
    struct i830_context *i830 = i830_context(&intel->ctx);
    i830->state.emitted = 0;
 
+   /* Signal that we should put new vertices into a new vertex buffer. */
+   intel->prim.needs_new_vb = GL_TRUE;
+
    /* Check that we didn't just wrap our batchbuffer at a bad time. */
    assert(!intel->no_batch_wrap);
 }
diff --git a/src/mesa/drivers/dri/i915/i915_reg.h b/src/mesa/drivers/dri/i915/i915_reg.h
index b718b8610c7..8891e11c6fd 100644
--- a/src/mesa/drivers/dri/i915/i915_reg.h
+++ b/src/mesa/drivers/dri/i915/i915_reg.h
@@ -325,118 +325,6 @@
 #define SCISSOR_RECT_0_YMAX(x)		((x)<<16)
 #define SCISSOR_RECT_0_XMAX(x)		(x)
 
-/* p189 */
-#define _3DSTATE_LOAD_STATE_IMMEDIATE_1   ((0x3<<29)|(0x1d<<24)|(0x04<<16))
-#define I1_LOAD_S(n)                      (1<<(4+n))
-
-#define S0_VB_OFFSET_MASK              0xffffffc
-#define S0_AUTO_CACHE_INV_DISABLE      (1<<0)
-
-#define S1_VERTEX_WIDTH_SHIFT          24
-#define S1_VERTEX_WIDTH_MASK           (0x3f<<24)
-#define S1_VERTEX_PITCH_SHIFT          16
-#define S1_VERTEX_PITCH_MASK           (0x3f<<16)
-
-#define TEXCOORDFMT_2D                 0x0
-#define TEXCOORDFMT_3D                 0x1
-#define TEXCOORDFMT_4D                 0x2
-#define TEXCOORDFMT_1D                 0x3
-#define TEXCOORDFMT_2D_16              0x4
-#define TEXCOORDFMT_4D_16              0x5
-#define TEXCOORDFMT_NOT_PRESENT        0xf
-#define S2_TEXCOORD_FMT0_MASK            0xf
-#define S2_TEXCOORD_FMT1_SHIFT           4
-#define S2_TEXCOORD_FMT(unit, type)    ((type)<<(unit*4))
-#define S2_TEXCOORD_NONE               (~0)
-
-/* S3 not interesting */
-
-#define S4_POINT_WIDTH_SHIFT           23
-#define S4_POINT_WIDTH_MASK            (0x1ff<<23)
-#define S4_LINE_WIDTH_SHIFT            19
-#define S4_LINE_WIDTH_ONE              (0x2<<19)
-#define S4_LINE_WIDTH_MASK             (0xf<<19)
-#define S4_FLATSHADE_ALPHA             (1<<18)
-#define S4_FLATSHADE_FOG               (1<<17)
-#define S4_FLATSHADE_SPECULAR          (1<<16)
-#define S4_FLATSHADE_COLOR             (1<<15)
-#define S4_CULLMODE_BOTH	       (0<<13)
-#define S4_CULLMODE_NONE	       (1<<13)
-#define S4_CULLMODE_CW		       (2<<13)
-#define S4_CULLMODE_CCW		       (3<<13)
-#define S4_CULLMODE_MASK	       (3<<13)
-#define S4_VFMT_POINT_WIDTH            (1<<12)
-#define S4_VFMT_SPEC_FOG               (1<<11)
-#define S4_VFMT_COLOR                  (1<<10)
-#define S4_VFMT_DEPTH_OFFSET           (1<<9)
-#define S4_VFMT_XYZ     	       (1<<6)
-#define S4_VFMT_XYZW     	       (2<<6)
-#define S4_VFMT_XY     		       (3<<6)
-#define S4_VFMT_XYW     	       (4<<6)
-#define S4_VFMT_XYZW_MASK              (7<<6)
-#define S4_FORCE_DEFAULT_DIFFUSE       (1<<5)
-#define S4_FORCE_DEFAULT_SPECULAR      (1<<4)
-#define S4_LOCAL_DEPTH_OFFSET_ENABLE   (1<<3)
-#define S4_VFMT_FOG_PARAM              (1<<2)
-#define S4_SPRITE_POINT_ENABLE         (1<<1)
-#define S4_LINE_ANTIALIAS_ENABLE       (1<<0)
-
-#define S4_VFMT_MASK (S4_VFMT_POINT_WIDTH   | 	\
-		      S4_VFMT_SPEC_FOG      |	\
-		      S4_VFMT_COLOR         |	\
-		      S4_VFMT_DEPTH_OFFSET  |	\
-		      S4_VFMT_XYZW_MASK     |	\
-		      S4_VFMT_FOG_PARAM)
-
-
-#define S5_WRITEDISABLE_ALPHA          (1<<31)
