Merge branch 'gallium-vertexelementcso'

Conflicts:
	src/gallium/auxiliary/cso_cache/cso_context.c
	src/gallium/auxiliary/util/u_blit.c
	src/gallium/auxiliary/util/u_gen_mipmap.c
	src/gallium/drivers/i915/i915_context.h
	src/gallium/drivers/i965/brw_draw_upload.c
	src/gallium/drivers/r300/r300_blit.c
	src/gallium/drivers/r300/r300_context.h
	src/gallium/drivers/r300/r300_render.c
	src/gallium/drivers/r300/r300_state_derived.c
	src/mesa/state_tracker/st_cb_bitmap.c
	src/mesa/state_tracker/st_cb_clear.c

79 files changed, 1254 insertions, 569 deletions

diff --git a/src/gallium/auxiliary/cso_cache/cso_cache.c b/src/gallium/auxiliary/cso_cache/cso_cache.c
index a6a07e72c2..900c64df4b 100644
--- a/src/gallium/auxiliary/cso_cache/cso_cache.c
+++ b/src/gallium/auxiliary/cso_cache/cso_cache.c
@@ -43,6 +43,7 @@ struct cso_cache {
    struct cso_hash *vs_hash;
    struct cso_hash *rasterizer_hash;
    struct cso_hash *sampler_hash;
+   struct cso_hash *velements_hash;
    int    max_size;
 
    cso_sanitize_callback sanitize_cb;
@@ -108,6 +109,9 @@ static struct cso_hash *_cso_hash_for_type(struct cso_cache *sc, enum cso_cache_
    case CSO_VERTEX_SHADER:
       hash = sc->vs_hash;
       break;
+   case CSO_VELEMENTS:
+      hash = sc->velements_hash;
+      break;
    }
 
    return hash;
@@ -161,6 +165,13 @@ static void delete_vs_state(void *state, void *data)
    FREE(state);
 }
 
+static void delete_velements(void *state, void *data)
+{
+   struct cso_velements *cso = (struct cso_velements *)state;
+   if (cso->delete_state)
+      cso->delete_state(cso->context, cso->data);
+   FREE(state);
+}
 
 static INLINE void delete_cso(void *state, enum cso_cache_type type)
 {
@@ -183,6 +194,9 @@ static INLINE void delete_cso(void *state, enum cso_cache_type type)
    case CSO_VERTEX_SHADER:
       delete_vs_state(state, 0);
       break;
+   case CSO_VELEMENTS:
+      delete_velements(state, 0);
+      break;
    default:
       assert(0);
       FREE(state);
@@ -294,6 +308,7 @@ struct cso_cache *cso_cache_create(void)
    sc->rasterizer_hash    = cso_hash_create();
    sc->fs_hash            = cso_hash_create();
    sc->vs_hash            = cso_hash_create();
+   sc->velements_hash     = cso_hash_create();
    sc->sanitize_cb        = sanitize_cb;
    sc->sanitize_data      = 0;
 
@@ -325,6 +340,9 @@ void cso_for_each_state(struct cso_cache *sc, enum cso_cache_type type,
    case CSO_VERTEX_SHADER:
       hash = sc->vs_hash;
       break;
+   case CSO_VELEMENTS:
+      hash = sc->velements_hash;
+      break;
    }
 
    iter = cso_hash_first_node(hash);
@@ -351,6 +369,7 @@ void cso_cache_delete(struct cso_cache *sc)
    cso_for_each_state(sc, CSO_VERTEX_SHADER, delete_vs_state, 0);
    cso_for_each_state(sc, CSO_RASTERIZER, delete_rasterizer_state, 0);
    cso_for_each_state(sc, CSO_SAMPLER, delete_sampler_state, 0);
+   cso_for_each_state(sc, CSO_VELEMENTS, delete_velements, 0);
 
    cso_hash_delete(sc->blend_hash);
    cso_hash_delete(sc->sampler_hash);
@@ -358,6 +377,7 @@ void cso_cache_delete(struct cso_cache *sc)
    cso_hash_delete(sc->rasterizer_hash);
    cso_hash_delete(sc->fs_hash);
    cso_hash_delete(sc->vs_hash);
+   cso_hash_delete(sc->velements_hash);
    FREE(sc);
 }
 
@@ -372,6 +392,7 @@ void cso_set_maximum_cache_size(struct cso_cache *sc, int number)
    sanitize_hash(sc, sc->vs_hash, CSO_VERTEX_SHADER, sc->max_size);
    sanitize_hash(sc, sc->rasterizer_hash, CSO_RASTERIZER, sc->max_size);
    sanitize_hash(sc, sc->sampler_hash, CSO_SAMPLER, sc->max_size);
+   sanitize_hash(sc, sc->velements_hash, CSO_VELEMENTS, sc->max_size);
 }
 
 int cso_maximum_cache_size(const struct cso_cache *sc)
diff --git a/src/gallium/auxiliary/cso_cache/cso_cache.h b/src/gallium/auxiliary/cso_cache/cso_cache.h
index eea60b940b..fb09b83c62 100644
--- a/src/gallium/auxiliary/cso_cache/cso_cache.h
+++ b/src/gallium/auxiliary/cso_cache/cso_cache.h
@@ -53,6 +53,7 @@
   * - rasterizer (old setup)
   * - sampler
   * - vertex shader
+  * - vertex elements
   *
   * Things that are not constant state objects include:
   * - blend_color
@@ -90,7 +91,8 @@ enum cso_cache_type {
    CSO_DEPTH_STENCIL_ALPHA,
    CSO_RASTERIZER,
    CSO_FRAGMENT_SHADER,
-   CSO_VERTEX_SHADER
+   CSO_VERTEX_SHADER,
+   CSO_VELEMENTS
 };
 
 typedef void (*cso_state_callback)(void *ctx, void *obj);
@@ -144,6 +146,18 @@ struct cso_sampler {
    struct pipe_context *context;
 };
 
+struct cso_velems_state {
+   unsigned count;
+   struct pipe_vertex_element velems[PIPE_MAX_ATTRIBS];
+};
+
+struct cso_velements {
+   struct cso_velems_state state;
+   void *data;
+   cso_state_callback delete_state;
+   struct pipe_context *context;
+};
+
 unsigned cso_construct_key(void *item, int item_size);
 
 struct cso_cache *cso_cache_create(void);
diff --git a/src/gallium/auxiliary/cso_cache/cso_context.c b/src/gallium/auxiliary/cso_cache/cso_context.c
index a7335c340c..292e489312 100644
--- a/src/gallium/auxiliary/cso_cache/cso_context.c
+++ b/src/gallium/auxiliary/cso_cache/cso_context.c
@@ -89,6 +89,7 @@ struct cso_context {
    void *rasterizer, *rasterizer_saved;
    void *fragment_shader, *fragment_shader_saved, *geometry_shader;
    void *vertex_shader, *vertex_shader_saved, *geometry_shader_saved;
+   void *velements, *velements_saved;
 
    struct pipe_clip_state clip;
    struct pipe_clip_state clip_saved;
@@ -174,6 +175,20 @@ static boolean delete_vs_state(struct cso_context *ctx, void *state)
    return FALSE;
 }
 
+static boolean delete_vertex_elements(struct cso_context *ctx,
+                                      void *state)
+{
+   struct cso_velements *cso = (struct cso_velements *)state;
+
+   if (ctx->velements == cso->data)
+      return FALSE;
+
+   if (cso->delete_state)
+      cso->delete_state(cso->context, cso->data);
+   FREE(state);
+   return TRUE;
+}
+
 
 static INLINE boolean delete_cso(struct cso_context *ctx,
                                  void *state, enum cso_cache_type type)
@@ -197,6 +212,9 @@ static INLINE boolean delete_cso(struct cso_context *ctx,
    case CSO_VERTEX_SHADER:
       return delete_vs_state(ctx, state);
       break;
+   case CSO_VELEMENTS:
+      return delete_vertex_elements(ctx, state);
+      break;
    default:
       assert(0);
       FREE(state);
@@ -1130,7 +1148,6 @@ void cso_restore_geometry_shader(struct cso_context *ctx)
    ctx->geometry_shader_saved = NULL;
 }
 
