Merge commit 'origin/master' into gallium-sw-api-2

12 files changed, 970 insertions, 1198 deletions

diff --git a/src/gallium/drivers/nv50/Makefile b/src/gallium/drivers/nv50/Makefile
index 612aea28a3..5d622e1c13 100644
--- a/src/gallium/drivers/nv50/Makefile
+++ b/src/gallium/drivers/nv50/Makefile
@@ -16,6 +16,7 @@ C_SOURCES = \
 	nv50_surface.c \
 	nv50_tex.c \
 	nv50_transfer.c \
-	nv50_vbo.c
+	nv50_vbo.c \
+	nv50_push.c
 
 include ../../Makefile.template
diff --git a/src/gallium/drivers/nv50/nv50_clear.c b/src/gallium/drivers/nv50/nv50_clear.c
index e0b2d2880b..8afc95c9fc 100644
--- a/src/gallium/drivers/nv50/nv50_clear.c
+++ b/src/gallium/drivers/nv50/nv50_clear.c
@@ -36,7 +36,7 @@ nv50_clear(struct pipe_context *pipe, unsigned buffers,
 	struct pipe_framebuffer_state *fb = &nv50->framebuffer;
 	unsigned mode = 0, i;
 
-	if (!nv50_state_validate(nv50))
+	if (!nv50_state_validate(nv50, 64))
 		return;
 
 	if (buffers & PIPE_CLEAR_COLOR && fb->nr_cbufs) {
diff --git a/src/gallium/drivers/nv50/nv50_context.c b/src/gallium/drivers/nv50/nv50_context.c
index 7be12fcdef..0eb42f323f 100644
--- a/src/gallium/drivers/nv50/nv50_context.c
+++ b/src/gallium/drivers/nv50/nv50_context.c
@@ -46,43 +46,13 @@ static void
 nv50_destroy(struct pipe_context *pipe)
 {
 	struct nv50_context *nv50 = nv50_context(pipe);
+	int i;
 
-        if (nv50->state.fb)
-		so_ref(NULL, &nv50->state.fb);
-	if (nv50->state.blend)
-		so_ref(NULL, &nv50->state.blend);
-	if (nv50->state.blend_colour)
-		so_ref(NULL, &nv50->state.blend_colour);
-	if (nv50->state.zsa)
-		so_ref(NULL, &nv50->state.zsa);
-	if (nv50->state.rast)
-		so_ref(NULL, &nv50->state.rast);
-	if (nv50->state.stipple)
-		so_ref(NULL, &nv50->state.stipple);
-	if (nv50->state.scissor)
-		so_ref(NULL, &nv50->state.scissor);
-	if (nv50->state.viewport)
-		so_ref(NULL, &nv50->state.viewport);
-	if (nv50->state.tsc_upload)
-		so_ref(NULL, &nv50->state.tsc_upload);
-	if (nv50->state.tic_upload)
-		so_ref(NULL, &nv50->state.tic_upload);
-	if (nv50->state.vertprog)
-		so_ref(NULL, &nv50->state.vertprog);
-	if (nv50->state.fragprog)
-		so_ref(NULL, &nv50->state.fragprog);
-	if (nv50->state.geomprog)
-		so_ref(NULL, &nv50->state.geomprog);
-	if (nv50->state.fp_linkage)
-		so_ref(NULL, &nv50->state.fp_linkage);
-	if (nv50->state.gp_linkage)
-		so_ref(NULL, &nv50->state.gp_linkage);
-	if (nv50->state.vtxfmt)
-		so_ref(NULL, &nv50->state.vtxfmt);
-	if (nv50->state.vtxbuf)
-		so_ref(NULL, &nv50->state.vtxbuf);
-	if (nv50->state.vtxattr)
-		so_ref(NULL, &nv50->state.vtxattr);
+	for (i = 0; i < 64; i++) {
+		if (!nv50->state.hw[i])
+			continue;
+		so_ref(NULL, &nv50->state.hw[i]);
+	}
 
 	draw_destroy(nv50->draw);
 
@@ -123,7 +93,6 @@ nv50_create(struct pipe_screen *pscreen, void *priv)
 	nv50->pipe.is_buffer_referenced = nouveau_is_buffer_referenced;
 
 	screen->base.channel->user_private = nv50;
-	screen->base.channel->flush_notify = nv50_state_flush_notify;
 
 	nv50_init_surface_functions(nv50);
 	nv50_init_state_functions(nv50);
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index 044437e75f..8793c2aac5 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -72,6 +72,12 @@ struct nv50_sampler_stateobj {
 	unsigned tsc[8];
 };
 
+struct nv50_vtxelt_stateobj {
+	struct pipe_vertex_element pipe[16];
+	unsigned num_elements;
+	uint32_t hw[16];
+};
+
 static INLINE unsigned
 get_tile_height(uint32_t tile_mode)
 {
@@ -117,30 +123,12 @@ nv50_surface(struct pipe_surface *pt)
 }
 
 struct nv50_state {
-	unsigned dirty;
+	struct nouveau_stateobj *hw[64];
+	uint64_t hw_dirty;
 
-	struct nouveau_stateobj *fb;
-	struct nouveau_stateobj *blend;
-	struct nouveau_stateobj *blend_colour;
-	struct nouveau_stateobj *zsa;
-	struct nouveau_stateobj *stencil_ref;
-	struct nouveau_stateobj *rast;
-	struct nouveau_stateobj *stipple;
-	struct nouveau_stateobj *scissor;
-	unsigned scissor_enabled;
-	struct nouveau_stateobj *viewport;
-	struct nouveau_stateobj *tsc_upload;
-	struct nouveau_stateobj *tic_upload;
 	unsigned miptree_nr[PIPE_SHADER_TYPES];
-	struct nouveau_stateobj *vertprog;
-	struct nouveau_stateobj *fragprog;
-	struct nouveau_stateobj *geomprog;
-	struct nouveau_stateobj *fp_linkage;
-	struct nouveau_stateobj *gp_linkage;
-	struct nouveau_stateobj *vtxfmt;
 	struct nouveau_stateobj *vtxbuf;
 	struct nouveau_stateobj *vtxattr;
-	struct nouveau_stateobj *instbuf;
 	unsigned vtxelt_nr;
 };
 
@@ -169,14 +157,13 @@ struct nv50_context {
 	struct pipe_buffer *constbuf[PIPE_SHADER_TYPES];
 	struct pipe_vertex_buffer vtxbuf[PIPE_MAX_ATTRIBS];
 	unsigned vtxbuf_nr;
-	struct pipe_vertex_element vtxelt[PIPE_MAX_ATTRIBS];
-	unsigned vtxelt_nr;
+	struct nv50_vtxelt_stateobj *vtxelt;
 	struct nv50_sampler_stateobj *sampler[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS];
 	unsigned sampler_nr[PIPE_SHADER_TYPES];
 	struct nv50_miptree *miptree[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS];
 	unsigned miptree_nr[PIPE_SHADER_TYPES];
 
-	uint16_t vbo_fifo;
+	unsigned vbo_fifo;
 };
 
 static INLINE struct nv50_context *
@@ -218,24 +205,36 @@ extern void nv50_draw_elements_instanced(struct pipe_context *pipe,
 					 unsigned count,
 					 unsigned startInstance,
 					 unsigned instanceCount);
-extern void nv50_vbo_validate(struct nv50_context *nv50);
+extern void nv50_vtxelt_construct(struct nv50_vtxelt_stateobj *cso);
+extern struct nouveau_stateobj *nv50_vbo_validate(struct nv50_context *nv50);
+
+/* nv50_push.c */
+extern void
+nv50_push_elements_instanced(struct pipe_context *, struct pipe_buffer *,
+			     unsigned idxsize, unsigned mode, unsigned start,
+			     unsigned count, unsigned i_start,
+			     unsigned i_count);
 
 /* nv50_clear.c */
 extern void nv50_clear(struct pipe_context *pipe, unsigned buffers,
 		       const float *rgba, double depth, unsigned stencil);
 
 /* nv50_program.c */
-extern void nv50_vertprog_validate(struct nv50_context *nv50);
-extern void nv50_fragprog_validate(struct nv50_context *nv50);
-extern void nv50_geomprog_validate(struct nv50_context *nv50);
-extern void nv50_fp_linkage_validate(struct nv50_context *nv50);
-extern void nv50_gp_linkage_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_vertprog_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_fragprog_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_geomprog_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_fp_linkage_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_gp_linkage_validate(struct nv50_context *nv50);
 extern void nv50_program_destroy(struct nv50_context *nv50,
 				 struct nv50_program *p);
 
 /* nv50_state_validate.c */
-extern boolean nv50_state_validate(struct nv50_context *nv50);
-extern void nv50_state_flush_notify(struct nouveau_channel *chan);
+extern boolean nv50_state_validate(struct nv50_context *nv50, unsigned dwords);
 
 extern void nv50_so_init_sifc(struct nv50_context *nv50,
 			      struct nouveau_stateobj *so,
@@ -243,7 +242,8 @@ extern void nv50_so_init_sifc(struct nv50_context *nv50,
 			      unsigned offset, unsigned size);
 
 /* nv50_tex.c */
-extern void nv50_tex_validate(struct nv50_context *);
+extern void nv50_tex_relocs(struct nv50_context *);
+extern struct nouveau_stateobj *nv50_tex_validate(struct nv50_context *);
 
 /* nv50_transfer.c */
 extern void
@@ -257,4 +257,35 @@ nv50_upload_sifc(struct nv50_context *nv50,
 struct pipe_context *
 nv50_create(struct pipe_screen *pscreen, void *priv);
 
+static INLINE unsigned
+nv50_prim(unsigned mode)
+{
+	switch (mode) {
+	case PIPE_PRIM_POINTS: return NV50TCL_VERTEX_BEGIN_POINTS;
+	case PIPE_PRIM_LINES: return NV50TCL_VERTEX_BEGIN_LINES;
+	case PIPE_PRIM_LINE_LOOP: return NV50TCL_VERTEX_BEGIN_LINE_LOOP;
+	case PIPE_PRIM_LINE_STRIP: return NV50TCL_VERTEX_BEGIN_LINE_STRIP;
+	case PIPE_PRIM_TRIANGLES: return NV50TCL_VERTEX_BEGIN_TRIANGLES;
+	case PIPE_PRIM_TRIANGLE_STRIP:
+		return NV50TCL_VERTEX_BEGIN_TRIANGLE_STRIP;
+	case PIPE_PRIM_TRIANGLE_FAN: return NV50TCL_VERTEX_BEGIN_TRIANGLE_FAN;
+	case PIPE_PRIM_QUADS: return NV50TCL_VERTEX_BEGIN_QUADS;
+	case PIPE_PRIM_QUAD_STRIP: return NV50TCL_VERTEX_BEGIN_QUAD_STRIP;
+	case PIPE_PRIM_POLYGON: return NV50TCL_VERTEX_BEGIN_POLYGON;
+	case PIPE_PRIM_LINES_ADJACENCY:
+		return NV50TCL_VERTEX_BEGIN_LINES_ADJACENCY;
+	case PIPE_PRIM_LINE_STRIP_ADJACENCY:
+		return NV50TCL_VERTEX_BEGIN_LINE_STRIP_ADJACENCY;
+	case PIPE_PRIM_TRIANGLES_ADJACENCY:
+		return NV50TCL_VERTEX_BEGIN_TRIANGLES_ADJACENCY;
+	case PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY:
+		return NV50TCL_VERTEX_BEGIN_TRIANGLE_STRIP_ADJACENCY;
+	default:
+		break;
+	}
+
+	NOUVEAU_ERR("invalid primitive type %d\n", mode);
+	return NV50TCL_VERTEX_BEGIN_POINTS;
+}
+
 #endif
diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index 2372cbbef6..c857816b31 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -4270,7 +4270,7 @@ nv50_program_validate_code(struct nv50_context *nv50, struct nv50_program *p)
 	FREE(up);
 }
 
-void
+struct nouveau_stateobj *
 nv50_vertprog_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4286,6 +4286,9 @@ nv50_vertprog_validate(struct nv50_context *nv50)
 	nv50_program_validate_data(nv50, p);
 	nv50_program_validate_code(nv50, p);
 
+	if (!(nv50->dirty & NV50_NEW_VERTPROG))
+		return NULL;
+
 	so = so_new(5, 7, 2);
 	so_method(so, tesla, NV50TCL_VP_ADDRESS_HIGH, 2);
 	so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
@@ -4301,11 +4304,10 @@ nv50_vertprog_validate(struct nv50_context *nv50)
 	so_data  (so, p->cfg.high_temp);
 	so_method(so, tesla, NV50TCL_VP_START_ID, 1);
 	so_data  (so, 0); /* program start offset */
-	so_ref(so, &nv50->state.vertprog);
-	so_ref(NULL, &so);
+	return so;
 }
 
-void
+struct nouveau_stateobj *
 nv50_fragprog_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4321,6 +4323,9 @@ nv50_fragprog_validate(struct nv50_context *nv50)
 	nv50_program_validate_data(nv50, p);
 	nv50_program_validate_code(nv50, p);
 