-#define S5_WRITEDISABLE_RED            (1<<30)
-#define S5_WRITEDISABLE_GREEN          (1<<29)
-#define S5_WRITEDISABLE_BLUE           (1<<28)
-#define S5_WRITEDISABLE_MASK           (0xf<<28)
-#define S5_FORCE_DEFAULT_POINT_SIZE    (1<<27)
-#define S5_LAST_PIXEL_ENABLE           (1<<26)
-#define S5_GLOBAL_DEPTH_OFFSET_ENABLE  (1<<25)
-#define S5_FOG_ENABLE                  (1<<24)
-#define S5_STENCIL_REF_SHIFT           16
-#define S5_STENCIL_REF_MASK            (0xff<<16)
-#define S5_STENCIL_TEST_FUNC_SHIFT     13
-#define S5_STENCIL_TEST_FUNC_MASK      (0x7<<13)
-#define S5_STENCIL_FAIL_SHIFT          10
-#define S5_STENCIL_FAIL_MASK           (0x7<<10)
-#define S5_STENCIL_PASS_Z_FAIL_SHIFT   7
-#define S5_STENCIL_PASS_Z_FAIL_MASK    (0x7<<7)
-#define S5_STENCIL_PASS_Z_PASS_SHIFT   4
-#define S5_STENCIL_PASS_Z_PASS_MASK    (0x7<<4)
-#define S5_STENCIL_WRITE_ENABLE        (1<<3)
-#define S5_STENCIL_TEST_ENABLE         (1<<2)
-#define S5_COLOR_DITHER_ENABLE         (1<<1)
-#define S5_LOGICOP_ENABLE              (1<<0)
-
-
-#define S6_ALPHA_TEST_ENABLE           (1<<31)
-#define S6_ALPHA_TEST_FUNC_SHIFT       28
-#define S6_ALPHA_TEST_FUNC_MASK        (0x7<<28)
-#define S6_ALPHA_REF_SHIFT             20
-#define S6_ALPHA_REF_MASK              (0xff<<20)
-#define S6_DEPTH_TEST_ENABLE           (1<<19)
-#define S6_DEPTH_TEST_FUNC_SHIFT       16
-#define S6_DEPTH_TEST_FUNC_MASK        (0x7<<16)
-#define S6_CBUF_BLEND_ENABLE           (1<<15)
-#define S6_CBUF_BLEND_FUNC_SHIFT       12
-#define S6_CBUF_BLEND_FUNC_MASK        (0x7<<12)
-#define S6_CBUF_SRC_BLEND_FACT_SHIFT   8
-#define S6_CBUF_SRC_BLEND_FACT_MASK    (0xf<<8)
-#define S6_CBUF_DST_BLEND_FACT_SHIFT   4
-#define S6_CBUF_DST_BLEND_FACT_MASK    (0xf<<4)
-#define S6_DEPTH_WRITE_ENABLE          (1<<3)
-#define S6_COLOR_WRITE_ENABLE          (1<<2)
-#define S6_TRISTRIP_PV_SHIFT           0
-#define S6_TRISTRIP_PV_MASK            (0x3<<0)
-
-#define S7_DEPTH_OFFSET_CONST_MASK     ~0
-
-
 /* Helper macros for blend factors
  */
 #define DST_BLND_FACT(f) ((f)<<S6_CBUF_DST_BLEND_FACT_SHIFT)
diff --git a/src/mesa/drivers/dri/i915/i915_vtbl.c b/src/mesa/drivers/dri/i915/i915_vtbl.c
index 43f5703d9e0..23d63fb47a5 100644
--- a/src/mesa/drivers/dri/i915/i915_vtbl.c
+++ b/src/mesa/drivers/dri/i915/i915_vtbl.c
@@ -39,6 +39,7 @@
 #include "intel_batchbuffer.h"
 #include "intel_tex.h"
 #include "intel_regions.h"
+#include "intel_tris.h"
 
 #include "i915_reg.h"
 #include "i915_context.h"
@@ -313,7 +314,8 @@ i915_emit_state(struct intel_context *intel)
     * Set the space as LOOP_CLIPRECTS now, since that's what our primitives
     * will be emitted under.
     */
-   intel_batchbuffer_require_space(intel->batch, get_state_size(state) + 8,
+   intel_batchbuffer_require_space(intel->batch,
+				   get_state_size(state) + INTEL_PRIM_EMIT_SIZE,
 				   LOOP_CLIPRECTS);
    count = 0;
  again:
@@ -587,6 +589,8 @@ i915_new_batch(struct intel_context *intel)
     * difficulties associated with them (physical address requirements).
     */
    i915->state.emitted = 0;
+   /* Signal that we should put new vertices into a new vertex buffer. */
+   intel->prim.needs_new_vb = GL_TRUE;
 