-
 /* clip state */
 
 static INLINE void
@@ -1180,3 +1197,66 @@ cso_restore_clip(struct cso_context *ctx)
       ctx->pipe->set_clip_state(ctx->pipe, &ctx->clip_saved);
    }
 }
+
+enum pipe_error cso_set_vertex_elements(struct cso_context *ctx,
+                                        unsigned count,
+                                        const struct pipe_vertex_element *states)
+{
+   unsigned key_size, hash_key;
+   struct cso_hash_iter iter;
+   void *handle;
+   struct cso_velems_state velems_state;
+
+   /* need to include the count into the stored state data too.
+      Otherwise first few count pipe_vertex_elements could be identical even if count
+      is different, and there's no guarantee the hash would be different in that
+      case neither */
+   key_size = sizeof(struct pipe_vertex_element) * count + sizeof(unsigned);
+   velems_state.count = count;
+   memcpy(velems_state.velems, states, sizeof(struct pipe_vertex_element) * count);
+   hash_key = cso_construct_key((void*)&velems_state, key_size);
+   iter = cso_find_state_template(ctx->cache, hash_key, CSO_VELEMENTS, (void*)&velems_state, key_size);
+
+   if (cso_hash_iter_is_null(iter)) {
+      struct cso_velements *cso = MALLOC(sizeof(struct cso_velements));
+      if (!cso)
+         return PIPE_ERROR_OUT_OF_MEMORY;
+
+      memcpy(&cso->state, &velems_state, key_size);
+      cso->data = ctx->pipe->create_vertex_elements_state(ctx->pipe, count, &cso->state.velems[0]);
+      cso->delete_state = (cso_state_callback)ctx->pipe->delete_vertex_elements_state;
+      cso->context = ctx->pipe;
+
+      iter = cso_insert_state(ctx->cache, hash_key, CSO_VELEMENTS, cso);
+      if (cso_hash_iter_is_null(iter)) {
+         FREE(cso);
+         return PIPE_ERROR_OUT_OF_MEMORY;
+      }
+
+      handle = cso->data;
+   }
+   else {
+      handle = ((struct cso_velements *)cso_hash_iter_data(iter))->data;
+   }
+
+   if (ctx->velements != handle) {
+      ctx->velements = handle;
+      ctx->pipe->bind_vertex_elements_state(ctx->pipe, handle);
+   }
+   return PIPE_OK;
+}
+
+void cso_save_vertex_elements(struct cso_context *ctx)
+{
+   assert(!ctx->velements);
+   ctx->velements_saved = ctx->velements;
+}
+
+void cso_restore_vertex_elements(struct cso_context *ctx)
+{
+   if (ctx->velements != ctx->velements_saved) {
+      ctx->velements = ctx->velements_saved;
+      ctx->pipe->bind_vertex_elements_state(ctx->pipe, ctx->velements_saved);
+   }
+   ctx->velements_saved = NULL;
+}
diff --git a/src/gallium/auxiliary/cso_cache/cso_context.h b/src/gallium/auxiliary/cso_cache/cso_context.h
index 251a9a644f..9c16abd28d 100644
--- a/src/gallium/auxiliary/cso_cache/cso_context.h
+++ b/src/gallium/auxiliary/cso_cache/cso_context.h
@@ -122,6 +122,12 @@ void
 cso_restore_vertex_sampler_textures(struct cso_context *cso);
 
 
+enum pipe_error cso_set_vertex_elements(struct cso_context *ctx,
+                                        unsigned count,
+                                        const struct pipe_vertex_element *states);
+void cso_save_vertex_elements(struct cso_context *ctx);
+void cso_restore_vertex_elements(struct cso_context *ctx);
+
 
 /* These aren't really sensible -- most of the time the api provides
  * object semantics for shaders anyway, and the cases where it doesn't
@@ -157,7 +163,6 @@ void cso_save_geometry_shader(struct cso_context *cso);
 void cso_restore_geometry_shader(struct cso_context *cso);
 
 
-
 enum pipe_error cso_set_framebuffer(struct cso_context *cso,
                                     const struct pipe_framebuffer_state *fb);
 void cso_save_framebuffer(struct cso_context *cso);
diff --git a/src/gallium/auxiliary/draw/draw_pt.c b/src/gallium/auxiliary/draw/draw_pt.c
index 6d90a6c42f..a8cdc57ad9 100644
--- a/src/gallium/auxiliary/draw/draw_pt.c
+++ b/src/gallium/auxiliary/draw/draw_pt.c
@@ -307,9 +307,8 @@ draw_arrays_instanced(struct draw_context *draw,
       tgsi_dump(draw->vs.vertex_shader->state.tokens, 0);
       debug_printf("Elements:\n");
       for (i = 0; i < draw->pt.nr_vertex_elements; i++) {
-         debug_printf("  format=%s comps=%u\n",
-                      util_format_name(draw->pt.vertex_element[i].src_format),
-                      draw->pt.vertex_element[i].nr_components);
+         debug_printf("  format=%s\n",
+                      util_format_name(draw->pt.vertex_element[i].src_format));
       }
       debug_printf("Buffers:\n");
       for (i = 0; i < draw->pt.nr_vertex_buffers; i++) {
diff --git a/src/gallium/auxiliary/util/u_blit.c b/src/gallium/auxiliary/util/u_blit.c
index 0b263a9db5..4d0737ccd3 100644
--- a/src/gallium/auxiliary/util/u_blit.c
+++ b/src/gallium/auxiliary/util/u_blit.c
@@ -63,6 +63,7 @@ struct blit_state
    struct pipe_sampler_state sampler;
    struct pipe_viewport_state viewport;
    struct pipe_clip_state clip;
+   struct pipe_vertex_element velem[2];
 
    void *vs;
    void *fs[TGSI_WRITEMASK_XYZW + 1];
@@ -114,6 +115,15 @@ util_create_blit(struct pipe_context *pipe, struct cso_context *cso)
    ctx->sampler.mag_img_filter = 0; /* set later */
    ctx->sampler.normalized_coords = 1;
 
+   /* vertex elements state */
+   memset(&ctx->velem[0], 0, sizeof(ctx->velem[0]) * 2);
+   for (i = 0; i < 2; i++) {
+      ctx->velem[i].src_offset = i * 4 * sizeof(float);
+      ctx->velem[i].instance_divisor = 0;
+      ctx->velem[i].vertex_buffer_index = 0;
+      ctx->velem[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+   }
+
    /* vertex shader - still required to provide the linkage between
     * fragment shader input semantics and vertex_element/buffers.
     */
@@ -410,12 +420,14 @@ util_blit_pixels_writemask(struct blit_state *ctx,
    cso_save_fragment_shader(ctx->cso);
    cso_save_vertex_shader(ctx->cso);
    cso_save_clip(ctx->cso);
+   cso_save_vertex_elements(ctx->cso);
 
    /* set misc state we care about */
    cso_set_blend(ctx->cso, &ctx->blend);
    cso_set_depth_stencil_alpha(ctx->cso, &ctx->depthstencil);
    cso_set_rasterizer(ctx->cso, &ctx->rasterizer);
    cso_set_clip(ctx->cso, &ctx->clip);
+   cso_set_vertex_elements(ctx->cso, 2, ctx->velem);
 
    /* sampler */
    ctx->sampler.min_img_filter = filter;
@@ -480,6 +492,7 @@ util_blit_pixels_writemask(struct blit_state *ctx,
    cso_restore_fragment_shader(ctx->cso);
    cso_restore_vertex_shader(ctx->cso);
    cso_restore_clip(ctx->cso);
+   cso_restore_vertex_elements(ctx->cso);
 
    pipe_texture_reference(&tex, NULL);
 }
@@ -564,12 +577,14 @@ util_blit_pixels_tex(struct blit_state *ctx,
    cso_save_fragment_shader(ctx->cso);
    cso_save_vertex_shader(ctx->cso);
    cso_save_clip(ctx->cso);
+   cso_save_vertex_elements(ctx->cso);
 
    /* set misc state we care about */
    cso_set_blend(ctx->cso, &ctx->blend);
    cso_set_depth_stencil_alpha(ctx->cso, &ctx->depthstencil);
    cso_set_rasterizer(ctx->cso, &ctx->rasterizer);
    cso_set_clip(ctx->cso, &ctx->clip);
+   cso_set_vertex_elements(ctx->cso, 2, ctx->velem);
 
    /* sampler */
    ctx->sampler.min_img_filter = filter;
@@ -628,4 +643,5 @@ util_blit_pixels_tex(struct blit_state *ctx,
    cso_restore_fragment_shader(ctx->cso);
    cso_restore_vertex_shader(ctx->cso);
    cso_restore_clip(ctx->cso);
+   cso_restore_vertex_elements(ctx->cso);
 }
diff --git a/src/gallium/auxiliary/util/u_blitter.c b/src/gallium/auxiliary/util/u_blitter.c
index 0ba09d33bf..33d09085f0 100644
--- a/src/gallium/auxiliary/util/u_blitter.c
+++ b/src/gallium/auxiliary/util/u_blitter.c
@@ -88,6 +88,8 @@ struct blitter_context_priv
    void *dsa_write_depth_keep_stencil;
    void *dsa_keep_depth_stencil;
 
+   void *velem_state;
+
    /* Sampler state for clamping to a miplevel. */
    void *sampler_state[PIPE_MAX_TEXTURE_LEVELS];
 