+	if (!(nv50->dirty & NV50_NEW_FRAGPROG))
+		return NULL;
+
 	so = so_new(6, 7, 2);
 	so_method(so, tesla, NV50TCL_FP_ADDRESS_HIGH, 2);
 	so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
@@ -4337,11 +4342,10 @@ nv50_fragprog_validate(struct nv50_context *nv50)
 	so_data  (so, p->cfg.regs[3]);
 	so_method(so, tesla, NV50TCL_FP_START_ID, 1);
 	so_data  (so, 0); /* program start offset */
-	so_ref(so, &nv50->state.fragprog);
-	so_ref(NULL, &so);
+	return so;
 }
 
-void
+struct nouveau_stateobj *
 nv50_geomprog_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4357,6 +4361,9 @@ nv50_geomprog_validate(struct nv50_context *nv50)
 	nv50_program_validate_data(nv50, p);
 	nv50_program_validate_code(nv50, p);
 
+	if (!(nv50->dirty & NV50_NEW_GEOMPROG))
+		return NULL;
+
 	so = so_new(6, 7, 2);
 	so_method(so, tesla, NV50TCL_GP_ADDRESS_HIGH, 2);
 	so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
@@ -4373,8 +4380,7 @@ nv50_geomprog_validate(struct nv50_context *nv50)
 	so_data  (so, p->cfg.vert_count);
 	so_method(so, tesla, NV50TCL_GP_START_ID, 1);
 	so_data  (so, 0);
-	so_ref(so, &nv50->state.geomprog);
-	so_ref(NULL, &so);
+	return so;
 }
 
 static uint32_t
@@ -4454,7 +4460,7 @@ nv50_vec4_map(uint32_t *map32, int mid, uint8_t zval, uint32_t lin[4],
 	return mid;
 }
 
-void
+struct nouveau_stateobj *
 nv50_fp_linkage_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4580,8 +4586,7 @@ nv50_fp_linkage_validate(struct nv50_context *nv50)
 	so_method(so, tesla, NV50TCL_GP_ENABLE, 1);
 	so_data  (so, (vp->type == PIPE_SHADER_GEOMETRY) ? 1 : 0);
 
-	so_ref(so, &nv50->state.fp_linkage);
-	so_ref(NULL, &so);
+	return so;
 }
 
 static int
@@ -4615,7 +4620,7 @@ construct_vp_gp_mapping(uint32_t *map32, int m,
 	return m;
 }
 
-void
+struct nouveau_stateobj *
 nv50_gp_linkage_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4625,10 +4630,8 @@ nv50_gp_linkage_validate(struct nv50_context *nv50)
 	uint32_t map[16];
 	int m = 0;
 
-	if (!gp) {
-		so_ref(NULL, &nv50->state.gp_linkage);
-		return;
-	}
+	if (!gp)
+		return NULL;
 	memset(map, 0, sizeof(map));
 
 	m = construct_vp_gp_mapping(map, m, vp, gp);
@@ -4646,8 +4649,7 @@ nv50_gp_linkage_validate(struct nv50_context *nv50)
 	so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), m);
 	so_datap (so, map, m);
 
-	so_ref(so, &nv50->state.gp_linkage);
-	so_ref(NULL, &so);
+	return so;
 }
 
 void
diff --git a/src/gallium/drivers/nv50/nv50_push.c b/src/gallium/drivers/nv50/nv50_push.c
new file mode 100644
index 0000000000..96a1f32d30
--- /dev/null
+++ b/src/gallium/drivers/nv50/nv50_push.c
@@ -0,0 +1,326 @@
+#include "pipe/p_context.h"
+#include "pipe/p_state.h"
+#include "util/u_inlines.h"
+#include "util/u_format.h"
+
+#include "nouveau/nouveau_util.h"
+#include "nv50_context.h"
+
+struct push_context {
+   struct nv50_context *nv50;
+
+   unsigned vtx_size;
+
+   void *idxbuf;
+   unsigned idxsize;
+
+   float edgeflag;
+   int edgeflag_attr;
+
+   struct {
+      void *map;
+      unsigned stride;
+      unsigned divisor;
+      unsigned step;
+      void (*push)(struct nouveau_channel *, void *);
+   } attr[16];
+   unsigned attr_nr;
+};
+
+static void
+emit_b32_1(struct nouveau_channel *chan, void *data)
+{
+   uint32_t *v = data;
+
+   OUT_RING(chan, v[0]);
+}
+
+static void
+emit_b32_2(struct nouveau_channel *chan, void *data)
+{
+   uint32_t *v = data;
+
+   OUT_RING(chan, v[0]);
+   OUT_RING(chan, v[1]);
+}
+
+static void
+emit_b32_3(struct nouveau_channel *chan, void *data)
+{
+   uint32_t *v = data;
+
+   OUT_RING(chan, v[0]);
+   OUT_RING(chan, v[1]);
+   OUT_RING(chan, v[2]);
+}
+
+static void
+emit_b32_4(struct nouveau_channel *chan, void *data)
+{
+   uint32_t *v = data;
+
+   OUT_RING(chan, v[0]);
+   OUT_RING(chan, v[1]);
+   OUT_RING(chan, v[2]);
+   OUT_RING(chan, v[3]);
+}
+
+static void
+emit_b16_1(struct nouveau_channel *chan, void *data)
+{
+   uint16_t *v = data;
+
+   OUT_RING(chan, v[0]);
+}
+
+static void
+emit_b16_3(struct nouveau_channel *chan, void *data)
+{
+   uint16_t *v = data;
+
+   OUT_RING(chan, (v[1] << 16) | v[0]);
+   OUT_RING(chan, v[2]);
+}
+
+static void
+emit_b08_1(struct nouveau_channel *chan, void *data)
+{
+   uint8_t *v = data;
+
+   OUT_RING(chan, v[0]);
+}
+
+static void
+emit_b08_3(struct nouveau_channel *chan, void *data)
+{
+   uint8_t *v = data;
+
+   OUT_RING(chan, (v[2] << 16) | (v[1] << 8) | v[0]);
+}
+
+static INLINE void
+emit_vertex(struct push_context *ctx, unsigned n)
+{
+   struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
+   struct nouveau_channel *chan = tesla->channel;
+   int i;
+
+   if (ctx->edgeflag_attr < 16) {
+      float *edgeflag = ctx->attr[ctx->edgeflag_attr].map +
+                        ctx->attr[ctx->edgeflag_attr].stride * n;
+
+      if (*edgeflag != ctx->edgeflag) {
+         BEGIN_RING(chan, tesla, NV50TCL_EDGEFLAG_ENABLE, 1);
+         OUT_RING  (chan, *edgeflag ? 1 : 0);
+         ctx->edgeflag = *edgeflag;
+      }
+   }
+
+   BEGIN_RING_NI(chan, tesla, NV50TCL_VERTEX_DATA, ctx->vtx_size);
+   for (i = 0; i < ctx->attr_nr; i++)
+      ctx->attr[i].push(chan, ctx->attr[i].map + ctx->attr[i].stride * n);
+}
+
+static void
+emit_edgeflag(void *priv, boolean enabled)
+{
+   struct push_context *ctx = priv;
+   struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
+   struct nouveau_channel *chan = tesla->channel;
+
+   BEGIN_RING(chan, tesla, NV50TCL_EDGEFLAG_ENABLE, 1);
+   OUT_RING  (chan, enabled ? 1 : 0);
+}
+
+static void
+emit_elt08(void *priv, unsigned start, unsigned count)
+{
+   struct push_context *ctx = priv;
+   uint8_t *idxbuf = ctx->idxbuf;
+
+   while (count--)
+      emit_vertex(ctx, idxbuf[start++]);
+}
+
+static void
+emit_elt16(void *priv, unsigned start, unsigned count)
+{
+   struct push_context *ctx = priv;
+   uint16_t *idxbuf = ctx->idxbuf;
+
+   while (count--)
+      emit_vertex(ctx, idxbuf[start++]);
+}
+
+static void
+emit_elt32(void *priv, unsigned start, unsigned count)
+{
+   struct push_context *ctx = priv;
+   uint32_t *idxbuf = ctx->idxbuf;
+
+   while (count--)
+      emit_vertex(ctx, idxbuf[start++]);
+}
+
+static void
+emit_verts(void *priv, unsigned start, unsigned count)
+{
+   while (count--)
+      emit_vertex(priv, start++);
+}
+
+void
+nv50_push_elements_instanced(struct pipe_context *pipe,
+                             struct pipe_buffer *idxbuf, unsigned idxsize,
+                             unsigned mode, unsigned start, unsigned count,
+                             unsigned i_start, unsigned i_count)
+{
+   struct nv50_context *nv50 = nv50_context(pipe);
+   struct nouveau_grobj *tesla = nv50->screen->tesla;
+   struct nouveau_channel *chan = tesla->channel;
+   struct push_context ctx;
+   const unsigned p_overhead = 4 + /* begin/end */
+                               4; /* potential edgeflag enable/disable */
+   const unsigned v_overhead = 1 + /* VERTEX_DATA packet header */
+                               2; /* potential edgeflag modification */
+   struct u_split_prim s;
+   unsigned vtx_size;
+   boolean nzi = FALSE;
+   int i;
+
+   ctx.nv50 = nv50;
+   ctx.attr_nr = 0;
+   ctx.idxbuf = NULL;
+   ctx.vtx_size = 0;
+   ctx.edgeflag = 0.5f;
+   ctx.edgeflag_attr = nv50->vertprog->cfg.edgeflag_in;
+
+   /* map vertex buffers, determine vertex size */
+   for (i = 0; i < nv50->vtxelt->num_elements; i++) {
+      struct pipe_vertex_element *ve = &nv50->vtxelt->pipe[i];
+      struct pipe_vertex_buffer *vb = &nv50->vtxbuf[ve->vertex_buffer_index];
+      struct nouveau_bo *bo = nouveau_bo(vb->buffer);
+      unsigned size, nr_components, n;
+
+      if (!(nv50->vbo_fifo & (1 << i)))
+         continue;
+      n = ctx.attr_nr++;
+
+      if (nouveau_bo_map(bo, NOUVEAU_BO_RD)) {
+         assert(bo->map);
+         return;
+      }
+      ctx.attr[n].map = bo->map + vb->buffer_offset + ve->src_offset;
+      nouveau_bo_unmap(bo);
+
+      ctx.attr[n].stride = vb->stride;
+      ctx.attr[n].divisor = ve->instance_divisor;
+      if (ctx.attr[n].divisor) {
+         ctx.attr[n].step = i_start % ve->instance_divisor;
+         ctx.attr[n].map += i_start * vb->stride;
+      }
+
+      size = util_format_get_component_bits(ve->src_format,
+                                            UTIL_FORMAT_COLORSPACE_RGB, 0);
+      nr_components = util_format_get_nr_components(ve->src_format);
+      switch (size) {
+      case 8:
+         switch (nr_components) {
+         case 1: ctx.attr[n].push = emit_b08_1; break;
+         case 2: ctx.attr[n].push = emit_b16_1; break;
+         case 3: ctx.attr[n].push = emit_b08_3; break;
+         case 4: ctx.attr[n].push = emit_b32_1; break;
+         }
+         ctx.vtx_size++;
+         break;
+      case 16:
+         switch (nr_components) {
+         case 1: ctx.attr[n].push = emit_b16_1; break;
+         case 2: ctx.attr[n].push = emit_b32_1; break;
+         case 3: ctx.attr[n].push = emit_b16_3; break;
+         case 4: ctx.attr[n].push = emit_b32_2; break;
+         }
+         ctx.vtx_size += (nr_components + 1) >> 1;
+         break;
+      case 32:
+         switch (nr_components) {
+         case 1: ctx.attr[n].push = emit_b32_1; break;
+         case 2: ctx.attr[n].push = emit_b32_2; break;
+         case 3: ctx.attr[n].push = emit_b32_3; break;
+         case 4: ctx.attr[n].push = emit_b32_4; break;
+         }
+         ctx.vtx_size += nr_components;
+         break;
+      default:
+         assert(0);
+         return;
+      }
+   }
+   vtx_size = ctx.vtx_size + v_overhead;
+
+   /* map index buffer, if present */
+   if (idxbuf) {
+      struct nouveau_bo *bo = nouveau_bo(idxbuf);
+
+      if (nouveau_bo_map(bo, NOUVEAU_BO_RD)) {
+         assert(bo->map);
+         return;
+      }
+      ctx.idxbuf = bo->map;
+      ctx.idxsize = idxsize;
+      nouveau_bo_unmap(bo);
+   }
+
+   s.priv = &ctx;
+   s.edge = emit_edgeflag;
+   if (idxbuf) {
+      if (idxsize == 1)
+         s.emit = emit_elt08;
+      else
+      if (idxsize == 2)
+         s.emit = emit_elt16;
+      else
+         s.emit = emit_elt32;
+   } else
+      s.emit = emit_verts;
+
+   /* per-instance loop */
+   BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
+   OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
+   OUT_RING  (chan, i_start);
+   while (i_count--) {
+      unsigned max_verts;
+      boolean done;
+
+      for (i = 0; i < ctx.attr_nr; i++) {
+         if (!ctx.attr[i].divisor ||
+              ctx.attr[i].divisor != ++ctx.attr[i].step)
+            continue;
+         ctx.attr[i].step = 0;
+         ctx.attr[i].map += ctx.attr[i].stride;
+      }
+
+      u_split_prim_init(&s, mode, start, count);
+      do {
+         if (AVAIL_RING(chan) < p_overhead + (6 * vtx_size)) {
+            FIRE_RING(chan);
+            if (!nv50_state_validate(nv50, p_overhead + (6 * vtx_size))) {
+               assert(0);
+               return;
+            }
+         }
+
+         max_verts  = AVAIL_RING(chan);
+         max_verts -= p_overhead;
+         max_verts /= vtx_size;
+
+         BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
+         OUT_RING  (chan, nv50_prim(s.mode) | (nzi ? (1 << 28) : 0));
+         done = u_split_prim_next(&s, max_verts);
+         BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
+         OUT_RING  (chan, 0);
+      } while (!done);
+
+      nzi = TRUE;
+   }
+}
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index eed6031eaf..7e2e8aa336 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -95,6 +95,8 @@ nv50_screen_is_format_supported(struct pipe_screen *pscreen,
 static int
 nv50_screen_get_param(struct pipe_screen *pscreen, int param)
 {
+	struct nv50_screen *screen = nv50_screen(pscreen);
+
 	switch (param) {
 	case PIPE_CAP_MAX_TEXTURE_IMAGE_UNITS:
 		return 32;
@@ -132,9 +134,9 @@ nv50_screen_get_param(struct pipe_screen *pscreen, int param)
 	case PIPE_CAP_BLEND_EQUATION_SEPARATE:
 		return 1;
 	case NOUVEAU_CAP_HW_VTXBUF:
-		return 1;
+		return screen->force_push ? 0 : 1;
 	case NOUVEAU_CAP_HW_IDXBUF:
-		return 1;
+		return screen->force_push ? 0 : 1;
 	case PIPE_CAP_INDEP_BLEND_ENABLE:
 		return 1;
 	case PIPE_CAP_INDEP_BLEND_FUNC:
@@ -202,28 +204,6 @@ nv50_screen_destroy(struct pipe_screen *pscreen)
 	FREE(screen);
 }
 