    /* Check that we didn't just wrap our batchbuffer at a bad time. */
    assert(!intel->no_batch_wrap);
diff --git a/src/mesa/drivers/dri/i915/intel_render.c b/src/mesa/drivers/dri/i915/intel_render.c
index 5e6500cfa1b..838d450378c 100644
--- a/src/mesa/drivers/dri/i915/intel_render.c
+++ b/src/mesa/drivers/dri/i915/intel_render.c
@@ -67,7 +67,7 @@
 
 #define HAVE_ELTS        0
 
-static GLuint hw_prim[GL_POLYGON + 1] = {
+static uint32_t hw_prim[GL_POLYGON + 1] = {
    0,
    PRIM3D_LINELIST,
    PRIM3D_LINESTRIP,
@@ -114,7 +114,7 @@ intelDmaPrimitive(struct intel_context *intel, GLenum prim)
       fprintf(stderr, "%s %s\n", __FUNCTION__, _mesa_lookup_enum_by_nr(prim));
    INTEL_FIREVERTICES(intel);
    intel->vtbl.reduced_primitive_state(intel, reduced_prim[prim]);
-   intelStartInlinePrimitive(intel, hw_prim[prim], LOOP_CLIPRECTS);
+   intel_set_prim(intel, hw_prim[prim]);
 }
 
 
@@ -126,12 +126,11 @@ do {						\
 
 #define FLUSH() INTEL_FIREVERTICES(intel)
 
-#define GET_SUBSEQUENT_VB_MAX_VERTS() \
-  ((intel->batch->size - 1500) / (intel->vertex_size*4))
-#define GET_CURRENT_VB_MAX_VERTS() GET_SUBSEQUENT_VB_MAX_VERTS()
+#define GET_SUBSEQUENT_VB_MAX_VERTS() (INTEL_VB_SIZE / (intel->vertex_size * 4))
+#define GET_CURRENT_VB_MAX_VERTS() \
+   ((INTEL_VB_SIZE - intel->prim.current_offset) / (intel->vertex_size * 4))
 
-#define ALLOC_VERTS( nr ) \
-   intelExtendInlinePrimitive( intel, (nr) * intel->vertex_size )
+#define ALLOC_VERTS(nr) intel_get_prim_space(intel, nr)
 
 #define EMIT_VERTS( ctx, j, nr, buf ) \
   _tnl_emit_vertices_to_buffer(ctx, j, (j)+(nr), buf )
diff --git a/src/mesa/drivers/dri/i915/intel_tris.c b/src/mesa/drivers/dri/i915/intel_tris.c
index bbb4e0f3cda..a1121925cb9 100644
--- a/src/mesa/drivers/dri/i915/intel_tris.c
+++ b/src/mesa/drivers/dri/i915/intel_tris.c
@@ -25,6 +25,12 @@
  * 
  **************************************************************************/
 
+/** @file intel_tris.c
+ *
+ * This file contains functions for managing the vertex buffer and emitting
+ * primitives into it.
+ */
+
 #include "glheader.h"
 #include "context.h"
 #include "macros.h"
@@ -47,111 +53,159 @@
 #include "intel_reg.h"
 #include "intel_span.h"
 #include "intel_tex.h"
+#include "intel_chipset.h"
+#include "i830_context.h"
+#include "i830_reg.h"
 
 static void intelRenderPrimitive(GLcontext * ctx, GLenum prim);
 static void intelRasterPrimitive(GLcontext * ctx, GLenum rprim,
                                  GLuint hwprim);
 
-/*
- */
-static void
-intel_flush_inline_primitive(struct intel_context *intel)
+/** Sets the primitive type for a primitive sequence, flushing as needed. */
+void intel_set_prim(struct intel_context *intel, uint32_t prim)
 {
-   GLuint used = intel->batch->ptr - intel->prim.start_ptr;
+   if (prim != intel->prim.primitive) {
+      INTEL_FIREVERTICES(intel);
+      intel->prim.primitive = prim;
+   }
+}
 
-   assert(intel->prim.primitive != ~0);
+/** Returns mapped VB space for the given number of vertices */
+uint32_t *intel_get_prim_space(struct intel_context *intel, unsigned int count)
+{
+   uint32_t *addr;
+
+   /* Check for space in the existing VB */
+   if (intel->prim.vb_bo == NULL ||
+       intel->prim.needs_new_vb ||
+       (intel->prim.current_offset +
+	count * intel->vertex_size * 4) > INTEL_VB_SIZE ||
+       (intel->prim.count + count) >= (1 << 16)) {
+      /* Flush existing prim if any */
+      INTEL_FIREVERTICES(intel);
 
