diff --git a/src/gallium/drivers/r300/r300_cs_inlines.h b/src/gallium/drivers/r300/r300_cs_inlines.h
index 98f9ee04511..16e412f586c 100644
--- a/src/gallium/drivers/r300/r300_cs_inlines.h
+++ b/src/gallium/drivers/r300/r300_cs_inlines.h
@@ -40,5 +40,10 @@
         (1 << 18) | (1 << 31)); \
 } while (0)
 
+#define R300_SCREENDOOR do { \
+    OUT_CS_REG(R300_SC_SCREENDOOR, 0x0); \
+    R300_PACIFY; \
+    OUT_CS_REG(R300_SC_SCREENDOOR, 0xffffff); \
+} while (0)
 
 #endif /* R300_CS_H */
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index 75864c0ef62..c6464e3eb63 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -115,7 +115,7 @@ void r500_emit_fragment_shader(struct r300_context* r300,
 {
     CS_LOCALS(r300);
     int i = 0;
-    BEGIN_CS(11 + (fs->instruction_count * 6));
+    BEGIN_CS(9 + (fs->instruction_count * 6));
     OUT_CS_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO);
     OUT_CS_REG(R500_US_PIXSIZE, fs->shader.stack_size);
     OUT_CS_REG(R500_US_CODE_ADDR, R500_US_CODE_START_ADDR(0) |
@@ -132,7 +132,6 @@ void r500_emit_fragment_shader(struct r300_context* r300,
         OUT_CS(fs->instructions[i].inst4);
         OUT_CS(fs->instructions[i].inst5);
     }
-    R300_PACIFY;
     END_CS;
 }
 
@@ -159,7 +158,7 @@ void r300_emit_fb_state(struct r300_context* r300,
     struct r300_texture* tex;
     int i;
 
-    BEGIN_CS((5 * fb->nr_cbufs) + (fb->zsbuf ? 5 : 0) + 6);
+    BEGIN_CS((5 * fb->nr_cbufs) + (fb->zsbuf ? 5 : 0) + 4);
     for (i = 0; i < fb->nr_cbufs; i++) {
         tex = (struct r300_texture*)fb->cbufs[i]->texture;
         OUT_CS_REG_SEQ(R300_RB3D_COLOROFFSET0 + (4 * i), 1);
@@ -187,7 +186,6 @@ void r300_emit_fb_state(struct r300_context* r300,
     OUT_CS_REG(R300_ZB_ZCACHE_CTLSTAT,
         R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE |
         R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE);
-    R300_PACIFY;
     END_CS;
 }
 
diff --git a/src/gallium/drivers/r300/r300_surface.c b/src/gallium/drivers/r300/r300_surface.c
index 30a0d06199b..392c7d318dc 100644
--- a/src/gallium/drivers/r300/r300_surface.c
+++ b/src/gallium/drivers/r300/r300_surface.c
@@ -54,8 +54,7 @@ static void r300_surface_fill(struct pipe_context* pipe,
         return;
     }
 
-    BEGIN_CS(164 + (caps->is_r500 ? 22 : 14) + (caps->has_tcl ? 4 : 2));
-    R300_PACIFY;
+    BEGIN_CS(158 + (caps->is_r500 ? 22 : 14) + (caps->has_tcl ? 4 : 2));
     /* Flush PVS. */
     OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x0);
 
@@ -127,7 +126,6 @@ static void r300_surface_fill(struct pipe_context* pipe,
     OUT_CS_REG(R300_SU_DEPTH_OFFSET, 0x00000000);
     OUT_CS_REG(R300_SC_HYPERZ, 0x0000001C);
     OUT_CS_REG(R300_SC_EDGERULE, 0x2DA49525);
-    OUT_CS_REG(R300_SC_SCREENDOOR, 0x00FFFFFF);
     OUT_CS_REG(R300_FG_FOG_BLEND, 0x00000002);
     OUT_CS_REG(R300_FG_FOG_COLOR_R, 0x00000000);
     OUT_CS_REG(R300_FG_FOG_COLOR_G, 0x00000000);
@@ -156,7 +154,6 @@ static void r300_surface_fill(struct pipe_context* pipe,
     OUT_CS_REG(0x4F34, 0x00000000); */
     OUT_CS_REG(R300_ZB_HIZ_OFFSET, 0x00000000);
     OUT_CS_REG(R300_ZB_HIZ_PITCH, 0x00000000);
-    R300_PACIFY;
     if (caps->has_tcl) {
         OUT_CS_REG(R300_VAP_PROG_STREAM_CNTL_0,
             (R300_DATA_TYPE_FLOAT_4 << R300_DATA_TYPE_0_SHIFT) |
@@ -233,7 +230,7 @@ static void r300_surface_fill(struct pipe_context* pipe,
         r300_emit_fragment_shader(r300, &r300_passthrough_fragment_shader);
     }
 