-static int
-nv50_pre_pipebuffer_map(struct pipe_screen *pscreen, struct pipe_buffer *pb,
-	unsigned usage)
-{
-	struct nv50_screen *screen = nv50_screen(pscreen);
-	struct nv50_context *ctx = screen->cur_ctx;
-
-	if (!(pb->usage & PIPE_BUFFER_USAGE_VERTEX))
-		return 0;
-
-	/* Our vtxbuf got mapped, it can no longer be considered part of current
-	 * state, remove it to avoid emitting reloc markers.
-	 */
-	if (ctx && ctx->state.vtxbuf && so_bo_is_reloc(ctx->state.vtxbuf,
-			nouveau_bo(pb))) {
-		so_ref(NULL, &ctx->state.vtxbuf);
-		ctx->dirty |= NV50_NEW_ARRAYS;
-	}
-
-	return 0;
-}
-
 struct pipe_screen *
 nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
 {
@@ -252,7 +232,6 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
 	pscreen->get_paramf = nv50_screen_get_paramf;
 	pscreen->is_format_supported = nv50_screen_is_format_supported;
 	pscreen->context_create = nv50_create;
-	screen->base.pre_pipebuffer_map_callback = nv50_pre_pipebuffer_map;
 
 	nv50_screen_init_miptree_functions(pscreen);
 	nv50_transfer_init_screen_functions(pscreen);
@@ -508,10 +487,6 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
 	so_method(so, screen->tesla, NV50TCL_LINKED_TSC, 1);
 	so_data  (so, 1);
 
-	/* activate first scissor rectangle */
-	so_method(so, screen->tesla, NV50TCL_SCISSOR_ENABLE(0), 1);
-	so_data  (so, 1);
-
 	so_method(so, screen->tesla, NV50TCL_EDGEFLAG_ENABLE, 1);
 	so_data  (so, 1); /* default edgeflag to TRUE */
 
@@ -520,6 +495,7 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
 	so_ref (NULL, &so);
 	nouveau_pushbuf_flush(chan, 0);
 
+	screen->force_push = debug_get_bool_option("NV50_ALWAYS_PUSH", FALSE);
 	return pscreen;
 }
 
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 2687b72127..d1bc80cb9e 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -28,6 +28,8 @@ struct nv50_screen {
 	struct nouveau_bo *tsc;
 
 	struct nouveau_stateobj *static_init;
+
+	boolean force_push;
 };
 
 static INLINE struct nv50_screen *
diff --git a/src/gallium/drivers/nv50/nv50_state.c b/src/gallium/drivers/nv50/nv50_state.c
index 7d304907b6..b0e5552eff 100644
--- a/src/gallium/drivers/nv50/nv50_state.c
+++ b/src/gallium/drivers/nv50/nv50_state.c
@@ -302,7 +302,7 @@ static void *
 nv50_rasterizer_state_create(struct pipe_context *pipe,
 			     const struct pipe_rasterizer_state *cso)
 {
-	struct nouveau_stateobj *so = so_new(15, 21, 0);
+	struct nouveau_stateobj *so = so_new(16, 22, 0);
 	struct nouveau_grobj *tesla = nv50_context(pipe)->screen->tesla;
 	struct nv50_rasterizer_stateobj *rso =
 		CALLOC_STRUCT(nv50_rasterizer_stateobj);
@@ -314,6 +314,9 @@ nv50_rasterizer_state_create(struct pipe_context *pipe,
 	 * 	- point_sprite / sprite_coord_mode
 	 */
 
+	so_method(so, tesla, NV50TCL_SCISSOR_ENABLE(0), 1);
+	so_data  (so, cso->scissor);
+
 	so_method(so, tesla, NV50TCL_SHADE_MODEL, 1);
 	so_data  (so, cso->flatshade ? NV50TCL_SHADE_MODEL_FLAT :
 				       NV50TCL_SHADE_MODEL_SMOOTH);
@@ -720,15 +723,34 @@ nv50_set_vertex_buffers(struct pipe_context *pipe, unsigned count,
 	nv50->dirty |= NV50_NEW_ARRAYS;
 }
 
+static void *
+nv50_vtxelts_state_create(struct pipe_context *pipe,
+			  unsigned num_elements,
+			  const struct pipe_vertex_element *elements)
+{
+	struct nv50_vtxelt_stateobj *cso = CALLOC_STRUCT(nv50_vtxelt_stateobj);
+
+	assert(num_elements < 16); /* not doing fallbacks yet */
+	cso->num_elements = num_elements;
+	memcpy(cso->pipe, elements, num_elements * sizeof(*elements));
+
+	nv50_vtxelt_construct(cso);
+
+	return (void *)cso;
+}
+
 static void
-nv50_set_vertex_elements(struct pipe_context *pipe, unsigned count,
-			 const struct pipe_vertex_element *ve)
+nv50_vtxelts_state_delete(struct pipe_context *pipe, void *hwcso)
 {
-	struct nv50_context *nv50 = nv50_context(pipe);
+	FREE(hwcso);
+}
 
-	memcpy(nv50->vtxelt, ve, sizeof(*ve) * count);
-	nv50->vtxelt_nr = count;
+static void
+nv50_vtxelts_state_bind(struct pipe_context *pipe, void *hwcso)
+{
+	struct nv50_context *nv50 = nv50_context(pipe);
 
+	nv50->vtxelt = hwcso;
 	nv50->dirty |= NV50_NEW_ARRAYS;
 }
 
@@ -778,7 +800,10 @@ nv50_init_state_functions(struct nv50_context *nv50)
 	nv50->pipe.set_scissor_state = nv50_set_scissor_state;
 	nv50->pipe.set_viewport_state = nv50_set_viewport_state;
 
+	nv50->pipe.create_vertex_elements_state = nv50_vtxelts_state_create;
+	nv50->pipe.delete_vertex_elements_state = nv50_vtxelts_state_delete;
+	nv50->pipe.bind_vertex_elements_state = nv50_vtxelts_state_bind;
+
 	nv50->pipe.set_vertex_buffers = nv50_set_vertex_buffers;
-	nv50->pipe.set_vertex_elements = nv50_set_vertex_elements;
 }
 
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index c974cc92dc..2c8e7ca798 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -25,8 +25,8 @@
 #include "nv50_context.h"
 #include "nouveau/nouveau_stateobj.h"
 
-static void
-nv50_state_validate_fb(struct nv50_context *nv50)
+static struct nouveau_stateobj *
+validate_fb(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
 	struct nouveau_stateobj *so = so_new(32, 79, 18);
@@ -167,12 +167,7 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 	so_data  (so, w << 16);
 	so_data  (so, h << 16);
 
-	/* we set scissors to framebuffer size when they're 'turned off' */
-	nv50->dirty |= NV50_NEW_SCISSOR;
-	so_ref(NULL, &nv50->state.scissor);
-
-	so_ref(so, &nv50->state.fb);
-	so_ref(NULL, &so);
+	return so;
 }
 
 static void
@@ -199,263 +194,256 @@ nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
 	}
 }
 
-static void
-nv50_state_emit(struct nv50_context *nv50)
+static struct nouveau_stateobj *
+validate_blend(struct nv50_context *nv50)
 {
-	struct nv50_screen *screen = nv50->screen;
-	struct nouveau_channel *chan = screen->base.channel;
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->blend->so, &so);
+	return so;
+}
 
-	/* XXX: this is racy for multiple contexts active on separate
-	 * threads.
-	 */
-	if (screen->cur_ctx != nv50) {
-		if (nv50->state.fb)
-			nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
-		if (nv50->state.blend)
-			nv50->state.dirty |= NV50_NEW_BLEND;
-		if (nv50->state.zsa)
-			nv50->state.dirty |= NV50_NEW_ZSA;
-		if (nv50->state.vertprog)
-			nv50->state.dirty |= NV50_NEW_VERTPROG;
-		if (nv50->state.fragprog)
-			nv50->state.dirty |= NV50_NEW_FRAGPROG;
-		if (nv50->state.geomprog)
-			nv50->state.dirty |= NV50_NEW_GEOMPROG;
-		if (nv50->state.rast)
-			nv50->state.dirty |= NV50_NEW_RASTERIZER;
-		if (nv50->state.blend_colour)
-			nv50->state.dirty |= NV50_NEW_BLEND_COLOUR;
-		if (nv50->state.stencil_ref)
-			nv50->state.dirty |= NV50_NEW_STENCIL_REF;
-		if (nv50->state.stipple)
-			nv50->state.dirty |= NV50_NEW_STIPPLE;
-		if (nv50->state.scissor)
-			nv50->state.dirty |= NV50_NEW_SCISSOR;
-		if (nv50->state.viewport)
-			nv50->state.dirty |= NV50_NEW_VIEWPORT;
-		if (nv50->state.tsc_upload)
-			nv50->state.dirty |= NV50_NEW_SAMPLER;
-		if (nv50->state.tic_upload)
-			nv50->state.dirty |= NV50_NEW_TEXTURE;
-		if (nv50->state.vtxfmt && nv50->state.vtxbuf)
-			nv50->state.dirty |= NV50_NEW_ARRAYS;
-		screen->cur_ctx = nv50;
-	}
+static struct nouveau_stateobj *
+validate_zsa(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->zsa->so, &so);
+	return so;
+}
 
-	if (nv50->state.dirty & NV50_NEW_FRAMEBUFFER)
-		so_emit(chan, nv50->state.fb);
-	if (nv50->state.dirty & NV50_NEW_BLEND)
-		so_emit(chan, nv50->state.blend);
-	if (nv50->state.dirty & NV50_NEW_ZSA)
-		so_emit(chan, nv50->state.zsa);
-	if (nv50->state.dirty & NV50_NEW_VERTPROG)
-		so_emit(chan, nv50->state.vertprog);
-	if (nv50->state.dirty & NV50_NEW_FRAGPROG)
-		so_emit(chan, nv50->state.fragprog);
-	if (nv50->state.dirty & NV50_NEW_GEOMPROG && nv50->state.geomprog)
-		so_emit(chan, nv50->state.geomprog);
-	if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
-				 NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
-		so_emit(chan, nv50->state.fp_linkage);
-	if ((nv50->state.dirty & (NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG))
-	    && nv50->state.gp_linkage)
-		so_emit(chan, nv50->state.gp_linkage);
-	if (nv50->state.dirty & NV50_NEW_RASTERIZER)
-		so_emit(chan, nv50->state.rast);
-	if (nv50->state.dirty & NV50_NEW_BLEND_COLOUR)
-		so_emit(chan, nv50->state.blend_colour);
-	if (nv50->state.dirty & NV50_NEW_STENCIL_REF)
-		so_emit(chan, nv50->state.stencil_ref);
-	if (nv50->state.dirty & NV50_NEW_STIPPLE)
-		so_emit(chan, nv50->state.stipple);
-	if (nv50->state.dirty & NV50_NEW_SCISSOR)
-		so_emit(chan, nv50->state.scissor);
-	if (nv50->state.dirty & NV50_NEW_VIEWPORT)
-		so_emit(chan, nv50->state.viewport);
-	if (nv50->state.dirty & NV50_NEW_SAMPLER)
-		so_emit(chan, nv50->state.tsc_upload);
-	if (nv50->state.dirty & NV50_NEW_TEXTURE)
-		so_emit(chan, nv50->state.tic_upload);
-	if (nv50->state.dirty & NV50_NEW_ARRAYS) {
-		so_emit(chan, nv50->state.vtxfmt);
-		so_emit(chan, nv50->state.vtxbuf);
-		if (nv50->state.vtxattr)
-			so_emit(chan, nv50->state.vtxattr);
-	}
-	nv50->state.dirty = 0;
+static struct nouveau_stateobj *
+validate_rast(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->rasterizer->so, &so);
+	return so;
 }
 