-/*    _mesa_printf("/\n"); */
+      /* Start a new VB */
+      dri_bo_unreference(intel->prim.vb_bo);
+      intel->prim.vb_bo = dri_bo_alloc(intel->bufmgr, "vb",
+				       INTEL_VB_SIZE, 4);
+      intel->prim.start_offset = 0;
+      intel->prim.current_offset = 0;
 
-   if (used < 8)
-      goto do_discard;
+      dri_bufmgr_check_aperture_space(intel->prim.vb_bo);
 
-   *(int *) intel->prim.start_ptr = (_3DPRIMITIVE |
-                                     intel->prim.primitive | (used / 4 - 2));
+      intel->prim.needs_new_vb = GL_FALSE;
 
-   goto finished;
+      dri_bo_map(intel->prim.vb_bo, GL_TRUE);
+   }
 
- do_discard:
-   intel->batch->ptr -= used;
+   intel->prim.flush = intel_flush_prim;
 
- finished:
-   intel->prim.primitive = ~0;
-   intel->prim.start_ptr = 0;
-   intel->prim.flush = 0;
-}
+   addr = (uint32_t *)((char *)intel->prim.vb_bo->virtual +
+		       intel->prim.current_offset);
+   intel->prim.current_offset += intel->vertex_size * 4 * count;
+   intel->prim.count += count;
 
+   return addr;
+}
 
-/* Emit a primitive referencing vertices in a vertex buffer.
- */
-void
-intelStartInlinePrimitive(struct intel_context *intel,
-                          GLuint prim, GLuint batch_flags)
+/** Dispatches the accumulated primitive to the batchbuffer. */
+void intel_flush_prim(struct intel_context *intel)
 {
    BATCH_LOCALS;
 
+   /* Must be called after an intel_start_prim. */
+   assert(intel->prim.primitive != ~0);
+
+   if (intel->prim.count == 0)
+      return;
+
    intel_wait_flips(intel);
 
+   dri_bo_unmap(intel->prim.vb_bo);
+
    intel->vtbl.emit_state(intel);
 
+   /* Ensure that we don't start a new batch for the following emit, which
+    * depends on the state just emitted. emit_state should be making sure we
+    * have the space for this.
+    */
    intel->no_batch_wrap = GL_TRUE;
 
-/*    _mesa_printf("%s *", __progname); */
-
-   /* Emit a slot which will be filled with the inline primitive
-    * command later.
+   /* Check that we actually emitted the state into this batch, using the
+    * UPLOAD_CTX bit as the signal.
     */
-   BEGIN_BATCH(2, batch_flags);
-   OUT_BATCH(0);
-
    assert((intel->batch->dirty_state & (1<<1)) == 0);
 
-   intel->prim.start_ptr = intel->batch->ptr;
-   intel->prim.primitive = prim;
-   intel->prim.flush = intel_flush_inline_primitive;
+#if 0
+   printf("emitting %d..%d=%d vertices size %d\n", intel->prim.start_offset,
+	  intel->prim.current_offset, intel->prim.count,
+	  intel->vertex_size * 4);
+#endif
 