-    BEGIN_CS(8 + (caps->has_tcl ? 23 : 2));
+    BEGIN_CS(8 + (caps->has_tcl ? 20 : 2));
     OUT_CS_REG_SEQ(R300_US_OUT_FMT_0, 4);
     OUT_CS(R300_C0_SEL_B | R300_C1_SEL_G | R300_C2_SEL_R | R300_C3_SEL_A);
     OUT_CS(R300_US_OUT_FMT_UNUSED);
@@ -250,7 +247,6 @@ static void r300_surface_fill(struct pipe_context* pipe,
         OUT_CS_REG(R300_VAP_PVS_CODE_CNTL_0, 0x00100000);
         OUT_CS_REG(R300_VAP_PVS_CONST_CNTL, 0x00000000);
         OUT_CS_REG(R300_VAP_PVS_CODE_CNTL_1, 0x00000001);
-        R300_PACIFY;
         /* XXX translate these back into normal instructions */
         OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x1);
         OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0x0);
@@ -269,15 +265,13 @@ static void r300_surface_fill(struct pipe_context* pipe,
             (0x5 << R300_VF_MAX_VTX_NUM_SHIFT) |
             (caps->num_vert_fpus << R300_PVS_NUM_FPUS_SHIFT));
     }
-    R300_PACIFY;
     END_CS;
 
     r300_emit_blend_state(r300, &blend_clear_state);
     r300_emit_blend_color_state(r300, &blend_color_clear_state);
     r300_emit_dsa_state(r300, &dsa_clear_state);
 
-    BEGIN_CS(32);
-    R300_PACIFY;
+    BEGIN_CS(24);
     /* Flush colorbuffer and blend caches. */
     OUT_CS_REG(R300_RB3D_DSTCACHE_CTLSTAT,
         R300_RB3D_DSTCACHE_CTLSTAT_DC_FLUSH_FLUSH_DIRTY_3D |
@@ -312,12 +306,10 @@ static void r300_surface_fill(struct pipe_context* pipe,
     /* XXX OUT_CS_REG(R300_ZB_ZCACHE_CTLSTAT,
         R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE |
         R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE); */
-    R300_SCREENDOOR;
 
     END_CS;
-    FLUSH_CS;
 
-    r300->dirty_state = R300_NEW_KITCHEN_SINK;
+    r300->dirty_hw++;
 }
 
 void r300_init_surface_functions(struct r300_context* r300)
diff --git a/src/gallium/drivers/r300/r300_swtcl_emit.c b/src/gallium/drivers/r300/r300_swtcl_emit.c
index f6e98d23e98..e51ac2c28d9 100644
--- a/src/gallium/drivers/r300/r300_swtcl_emit.c
+++ b/src/gallium/drivers/r300/r300_swtcl_emit.c
@@ -41,10 +41,44 @@ static INLINE struct swtcl_stage* swtcl_stage(struct draw_stage* draw) {
     return (struct swtcl_stage*)draw;
 }
 
-static void r300_emit_vertex(struct r300_context* r300,
-                             const struct vertex_header* vertex)
+static INLINE void r300_emit_vertex(struct r300_context* r300,
+                                    const struct vertex_header* vertex)
 {
-    /* XXX */
+    struct vertex_info* vinfo = &r300->vertex_info;
+    CS_LOCALS(r300);
+    uint i, j;
+
+    BEGIN_CS(vinfo->size);
+
+    for (i = 0; i < vinfo->num_attribs; i++) {
+        j = vinfo->attrib[i].src_index;
+        switch (vinfo->attrib[i].emit) {
+            case EMIT_1F:
+                CS_OUT_32F(vertex->data[j][0]);
+                break;
+            case EMIT_2F:
+                CS_OUT_32F(vertex->data[j][0]);
+                CS_OUT_32F(vertex->data[j][1]);
+                break;
+            case EMIT_3F:
+                CS_OUT_32F(vertex->data[j][0]);
+                CS_OUT_32F(vertex->data[j][1]);
+                CS_OUT_32F(vertex->data[j][2]);
+                break;
+            case EMIT_4F:
+                CS_OUT_32F(vertex->data[j][0]);
+                CS_OUT_32F(vertex->data[j][1]);
+                CS_OUT_32F(vertex->data[j][2]);
+                CS_OUT_32F(vertex->data[j][3]);
+                break;
+            default:
+                debug_printf("r300: Unknown emit value %d\n",
+                    vinfo->attrib[i].emit);
+                break;
+        }
+    }
+
+    END_CS;
 }
 
 static INLINE void r300_emit_prim(struct draw_stage* draw,
@@ -58,8 +92,12 @@ static INLINE void r300_emit_prim(struct draw_stage* draw,
 
     r300_emit_dirty_state(r300);
 
-    /* XXX should be count * vtx size */
-    BEGIN_CS(2 + count + 6);
+    BEGIN_CS(3);
+    OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
+    OUT_CS(r300->vertex_info.hwfmt[0]);
+    OUT_CS(r300->vertex_info.hwfmt[1]);
+
+    BEGIN_CS(2 + (count * r300->vertex_info.size) + 2);
     OUT_CS(CP_PACKET3(R200_3D_DRAW_IMMD_2, count));
     OUT_CS(hwprim | R300_PRIM_WALK_RING |
         (count << R300_PRIM_NUM_VERTICES_SHIFT));
@@ -67,7 +105,7 @@ static INLINE void r300_emit_prim(struct draw_stage* draw,
     for (i = 0; i < count; i++) {
         r300_emit_vertex(r300, prim->v[i]);
     }
-    R300_PACIFY;
+
     END_CS;
 }