-void
-nv50_state_flush_notify(struct nouveau_channel *chan)
+static struct nouveau_stateobj *
+validate_blend_colour(struct nv50_context *nv50)
 {
-	struct nv50_context *nv50 = chan->user_private;
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so_new(1, 4, 0);
+
+	so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
+	so_data  (so, fui(nv50->blend_colour.color[0]));
+	so_data  (so, fui(nv50->blend_colour.color[1]));
+	so_data  (so, fui(nv50->blend_colour.color[2]));
+	so_data  (so, fui(nv50->blend_colour.color[3]));
+	return so;
+}
 
-	if (nv50->state.tic_upload && !(nv50->dirty & NV50_NEW_TEXTURE))
-		so_emit(chan, nv50->state.tic_upload);
+static struct nouveau_stateobj *
+validate_stencil_ref(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so = so_new(2, 2, 0);
 
-	so_emit_reloc_markers(chan, nv50->state.fb);
-	so_emit_reloc_markers(chan, nv50->state.vertprog);
-	so_emit_reloc_markers(chan, nv50->state.fragprog);
-	so_emit_reloc_markers(chan, nv50->state.vtxbuf);
-	so_emit_reloc_markers(chan, nv50->screen->static_init);
+	so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
+	so_data  (so, nv50->stencil_ref.ref_value[0]);
+	so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
+	so_data  (so, nv50->stencil_ref.ref_value[1]);
+	return so;
+}
 
-	if (nv50->state.instbuf)
-		so_emit_reloc_markers(chan, nv50->state.instbuf);
+static struct nouveau_stateobj *
+validate_stipple(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so_new(1, 32, 0);
+	int i;
+
+	so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
+	for (i = 0; i < 32; i++)
+		so_data(so, util_bswap32(nv50->stipple.stipple[i]));
+	return so;
 }
 
-boolean
-nv50_state_validate(struct nv50_context *nv50)
+static struct nouveau_stateobj *
+validate_scissor(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
+        struct pipe_scissor_state *s = &nv50->scissor;
 	struct nouveau_stateobj *so;
-	unsigned i;
 
-	if (nv50->dirty & NV50_NEW_FRAMEBUFFER)
-		nv50_state_validate_fb(nv50);
+	so = so_new(1, 2, 0);
+	so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
+	so_data  (so, (s->maxx << 16) | s->minx);
+	so_data  (so, (s->maxy << 16) | s->miny);
+	return so;
+}
+
+static struct nouveau_stateobj *
+validate_viewport(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so_new(5, 9, 0);
+
+	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
+	so_data  (so, fui(nv50->viewport.translate[0]));
+	so_data  (so, fui(nv50->viewport.translate[1]));
+	so_data  (so, fui(nv50->viewport.translate[2]));
+	so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
+	so_data  (so, fui(nv50->viewport.scale[0]));
+	so_data  (so, fui(nv50->viewport.scale[1]));
+	so_data  (so, fui(nv50->viewport.scale[2]));
+
+	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
+	so_data  (so, 1);
+	/* 0x0000 = remove whole primitive only (xyz)
+	 * 0x1018 = remove whole primitive only (xy), clamp z
+	 * 0x1080 = clip primitive (xyz)
+	 * 0x1098 = clip primitive (xy), clamp z
+	 */
+	so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
+	so_data  (so, 0x1080);
+	/* no idea what 0f90 does */
+	so_method(so, tesla, 0x0f90, 1);
+	so_data  (so, 0);
+
+	return so;
+}
 
-	if (nv50->dirty & NV50_NEW_BLEND)
-		so_ref(nv50->blend->so, &nv50->state.blend);
+static struct nouveau_stateobj *
+validate_sampler(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so;
+	unsigned nr = 0, i;
 
-	if (nv50->dirty & NV50_NEW_ZSA)
-		so_ref(nv50->zsa->so, &nv50->state.zsa);
+	for (i = 0; i < PIPE_SHADER_TYPES; ++i)
+		nr += nv50->sampler_nr[i];
 
-	if (nv50->dirty & (NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB))
-		nv50_vertprog_validate(nv50);
+	so = so_new(1 + 5 * PIPE_SHADER_TYPES,
+		    1 + 19 * PIPE_SHADER_TYPES + nr * 8,
+		    PIPE_SHADER_TYPES * 2);
 
-	if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
-		nv50_fragprog_validate(nv50);
+	nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
+	nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
 
-	if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB))
-		nv50_geomprog_validate(nv50);
+	so_method(so, tesla, 0x1334, 1); /* flush TSC */
+	so_data  (so, 0);
 
-	if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
-			   NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
-		nv50_fp_linkage_validate(nv50);
+	return so;
+}
 
-	if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_VERTPROG))
-		nv50_gp_linkage_validate(nv50);
+static struct nouveau_stateobj *
+validate_vtxbuf(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->state.vtxbuf, &so);
+	return so;
+}
 
-	if (nv50->dirty & NV50_NEW_RASTERIZER)
-		so_ref(nv50->rasterizer->so, &nv50->state.rast);
+static struct nouveau_stateobj *
+validate_vtxattr(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->state.vtxattr, &so);
+	return so;
+}
 
-	if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
-		so = so_new(1, 4, 0);
-		so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
-		so_data  (so, fui(nv50->blend_colour.color[0]));
-		so_data  (so, fui(nv50->blend_colour.color[1]));
-		so_data  (so, fui(nv50->blend_colour.color[2]));
-		so_data  (so, fui(nv50->blend_colour.color[3]));
-		so_ref(so, &nv50->state.blend_colour);
-		so_ref(NULL, &so);
-	}
+struct state_validate {
+	struct nouveau_stateobj *(*func)(struct nv50_context *nv50);
+	unsigned states;
+} validate_list[] = {
+	{ validate_fb             , NV50_NEW_FRAMEBUFFER                      },
+	{ validate_blend          , NV50_NEW_BLEND                            },
+	{ validate_zsa            , NV50_NEW_ZSA                              },
+	{ nv50_vertprog_validate  , NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB  },
+	{ nv50_fragprog_validate  , NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB  },
+	{ nv50_geomprog_validate  , NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB  },
+	{ nv50_fp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG |
+				    NV50_NEW_FRAGPROG | NV50_NEW_RASTERIZER   },
+	{ nv50_gp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG     },
+	{ validate_rast           , NV50_NEW_RASTERIZER                       },
+	{ validate_blend_colour   , NV50_NEW_BLEND_COLOUR                     },
+	{ validate_stencil_ref    , NV50_NEW_STENCIL_REF                      },
+	{ validate_stipple        , NV50_NEW_STIPPLE                          },
+	{ validate_scissor        , NV50_NEW_SCISSOR                          },
+	{ validate_viewport       , NV50_NEW_VIEWPORT                         },
+	{ validate_sampler        , NV50_NEW_SAMPLER                          },
+	{ nv50_tex_validate       , NV50_NEW_TEXTURE | NV50_NEW_SAMPLER       },
+	{ nv50_vbo_validate       , NV50_NEW_ARRAYS                           },
+	{ validate_vtxbuf         , NV50_NEW_ARRAYS                           },
+	{ validate_vtxattr        , NV50_NEW_ARRAYS                           },
+	{}
+};
+#define validate_list_len (sizeof(validate_list) / sizeof(validate_list[0]))
 
-	if (nv50->dirty & NV50_NEW_STENCIL_REF) {
-		so = so_new(2, 2, 0);
-		so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
-		so_data  (so, nv50->stencil_ref.ref_value[0]);
-		so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
-		so_data  (so, nv50->stencil_ref.ref_value[1]);
-		so_ref(so, &nv50->state.stencil_ref);
-		so_ref(NULL, &so);
-	}
+boolean
+nv50_state_validate(struct nv50_context *nv50, unsigned wait_dwords)
+{
+	struct nouveau_channel *chan = nv50->screen->base.channel;
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	unsigned nr_relocs = 128, nr_dwords = wait_dwords + 128 + 4;
+	int ret, i;
 
-	if (nv50->dirty & NV50_NEW_STIPPLE) {
-		so = so_new(1, 32, 0);
-		so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
-		for (i = 0; i < 32; i++)
-			so_data(so, util_bswap32(nv50->stipple.stipple[i]));
-		so_ref(so, &nv50->state.stipple);
-		so_ref(NULL, &so);
-	}
+	for (i = 0; i < validate_list_len; i++) {
+		struct state_validate *validate = &validate_list[i];
+		struct nouveau_stateobj *so;
 
-	if (nv50->dirty & (NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER)) {
-		struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
-		struct pipe_scissor_state *s = &nv50->scissor;
+		if (!(nv50->dirty & validate->states))
+			continue;
 
-		if (nv50->state.scissor &&
-		    (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
-			goto scissor_uptodate;
-		nv50->state.scissor_enabled = rast->scissor;
+		so = validate->func(nv50);
+		if (!so)
+			continue;
 
-		so = so_new(1, 2, 0);
-		so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
-		if (nv50->state.scissor_enabled) {
-			so_data(so, (s->maxx << 16) | s->minx);
-			so_data(so, (s->maxy << 16) | s->miny);
-		} else {
-			so_data(so, (nv50->framebuffer.width << 16));
-			so_data(so, (nv50->framebuffer.height << 16));
-		}
-		so_ref(so, &nv50->state.scissor);
-		so_ref(NULL, &so);
-		nv50->state.dirty |= NV50_NEW_SCISSOR;
-	}
-scissor_uptodate:
-
-	if (nv50->dirty & (NV50_NEW_VIEWPORT | NV50_NEW_RASTERIZER)) {
-		if (nv50->state.viewport &&
-		    !(nv50->dirty & NV50_NEW_VIEWPORT))
-			goto viewport_uptodate;
-
-		so = so_new(5, 9, 0);
-		so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
-		so_data  (so, fui(nv50->viewport.translate[0]));
-		so_data  (so, fui(nv50->viewport.translate[1]));
-		so_data  (so, fui(nv50->viewport.translate[2]));
-		so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
-		so_data  (so, fui(nv50->viewport.scale[0]));
-		so_data  (so, fui(nv50->viewport.scale[1]));
-		so_data  (so, fui(nv50->viewport.scale[2]));
-
-		so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
-		so_data  (so, 1);
-		/* 0x0000 = remove whole primitive only (xyz)
-		 * 0x1018 = remove whole primitive only (xy), clamp z
-		 * 0x1080 = clip primitive (xyz)
-		 * 0x1098 = clip primitive (xy), clamp z
-		 */
-		so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
-		so_data  (so, 0x1080);
-		/* no idea what 0f90 does */
-		so_method(so, tesla, 0x0f90, 1);
-		so_data  (so, 0);
+		nr_dwords += (so->total + so->cur);
+		nr_relocs += so->cur_reloc;
 
-		so_ref(so, &nv50->state.viewport);
+		so_ref(so, &nv50->state.hw[i]);
 		so_ref(NULL, &so);
-		nv50->state.dirty |= NV50_NEW_VIEWPORT;
+		nv50->state.hw_dirty |= (1 << i);
 	}
-viewport_uptodate:
-
-	if (nv50->dirty & NV50_NEW_SAMPLER) {
-		unsigned nr = 0;
-
-		for (i = 0; i < PIPE_SHADER_TYPES; ++i)
-			nr += nv50->sampler_nr[i];
+	nv50->dirty = 0;
 
-		so = so_new(1 + 5 * PIPE_SHADER_TYPES,
-			    1 + 19 * PIPE_SHADER_TYPES + nr * 8,
-			    PIPE_SHADER_TYPES * 2);
+	if (nv50->screen->cur_ctx != nv50) {
+		for (i = 0; i < validate_list_len; i++) {
+			if (!nv50->state.hw[i] ||
+			    (nv50->state.hw_dirty & (1 << i)))
+				continue;
 
-		nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
-		nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
+			nr_dwords += (nv50->state.hw[i]->total +
+				      nv50->state.hw[i]->cur);
+			nr_relocs += nv50->state.hw[i]->cur_reloc;
+			nv50->state.hw_dirty |= (1 << i);
+		}
 