@@ -108,6 +110,7 @@ struct blitter_context *util_blitter_create(struct pipe_context *pipe)
    struct pipe_depth_stencil_alpha_state dsa = { { 0 } };
    struct pipe_rasterizer_state rs_state = { 0 };
    struct pipe_sampler_state *sampler_state;
+   struct pipe_vertex_element velem[2];
    unsigned i;
 
    ctx = CALLOC_STRUCT(blitter_context_priv);
@@ -122,6 +125,7 @@ struct blitter_context *util_blitter_create(struct pipe_context *pipe)
    ctx->blitter.saved_rs_state = INVALID_PTR;
    ctx->blitter.saved_fs = INVALID_PTR;
    ctx->blitter.saved_vs = INVALID_PTR;
+   ctx->blitter.saved_velem_state = INVALID_PTR;
    ctx->blitter.saved_fb_state.nr_cbufs = ~0;
    ctx->blitter.saved_num_textures = ~0;
    ctx->blitter.saved_num_sampler_states = ~0;
@@ -170,6 +174,16 @@ struct blitter_context *util_blitter_create(struct pipe_context *pipe)
    rs_state.flatshade = 1;
    ctx->rs_state = pipe->create_rasterizer_state(pipe, &rs_state);
 
+   /* vertex elements state */
+   memset(&velem[0], 0, sizeof(velem[0]) * 2);
+   for (i = 0; i < 2; i++) {
+      velem[i].src_offset = i * 4 * sizeof(float);
+      velem[i].instance_divisor = 0;
+      velem[i].vertex_buffer_index = 0;
+      velem[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+   }
+   ctx->velem_state = pipe->create_vertex_elements_state(pipe, 2, &velem[0]);
+
    /* fragment shaders are created on-demand */
 
    /* vertex shaders */
@@ -219,6 +233,7 @@ void util_blitter_destroy(struct blitter_context *blitter)
    pipe->delete_rasterizer_state(pipe, ctx->rs_state);
    pipe->delete_vs_state(pipe, ctx->vs_col);
    pipe->delete_vs_state(pipe, ctx->vs_tex);
+   pipe->delete_vertex_elements_state(pipe, ctx->velem_state);
 
    for (i = 0; i < PIPE_MAX_TEXTURE_TYPES; i++) {
       if (ctx->fs_texfetch_col[i])
@@ -246,7 +261,8 @@ static void blitter_check_saved_CSOs(struct blitter_context_priv *ctx)
           ctx->blitter.saved_dsa_state != INVALID_PTR &&
           ctx->blitter.saved_rs_state != INVALID_PTR &&
           ctx->blitter.saved_fs != INVALID_PTR &&
-          ctx->blitter.saved_vs != INVALID_PTR);
+          ctx->blitter.saved_vs != INVALID_PTR &&
+          ctx->blitter.saved_velem_state != INVALID_PTR);
 }
 
 static void blitter_restore_CSOs(struct blitter_context_priv *ctx)
@@ -259,12 +275,14 @@ static void blitter_restore_CSOs(struct blitter_context_priv *ctx)
    pipe->bind_rasterizer_state(pipe, ctx->blitter.saved_rs_state);
    pipe->bind_fs_state(pipe, ctx->blitter.saved_fs);
    pipe->bind_vs_state(pipe, ctx->blitter.saved_vs);
+   pipe->bind_vertex_elements_state(pipe, ctx->blitter.saved_velem_state);
 
    ctx->blitter.saved_blend_state = INVALID_PTR;
    ctx->blitter.saved_dsa_state = INVALID_PTR;
    ctx->blitter.saved_rs_state = INVALID_PTR;
    ctx->blitter.saved_fs = INVALID_PTR;
    ctx->blitter.saved_vs = INVALID_PTR;
+   ctx->blitter.saved_velem_state = INVALID_PTR;
 
    pipe->set_stencil_ref(pipe, &ctx->blitter.saved_stencil_ref);
 
@@ -569,6 +587,7 @@ void util_blitter_clear(struct blitter_context *blitter,
       pipe->bind_depth_stencil_alpha_state(pipe, ctx->dsa_keep_depth_stencil);
 
    pipe->bind_rasterizer_state(pipe, ctx->rs_state);
+   pipe->bind_vertex_elements_state(pipe, ctx->velem_state);
    pipe->bind_fs_state(pipe, blitter_get_fs_col(ctx, num_cbufs));
    pipe->bind_vs_state(pipe, ctx->vs_col);
 
@@ -634,6 +653,7 @@ static void util_blitter_do_copy(struct blitter_context *blitter,
    pipe->bind_vs_state(pipe, ctx->vs_tex);
    pipe->bind_fragment_sampler_states(pipe, 1,
       blitter_get_sampler_state(ctx, src->level));
+   pipe->bind_vertex_elements_state(pipe, ctx->velem_state);
    pipe->set_fragment_sampler_textures(pipe, 1, &src->texture);
    pipe->set_framebuffer_state(pipe, &fb_state);
 
@@ -807,6 +827,7 @@ void util_blitter_fill(struct blitter_context *blitter,
    pipe->bind_rasterizer_state(pipe, ctx->rs_state);
    pipe->bind_fs_state(pipe, blitter_get_fs_col(ctx, 1));
    pipe->bind_vs_state(pipe, ctx->vs_col);
+   pipe->bind_vertex_elements_state(pipe, ctx->velem_state);
 
    /* set a framebuffer state */
    fb_state.width = dst->width;
diff --git a/src/gallium/auxiliary/util/u_blitter.h b/src/gallium/auxiliary/util/u_blitter.h
index 92008fce99..ecafdabafa 100644
--- a/src/gallium/auxiliary/util/u_blitter.h
+++ b/src/gallium/auxiliary/util/u_blitter.h
@@ -43,6 +43,7 @@ struct blitter_context
    /* Private members, really. */
    void *saved_blend_state;   /**< blend state */
    void *saved_dsa_state;     /**< depth stencil alpha state */
+   void *saved_velem_state;   /**< vertex elements state */
    void *saved_rs_state;      /**< rasterizer state */
    void *saved_fs, *saved_vs; /**< fragment shader, vertex shader */
 
@@ -173,6 +174,13 @@ void util_blitter_save_depth_stencil_alpha(struct blitter_context *blitter,
 }
 
 static INLINE
+void util_blitter_save_vertex_elements(struct blitter_context *blitter,
+                                       void *state)
+{
+   blitter->saved_velem_state = state;
+}
+
+static INLINE
 void util_blitter_save_stencil_ref(struct blitter_context *blitter,
                                    const struct pipe_stencil_ref *state)
 {
diff --git a/src/gallium/auxiliary/util/u_draw_quad.c b/src/gallium/auxiliary/util/u_draw_quad.c
index 14506e8451..8c194102bf 100644
--- a/src/gallium/auxiliary/util/u_draw_quad.c
+++ b/src/gallium/auxiliary/util/u_draw_quad.c
@@ -45,8 +45,6 @@ util_draw_vertex_buffer(struct pipe_context *pipe,
                         uint num_attribs)
 {
    struct pipe_vertex_buffer vbuffer;
-   struct pipe_vertex_element velements[PIPE_MAX_ATTRIBS];
-   uint i;
 
    assert(num_attribs <= PIPE_MAX_ATTRIBS);
 
@@ -58,15 +56,7 @@ util_draw_vertex_buffer(struct pipe_context *pipe,
    vbuffer.max_index = num_verts - 1;
    pipe->set_vertex_buffers(pipe, 1, &vbuffer);
 
-   /* tell pipe about the vertex attributes */
-   for (i = 0; i < num_attribs; i++) {
-      velements[i].src_offset = i * 4 * sizeof(float);
-      velements[i].instance_divisor = 0;
-      velements[i].vertex_buffer_index = 0;
-      velements[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
-      velements[i].nr_components = 4;
-   }
-   pipe->set_vertex_elements(pipe, num_attribs, velements);
+   /* note: vertex elements already set by caller */
 
    /* draw */
    pipe->draw_arrays(pipe, prim_type, 0, num_verts);
diff --git a/src/gallium/auxiliary/util/u_dump_state.c b/src/gallium/auxiliary/util/u_dump_state.c
index ae7afd7311..52cf3ef4ce 100644
--- a/src/gallium/auxiliary/util/u_dump_state.c
+++ b/src/gallium/auxiliary/util/u_dump_state.c
@@ -700,7 +700,6 @@ util_dump_vertex_element(struct os_stream *stream, const struct pipe_vertex_elem
    util_dump_member(stream, uint, state, src_offset);
 
    util_dump_member(stream, uint, state, vertex_buffer_index);
-   util_dump_member(stream, uint, state, nr_components);
 
    util_dump_member(stream, format, state, src_format);
 
diff --git a/src/gallium/auxiliary/util/u_format.h b/src/gallium/auxiliary/util/u_format.h
index b2aa5bfb18..20fdaca007 100644
--- a/src/gallium/auxiliary/util/u_format.h
+++ b/src/gallium/auxiliary/util/u_format.h
@@ -415,6 +415,23 @@ util_format_has_alpha(enum pipe_format format)
    }
 }
 
+/**
+ * Return the number of components stored.
+ * Formats with block size != 1x1 will always have 1 component (the block).
+ */
+static INLINE unsigned
+util_format_get_nr_components(enum pipe_format format)
+{
+   const struct util_format_description *desc = util_format_description(format);
+   unsigned chan;
+   unsigned nr_comp = 0;
+   for (chan = 0; chan < 4; chan++) {
+      if (desc->channel[chan].size != 0) {
+         nr_comp++;
+      }
+   }
+   return nr_comp;
+}
 