-   OUT_BATCH(0);
-   ADVANCE_BATCH();
+   if (IS_9XX(intel->intelScreen->deviceID)) {
+      BEGIN_BATCH(5, LOOP_CLIPRECTS);
+      OUT_BATCH(_3DSTATE_LOAD_STATE_IMMEDIATE_1 |
+		I1_LOAD_S(0) | I1_LOAD_S(1) | 1);
+      assert((intel->prim.start_offset & !S0_VB_OFFSET_MASK) == 0);
+      OUT_RELOC(intel->prim.vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
+		intel->prim.start_offset);
+      OUT_BATCH((intel->vertex_size << S1_VERTEX_WIDTH_SHIFT) |
+		(intel->vertex_size << S1_VERTEX_PITCH_SHIFT));
+
+      OUT_BATCH(_3DPRIMITIVE |
+		PRIM_INDIRECT |
+		PRIM_INDIRECT_SEQUENTIAL |
+		intel->prim.primitive |
+		intel->prim.count);
+      OUT_BATCH(0); /* Beginning vertex index */
+      ADVANCE_BATCH();
+   } else {
+      struct i830_context *i830 = i830_context(&intel->ctx);
+
+      BEGIN_BATCH(5, LOOP_CLIPRECTS);
+      OUT_BATCH(_3DSTATE_LOAD_STATE_IMMEDIATE_1 |
+		I1_LOAD_S(0) | I1_LOAD_S(2) | 1);
+      /* S0 */
+      assert((intel->prim.start_offset & !S0_VB_OFFSET_MASK_830) == 0);
+      OUT_RELOC(intel->prim.vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
+		intel->prim.start_offset |
+		(intel->vertex_size << S0_VB_PITCH_SHIFT_830) |
+		S0_VB_ENABLE_830);
+      /* S1
+       * This is somewhat unfortunate -- VB width is tied up with
+       * vertex format data that we've already uploaded through
+       * _3DSTATE_VFT[01]_CMD.  We may want to replace emits of VFT state with
+       * STATE_IMMEDIATE_1 like this to avoid duplication.
+       */
+      OUT_BATCH((i830->state.Ctx[I830_CTXREG_VF] & VFT0_TEX_COUNT_MASK) >>
+		VFT0_TEX_COUNT_SHIFT << S2_TEX_COUNT_SHIFT_830 |
+		(i830->state.Ctx[I830_CTXREG_VF2] << 16) |
+		intel->vertex_size << S2_VERTEX_0_WIDTH_SHIFT_830);
+
+      OUT_BATCH(_3DPRIMITIVE |
+		PRIM_INDIRECT |
+		PRIM_INDIRECT_SEQUENTIAL |
+		intel->prim.primitive |
+		intel->prim.count);
+      OUT_BATCH(0); /* Beginning vertex index */
+      ADVANCE_BATCH();
+   }
 
    intel->no_batch_wrap = GL_FALSE;
 
-/*    _mesa_printf(">"); */
-}
-
-
-void
-intelWrapInlinePrimitive(struct intel_context *intel)
-{
-   GLuint prim = intel->prim.primitive;
-   enum cliprect_mode cliprect_mode = intel->batch->cliprect_mode;
-
-   intel_flush_inline_primitive(intel);
-   intel_batchbuffer_flush(intel->batch);
-   intelStartInlinePrimitive(intel, prim, cliprect_mode);  /* ??? */
-}
-
-GLuint *
-intelExtendInlinePrimitive(struct intel_context *intel, GLuint dwords)
-{
-   GLuint sz = dwords * sizeof(GLuint);
-   GLuint *ptr;
-
-   assert(intel->prim.flush == intel_flush_inline_primitive);
-
-   if (intel_batchbuffer_space(intel->batch) < sz)
-      intelWrapInlinePrimitive(intel);
-
-/*    _mesa_printf("."); */
-
-   intel->vtbl.assert_not_dirty(intel);
-
-   ptr = (GLuint *) intel->batch->ptr;
-   intel->batch->ptr += sz;
-
-   return ptr;
+   /* If we're going to keep using this VB for more primitives, map it
+    * again.
+    */
+   if (!intel->prim.needs_new_vb)
+      dri_bo_map(intel->prim.vb_bo, GL_TRUE);
+
+   intel->prim.flush = NULL;
+   intel->prim.start_offset = intel->prim.current_offset;
+   if (!IS_9XX(intel->intelScreen->deviceID))
+      intel->prim.start_offset = ALIGN(intel->prim.start_offset, 128);
+   intel->prim.count = 0;
 }
 
-
-
 /***********************************************************************
  *                    Emit primitives as inline vertices               *
  ***********************************************************************/
@@ -182,7 +236,7 @@ intel_draw_quad(struct intel_context *intel,
                 intelVertexPtr v1, intelVertexPtr v2, intelVertexPtr v3)
 {
    GLuint vertsize = intel->vertex_size;
-   GLuint *vb = intelExtendInlinePrimitive(intel, 6 * vertsize);
+   GLuint *vb = intel_get_prim_space(intel, 6);
    int j;
 
    COPY_DWORDS(j, vb, vertsize, v0);
@@ -210,7 +264,7 @@ intel_draw_triangle(struct intel_context *intel,
                     intelVertexPtr v0, intelVertexPtr v1, intelVertexPtr v2)
 {
    GLuint vertsize = intel->vertex_size;
-   GLuint *vb = intelExtendInlinePrimitive(intel, 3 * vertsize);
+   GLuint *vb = intel_get_prim_space(intel, 3);
    int j;
 
    COPY_DWORDS(j, vb, vertsize, v0);
@@ -224,7 +278,7 @@ intel_draw_line(struct intel_context *intel,
                 intelVertexPtr v0, intelVertexPtr v1)
 {
    GLuint vertsize = intel->vertex_size;
-   GLuint *vb = intelExtendInlinePrimitive(intel, 2 * vertsize);
+   GLuint *vb = intel_get_prim_space(intel, 2);
    int j;
 