-		so_method(so, tesla, 0x1334, 1); /* flush TSC */
-		so_data  (so, 0);
+		nv50->screen->cur_ctx = nv50;
+	}
 
-		so_ref(so, &nv50->state.tsc_upload);
-		so_ref(NULL, &so);
+	ret = MARK_RING(chan, nr_dwords, nr_relocs);
+	if (ret) {
+		debug_printf("MARK_RING(%d, %d) failed: %d\n",
+			     nr_dwords, nr_relocs, ret);
+		return FALSE;
 	}
 
-	if (nv50->dirty & (NV50_NEW_TEXTURE | NV50_NEW_SAMPLER))
-		nv50_tex_validate(nv50);
+	while (nv50->state.hw_dirty) {
+		i = ffs(nv50->state.hw_dirty) - 1;
+		nv50->state.hw_dirty &= ~(1 << i);
 
-	if (nv50->dirty & NV50_NEW_ARRAYS)
-		nv50_vbo_validate(nv50);
+		so_emit(chan, nv50->state.hw[i]);
+	}
 
-	nv50->state.dirty |= nv50->dirty;
-	nv50->dirty = 0;
-	nv50_state_emit(nv50);
+	/* Yes, really, we need to do this.  If a buffer that is referenced
+	 * on the hardware isn't part of changed state above, without doing
+	 * this the kernel is given no clue that the buffer is being used
+	 * still.  This can cause all sorts of fun issues.
+	 */
+	nv50_tex_relocs(nv50);
+	so_emit_reloc_markers(chan, nv50->state.hw[0]); /* fb */
+	so_emit_reloc_markers(chan, nv50->state.hw[3]); /* vp */
+	so_emit_reloc_markers(chan, nv50->state.hw[4]); /* fp */
+	so_emit_reloc_markers(chan, nv50->state.hw[17]); /* vb */
+	so_emit_reloc_markers(chan, nv50->screen->static_init);
 
+	/* No idea.. */
+	BEGIN_RING(chan, tesla, 0x142c, 1);
+	OUT_RING  (chan, 0);
+	BEGIN_RING(chan, tesla, 0x142c, 1);
+	OUT_RING  (chan, 0);
 	return TRUE;
 }
 
diff --git a/src/gallium/drivers/nv50/nv50_tex.c b/src/gallium/drivers/nv50/nv50_tex.c
index de0560e20c..4c48b12cd8 100644
--- a/src/gallium/drivers/nv50/nv50_tex.c
+++ b/src/gallium/drivers/nv50/nv50_tex.c
@@ -24,6 +24,7 @@
 #include "nv50_texture.h"
 
 #include "nouveau/nouveau_stateobj.h"
+#include "nouveau/nouveau_reloc.h"
 
 #include "util/u_format.h"
 
@@ -195,6 +196,35 @@ nv50_validate_textures(struct nv50_context *nv50, struct nouveau_stateobj *so,
 }
 
 void
+nv50_tex_relocs(struct nv50_context *nv50)
+{
+	struct nouveau_channel *chan = nv50->screen->tesla->channel;
+	int p, unit;
+
+	p = PIPE_SHADER_FRAGMENT;
+	for (unit = 0; unit < nv50->miptree_nr[p]; unit++) {
+		if (!nv50->miptree[p][unit])
+			continue;
+		nouveau_reloc_emit(chan, nv50->screen->tic,
+				   ((p * 32) + unit) * 32, NULL,
+				   nv50->miptree[p][unit]->base.bo, 0, 0,
+				   NOUVEAU_BO_VRAM | NOUVEAU_BO_LOW |
+				   NOUVEAU_BO_RD, 0, 0);
+	}
+
+	p = PIPE_SHADER_VERTEX;
+	for (unit = 0; unit < nv50->miptree_nr[p]; unit++) {
+		if (!nv50->miptree[p][unit])
+			continue;
+		nouveau_reloc_emit(chan, nv50->screen->tic,
+				   ((p * 32) + unit) * 32, NULL,
+				   nv50->miptree[p][unit]->base.bo, 0, 0,
+				   NOUVEAU_BO_VRAM | NOUVEAU_BO_LOW |
+				   NOUVEAU_BO_RD, 0, 0);
+	}
+}
+
+struct nouveau_stateobj *
 nv50_tex_validate(struct nv50_context *nv50)
 {
 	struct nouveau_stateobj *so;
@@ -217,12 +247,11 @@ nv50_tex_validate(struct nv50_context *nv50)
 		so_ref(NULL, &so);
 
 		NOUVEAU_ERR("failed tex validate\n");
-		return;
+		return NULL;
 	}
 
 	so_method(so, tesla, 0x1330, 1); /* flush TIC */
 	so_data  (so, 0);
 
-	so_ref(so, &nv50->state.tic_upload);
-	so_ref(NULL, &so);
+	return so;
 }
diff --git a/src/gallium/drivers/nv50/nv50_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index 1c8ee0b9ad..6b9c1ee231 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -25,53 +25,9 @@
 #include "util/u_inlines.h"
 #include "util/u_format.h"
 
+#include "nouveau/nouveau_util.h"
 #include "nv50_context.h"
 
-static boolean
-nv50_push_elements_u08(struct nv50_context *, uint8_t *, unsigned);
-
-static boolean
-nv50_push_elements_u16(struct nv50_context *, uint16_t *, unsigned);
-
-static boolean
-nv50_push_elements_u32(struct nv50_context *, uint32_t *, unsigned);
-
-static boolean
-nv50_push_arrays(struct nv50_context *, unsigned, unsigned);
-
-#define NV50_USING_LOATHED_EDGEFLAG(ctx) ((ctx)->vertprog->cfg.edgeflag_in < 16)
-
-static INLINE unsigned
-nv50_prim(unsigned mode)
-{
-	switch (mode) {
-	case PIPE_PRIM_POINTS: return NV50TCL_VERTEX_BEGIN_POINTS;
-	case PIPE_PRIM_LINES: return NV50TCL_VERTEX_BEGIN_LINES;
-	case PIPE_PRIM_LINE_LOOP: return NV50TCL_VERTEX_BEGIN_LINE_LOOP;
-	case PIPE_PRIM_LINE_STRIP: return NV50TCL_VERTEX_BEGIN_LINE_STRIP;
-	case PIPE_PRIM_TRIANGLES: return NV50TCL_VERTEX_BEGIN_TRIANGLES;
-	case PIPE_PRIM_TRIANGLE_STRIP:
-		return NV50TCL_VERTEX_BEGIN_TRIANGLE_STRIP;
-	case PIPE_PRIM_TRIANGLE_FAN: return NV50TCL_VERTEX_BEGIN_TRIANGLE_FAN;
-	case PIPE_PRIM_QUADS: return NV50TCL_VERTEX_BEGIN_QUADS;
-	case PIPE_PRIM_QUAD_STRIP: return NV50TCL_VERTEX_BEGIN_QUAD_STRIP;
-	case PIPE_PRIM_POLYGON: return NV50TCL_VERTEX_BEGIN_POLYGON;
-	case PIPE_PRIM_LINES_ADJACENCY:
-		return NV50TCL_VERTEX_BEGIN_LINES_ADJACENCY;
-	case PIPE_PRIM_LINE_STRIP_ADJACENCY:
-		return NV50TCL_VERTEX_BEGIN_LINE_STRIP_ADJACENCY;
-	case PIPE_PRIM_TRIANGLES_ADJACENCY:
-		return NV50TCL_VERTEX_BEGIN_TRIANGLES_ADJACENCY;
-	case PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY:
-		return NV50TCL_VERTEX_BEGIN_TRIANGLE_STRIP_ADJACENCY;
-	default:
-		break;
-	}
-
-	NOUVEAU_ERR("invalid primitive type %d\n", mode);
-	return NV50TCL_VERTEX_BEGIN_POINTS;
-}
-
 static INLINE uint32_t
 nv50_vbo_type_to_hw(enum pipe_format format)
 {
@@ -139,15 +95,16 @@ nv50_vbo_vtxelt_to_hw(struct pipe_vertex_element *ve)
 	uint32_t hw_type, hw_size;
 	enum pipe_format pf = ve->src_format;
 	const struct util_format_description *desc;
-	unsigned size;
+	unsigned size, nr_components;
 
 	desc = util_format_description(pf);
 	assert(desc);
 
 	size = util_format_get_component_bits(pf, UTIL_FORMAT_COLORSPACE_RGB, 0);
+	nr_components = util_format_get_nr_components(pf);
 
 	hw_type = nv50_vbo_type_to_hw(pf);
-	hw_size = nv50_vbo_size_to_hw(size, ve->nr_components);
+	hw_size = nv50_vbo_size_to_hw(size, nr_components);
 
 	if (!hw_type || !hw_size) {
 		NOUVEAU_ERR("unsupported vbo format: %s\n", util_format_name(pf));
@@ -161,250 +118,58 @@ nv50_vbo_vtxelt_to_hw(struct pipe_vertex_element *ve)
 	return (hw_type | hw_size);
 }
 
-/* For instanced drawing from user buffers, hitting the FIFO repeatedly
- * with the same vertex data is probably worse than uploading all data.
- */
-static boolean
-nv50_upload_vtxbuf(struct nv50_context *nv50, unsigned i)
-{
-	struct nv50_screen *nscreen = nv50->screen;
-	struct pipe_screen *pscreen = &nscreen->base.base;
-	struct pipe_buffer *buf = nscreen->strm_vbuf[i];
-	struct pipe_vertex_buffer *vb = &nv50->vtxbuf[i];
-	uint8_t *src;
-	unsigned size = align(vb->buffer->size, 4096);
-
-	if (buf && buf->size < size)
-		pipe_buffer_reference(&nscreen->strm_vbuf[i], NULL);
-
-	if (!nscreen->strm_vbuf[i]) {
-		nscreen->strm_vbuf[i] = pipe_buffer_create(
-			pscreen, 0, PIPE_BUFFER_USAGE_VERTEX, size);
-		buf = nscreen->strm_vbuf[i];
-	}
-
-	src = pipe_buffer_map(pscreen, vb->buffer, PIPE_BUFFER_USAGE_CPU_READ);
-	if (!src)
-		return FALSE;
-	src += vb->buffer_offset;
-
-	size = (vb->max_index + 1) * vb->stride + 16; /* + 16 is for stride 0 */
-	if (vb->buffer_offset + size > vb->buffer->size)
-		size = vb->buffer->size - vb->buffer_offset;
-
-	pipe_buffer_write(pscreen, buf, vb->buffer_offset, size, src);
-	pipe_buffer_unmap(pscreen, vb->buffer);
-
-	vb->buffer = buf; /* don't pipe_reference, this is a private copy */
-	return TRUE;
-}
-
-static void
-nv50_upload_user_vbufs(struct nv50_context *nv50)
-{
-	unsigned i;
-
-	if (nv50->vbo_fifo)
-		nv50->dirty |= NV50_NEW_ARRAYS;
-	if (!(nv50->dirty & NV50_NEW_ARRAYS))
-		return;
-
-	for (i = 0; i < nv50->vtxbuf_nr; ++i) {
-		if (nv50->vtxbuf[i].buffer->usage & PIPE_BUFFER_USAGE_VERTEX)
-			continue;
-		nv50_upload_vtxbuf(nv50, i);
-	}
-}
-
-static void
-nv50_set_static_vtxattr(struct nv50_context *nv50, unsigned i, void *data)
-{
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nouveau_channel *chan = tesla->channel;
-	float v[4];
-
-	util_format_read_4f(nv50->vtxelt[i].src_format,
-			    v, 0, data, 0, 0, 0, 1, 1);
-
-	switch (nv50->vtxelt[i].nr_components) {
-	case 4:
-		BEGIN_RING(chan, tesla, NV50TCL_VTX_ATTR_4F_X(i), 4);
-		OUT_RINGf (chan, v[0]);
-		OUT_RINGf (chan, v[1]);
-		OUT_RINGf (chan, v[2]);
-		OUT_RINGf (chan, v[3]);
-		break;
-	case 3:
-		BEGIN_RING(chan, tesla, NV50TCL_VTX_ATTR_3F_X(i), 3);
-		OUT_RINGf (chan, v[0]);
-		OUT_RINGf (chan, v[1]);
-		OUT_RINGf (chan, v[2]);
-		break;
-	case 2:
-		BEGIN_RING(chan, tesla, NV50TCL_VTX_ATTR_2F_X(i), 2);
-		OUT_RINGf (chan, v[0]);
-		OUT_RINGf (chan, v[1]);
-		break;
-	case 1:
-		BEGIN_RING(chan, tesla, NV50TCL_VTX_ATTR_1F(i), 1);
-		OUT_RINGf (chan, v[0]);
-		break;
-	default:
-		assert(0);
-		break;
-	}
-}
-
-static unsigned
-init_per_instance_arrays_immd(struct nv50_context *nv50,
-			      unsigned startInstance,
-			      unsigned pos[16], unsigned step[16])
-{
-	struct nouveau_bo *bo;
-	unsigned i, b, count = 0;
-
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
-			continue;
-		++count;
-		b = nv50->vtxelt[i].vertex_buffer_index;
-
-		pos[i] = nv50->vtxelt[i].src_offset +
-			nv50->vtxbuf[b].buffer_offset +
-			startInstance * nv50->vtxbuf[b].stride;
-		step[i] = startInstance % nv50->vtxelt[i].instance_divisor;
-
-		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
-		if (!bo->map)
-			nouveau_bo_map(bo, NOUVEAU_BO_RD);
-
-		nv50_set_static_vtxattr(nv50, i, (uint8_t *)bo->map + pos[i]);
-	}
-
-	return count;
-}
-
-static unsigned
-init_per_instance_arrays(struct nv50_context *nv50,
-			 unsigned startInstance,
-			 unsigned pos[16], unsigned step[16])
-{
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nouveau_channel *chan = tesla->channel;
+struct instance {
 	struct nouveau_bo *bo;
-	struct nouveau_stateobj *so;
-	unsigned i, b, count = 0;
-	const uint32_t rl = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
-
-	if (nv50->vbo_fifo)
-		return init_per_instance_arrays_immd(nv50, startInstance,
-						     pos, step);
-
-	so = so_new(nv50->vtxelt_nr, nv50->vtxelt_nr * 2, nv50->vtxelt_nr * 2);
-
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
-			continue;
-		++count;
-		b = nv50->vtxelt[i].vertex_buffer_index;
-
-		pos[i] = nv50->vtxelt[i].src_offset +
-			nv50->vtxbuf[b].buffer_offset +
-			startInstance * nv50->vtxbuf[b].stride;
-
-		if (!startInstance) {
-			step[i] = 0;
-			continue;
-		}
-		step[i] = startInstance % nv50->vtxelt[i].instance_divisor;
-
-		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
-
-		so_method(so, tesla, NV50TCL_VERTEX_ARRAY_START_HIGH(i), 2);
-		so_reloc (so, bo, pos[i], rl | NOUVEAU_BO_HIGH, 0, 0);
-		so_reloc (so, bo, pos[i], rl | NOUVEAU_BO_LOW, 0, 0);
-	}
-
-	if (count && startInstance) {
-		so_ref (so, &nv50->state.instbuf); /* for flush notify */
-		so_emit(chan, nv50->state.instbuf);
-	}
-	so_ref (NULL, &so);
-
-	return count;
-}
+	unsigned delta;
+	unsigned stride;
+	unsigned step;
+	unsigned divisor;
+};
 