 /*
  * Format access functions.
diff --git a/src/gallium/auxiliary/util/u_gen_mipmap.c b/src/gallium/auxiliary/util/u_gen_mipmap.c
index fc027e48e4..d421bee8ef 100644
--- a/src/gallium/auxiliary/util/u_gen_mipmap.c
+++ b/src/gallium/auxiliary/util/u_gen_mipmap.c
@@ -62,6 +62,7 @@ struct gen_mipmap_state
    struct pipe_rasterizer_state rasterizer;
    struct pipe_sampler_state sampler;
    struct pipe_clip_state clip;
+   struct pipe_vertex_element velem[2];
 
    void *vs;
    void *fs2d, *fsCube;
@@ -1307,6 +1308,15 @@ util_create_gen_mipmap(struct pipe_context *pipe,
    ctx->sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NEAREST;
    ctx->sampler.normalized_coords = 1;
 
+   /* vertex elements state */
+   memset(&ctx->velem[0], 0, sizeof(ctx->velem[0]) * 2);
+   for (i = 0; i < 2; i++) {
+      ctx->velem[i].src_offset = i * 4 * sizeof(float);
+      ctx->velem[i].instance_divisor = 0;
+      ctx->velem[i].vertex_buffer_index = 0;
+      ctx->velem[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+   }
+
    /* vertex shader - still needed to specify mapping from fragment
     * shader input semantics to vertex elements 
     */
@@ -1501,12 +1511,14 @@ util_gen_mipmap(struct gen_mipmap_state *ctx,
    cso_save_vertex_shader(ctx->cso);
    cso_save_viewport(ctx->cso);
    cso_save_clip(ctx->cso);
+   cso_save_vertex_elements(ctx->cso);
 
    /* bind our state */
    cso_set_blend(ctx->cso, &ctx->blend);
    cso_set_depth_stencil_alpha(ctx->cso, &ctx->depthstencil);
    cso_set_rasterizer(ctx->cso, &ctx->rasterizer);
    cso_set_clip(ctx->cso, &ctx->clip);
+   cso_set_vertex_elements(ctx->cso, 2, ctx->velem);
 
    cso_set_fragment_shader_handle(ctx->cso, fs);
    cso_set_vertex_shader_handle(ctx->cso, ctx->vs);
@@ -1593,4 +1605,5 @@ util_gen_mipmap(struct gen_mipmap_state *ctx,
    cso_restore_vertex_shader(ctx->cso);
    cso_restore_viewport(ctx->cso);
    cso_restore_clip(ctx->cso);
+   cso_restore_vertex_elements(ctx->cso);
 }
diff --git a/src/gallium/auxiliary/vl/vl_compositor.c b/src/gallium/auxiliary/vl/vl_compositor.c
index ba23435f69..6d461cb880 100644
--- a/src/gallium/auxiliary/vl/vl_compositor.c
+++ b/src/gallium/auxiliary/vl/vl_compositor.c
@@ -230,6 +230,7 @@ static bool
 init_pipe_state(struct vl_compositor *c)
 {
    struct pipe_sampler_state sampler;
+   struct pipe_vertex_element vertex_elems[2];
 
    assert(c);
 
@@ -251,15 +252,27 @@ init_pipe_state(struct vl_compositor *c)
    /*sampler.border_color[i] = ;*/
    /*sampler.max_anisotropy = ;*/
    c->sampler = c->pipe->create_sampler_state(c->pipe, &sampler);
-	
+
+   vertex_elems[0].src_offset = 0;
+   vertex_elems[0].instance_divisor = 0;
+   vertex_elems[0].vertex_buffer_index = 0;
+   vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT;
+   vertex_elems[1].src_offset = 0;
+   vertex_elems[1].instance_divisor = 0;
+   vertex_elems[1].vertex_buffer_index = 1;
+   vertex_elems[1].src_format = PIPE_FORMAT_R32G32_FLOAT;
+   c->vertex_elems = c->pipe->create_vertex_elements_state(c->pipe, 2, vertex_elems);
+
+
    return true;
 }
 
 static void cleanup_pipe_state(struct vl_compositor *c)
 {
    assert(c);
-	
+
    c->pipe->delete_sampler_state(c->pipe, c->sampler);
+   c->pipe->delete_vertex_elements_state(c->pipe, c->vertex_elems);
 }
 
 static bool
@@ -314,12 +327,6 @@ init_buffers(struct vl_compositor *c)
 
    pipe_buffer_unmap(c->pipe->screen, c->vertex_bufs[0].buffer);
 
-   c->vertex_elems[0].src_offset = 0;
-   c->vertex_elems[0].instance_divisor = 0;
-   c->vertex_elems[0].vertex_buffer_index = 0;
-   c->vertex_elems[0].nr_components = 2;
-   c->vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
    /*
     * Create our texcoord buffer and texcoord buffer element
     * Texcoord buffer contains the TCs for mapping the rendered surface to the 4 vertices
@@ -344,12 +351,6 @@ init_buffers(struct vl_compositor *c)
 
    pipe_buffer_unmap(c->pipe->screen, c->vertex_bufs[1].buffer);
 
-   c->vertex_elems[1].src_offset = 0;
-   c->vertex_elems[1].instance_divisor = 0;
-   c->vertex_elems[1].vertex_buffer_index = 1;
-   c->vertex_elems[1].nr_components = 2;
-   c->vertex_elems[1].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
    /*
     * Create our vertex shader's constant buffer
     * Const buffer contains scaling and translation vectors
@@ -483,7 +484,7 @@ void vl_compositor_render(struct vl_compositor          *compositor,
    compositor->pipe->bind_vs_state(compositor->pipe, compositor->vertex_shader);
    compositor->pipe->bind_fs_state(compositor->pipe, compositor->fragment_shader);
    compositor->pipe->set_vertex_buffers(compositor->pipe, 2, compositor->vertex_bufs);
-   compositor->pipe->set_vertex_elements(compositor->pipe, 2, compositor->vertex_elems);
+   compositor->pipe->bind_vertex_elements_state(compositor->pipe, compositor->vertex_elems);
    compositor->pipe->set_constant_buffer(compositor->pipe, PIPE_SHADER_VERTEX, 0, compositor->vs_const_buf);
    compositor->pipe->set_constant_buffer(compositor->pipe, PIPE_SHADER_FRAGMENT, 0, compositor->fs_const_buf);
 
diff --git a/src/gallium/auxiliary/vl/vl_compositor.h b/src/gallium/auxiliary/vl/vl_compositor.h
index 6a9a3fd7af..51755554da 100644
--- a/src/gallium/auxiliary/vl/vl_compositor.h
+++ b/src/gallium/auxiliary/vl/vl_compositor.h
@@ -43,10 +43,10 @@ struct vl_compositor
    void *sampler;
    void *vertex_shader;
    void *fragment_shader;
+   void *vertex_elems;
    struct pipe_viewport_state viewport;
    struct pipe_scissor_state scissor;
    struct pipe_vertex_buffer vertex_bufs[2];
-   struct pipe_vertex_element vertex_elems[2];
    struct pipe_buffer *vs_const_buf, *fs_const_buf;
 };
 
diff --git a/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.c b/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.c
index f323de0ea5..0763b5bb0e 100644
--- a/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.c
+++ b/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.c
@@ -708,6 +708,7 @@ static bool
 init_pipe_state(struct vl_mpeg12_mc_renderer *r)
 {
    struct pipe_sampler_state sampler;
+   struct pipe_vertex_element vertex_elems[8];
    unsigned filters[5];
    unsigned i;
 
@@ -771,6 +772,59 @@ init_pipe_state(struct vl_mpeg12_mc_renderer *r)
       r->samplers.all[i] = r->pipe->create_sampler_state(r->pipe, &sampler);
    }
 