    COPY_DWORDS(j, vb, vertsize, v0);
@@ -236,7 +290,7 @@ static void
 intel_draw_point(struct intel_context *intel, intelVertexPtr v0)
 {
    GLuint vertsize = intel->vertex_size;
-   GLuint *vb = intelExtendInlinePrimitive(intel, vertsize);
+   GLuint *vb = intel_get_prim_space(intel, 1);
    int j;
 
    /* Adjust for sub pixel position -- still required for conform. */
@@ -745,7 +799,7 @@ intelFastRenderClippedPoly(GLcontext * ctx, const GLuint * elts, GLuint n)
 {
    struct intel_context *intel = intel_context(ctx);
    const GLuint vertsize = intel->vertex_size;
-   GLuint *vb = intelExtendInlinePrimitive(intel, (n - 2) * 3 * vertsize);
+   GLuint *vb = intel_get_prim_space(intel, (n - 2) * 3);
    GLubyte *vertptr = (GLubyte *) intel->verts;
    const GLuint *start = (const GLuint *) V(elts[0]);
    int i, j;
@@ -950,7 +1004,7 @@ intelRasterPrimitive(GLcontext * ctx, GLenum rprim, GLuint hwprim)
    if (hwprim != intel->prim.primitive) {
       INTEL_FIREVERTICES(intel);
 
-      intelStartInlinePrimitive(intel, hwprim, LOOP_CLIPRECTS);
+      intel_set_prim(intel, hwprim);
    }
 }
 
@@ -1083,15 +1137,18 @@ intel_meta_draw_poly(struct intel_context *intel,
    union fi *vb;
    GLint i;
    GLboolean was_locked = intel->locked;
+   unsigned int saved_vertex_size = intel->vertex_size;
 
    if (!was_locked)
        LOCK_HARDWARE(intel);
 
+   intel->vertex_size = 6;
+
    /* All 3d primitives should be emitted with LOOP_CLIPRECTS,
     * otherwise the drawing origin (DR4) might not be set correctly.
     */
-   intelStartInlinePrimitive(intel, PRIM3D_TRIFAN, LOOP_CLIPRECTS);
-   vb = (union fi *) intelExtendInlinePrimitive(intel, n * 6);
+   intel_set_prim(intel, PRIM3D_TRIFAN);
+   vb = (union fi *) intel_get_prim_space(intel, n);
 
    for (i = 0; i < n; i++) {
       vb[0].f = xy[i][0];
@@ -1105,6 +1162,8 @@ intel_meta_draw_poly(struct intel_context *intel,
 
    INTEL_FIREVERTICES(intel);
 
+   intel->vertex_size = saved_vertex_size;
+
    if (!was_locked)
        UNLOCK_HARDWARE(intel);
 }
diff --git a/src/mesa/drivers/dri/i915/intel_tris.h b/src/mesa/drivers/dri/i915/intel_tris.h
index 021e5c64500..6b38cd6fbd3 100644
--- a/src/mesa/drivers/dri/i915/intel_tris.h
+++ b/src/mesa/drivers/dri/i915/intel_tris.h
@@ -30,7 +30,9 @@
 
 #include "mtypes.h"
 
-
+#define INTEL_VB_SIZE		(8 * 1024)
+/** 3 dwords of state_immediate and 2 of 3dprim, in intel_flush_prim */
+#define INTEL_PRIM_EMIT_SIZE	(5 * 4)
 
 #define _INTEL_NEW_RENDERSTATE (_DD_NEW_LINE_STIPPLE |		\
 			       _DD_NEW_TRI_UNFILLED |		\
@@ -44,11 +46,8 @@ extern void intelInitTriFuncs(GLcontext * ctx);
 
 extern void intelChooseRenderState(GLcontext * ctx);
 
-extern void intelStartInlinePrimitive(struct intel_context *intel,
-                                      GLuint prim, GLuint flags);
-extern void intelWrapInlinePrimitive(struct intel_context *intel);
-
-GLuint *intelExtendInlinePrimitive(struct intel_context *intel,
-                                   GLuint dwords);
+void intel_set_prim(struct intel_context *intel, uint32_t prim);
+GLuint *intel_get_prim_space(struct intel_context *intel, unsigned int count);
+void intel_flush_prim(struct intel_context *intel);
 