 static void
-step_per_instance_arrays_immd(struct nv50_context *nv50,
-			      unsigned pos[16], unsigned step[16])
+instance_init(struct nv50_context *nv50, struct instance *a, unsigned first)
 {
-	struct nouveau_bo *bo;
-	unsigned i, b;
+	int i;
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
-			continue;
-		if (++step[i] != nv50->vtxelt[i].instance_divisor)
-			continue;
-		b = nv50->vtxelt[i].vertex_buffer_index;
-		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
+	for (i = 0; i < nv50->vtxelt->num_elements; i++) {
+		struct pipe_vertex_element *ve = &nv50->vtxelt->pipe[i];
+		struct pipe_vertex_buffer *vb;
 
-		step[i] = 0;
-		pos[i] += nv50->vtxbuf[b].stride;
+		a[i].divisor = ve->instance_divisor;
+		if (a[i].divisor) {
+			vb = &nv50->vtxbuf[ve->vertex_buffer_index];
 
-		nv50_set_static_vtxattr(nv50, i, (uint8_t *)bo->map + pos[i]);
+			a[i].bo = nouveau_bo(vb->buffer);
+			a[i].stride = vb->stride;
+			a[i].step = first % a[i].divisor;
+			a[i].delta = vb->buffer_offset + ve->src_offset +
+				     (first * a[i].stride);
+		}
 	}
 }
 
 static void
-step_per_instance_arrays(struct nv50_context *nv50,
-			 unsigned pos[16], unsigned step[16])
+instance_step(struct nv50_context *nv50, struct instance *a)
 {
+	struct nouveau_channel *chan = nv50->screen->tesla->channel;
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nouveau_channel *chan = tesla->channel;
-	struct nouveau_bo *bo;
-	struct nouveau_stateobj *so;
-	unsigned i, b;
-	const uint32_t rl = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
-
-	if (nv50->vbo_fifo) {
-		step_per_instance_arrays_immd(nv50, pos, step);
-		return;
-	}
-
-	so = so_new(nv50->vtxelt_nr, nv50->vtxelt_nr * 2, nv50->vtxelt_nr * 2);
+	int i;
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
+	for (i = 0; i < nv50->vtxelt->num_elements; i++) {
+		if (!a[i].divisor)
 			continue;
-		b = nv50->vtxelt[i].vertex_buffer_index;
 
-		if (++step[i] == nv50->vtxelt[i].instance_divisor) {
-			step[i] = 0;
-			pos[i] += nv50->vtxbuf[b].stride;
+		BEGIN_RING(chan, tesla,
+			   NV50TCL_VERTEX_ARRAY_START_HIGH(i), 2);
+		OUT_RELOCh(chan, a[i].bo, a[i].delta, NOUVEAU_BO_RD |
+			   NOUVEAU_BO_VRAM | NOUVEAU_BO_GART);
+		OUT_RELOCl(chan, a[i].bo, a[i].delta, NOUVEAU_BO_RD |
+			   NOUVEAU_BO_VRAM | NOUVEAU_BO_GART);
+		if (++a[i].step == a[i].divisor) {
+			a[i].step = 0;
+			a[i].delta += a[i].stride;
 		}
-
-		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
-
-		so_method(so, tesla, NV50TCL_VERTEX_ARRAY_START_HIGH(i), 2);
-		so_reloc (so, bo, pos[i], rl | NOUVEAU_BO_HIGH, 0, 0);
-		so_reloc (so, bo, pos[i], rl | NOUVEAU_BO_LOW, 0, 0);
 	}
-
-	so_ref (so, &nv50->state.instbuf); /* for flush notify */
-	so_ref (NULL, &so);
-
-	so_emit(chan, nv50->state.instbuf);
-}
-
-static INLINE void
-nv50_unmap_vbufs(struct nv50_context *nv50)
-{
-        unsigned i;
-
-        for (i = 0; i < nv50->vtxbuf_nr; ++i)
-                if (nouveau_bo(nv50->vtxbuf[i].buffer)->map)
-                        nouveau_bo_unmap(nouveau_bo(nv50->vtxbuf[i].buffer));
 }
 
 void
@@ -415,198 +180,207 @@ nv50_draw_arrays_instanced(struct pipe_context *pipe,
 	struct nv50_context *nv50 = nv50_context(pipe);
 	struct nouveau_channel *chan = nv50->screen->tesla->channel;
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	unsigned i, nz_divisors;
-	unsigned step[16], pos[16];
-
-	if (!NV50_USING_LOATHED_EDGEFLAG(nv50))
-		nv50_upload_user_vbufs(nv50);
+	struct instance a[16];
+	unsigned prim = nv50_prim(mode);
 
-	nv50_state_validate(nv50);
+	instance_init(nv50, a, startInstance);
+	if (!nv50_state_validate(nv50, 10 + 16*3))
+		return;
 
-	nz_divisors = init_per_instance_arrays(nv50, startInstance, pos, step);
+	if (nv50->vbo_fifo) {
+		nv50_push_elements_instanced(pipe, NULL, 0, mode, start,
+					     count, startInstance,
+					     instanceCount);
+		return;
+	}
 
 	BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
 	OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
 	OUT_RING  (chan, startInstance);
+	while (instanceCount--) {
+		if (AVAIL_RING(chan) < (7 + 16*3)) {
+			FIRE_RING(chan);
+			if (!nv50_state_validate(nv50, 7 + 16*3)) {
+				assert(0);
+				return;
+			}
+		}
+		instance_step(nv50, a);
 
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-	OUT_RING  (chan, nv50_prim(mode));
-
-	if (nv50->vbo_fifo)
-		nv50_push_arrays(nv50, start, count);
-	else {
+		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
+		OUT_RING  (chan, prim);
 		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
 		OUT_RING  (chan, start);
 		OUT_RING  (chan, count);
-	}
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-	OUT_RING  (chan, 0);
-
-	for (i = 1; i < instanceCount; i++) {
-		if (nz_divisors) /* any non-zero array divisors ? */
-			step_per_instance_arrays(nv50, pos, step);
-
-		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-		OUT_RING  (chan, nv50_prim(mode) | (1 << 28));
-
-		if (nv50->vbo_fifo)
-			nv50_push_arrays(nv50, start, count);
-		else {
-			BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
-			OUT_RING  (chan, start);
-			OUT_RING  (chan, count);
-		}
 		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
 		OUT_RING  (chan, 0);
-	}
-	nv50_unmap_vbufs(nv50);
 
-	so_ref(NULL, &nv50->state.instbuf);
+		prim |= (1 << 28);
+	}
 }
 
 void
 nv50_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
 		 unsigned count)
 {
-	struct nv50_context *nv50 = nv50_context(pipe);
-	struct nouveau_channel *chan = nv50->screen->tesla->channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	boolean ret;
-
-	nv50_state_validate(nv50);
-
-	BEGIN_RING(chan, tesla, 0x142c, 1);
-	OUT_RING  (chan, 0);
-	BEGIN_RING(chan, tesla, 0x142c, 1);
-	OUT_RING  (chan, 0);
-
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-	OUT_RING  (chan, nv50_prim(mode));
-
-	if (nv50->vbo_fifo)
-		ret = nv50_push_arrays(nv50, start, count);
-	else {
-		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
-		OUT_RING  (chan, start);
-		OUT_RING  (chan, count);
-		ret = TRUE;
-	}
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-	OUT_RING  (chan, 0);
-
-	nv50_unmap_vbufs(nv50);
-
-        /* XXX: not sure what to do if ret != TRUE: flush and retry?
-         */
-        assert(ret);
+	nv50_draw_arrays_instanced(pipe, mode, start, count, 0, 1);
 }
 
-static INLINE boolean
-nv50_draw_elements_inline_u08(struct nv50_context *nv50, uint8_t *map,
-			      unsigned start, unsigned count)
-{
-	struct nouveau_channel *chan = nv50->screen->tesla->channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-
-	map += start;
+struct inline_ctx {
+	struct nv50_context *nv50;
+	void *map;
+};
 
-	if (nv50->vbo_fifo)
-		return nv50_push_elements_u08(nv50, map, count);
+static void
+inline_elt08(void *priv, unsigned start, unsigned count)
+{
+	struct inline_ctx *ctx = priv;
+	struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
+	struct nouveau_channel *chan = tesla->channel;
+	uint8_t *map = (uint8_t *)ctx->map + start;
 
 	if (count & 1) {
 		BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32, 1);
 		OUT_RING  (chan, map[0]);
 		map++;
-		count--;
+		count &= ~1;
 	}
 
-	while (count) {
-		unsigned nr = count > 2046 ? 2046 : count;
-		int i;
-
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U16, nr >> 1);
-		for (i = 0; i < nr; i += 2)
-			OUT_RING  (chan, (map[i + 1] << 16) | map[i]);
+	count >>= 1;
+	if (!count)
+		return;
 
-		count -= nr;
-		map += nr;
+	BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U16, count);
+	while (count--) {
+		OUT_RING(chan, (map[1] << 16) | map[0]);
+		map += 2;
 	}
-	return TRUE;
 }
 
-static INLINE boolean
-nv50_draw_elements_inline_u16(struct nv50_context *nv50, uint16_t *map,
-			      unsigned start, unsigned count)
+static void
+inline_elt16(void *priv, unsigned start, unsigned count)
 {
-	struct nouveau_channel *chan = nv50->screen->tesla->channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-
-	map += start;
-
-	if (nv50->vbo_fifo)
-		return nv50_push_elements_u16(nv50, map, count);
+	struct inline_ctx *ctx = priv;
+	struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
+	struct nouveau_channel *chan = tesla->channel;
+	uint16_t *map = (uint16_t *)ctx->map + start;
 
 	if (count & 1) {
 		BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32, 1);
 		OUT_RING  (chan, map[0]);
+		count &= ~1;
 		map++;
-		count--;
 	}
 
-	while (count) {
-		unsigned nr = count > 2046 ? 2046 : count;
-		int i;
-
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U16, nr >> 1);
-		for (i = 0; i < nr; i += 2)
-			OUT_RING  (chan, (map[i + 1] << 16) | map[i]);
+	count >>= 1;
+	if (!count)
+		return;
 
-		count -= nr;
-		map += nr;
+	BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U16, count);
+	while (count--) {
+		OUT_RING(chan, (map[1] << 16) | map[0]);
+		map += 2;
 	}
-	return TRUE;
 }
 