+   /* Position element */
+   vertex_elems[0].src_offset = 0;
+   vertex_elems[0].instance_divisor = 0;
+   vertex_elems[0].vertex_buffer_index = 0;
+   vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* Luma, texcoord element */
+   vertex_elems[1].src_offset = sizeof(struct vertex2f);
+   vertex_elems[1].instance_divisor = 0;
+   vertex_elems[1].vertex_buffer_index = 0;
+   vertex_elems[1].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* Chroma Cr texcoord element */
+   vertex_elems[2].src_offset = sizeof(struct vertex2f) * 2;
+   vertex_elems[2].instance_divisor = 0;
+   vertex_elems[2].vertex_buffer_index = 0;
+   vertex_elems[2].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* Chroma Cb texcoord element */
+   vertex_elems[3].src_offset = sizeof(struct vertex2f) * 3;
+   vertex_elems[3].instance_divisor = 0;
+   vertex_elems[3].vertex_buffer_index = 0;
+   vertex_elems[3].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* First ref surface top field texcoord element */
+   vertex_elems[4].src_offset = 0;
+   vertex_elems[4].instance_divisor = 0;
+   vertex_elems[4].vertex_buffer_index = 1;
+   vertex_elems[4].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* First ref surface bottom field texcoord element */
+   vertex_elems[5].src_offset = sizeof(struct vertex2f);
+   vertex_elems[5].instance_divisor = 0;
+   vertex_elems[5].vertex_buffer_index = 1;
+   vertex_elems[5].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* Second ref surface top field texcoord element */
+   vertex_elems[6].src_offset = 0;
+   vertex_elems[6].instance_divisor = 0;
+   vertex_elems[6].vertex_buffer_index = 2;
+   vertex_elems[6].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* Second ref surface bottom field texcoord element */
+   vertex_elems[7].src_offset = sizeof(struct vertex2f);
+   vertex_elems[7].instance_divisor = 0;
+   vertex_elems[7].vertex_buffer_index = 2;
+   vertex_elems[7].src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+   /* need versions with 4,6 and 8 vertex elems */
+   r->vertex_elems[0] = r->pipe->create_vertex_elements_state(r->pipe, 4, vertex_elems);
+   r->vertex_elems[1] = r->pipe->create_vertex_elements_state(r->pipe, 6, vertex_elems);
+   r->vertex_elems[2] = r->pipe->create_vertex_elements_state(r->pipe, 8, vertex_elems);
+
    return true;
 }
 
@@ -783,6 +837,8 @@ cleanup_pipe_state(struct vl_mpeg12_mc_renderer *r)
 
    for (i = 0; i < 5; ++i)
       r->pipe->delete_sampler_state(r->pipe, r->samplers.all[i]);
+   for (i = 0; i < 3; i++)
+      r->pipe->delete_vertex_elements_state(r->pipe, r->vertex_elems[i]);
 }
 
 static bool
@@ -888,62 +944,6 @@ init_buffers(struct vl_mpeg12_mc_renderer *r)
       );
    }
 
-   /* Position element */
-   r->vertex_elems[0].src_offset = 0;
-   r->vertex_elems[0].instance_divisor = 0;
-   r->vertex_elems[0].vertex_buffer_index = 0;
-   r->vertex_elems[0].nr_components = 2;
-   r->vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* Luma, texcoord element */
-   r->vertex_elems[1].src_offset = sizeof(struct vertex2f);
-   r->vertex_elems[1].instance_divisor = 0;
-   r->vertex_elems[1].vertex_buffer_index = 0;
-   r->vertex_elems[1].nr_components = 2;
-   r->vertex_elems[1].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* Chroma Cr texcoord element */
-   r->vertex_elems[2].src_offset = sizeof(struct vertex2f) * 2;
-   r->vertex_elems[2].instance_divisor = 0;
-   r->vertex_elems[2].vertex_buffer_index = 0;
-   r->vertex_elems[2].nr_components = 2;
-   r->vertex_elems[2].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* Chroma Cb texcoord element */
-   r->vertex_elems[3].src_offset = sizeof(struct vertex2f) * 3;
-   r->vertex_elems[3].instance_divisor = 0;
-   r->vertex_elems[3].vertex_buffer_index = 0;
-   r->vertex_elems[3].nr_components = 2;
-   r->vertex_elems[3].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* First ref surface top field texcoord element */
-   r->vertex_elems[4].src_offset = 0;
-   r->vertex_elems[4].instance_divisor = 0;
-   r->vertex_elems[4].vertex_buffer_index = 1;
-   r->vertex_elems[4].nr_components = 2;
-   r->vertex_elems[4].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* First ref surface bottom field texcoord element */
-   r->vertex_elems[5].src_offset = sizeof(struct vertex2f);
-   r->vertex_elems[5].instance_divisor = 0;
-   r->vertex_elems[5].vertex_buffer_index = 1;
-   r->vertex_elems[5].nr_components = 2;
-   r->vertex_elems[5].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* Second ref surface top field texcoord element */
-   r->vertex_elems[6].src_offset = 0;
-   r->vertex_elems[6].instance_divisor = 0;
-   r->vertex_elems[6].vertex_buffer_index = 2;
-   r->vertex_elems[6].nr_components = 2;
-   r->vertex_elems[6].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
-   /* Second ref surface bottom field texcoord element */
-   r->vertex_elems[7].src_offset = sizeof(struct vertex2f);
-   r->vertex_elems[7].instance_divisor = 0;
-   r->vertex_elems[7].vertex_buffer_index = 2;
-   r->vertex_elems[7].nr_components = 2;
-   r->vertex_elems[7].src_format = PIPE_FORMAT_R32G32_FLOAT;
-
    r->vs_const_buf = pipe_buffer_create
    (
       r->pipe->screen,
@@ -1307,7 +1307,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (num_macroblocks[MACROBLOCK_TYPE_INTRA] > 0) {
       r->pipe->set_vertex_buffers(r->pipe, 1, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 4, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[0]);
       r->pipe->set_fragment_sampler_textures(r->pipe, 3, r->textures.all);
       r->pipe->bind_fragment_sampler_states(r->pipe, 3, r->samplers.all);
       r->pipe->bind_vs_state(r->pipe, r->i_vs);
@@ -1320,7 +1320,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (num_macroblocks[MACROBLOCK_TYPE_FWD_FRAME_PRED] > 0) {
       r->pipe->set_vertex_buffers(r->pipe, 2, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 6, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[1]);
       r->textures.individual.ref[0] = r->past;
       r->pipe->set_fragment_sampler_textures(r->pipe, 4, r->textures.all);
       r->pipe->bind_fragment_sampler_states(r->pipe, 4, r->samplers.all);
@@ -1334,7 +1334,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (false /*num_macroblocks[MACROBLOCK_TYPE_FWD_FIELD_PRED] > 0 */ ) {
       r->pipe->set_vertex_buffers(r->pipe, 2, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 6, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[1]);
       r->textures.individual.ref[0] = r->past;
       r->pipe->set_fragment_sampler_textures(r->pipe, 4, r->textures.all);
       r->pipe->bind_fragment_sampler_states(r->pipe, 4, r->samplers.all);
@@ -1348,7 +1348,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (num_macroblocks[MACROBLOCK_TYPE_BKWD_FRAME_PRED] > 0) {
       r->pipe->set_vertex_buffers(r->pipe, 2, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 6, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[1]);
       r->textures.individual.ref[0] = r->future;
       r->pipe->set_fragment_sampler_textures(r->pipe, 4, r->textures.all);
       r->pipe->bind_fragment_sampler_states(r->pipe, 4, r->samplers.all);
@@ -1362,7 +1362,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (false /*num_macroblocks[MACROBLOCK_TYPE_BKWD_FIELD_PRED] > 0 */ ) {
       r->pipe->set_vertex_buffers(r->pipe, 2, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 6, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[1]);
       r->textures.individual.ref[0] = r->future;
       r->pipe->set_fragment_sampler_textures(r->pipe, 4, r->textures.all);
       r->pipe->bind_fragment_sampler_states(r->pipe, 4, r->samplers.all);
@@ -1376,7 +1376,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (num_macroblocks[MACROBLOCK_TYPE_BI_FRAME_PRED] > 0) {
       r->pipe->set_vertex_buffers(r->pipe, 3, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 8, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[2]);
       r->textures.individual.ref[0] = r->past;
       r->textures.individual.ref[1] = r->future;
       r->pipe->set_fragment_sampler_textures(r->pipe, 5, r->textures.all);
@@ -1391,7 +1391,7 @@ flush(struct vl_mpeg12_mc_renderer *r)
 
    if (false /*num_macroblocks[MACROBLOCK_TYPE_BI_FIELD_PRED] > 0 */ ) {
       r->pipe->set_vertex_buffers(r->pipe, 3, r->vertex_bufs.all);
-      r->pipe->set_vertex_elements(r->pipe, 8, r->vertex_elems);
+      r->pipe->bind_vertex_elements_state(r->pipe, r->vertex_elems[2]);
       r->textures.individual.ref[0] = r->past;
       r->textures.individual.ref[1] = r->future;
       r->pipe->set_fragment_sampler_textures(r->pipe, 5, r->textures.all);
diff --git a/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.h b/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.h
index f00b8c7b8b..a11a3e7307 100644
--- a/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.h
+++ b/src/gallium/auxiliary/vl/vl_mpeg12_mc_renderer.h
@@ -66,8 +66,8 @@ struct vl_mpeg12_mc_renderer
    struct pipe_buffer *vs_const_buf;
    struct pipe_buffer *fs_const_buf;
    struct pipe_framebuffer_state fb_state;
-   struct pipe_vertex_element vertex_elems[8];
-	
+   void *vertex_elems[3];
+
    union
    {
       void *all[5];
diff --git a/src/gallium/docs/source/context.rst b/src/gallium/docs/source/context.rst
index 9080addba4..4608e97adb 100644
--- a/src/gallium/docs/source/context.rst
+++ b/src/gallium/docs/source/context.rst
@@ -24,6 +24,7 @@ CSO objects handled by the context object:
 * :ref:`Depth, Stencil, & Alpha`: ``*_depth_stencil_alpha_state``
 * :ref:`Shader`: These have two sets of methods. ``*_fs_state`` is for
   fragment shaders, and ``*_vs_state`` is for vertex shaders.
+* :ref:`Vertex Elements`: ``*_vertex_elements_state``
 