 #endif
diff --git a/src/mesa/drivers/dri/intel/intel_context.h b/src/mesa/drivers/dri/intel/intel_context.h
index 579883437fc..1aa9c3d7117 100644
--- a/src/mesa/drivers/dri/intel/intel_context.h
+++ b/src/mesa/drivers/dri/intel/intel_context.h
@@ -182,9 +182,19 @@ struct intel_context
    struct
    {
       GLuint id;
-      GLuint primitive;
-      GLubyte *start_ptr;
+      uint32_t primitive;	/**< Current hardware primitive type */
       void (*flush) (struct intel_context *);
+      dri_bo *vb_bo;
+      unsigned int start_offset; /**< Byte offset of primitive sequence */
+      unsigned int current_offset; /**< Byte offset of next vertex */
+      unsigned int count;	/**< Number of vertices in current primitive */
+      /**
+       * Signals when a new VB should be started, regardless of remaining
+       * space.
+       *
+       * Used to avoid rewriting a VB that's being rendered from.
+       */
+      GLboolean needs_new_vb;
    } prim;
 
    GLuint stats_wm;
diff --git a/src/mesa/drivers/dri/intel/intel_reg.h b/src/mesa/drivers/dri/intel/intel_reg.h
index c12ccf4ae18..96af7e1a030 100644
--- a/src/mesa/drivers/dri/intel/intel_reg.h
+++ b/src/mesa/drivers/dri/intel/intel_reg.h
@@ -40,6 +40,131 @@
 #define MI_WAIT_FOR_PLANE_B_FLIP        (1<<6)
 #define MI_WAIT_FOR_PLANE_A_FLIP        (1<<2)
 