-static INLINE boolean
-nv50_draw_elements_inline_u32(struct nv50_context *nv50, uint32_t *map,
-			      unsigned start, unsigned count)
+static void
+inline_elt32(void *priv, unsigned start, unsigned count)
+{
+	struct inline_ctx *ctx = priv;
+	struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
+	struct nouveau_channel *chan = tesla->channel;
+
+	BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U32, count);
+	OUT_RINGp    (chan, (uint32_t *)ctx->map + start, count);
+}
+
+static void
+inline_edgeflag(void *priv, boolean enabled)
 {
+	struct inline_ctx *ctx = priv;
+	struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
+	struct nouveau_channel *chan = tesla->channel;
+
+	BEGIN_RING(chan, tesla, NV50TCL_EDGEFLAG_ENABLE, 1);
+	OUT_RING  (chan, enabled ? 1 : 0);
+}
+
+static void
+nv50_draw_elements_inline(struct pipe_context *pipe,
+			  struct pipe_buffer *indexBuffer, unsigned indexSize,
+			  unsigned mode, unsigned start, unsigned count,
+			  unsigned startInstance, unsigned instanceCount)
+{
+	struct pipe_screen *pscreen = pipe->screen;
+	struct nv50_context *nv50 = nv50_context(pipe);
 	struct nouveau_channel *chan = nv50->screen->tesla->channel;
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct instance a[16];
+	struct inline_ctx ctx;
+	struct u_split_prim s;
+	boolean nzi = FALSE;
+	unsigned overhead;
+
+	overhead = 16*3; /* potential instance adjustments */
+	overhead += 4; /* Begin()/End() */
+	overhead += 4; /* potential edgeflag disable/reenable */
+	overhead += 3; /* potentially 3 VTX_ELT_U16/U32 packet headers */
+
+	s.priv = &ctx;
+	if (indexSize == 1)
+		s.emit = inline_elt08;
+	else
+	if (indexSize == 2)
+		s.emit = inline_elt16;
+	else
+		s.emit = inline_elt32;
+	s.edge = inline_edgeflag;
+
+	ctx.nv50 = nv50;
+	ctx.map = pipe_buffer_map(pscreen, indexBuffer, PIPE_BUFFER_USAGE_CPU_READ);
+	assert(ctx.map);
+	if (!ctx.map)
+		return;
 
-	map += start;
+	instance_init(nv50, a, startInstance);
+	if (!nv50_state_validate(nv50, overhead + 6 + 3))
+		return;
 
-	if (nv50->vbo_fifo)
-		return nv50_push_elements_u32(nv50, map, count);
+	BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
+	OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
+	OUT_RING  (chan, startInstance);
+	while (instanceCount--) {
+		unsigned max_verts;
+		boolean done;
+
+		u_split_prim_init(&s, mode, start, count);
+		do {
+			if (AVAIL_RING(chan) < (overhead + 6)) {
+				FIRE_RING(chan);
+				if (!nv50_state_validate(nv50, (overhead + 6))) {
+					assert(0);
+					return;
+				}
+			}
 
-	while (count) {
-		unsigned nr = count > 2047 ? 2047 : count;
+			max_verts = AVAIL_RING(chan) - overhead;
+			if (max_verts > 2047)
+				max_verts = 2047;
+			if (indexSize != 4)
+				max_verts <<= 1;
+			instance_step(nv50, a);
 
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U32, nr);
-		OUT_RINGp (chan, map, nr);
+			BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
+			OUT_RING  (chan, nv50_prim(s.mode) | (nzi ? (1<<28) : 0));
+			done = u_split_prim_next(&s, max_verts);
+			BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
+			OUT_RING  (chan, 0);
+		} while (!done);
 
-		count -= nr;
-		map += nr;
+		nzi = TRUE;
 	}
-	return TRUE;
-}
 
-static INLINE void
-nv50_draw_elements_inline(struct nv50_context *nv50,
-			  void *map, unsigned indexSize,
-			  unsigned start, unsigned count)
-{
-	switch (indexSize) {
-	case 1:
-		nv50_draw_elements_inline_u08(nv50, map, start, count);
-		break;
-	case 2:
-		nv50_draw_elements_inline_u16(nv50, map, start, count);
-		break;
-	case 4:
-		nv50_draw_elements_inline_u32(nv50, map, start, count);
-		break;
-	}
+	pipe_buffer_unmap(pscreen, indexBuffer);
 }
 
 void
@@ -617,49 +391,68 @@ nv50_draw_elements_instanced(struct pipe_context *pipe,
 			     unsigned startInstance, unsigned instanceCount)
 {
 	struct nv50_context *nv50 = nv50_context(pipe);
+	struct nouveau_channel *chan = nv50->screen->tesla->channel;
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nouveau_channel *chan = tesla->channel;
-	struct pipe_screen *pscreen = pipe->screen;
-	void *map;
-	unsigned i, nz_divisors;
-	unsigned step[16], pos[16];
+	struct instance a[16];
+	unsigned prim = nv50_prim(mode);
 
-	map = pipe_buffer_map(pscreen, indexBuffer, PIPE_BUFFER_USAGE_CPU_READ);
-
-	if (!NV50_USING_LOATHED_EDGEFLAG(nv50))
-		nv50_upload_user_vbufs(nv50);
-
-	nv50_state_validate(nv50);
+	instance_init(nv50, a, startInstance);
+	if (!nv50_state_validate(nv50, 13 + 16*3))
+		return;
 
-	nz_divisors = init_per_instance_arrays(nv50, startInstance, pos, step);
+	if (nv50->vbo_fifo) {
+		nv50_push_elements_instanced(pipe, indexBuffer, indexSize,
+					     mode, start, count, startInstance,
+					     instanceCount);
+		return;
+	} else
+	if (!(indexBuffer->usage & PIPE_BUFFER_USAGE_INDEX) || indexSize == 1) {
+		nv50_draw_elements_inline(pipe, indexBuffer, indexSize,
+					  mode, start, count, startInstance,
+					  instanceCount);
+		return;
+	}
 
 	BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
 	OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
 	OUT_RING  (chan, startInstance);
-
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-	OUT_RING  (chan, nv50_prim(mode));
-
-	nv50_draw_elements_inline(nv50, map, indexSize, start, count);
-
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-	OUT_RING  (chan, 0);
-
-	for (i = 1; i < instanceCount; ++i) {
-		if (nz_divisors) /* any non-zero array divisors ? */
-			step_per_instance_arrays(nv50, pos, step);
+	while (instanceCount--) {
+		if (AVAIL_RING(chan) < (7 + 16*3)) {
+			FIRE_RING(chan);
+			if (!nv50_state_validate(nv50, 10 + 16*3)) {
+				assert(0);
+				return;
+			}
+		}
+		instance_step(nv50, a);
 
 		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-		OUT_RING  (chan, nv50_prim(mode) | (1 << 28));
-
-		nv50_draw_elements_inline(nv50, map, indexSize, start, count);
-
+		OUT_RING  (chan, prim);
+		if (indexSize == 4) {
+			BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32 | 0x30000, 0);
+			OUT_RING  (chan, count);
+			nouveau_pushbuf_submit(chan, nouveau_bo(indexBuffer),
+					       start << 2, count << 2);
+		} else
+		if (indexSize == 2) {
+			unsigned vb_start = (start & ~1);
+			unsigned vb_end = (start + count + 1) & ~1;
+			unsigned dwords = (vb_end - vb_start) >> 1;
+
+			BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16_SETUP, 1);
+			OUT_RING  (chan, ((start & 1) << 31) | count);
+			BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16 | 0x30000, 0);
+			OUT_RING  (chan, dwords);
+			nouveau_pushbuf_submit(chan, nouveau_bo(indexBuffer),
+					       vb_start << 1, dwords << 2);
+			BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16_SETUP, 1);
+			OUT_RING  (chan, 0);
+		}
 		BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
 		OUT_RING  (chan, 0);
-	}
-	nv50_unmap_vbufs(nv50);
 
-	so_ref(NULL, &nv50->state.instbuf);
+		prim |= (1 << 28);
+	}
 }
 
 void
@@ -667,51 +460,8 @@ nv50_draw_elements(struct pipe_context *pipe,
 		   struct pipe_buffer *indexBuffer, unsigned indexSize,
 		   unsigned mode, unsigned start, unsigned count)
 {
-	struct nv50_context *nv50 = nv50_context(pipe);
-	struct nouveau_channel *chan = nv50->screen->tesla->channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct pipe_screen *pscreen = pipe->screen;
-	void *map;
-	
-	nv50_state_validate(nv50);
-
-	BEGIN_RING(chan, tesla, 0x142c, 1);
-	OUT_RING  (chan, 0);
-	BEGIN_RING(chan, tesla, 0x142c, 1);
-	OUT_RING  (chan, 0);
-
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-	OUT_RING  (chan, nv50_prim(mode));
-
-	if (!nv50->vbo_fifo && indexSize == 4) {
-		BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32 | 0x30000, 0);
-		OUT_RING  (chan, count);
-		nouveau_pushbuf_submit(chan, nouveau_bo(indexBuffer),
-				       start << 2, count << 2);
-	} else
-	if (!nv50->vbo_fifo && indexSize == 2) {
-		unsigned vb_start = (start & ~1);
-		unsigned vb_end = (start + count + 1) & ~1;
-		unsigned dwords = (vb_end - vb_start) >> 1;
-
-		BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16_SETUP, 1);
-		OUT_RING  (chan, ((start & 1) << 31) | count);
-		BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16 | 0x30000, 0);
-		OUT_RING  (chan, dwords);
-		nouveau_pushbuf_submit(chan, nouveau_bo(indexBuffer),
-				       vb_start << 1, dwords << 2);
-		BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16_SETUP, 1);
-		OUT_RING  (chan, 0);
-	} else {
-		map = pipe_buffer_map(pscreen, indexBuffer,
-				      PIPE_BUFFER_USAGE_CPU_READ);
-		nv50_draw_elements_inline(nv50, map, indexSize, start, count);
-		nv50_unmap_vbufs(nv50);
-		pipe_buffer_unmap(pscreen, indexBuffer);
-	}
-
-	BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-	OUT_RING  (chan, 0);
+	nv50_draw_elements_instanced(pipe, indexBuffer, indexSize,
+				     mode, start, count, 0, 1);
 }
 
 static INLINE boolean
@@ -726,6 +476,7 @@ nv50_vbo_static_attrib(struct nv50_context *nv50, unsigned attrib,
 	struct nouveau_bo *bo = nouveau_bo(vb->buffer);
 	float v[4];
 	int ret;
+	unsigned nr_components = util_format_get_nr_components(ve->src_format);
 
 	ret = nouveau_bo_map(bo, NOUVEAU_BO_RD);
 	if (ret)
@@ -736,9 +487,10 @@ nv50_vbo_static_attrib(struct nv50_context *nv50, unsigned attrib,
 			    0, 0, 1, 1);
 	so = *pso;
 	if (!so)
-		*pso = so = so_new(nv50->vtxelt_nr, nv50->vtxelt_nr * 4, 0);
+		*pso = so = so_new(nv50->vtxelt->num_elements,
+				   nv50->vtxelt->num_elements * 4, 0);
 
-	switch (ve->nr_components) {
+	switch (nr_components) {
 	case 4:
 		so_method(so, tesla, NV50TCL_VTX_ATTR_4F_X(attrib), 4);
 		so_data  (so, fui(v[0]));
@@ -775,6 +527,18 @@ nv50_vbo_static_attrib(struct nv50_context *nv50, unsigned attrib,
 }
 
 void
+nv50_vtxelt_construct(struct nv50_vtxelt_stateobj *cso)
+{
+	unsigned i;
+
+	for (i = 0; i < cso->num_elements; ++i) {
+		struct pipe_vertex_element *ve = &cso->pipe[i];
+
+		cso->hw[i] = nv50_vbo_vtxelt_to_hw(ve);
+	}
+}
+
+struct nouveau_stateobj *
 nv50_vbo_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -783,30 +547,31 @@ nv50_vbo_validate(struct nv50_context *nv50)
 
 	/* don't validate if Gallium took away our buffers */
 	if (nv50->vtxbuf_nr == 0)
-		return;
-	nv50->vbo_fifo = 0;
+		return NULL;
+
+	if (nv50->screen->force_push ||
+	    nv50->vertprog->cfg.edgeflag_in < 16)
+		nv50->vbo_fifo = 0xffff;
 
-	for (i = 0; i < nv50->vtxbuf_nr; ++i)
+	for (i = 0; i < nv50->vtxbuf_nr; i++) {
 		if (nv50->vtxbuf[i].stride &&
 		    !(nv50->vtxbuf[i].buffer->usage & PIPE_BUFFER_USAGE_VERTEX))
 			nv50->vbo_fifo = 0xffff;
+	}
 
-	if (NV50_USING_LOATHED_EDGEFLAG(nv50))
-		nv50->vbo_fifo = 0xffff; /* vertprog can't set edgeflag */
-
-	n_ve = MAX2(nv50->vtxelt_nr, nv50->state.vtxelt_nr);
+	n_ve = MAX2(nv50->vtxelt->num_elements, nv50->state.vtxelt_nr);
 
 	vtxattr = NULL;
-	vtxbuf = so_new(n_ve * 2, n_ve * 5, nv50->vtxelt_nr * 4);
+	vtxbuf = so_new(n_ve * 2, n_ve * 5, nv50->vtxelt->num_elements * 4);
 	vtxfmt = so_new(1, n_ve, 0);
 	so_method(vtxfmt, tesla, NV50TCL_VERTEX_ARRAY_ATTRIB(0), n_ve);
 