 
 Resource Binding State
@@ -60,7 +61,6 @@ objects. They all follow simple, one-method binding calls, e.g.
   not have the scissor test enabled, then the scissor bounds never need to
   be set since they will not be used.
 * ``set_viewport_state``
-* ``set_vertex_elements``
 
 
 Clearing
diff --git a/src/gallium/docs/source/cso/velems.rst b/src/gallium/docs/source/cso/velems.rst
new file mode 100644
index 0000000000..8e758fae10
--- /dev/null
+++ b/src/gallium/docs/source/cso/velems.rst
@@ -0,0 +1,24 @@
+.. _vertex,elements
+
+Vertex Elements
+===============
+
+This state controls format etc. of the input attributes contained
+in the pipe_vertex_buffer(s). There's one pipe_vertex_element array member
+for each input attribute.
+
+Members
+-------
+
+src_offset
+    The byte offset of the attribute in the buffer given by
+    vertex_buffer_index for the first vertex.
+instance_divisor
+    The instance data rate divisor, used for instancing.
+    0 means this is per-vertex data, n means per-instance data used for
+    n consecutive instances (n > 0).
+vertex_buffer_index
+    The vertex buffer this attribute lives in. Several attributes may
+    live in the same vertex buffer.
+src_format
+    The format of the attribute data. One of the PIPE_FORMAT tokens.
diff --git a/src/gallium/drivers/cell/ppu/cell_context.h b/src/gallium/drivers/cell/ppu/cell_context.h
index a77cc5b906..584f355804 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.h
+++ b/src/gallium/drivers/cell/ppu/cell_context.h
@@ -93,6 +93,11 @@ struct cell_buffer_list
    struct cell_buffer_node *head;
 };
 
+struct cell_velems_state
+{
+   unsigned count;
+   struct pipe_vertex_element velem[PIPE_MAX_ATTRIBS];
+}
 
 /**
  * Per-context state, subclass of pipe_context.
@@ -110,6 +115,7 @@ struct cell_context
    const struct pipe_rasterizer_state *rasterizer;
    const struct cell_vertex_shader_state *vs;
    const struct cell_fragment_shader_state *fs;
+   const struct cell_velems_state *velems;
 
    struct spe_function logic_op;
 
@@ -125,8 +131,6 @@ struct cell_context
    struct pipe_viewport_state viewport;
    struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS];
    uint num_vertex_buffers;
-   struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
-   uint num_vertex_elements;
 
    ubyte *cbuf_map[PIPE_MAX_COLOR_BUFS];
    ubyte *zsbuf_map;
diff --git a/src/gallium/drivers/cell/ppu/cell_state_vertex.c b/src/gallium/drivers/cell/ppu/cell_state_vertex.c
index fbe55c8472..d3efb8ecea 100644
--- a/src/gallium/drivers/cell/ppu/cell_state_vertex.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_vertex.c
@@ -32,24 +32,43 @@
 #include "cell_context.h"
 #include "cell_state.h"
 
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 
 
-static void
-cell_set_vertex_elements(struct pipe_context *pipe,
-                         unsigned count,
-                         const struct pipe_vertex_element *elements)
+void *
+cell_create_vertex_elements_state(struct pipe_context *pipe,
+                                  unsigned count,
+                                  const struct pipe_vertex_element *attribs)
 {
-   struct cell_context *cell = cell_context(pipe);
-
+   struct cell_velems_state *velems;
    assert(count <= PIPE_MAX_ATTRIBS);
+   velems = (struct cell_velems_state *) MALLOC(sizeof(struct cell_velems_state));
+   if (velems) {
+      velems->count = count;
+      memcpy(velems->velem, attribs, sizeof(*attribs) * count);
+   }
+   return velems;
+}
+
+void
+cell_bind_vertex_elements_state(struct pipe_context *pipe,
+                                void *velems)
+{
+   struct cell_context *cell = cell_context(pipe);
+   struct cell_velems_state *cell_velems = (struct cell_velems_state *) velems;
 
-   memcpy(cell->vertex_element, elements, count * sizeof(elements[0]));
-   cell->num_vertex_elements = count;
+   cell->velems = cell_velems;
 
    cell->dirty |= CELL_NEW_VERTEX;
 
-   draw_set_vertex_elements(cell->draw, count, elements);
+   draw_set_vertex_elements(cell->draw, cell_velems->count, cell_velems->velem);
+}
+
+void
+cell_delete_vertex_elements_state(struct pipe_context *pipe, void *velems)
+{
+   FREE( velems );
 }
 
 
@@ -75,5 +94,7 @@ void
 cell_init_vertex_functions(struct cell_context *cell)
 {
    cell->pipe.set_vertex_buffers = cell_set_vertex_buffers;
-   cell->pipe.set_vertex_elements = cell_set_vertex_elements;
+   cell->pipe.create_vertex_elements_state = cell_create_vertex_elements_state;
+   cell->pipe.bind_vertex_elements_state = cell_bind_vertex_elements_state;
+   cell->pipe.delete_vertex_elements_state = cell_delete_vertex_elements_state;
 }
diff --git a/src/gallium/drivers/failover/fo_context.h b/src/gallium/drivers/failover/fo_context.h
index bb1a168ea7..4a754465bb 100644
--- a/src/gallium/drivers/failover/fo_context.h
+++ b/src/gallium/drivers/failover/fo_context.h
@@ -78,6 +78,7 @@ struct failover_context {
    const struct fo_state     *rasterizer;
    const struct fo_state     *fragment_shader;
    const struct fo_state     *vertex_shader;
+   const struct fo_state     *vertex_elements;
 
    struct pipe_blend_color blend_color;
    struct pipe_stencil_ref stencil_ref;
@@ -89,10 +90,8 @@ struct failover_context {
    struct pipe_texture *vertex_textures[PIPE_MAX_VERTEX_SAMPLERS];
    struct pipe_viewport_state viewport;
    struct pipe_vertex_buffer vertex_buffers[PIPE_MAX_ATTRIBS];
-   struct pipe_vertex_element vertex_elements[PIPE_MAX_ATTRIBS];
 
    uint num_vertex_buffers;
-   uint num_vertex_elements;
 
    void *sw_sampler_state[PIPE_MAX_SAMPLERS];
    void *hw_sampler_state[PIPE_MAX_SAMPLERS];
diff --git a/src/gallium/drivers/failover/fo_state.c b/src/gallium/drivers/failover/fo_state.c
index 970606a3f5..0247fb803b 100644
--- a/src/gallium/drivers/failover/fo_state.c
+++ b/src/gallium/drivers/failover/fo_state.c
@@ -255,9 +255,52 @@ failover_delete_vs_state(struct pipe_context *pipe,
    free(state);
 }
 
+
+
+static void *
+failover_create_vertex_elements_state( struct pipe_context *pipe,
+                                       unsigned count,
+                                       const struct pipe_vertex_element *velems )
+{
+   struct fo_state *state = malloc(sizeof(struct fo_state));
+   struct failover_context *failover = failover_context(pipe);
+
+   state->sw_state = failover->sw->create_vertex_elements_state(failover->sw, count, velems);
+   state->hw_state = failover->hw->create_vertex_elements_state(failover->hw, count, velems);
+
+   return state;
+}
+
+static void
+failover_bind_vertex_elements_state(struct pipe_context *pipe,
+                                    void *velems )
+{
+   struct failover_context *failover = failover_context(pipe);
+   struct fo_state *state = (struct fo_state*)velems;
+
+   failover->vertex_elements = state;
+   failover->dirty |= FO_NEW_VERTEX_ELEMENT;
+   failover->sw->bind_vertex_elements_state( failover->sw, velems );
+   failover->hw->bind_vertex_elements_state( failover->hw, velems );
+}
+
+static void
+failover_delete_vertex_elements_state( struct pipe_context *pipe,
+                                       void *velems )
+{
+   struct fo_state *state = (struct fo_state*)velems;
+   struct failover_context *failover = failover_context(pipe);
+
+   failover->sw->delete_vertex_elements_state(failover->sw, state->sw_state);
+   failover->hw->delete_vertex_elements_state(failover->hw, state->hw_state);
+   state->sw_state = 0;
+   state->hw_state = 0;
+   free(state);
+}
+
 static void 
 failover_set_polygon_stipple( struct pipe_context *pipe,
-			      const struct pipe_poly_stipple *stipple )
+                              const struct pipe_poly_stipple *stipple )
 {
    struct failover_context *failover = failover_context(pipe);
 