+/* p189 */
+#define _3DSTATE_LOAD_STATE_IMMEDIATE_1   (CMD_3D | (0x1d<<24) | (0x04<<16))
+#define I1_LOAD_S(n)                      (1<<(4+n))
+
+/** @{
+ * 915 definitions
+ */
+#define S0_VB_OFFSET_MASK		0xffffffc
+#define S0_AUTO_CACHE_INV_DISABLE	(1<<0)
+/** @} */
+
+/** @{
+ * 830 definitions
+ */
+#define S0_VB_OFFSET_MASK_830		0xffffff8
+#define S0_VB_PITCH_SHIFT_830		1
+#define S0_VB_ENABLE_830		0
+/** @} */
+
+#define S1_VERTEX_WIDTH_SHIFT          24
+#define S1_VERTEX_WIDTH_MASK           (0x3f<<24)
+#define S1_VERTEX_PITCH_SHIFT          16
+#define S1_VERTEX_PITCH_MASK           (0x3f<<16)
+
+#define TEXCOORDFMT_2D                 0x0
+#define TEXCOORDFMT_3D                 0x1
+#define TEXCOORDFMT_4D                 0x2
+#define TEXCOORDFMT_1D                 0x3
+#define TEXCOORDFMT_2D_16              0x4
+#define TEXCOORDFMT_4D_16              0x5
+#define TEXCOORDFMT_NOT_PRESENT        0xf
+#define S2_TEXCOORD_FMT0_MASK            0xf
+#define S2_TEXCOORD_FMT1_SHIFT           4
+#define S2_TEXCOORD_FMT(unit, type)    ((type)<<(unit*4))
+#define S2_TEXCOORD_NONE               (~0)
+#define S2_TEX_COUNT_SHIFT_830		12
+#define S2_VERTEX_0_WIDTH_SHIFT_830	0
+#define S2_VERTEX_1_WIDTH_SHIFT_830	6
+/* S3 not interesting */
+
+#define S4_POINT_WIDTH_SHIFT           23
+#define S4_POINT_WIDTH_MASK            (0x1ff<<23)
+#define S4_LINE_WIDTH_SHIFT            19
+#define S4_LINE_WIDTH_ONE              (0x2<<19)
+#define S4_LINE_WIDTH_MASK             (0xf<<19)
+#define S4_FLATSHADE_ALPHA             (1<<18)
+#define S4_FLATSHADE_FOG               (1<<17)
+#define S4_FLATSHADE_SPECULAR          (1<<16)
+#define S4_FLATSHADE_COLOR             (1<<15)
+#define S4_CULLMODE_BOTH	       (0<<13)
+#define S4_CULLMODE_NONE	       (1<<13)
+#define S4_CULLMODE_CW		       (2<<13)
+#define S4_CULLMODE_CCW		       (3<<13)
+#define S4_CULLMODE_MASK	       (3<<13)
+#define S4_VFMT_POINT_WIDTH            (1<<12)
+#define S4_VFMT_SPEC_FOG               (1<<11)
+#define S4_VFMT_COLOR                  (1<<10)
+#define S4_VFMT_DEPTH_OFFSET           (1<<9)
+#define S4_VFMT_XYZ     	       (1<<6)
+#define S4_VFMT_XYZW     	       (2<<6)
+#define S4_VFMT_XY     		       (3<<6)
+#define S4_VFMT_XYW     	       (4<<6)
+#define S4_VFMT_XYZW_MASK              (7<<6)
+#define S4_FORCE_DEFAULT_DIFFUSE       (1<<5)
+#define S4_FORCE_DEFAULT_SPECULAR      (1<<4)
+#define S4_LOCAL_DEPTH_OFFSET_ENABLE   (1<<3)
+#define S4_VFMT_FOG_PARAM              (1<<2)
+#define S4_SPRITE_POINT_ENABLE         (1<<1)
+#define S4_LINE_ANTIALIAS_ENABLE       (1<<0)
+
+#define S4_VFMT_MASK (S4_VFMT_POINT_WIDTH   | 	\
+		      S4_VFMT_SPEC_FOG      |	\
+		      S4_VFMT_COLOR         |	\
+		      S4_VFMT_DEPTH_OFFSET  |	\
+		      S4_VFMT_XYZW_MASK     |	\
+		      S4_VFMT_FOG_PARAM)
+
+
+#define S5_WRITEDISABLE_ALPHA          (1<<31)
+#define S5_WRITEDISABLE_RED            (1<<30)
+#define S5_WRITEDISABLE_GREEN          (1<<29)
+#define S5_WRITEDISABLE_BLUE           (1<<28)
+#define S5_WRITEDISABLE_MASK           (0xf<<28)
+#define S5_FORCE_DEFAULT_POINT_SIZE    (1<<27)
+#define S5_LAST_PIXEL_ENABLE           (1<<26)
+#define S5_GLOBAL_DEPTH_OFFSET_ENABLE  (1<<25)
+#define S5_FOG_ENABLE                  (1<<24)
+#define S5_STENCIL_REF_SHIFT           16
+#define S5_STENCIL_REF_MASK            (0xff<<16)
+#define S5_STENCIL_TEST_FUNC_SHIFT     13
+#define S5_STENCIL_TEST_FUNC_MASK      (0x7<<13)
+#define S5_STENCIL_FAIL_SHIFT          10
+#define S5_STENCIL_FAIL_MASK           (0x7<<10)
+#define S5_STENCIL_PASS_Z_FAIL_SHIFT   7
+#define S5_STENCIL_PASS_Z_FAIL_MASK    (0x7<<7)
+#define S5_STENCIL_PASS_Z_PASS_SHIFT   4
+#define S5_STENCIL_PASS_Z_PASS_MASK    (0x7<<4)
+#define S5_STENCIL_WRITE_ENABLE        (1<<3)
+#define S5_STENCIL_TEST_ENABLE         (1<<2)
+#define S5_COLOR_DITHER_ENABLE         (1<<1)
+#define S5_LOGICOP_ENABLE              (1<<0)
+
+
+#define S6_ALPHA_TEST_ENABLE           (1<<31)
+#define S6_ALPHA_TEST_FUNC_SHIFT       28
+#define S6_ALPHA_TEST_FUNC_MASK        (0x7<<28)
+#define S6_ALPHA_REF_SHIFT             20
+#define S6_ALPHA_REF_MASK              (0xff<<20)
+#define S6_DEPTH_TEST_ENABLE           (1<<19)
+#define S6_DEPTH_TEST_FUNC_SHIFT       16
+#define S6_DEPTH_TEST_FUNC_MASK        (0x7<<16)
+#define S6_CBUF_BLEND_ENABLE           (1<<15)
+#define S6_CBUF_BLEND_FUNC_SHIFT       12
+#define S6_CBUF_BLEND_FUNC_MASK        (0x7<<12)
+#define S6_CBUF_SRC_BLEND_FACT_SHIFT   8
+#define S6_CBUF_SRC_BLEND_FACT_MASK    (0xf<<8)
+#define S6_CBUF_DST_BLEND_FACT_SHIFT   4
+#define S6_CBUF_DST_BLEND_FACT_MASK    (0xf<<4)
+#define S6_DEPTH_WRITE_ENABLE          (1<<3)
+#define S6_COLOR_WRITE_ENABLE          (1<<2)
+#define S6_TRISTRIP_PV_SHIFT           0
+#define S6_TRISTRIP_PV_MASK            (0x3<<0)
+
+#define S7_DEPTH_OFFSET_CONST_MASK     ~0
+
 /* Primitive dispatch on 830-945 */
 #define _3DPRIMITIVE			(CMD_3D | (0x1f << 24))
 #define PRIM_INDIRECT            (1<<23)