-	for (i = 0; i < nv50->vtxelt_nr; i++) {
-		struct pipe_vertex_element *ve = &nv50->vtxelt[i];
+	for (i = 0; i < nv50->vtxelt->num_elements; i++) {
+		struct pipe_vertex_element *ve = &nv50->vtxelt->pipe[i];
 		struct pipe_vertex_buffer *vb =
 			&nv50->vtxbuf[ve->vertex_buffer_index];
 		struct nouveau_bo *bo = nouveau_bo(vb->buffer);
-		uint32_t hw = nv50_vbo_vtxelt_to_hw(ve);
+		uint32_t hw = nv50->vtxelt->hw[i];
 
 		if (!vb->stride &&
 		    nv50_vbo_static_attrib(nv50, i, &vtxattr, ve, vb)) {
@@ -821,13 +586,13 @@ nv50_vbo_validate(struct nv50_context *nv50)
 		}
 
 		if (nv50->vbo_fifo) {
-			so_data  (vtxfmt, hw |
-				  (ve->instance_divisor ? (1 << 4) : i));
+			so_data  (vtxfmt, hw | (ve->instance_divisor ? (1 << 4) : i));
 			so_method(vtxbuf, tesla,
 				  NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
 			so_data  (vtxbuf, 0);
 			continue;
 		}
+
 		so_data(vtxfmt, hw | i);
 
 		so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_FORMAT(i), 3);
@@ -855,355 +620,13 @@ nv50_vbo_validate(struct nv50_context *nv50)
 		so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
 		so_data  (vtxbuf, 0);
 	}
-	nv50->state.vtxelt_nr = nv50->vtxelt_nr;
+	nv50->state.vtxelt_nr = nv50->vtxelt->num_elements;
 
-	so_ref (vtxfmt, &nv50->state.vtxfmt);
 	so_ref (vtxbuf, &nv50->state.vtxbuf);
 	so_ref (vtxattr, &nv50->state.vtxattr);
 	so_ref (NULL, &vtxbuf);
-	so_ref (NULL, &vtxfmt);
 	so_ref (NULL, &vtxattr);
+	return vtxfmt;
 }
 
-typedef void (*pfn_push)(struct nouveau_channel *, void *);
-
-struct nv50_vbo_emitctx
-{
-	pfn_push push[16];
-	uint8_t *map[16];
-	unsigned stride[16];
-	unsigned nr_ve;
-	unsigned vtx_dwords;
-	unsigned vtx_max;
-
-	float edgeflag;
-	unsigned ve_edgeflag;
-};
-
-static INLINE void
-emit_vtx_next(struct nouveau_channel *chan, struct nv50_vbo_emitctx *emit)
-{
-	unsigned i;
-
-	for (i = 0; i < emit->nr_ve; ++i) {
-		emit->push[i](chan, emit->map[i]);
-		emit->map[i] += emit->stride[i];
-	}
-}
-
-static INLINE void
-emit_vtx(struct nouveau_channel *chan, struct nv50_vbo_emitctx *emit,
-	 uint32_t vi)
-{
-	unsigned i;
-
-	for (i = 0; i < emit->nr_ve; ++i)
-		emit->push[i](chan, emit->map[i] + emit->stride[i] * vi);
-}
-
-static INLINE boolean
-nv50_map_vbufs(struct nv50_context *nv50)
-{
-	int i;
-
-	for (i = 0; i < nv50->vtxbuf_nr; ++i) {
-		struct pipe_vertex_buffer *vb = &nv50->vtxbuf[i];
-		unsigned size = vb->stride * (vb->max_index + 1) + 16;
-
-		if (nouveau_bo(vb->buffer)->map)
-			continue;
-
-		size = vb->stride * (vb->max_index + 1) + 16;
-		size = MIN2(size, vb->buffer->size);
-		if (!size)
-			size = vb->buffer->size;
-
-		if (nouveau_bo_map_range(nouveau_bo(vb->buffer),
-					 0, size, NOUVEAU_BO_RD))
-			break;
-	}
-
-	if (i == nv50->vtxbuf_nr)
-		return TRUE;
-	for (; i >= 0; --i)
-		nouveau_bo_unmap(nouveau_bo(nv50->vtxbuf[i].buffer));
-	return FALSE;
-}
-
-static void
-emit_b32_1(struct nouveau_channel *chan, void *data)
-{
-	uint32_t *v = data;
-
-	OUT_RING(chan, v[0]);
-}
-
-static void
-emit_b32_2(struct nouveau_channel *chan, void *data)
-{
-	uint32_t *v = data;
-
-	OUT_RING(chan, v[0]);
-	OUT_RING(chan, v[1]);
-}
-
-static void
-emit_b32_3(struct nouveau_channel *chan, void *data)
-{
-	uint32_t *v = data;
-
-	OUT_RING(chan, v[0]);
-	OUT_RING(chan, v[1]);
-	OUT_RING(chan, v[2]);
-}
-
-static void
-emit_b32_4(struct nouveau_channel *chan, void *data)
-{
-	uint32_t *v = data;
-
-	OUT_RING(chan, v[0]);
-	OUT_RING(chan, v[1]);
-	OUT_RING(chan, v[2]);
-	OUT_RING(chan, v[3]);
-}
-
-static void
-emit_b16_1(struct nouveau_channel *chan, void *data)
-{
-	uint16_t *v = data;
-
-	OUT_RING(chan, v[0]);
-}
-
-static void
-emit_b16_3(struct nouveau_channel *chan, void *data)
-{
-	uint16_t *v = data;
-
-	OUT_RING(chan, (v[1] << 16) | v[0]);
-	OUT_RING(chan, v[2]);
-}
-
-static void
-emit_b08_1(struct nouveau_channel *chan, void *data)
-{
-	uint8_t *v = data;
-
-	OUT_RING(chan, v[0]);
-}
-
-static void
-emit_b08_3(struct nouveau_channel *chan, void *data)
-{
-	uint8_t *v = data;
-
-	OUT_RING(chan, (v[2] << 16) | (v[1] << 8) | v[0]);
-}
-
-static boolean
-emit_prepare(struct nv50_context *nv50, struct nv50_vbo_emitctx *emit,
-	     unsigned start)
-{
-	unsigned i;
-
-	if (nv50_map_vbufs(nv50) == FALSE)
-		return FALSE;
-
-	emit->ve_edgeflag = nv50->vertprog->cfg.edgeflag_in;
-
-	emit->edgeflag = 0.5f;
-	emit->nr_ve = 0;
-	emit->vtx_dwords = 0;
-
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		struct pipe_vertex_element *ve;
-		struct pipe_vertex_buffer *vb;
-		unsigned n, size;
-		const struct util_format_description *desc;
-
-		ve = &nv50->vtxelt[i];
-		vb = &nv50->vtxbuf[ve->vertex_buffer_index];
-		if (!(nv50->vbo_fifo & (1 << i)) || ve->instance_divisor)
-			continue;
-		n = emit->nr_ve++;
-
-		emit->stride[n] = vb->stride;
-		emit->map[n] = (uint8_t *)nouveau_bo(vb->buffer)->map +
-			vb->buffer_offset +
-			(start * vb->stride + ve->src_offset);
-
-		desc = util_format_description(ve->src_format);
-		assert(desc);
-
-		size = util_format_get_component_bits(
-			ve->src_format, UTIL_FORMAT_COLORSPACE_RGB, 0);
-
-		assert(ve->nr_components > 0 && ve->nr_components <= 4);
-
-		/* It shouldn't be necessary to push the implicit 1s
-		 * for case 3 and size 8 cases 1, 2, 3.
-		 */
-		switch (size) {
-		default:
-			NOUVEAU_ERR("unsupported vtxelt size: %u\n", size);
-			return FALSE;
-		case 32:
-			switch (ve->nr_components) {
-			case 1: emit->push[n] = emit_b32_1; break;
-			case 2: emit->push[n] = emit_b32_2; break;
-			case 3: emit->push[n] = emit_b32_3; break;
-			case 4: emit->push[n] = emit_b32_4; break;
-			}
-			emit->vtx_dwords += ve->nr_components;
-			break;
-		case 16:
-			switch (ve->nr_components) {
-			case 1: emit->push[n] = emit_b16_1; break;
-			case 2: emit->push[n] = emit_b32_1; break;
-			case 3: emit->push[n] = emit_b16_3; break;
-			case 4: emit->push[n] = emit_b32_2; break;
-			}
-			emit->vtx_dwords += (ve->nr_components + 1) >> 1;
-			break;
-		case 8:
-			switch (ve->nr_components) {
-			case 1: emit->push[n] = emit_b08_1; break;
-			case 2: emit->push[n] = emit_b16_1; break;
-			case 3: emit->push[n] = emit_b08_3; break;
-			case 4: emit->push[n] = emit_b32_1; break;
-			}
-			emit->vtx_dwords += 1;
-			break;
-		}
-	}
-
-	emit->vtx_max = 512 / emit->vtx_dwords;
-	if (emit->ve_edgeflag < 16)
-		emit->vtx_max = 1;
-
-	return TRUE;
-}
-
-static INLINE void
-set_edgeflag(struct nouveau_channel *chan,
-	     struct nouveau_grobj *tesla,
-	     struct nv50_vbo_emitctx *emit, uint32_t index)
-{
-	unsigned i = emit->ve_edgeflag;
-
-	if (i < 16) {
-		float f = *((float *)(emit->map[i] + index * emit->stride[i]));
-
-		if (emit->edgeflag != f) {
-			emit->edgeflag = f;
-
-			BEGIN_RING(chan, tesla, 0x15e4, 1);
-			OUT_RING  (chan, f ? 1 : 0);
-		}
-	}
-}
-
-static boolean
-nv50_push_arrays(struct nv50_context *nv50, unsigned start, unsigned count)
-{
-	struct nouveau_channel *chan = nv50->screen->base.channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nv50_vbo_emitctx emit;
 
-	if (emit_prepare(nv50, &emit, start) == FALSE)
-		return FALSE;
-
-	while (count) {
-		unsigned i, dw, nr = MIN2(count, emit.vtx_max);
-	        dw = nr * emit.vtx_dwords;
-
-		set_edgeflag(chan, tesla, &emit, 0); /* nr will be 1 */
-
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VERTEX_DATA, dw);
-		for (i = 0; i < nr; ++i)
-			emit_vtx_next(chan, &emit);
-
-		count -= nr;
-	}
-
-	return TRUE;
-}
-
-static boolean
-nv50_push_elements_u32(struct nv50_context *nv50, uint32_t *map, unsigned count)
-{
-	struct nouveau_channel *chan = nv50->screen->base.channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nv50_vbo_emitctx emit;
-
-	if (emit_prepare(nv50, &emit, 0) == FALSE)
-		return FALSE;
-
-	while (count) {
-		unsigned i, dw, nr = MIN2(count, emit.vtx_max);
-	        dw = nr * emit.vtx_dwords;
-
-		set_edgeflag(chan, tesla, &emit, *map);
-
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VERTEX_DATA, dw);
-		for (i = 0; i < nr; ++i)
-			emit_vtx(chan, &emit, *map++);
-
-		count -= nr;
-	}
-
-	return TRUE;
-}
-
-static boolean
-nv50_push_elements_u16(struct nv50_context *nv50, uint16_t *map, unsigned count)
-{
-	struct nouveau_channel *chan = nv50->screen->base.channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nv50_vbo_emitctx emit;
-
-	if (emit_prepare(nv50, &emit, 0) == FALSE)
-		return FALSE;
-
-	while (count) {
-		unsigned i, dw, nr = MIN2(count, emit.vtx_max);
-	        dw = nr * emit.vtx_dwords;
-
-		set_edgeflag(chan, tesla, &emit, *map);
-
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VERTEX_DATA, dw);
-		for (i = 0; i < nr; ++i)
-			emit_vtx(chan, &emit, *map++);
-
-		count -= nr;
-	}
-
-	return TRUE;
-}
-
-static boolean
-nv50_push_elements_u08(struct nv50_context *nv50, uint8_t *map, unsigned count)
-{
-	struct nouveau_channel *chan = nv50->screen->base.channel;
-	struct nouveau_grobj *tesla = nv50->screen->tesla;
-	struct nv50_vbo_emitctx emit;
-
-	if (emit_prepare(nv50, &emit, 0) == FALSE)
-		return FALSE;
-
-	while (count) {
-		unsigned i, dw, nr = MIN2(count, emit.vtx_max);
-	        dw = nr * emit.vtx_dwords;
-
-		set_edgeflag(chan, tesla, &emit, *map);
-
-		BEGIN_RING_NI(chan, tesla, NV50TCL_VERTEX_DATA, dw);
-		for (i = 0; i < nr; ++i)
-			emit_vtx(chan, &emit, *map++);
-
-		count -= nr;
-	}
-
-	return TRUE;
-}