@@ -490,22 +533,6 @@ failover_set_vertex_buffers(struct pipe_context *pipe,
 }
 
 
-static void
-failover_set_vertex_elements(struct pipe_context *pipe,
-                             unsigned count,
-                             const struct pipe_vertex_element *vertex_elements)
-{
-   struct failover_context *failover = failover_context(pipe);
-
-   memcpy(failover->vertex_elements, vertex_elements,
-          count * sizeof(vertex_elements[0]));
-
-   failover->dirty |= FO_NEW_VERTEX_ELEMENT;
-   failover->num_vertex_elements = count;
-   failover->sw->set_vertex_elements( failover->sw, count, vertex_elements );
-   failover->hw->set_vertex_elements( failover->hw, count, vertex_elements );
-}
-
 void
 failover_set_constant_buffer(struct pipe_context *pipe,
                              uint shader, uint index,
@@ -543,6 +570,9 @@ failover_init_state_functions( struct failover_context *failover )
    failover->pipe.create_vs_state = failover_create_vs_state;
    failover->pipe.bind_vs_state   = failover_bind_vs_state;
    failover->pipe.delete_vs_state = failover_delete_vs_state;
+   failover->pipe.create_vertex_elements_state = failover_create_vertex_elements_state;
+   failover->pipe.bind_vertex_elements_state = failover_bind_vertex_elements_state;
+   failover->pipe.delete_vertex_elements_state = failover_delete_vertex_elements_state;
 
    failover->pipe.set_blend_color = failover_set_blend_color;
    failover->pipe.set_stencil_ref = failover_set_stencil_ref;
@@ -554,6 +584,5 @@ failover_init_state_functions( struct failover_context *failover )
    failover->pipe.set_vertex_sampler_textures = failover_set_vertex_sampler_textures;
    failover->pipe.set_viewport_state = failover_set_viewport_state;
    failover->pipe.set_vertex_buffers = failover_set_vertex_buffers;
-   failover->pipe.set_vertex_elements = failover_set_vertex_elements;
    failover->pipe.set_constant_buffer = failover_set_constant_buffer;
 }
diff --git a/src/gallium/drivers/failover/fo_state_emit.c b/src/gallium/drivers/failover/fo_state_emit.c
index 5c00080842..09ca194497 100644
--- a/src/gallium/drivers/failover/fo_state_emit.c
+++ b/src/gallium/drivers/failover/fo_state_emit.c
@@ -81,6 +81,10 @@ failover_state_emit( struct failover_context *failover )
       failover->sw->bind_vs_state( failover->sw,
                                    failover->vertex_shader->sw_state );
 
+   if (failover->dirty & FO_NEW_VERTEX_ELEMENT)
+      failover->sw->bind_vertex_elements_state( failover->sw,
+                                                failover->vertex_elements->sw_state );
+
    if (failover->dirty & FO_NEW_STIPPLE)
       failover->sw->set_polygon_stipple( failover->sw, &failover->poly_stipple );
 
@@ -116,11 +120,5 @@ failover_state_emit( struct failover_context *failover )
                                         failover->vertex_buffers );
    }
 
-   if (failover->dirty & FO_NEW_VERTEX_ELEMENT) {
-      failover->sw->set_vertex_elements( failover->sw,
-                                         failover->num_vertex_elements,
-                                         failover->vertex_elements );
-   }
-
    failover->dirty = 0;
 }
diff --git a/src/gallium/drivers/i915/i915_context.h b/src/gallium/drivers/i915/i915_context.h
index 499a727314..4994537683 100644
--- a/src/gallium/drivers/i915/i915_context.h
+++ b/src/gallium/drivers/i915/i915_context.h
@@ -148,7 +148,7 @@ struct i915_state
 
    /** Describes the current hardware vertex layout */
    struct vertex_info vertex_info;
-   
+
    unsigned id;			/* track lost context events */
 };
 
@@ -187,6 +187,11 @@ struct i915_sampler_state {
    unsigned maxlod;
 };
 
+struct i915_velems_state {
+   unsigned count;
+   struct pipe_vertex_element velem[PIPE_MAX_ATTRIBS];
+};
+
 #define I915_MAX_TEXTURE_2D_LEVELS 11  /* max 1024x1024 */
 #define I915_MAX_TEXTURE_3D_LEVELS  8  /* max 128x128x128 */
 
@@ -250,7 +255,6 @@ struct i915_context
 
    unsigned num_samplers;
    unsigned num_textures;
-   unsigned num_vertex_elements;
    unsigned num_vertex_buffers;
 
    struct intel_batchbuffer *batch;
diff --git a/src/gallium/drivers/i915/i915_state.c b/src/gallium/drivers/i915/i915_state.c
index 62169918e2..8927dfc33d 100644
--- a/src/gallium/drivers/i915/i915_state.c
+++ b/src/gallium/drivers/i915/i915_state.c
@@ -742,21 +742,42 @@ static void i915_set_vertex_buffers(struct pipe_context *pipe,
    draw_set_vertex_buffers(i915->draw, count, buffers);
 }
 
-static void i915_set_vertex_elements(struct pipe_context *pipe,
-                                     unsigned count,
-                                     const struct pipe_vertex_element *elements)
+static void *
+i915_create_vertex_elements_state(struct pipe_context *pipe,
+                                  unsigned count,
+                                  const struct pipe_vertex_element *attribs)
+{
+   struct i915_velems_state *velems;
+   assert(count <= PIPE_MAX_ATTRIBS);
+   velems = (struct i915_velems_state *) MALLOC(sizeof(struct i915_velems_state));
+   if (velems) {
+      velems->count = count;
+      memcpy(velems->velem, attribs, sizeof(*attribs) * count);
+   }
+   return velems;
+}
+
+static void
+i915_bind_vertex_elements_state(struct pipe_context *pipe,
+                                void *velems)
 {
    struct i915_context *i915 = i915_context(pipe);
+   struct i915_velems_state *i915_velems = (struct i915_velems_state *) velems;
+
    /* Because we change state before the draw_set_vertex_buffers call
     * we need a flush here, just to be sure.
     */
    draw_flush(i915->draw);
 
-   i915->num_vertex_elements = count;
    /* pass-through to draw module */
-   draw_set_vertex_elements(i915->draw, count, elements);
+   draw_set_vertex_elements(i915->draw, i915_velems->count, i915_velems->velem);
 }
 
+static void
+i915_delete_vertex_elements_state(struct pipe_context *pipe, void *velems)
+{
+   FREE( velems );
+}
 
 void
 i915_init_state_functions( struct i915_context *i915 )
@@ -782,6 +803,9 @@ i915_init_state_functions( struct i915_context *i915 )
    i915->base.create_vs_state = i915_create_vs_state;
    i915->base.bind_vs_state = i915_bind_vs_state;
    i915->base.delete_vs_state = i915_delete_vs_state;
+   i915->base.create_vertex_elements_state = i915_create_vertex_elements_state;
+   i915->base.bind_vertex_elements_state = i915_bind_vertex_elements_state;
+   i915->base.delete_vertex_elements_state = i915_delete_vertex_elements_state;
 
    i915->base.set_blend_color = i915_set_blend_color;
    i915->base.set_stencil_ref = i915_set_stencil_ref;
@@ -794,5 +818,4 @@ i915_init_state_functions( struct i915_context *i915 )
    i915->base.set_fragment_sampler_textures = i915_set_sampler_textures;
    i915->base.set_viewport_state = i915_set_viewport_state;
    i915->base.set_vertex_buffers = i915_set_vertex_buffers;
-   i915->base.set_vertex_elements = i915_set_vertex_elements;
 }
diff --git a/src/gallium/drivers/i965/brw_context.h b/src/gallium/drivers/i965/brw_context.h
index 12cfa7b049..f5b1a06576 100644
--- a/src/gallium/drivers/i965/brw_context.h
+++ b/src/gallium/drivers/i965/brw_context.h
@@ -351,7 +351,7 @@ struct brw_vs_prog_data {
 
 /* Size == 0 if output either not written, or always [0,0,0,1]
  */
-struct brw_vs_ouput_sizes {
+struct brw_vs_output_sizes {
    GLubyte output_size[PIPE_MAX_SHADER_OUTPUTS];
 };
 
@@ -546,14 +546,13 @@ struct brw_context
       const struct brw_blend_state *blend;
       const struct brw_rasterizer_state *rast;
       const struct brw_depth_stencil_state *zstencil;
+      const struct brw_vertex_element_packet *velems;
 
       const struct brw_sampler *sampler[PIPE_MAX_SAMPLERS];
       unsigned num_samplers;
 
       struct pipe_texture *texture[PIPE_MAX_SAMPLERS];
       struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS];
-      struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
-      unsigned num_vertex_elements;
       unsigned num_textures;
       unsigned num_vertex_buffers;
 
diff --git a/src/gallium/drivers/i965/brw_draw_upload.c b/src/gallium/drivers/i965/brw_draw_upload.c
index 9f136eec71..0820ba20a0 100644
--- a/src/gallium/drivers/i965/brw_draw_upload.c
+++ b/src/gallium/drivers/i965/brw_draw_upload.c
@@ -42,141 +42,6 @@
 
 
 
-static unsigned brw_translate_surface_format( unsigned id )
-{
-   switch (id) {
-   case PIPE_FORMAT_R64_FLOAT:
-      return BRW_SURFACEFORMAT_R64_FLOAT;
-   case PIPE_FORMAT_R64G64_FLOAT:
-      return BRW_SURFACEFORMAT_R64G64_FLOAT;
-   case PIPE_FORMAT_R64G64B64_FLOAT:
-      return BRW_SURFACEFORMAT_R64G64B64_FLOAT;
-   case PIPE_FORMAT_R64G64B64A64_FLOAT:
-      return BRW_SURFACEFORMAT_R64G64B64A64_FLOAT;
-
-   case PIPE_FORMAT_R32_FLOAT:
-      return BRW_SURFACEFORMAT_R32_FLOAT;
-   case PIPE_FORMAT_R32G32_FLOAT:
-      return BRW_SURFACEFORMAT_R32G32_FLOAT;
-   case PIPE_FORMAT_R32G32B32_FLOAT:
-      return BRW_SURFACEFORMAT_R32G32B32_FLOAT;
-   case PIPE_FORMAT_R32G32B32A32_FLOAT:
-      return BRW_SURFACEFORMAT_R32G32B32A32_FLOAT;
-
-   case PIPE_FORMAT_R32_UNORM:
-      return BRW_SURFACEFORMAT_R32_UNORM;
-   case PIPE_FORMAT_R32G32_UNORM:
-      return BRW_SURFACEFORMAT_R32G32_UNORM;
-   case PIPE_FORMAT_R32G32B32_UNORM:
-      return BRW_SURFACEFORMAT_R32G32B32_UNORM;
-   case PIPE_FORMAT_R32G32B32A32_UNORM:
-      return BRW_SURFACEFORMAT_R32G32B32A32_UNORM;
-
-   case PIPE_FORMAT_R32_USCALED:
-      return BRW_SURFACEFORMAT_R32_USCALED;
-   case PIPE_FORMAT_R32G32_USCALED:
-      return BRW_SURFACEFORMAT_R32G32_USCALED;
-   case PIPE_FORMAT_R32G32B32_USCALED:
-      return BRW_SURFACEFORMAT_R32G32B32_USCALED;
-   case PIPE_FORMAT_R32G32B32A32_USCALED:
-      return BRW_SURFACEFORMAT_R32G32B32A32_USCALED;
-
-   case PIPE_FORMAT_R32_SNORM:
-      return BRW_SURFACEFORMAT_R32_SNORM;
-   case PIPE_FORMAT_R32G32_SNORM:
-      return BRW_SURFACEFORMAT_R32G32_SNORM;
-   case PIPE_FORMAT_R32G32B32_SNORM:
-      return BRW_SURFACEFORMAT_R32G32B32_SNORM;
-   case PIPE_FORMAT_R32G32B32A32_SNORM:
-      return BRW_SURFACEFORMAT_R32G32B32A32_SNORM;
-
-   case PIPE_FORMAT_R32_SSCALED:
-      return BRW_SURFACEFORMAT_R32_SSCALED;
-   case PIPE_FORMAT_R32G32_SSCALED:
-      return BRW_SURFACEFORMAT_R32G32_SSCALED;
-   case PIPE_FORMAT_R32G32B32_SSCALED:
-      return BRW_SURFACEFORMAT_R32G32B32_SSCALED;
-   case PIPE_FORMAT_R32G32B32A32_SSCALED:
-      return BRW_SURFACEFORMAT_R32G32B32A32_SSCALED;
-
-   case PIPE_FORMAT_R16_UNORM:
-      return BRW_SURFACEFORMAT_R16_UNORM;
-   case PIPE_FORMAT_R16G16_UNORM:
-      return BRW_SURFACEFORMAT_R16G16_UNORM;
-   case PIPE_FORMAT_R16G16B16_UNORM:
-      return BRW_SURFACEFORMAT_R16G16B16_UNORM;
-   case PIPE_FORMAT_R16G16B16A16_UNORM:
-      return BRW_SURFACEFORMAT_R16G16B16A16_UNORM;
-
-   case PIPE_FORMAT_R16_USCALED:
-      return BRW_SURFACEFORMAT_R16_USCALED;
-   case PIPE_FORMAT_R16G16_USCALED:
-      return BRW_SURFACEFORMAT_R16G16_USCALED;
-   case PIPE_FORMAT_R16G16B16_USCALED:
-      return BRW_SURFACEFORMAT_R16G16B16_USCALED;
-   case PIPE_FORMAT_R16G16B16A16_USCALED:
-      return BRW_SURFACEFORMAT_R16G16B16A16_USCALED;
-
-   case PIPE_FORMAT_R16_SNORM:
-      return BRW_SURFACEFORMAT_R16_SNORM;
-   case PIPE_FORMAT_R16G16_SNORM:
-      return BRW_SURFACEFORMAT_R16G16_SNORM;
-   case PIPE_FORMAT_R16G16B16_SNORM:
-      return BRW_SURFACEFORMAT_R16G16B16_SNORM;
-   case PIPE_FORMAT_R16G16B16A16_SNORM:
-      return BRW_SURFACEFORMAT_R16G16B16A16_SNORM;
-
-   case PIPE_FORMAT_R16_SSCALED:
-      return BRW_SURFACEFORMAT_R16_SSCALED;
-   case PIPE_FORMAT_R16G16_SSCALED:
-      return BRW_SURFACEFORMAT_R16G16_SSCALED;
-   case PIPE_FORMAT_R16G16B16_SSCALED:
-      return BRW_SURFACEFORMAT_R16G16B16_SSCALED;
-   case PIPE_FORMAT_R16G16B16A16_SSCALED:
-      return BRW_SURFACEFORMAT_R16G16B16A16_SSCALED;
-
-   case PIPE_FORMAT_R8_UNORM:
-      return BRW_SURFACEFORMAT_R8_UNORM;
-   case PIPE_FORMAT_R8G8_UNORM:
-      return BRW_SURFACEFORMAT_R8G8_UNORM;
-   case PIPE_FORMAT_R8G8B8_UNORM:
-      return BRW_SURFACEFORMAT_R8G8B8_UNORM;
-   case PIPE_FORMAT_R8G8B8A8_UNORM:
-      return BRW_SURFACEFORMAT_R8G8B8A8_UNORM;
-
-   case PIPE_FORMAT_R8_USCALED:
-      return BRW_SURFACEFORMAT_R8_USCALED;
-   case PIPE_FORMAT_R8G8_USCALED:
-      return BRW_SURFACEFORMAT_R8G8_USCALED;
-   case PIPE_FORMAT_R8G8B8_USCALED:
-      return BRW_SURFACEFORMAT_R8G8B8_USCALED;
-   case PIPE_FORMAT_R8G8B8A8_USCALED:
-      return BRW_SURFACEFORMAT_R8G8B8A8_USCALED;
-
-   case PIPE_FORMAT_R8_SNORM:
-      return BRW_SURFACEFORMAT_R8_SNORM;
-   case PIPE_FORMAT_R8G8_SNORM:
-      return BRW_SURFACEFORMAT_R8G8_SNORM;
-   case PIPE_FORMAT_R8G8B8_SNORM:
-      return BRW_SURFACEFORMAT_R8G8B8_SNORM;
-   case PIPE_FORMAT_R8G8B8A8_SNORM:
-      return BRW_SURFACEFORMAT_R8G8B8A8_SNORM;
-
-   case PIPE_FORMAT_R8_SSCALED:
-      return BRW_SURFACEFORMAT_R8_SSCALED;
-   case PIPE_FORMAT_R8G8_SSCALED:
-      return BRW_SURFACEFORMAT_R8G8_SSCALED;
-   case PIPE_FORMAT_R8G8B8_SSCALED:
-      return BRW_SURFACEFORMAT_R8G8B8_SSCALED;
-   case PIPE_FORMAT_R8G8B8A8_SSCALED:
-      return BRW_SURFACEFORMAT_R8G8B8A8_SSCALED;
-
-   default:
-      assert(0);
-      return 0;
-   }
-}
-
 static unsigned get_index_type(int type)
 {
    switch (type) {
@@ -315,75 +180,16 @@ static int brw_emit_vertex_buffers( struct brw_context *brw )
 
 
 
-
 static int brw_emit_vertex_elements(struct brw_context *brw)
 {
-   GLuint nr = brw->curr.num_vertex_elements;
-   GLuint i;
+   const struct brw_vertex_element_packet *brw_velems = brw->curr.velems;
+   unsigned size = brw_velems->header.length + 2;
 
+   /* why is this here */
    brw_emit_query_begin(brw);
 
-   /* If the VS doesn't read any inputs (calculating vertex position from
-    * a state variable for some reason, for example), emit a single pad
-    * VERTEX_ELEMENT struct and bail.
-    *
-    * The stale VB state stays in place, but they don't do anything unless
-    * a VE loads from them.
-    */
-   if (nr == 0) {
-      BEGIN_BATCH(3, IGNORE_CLIPRECTS);
-      OUT_BATCH((CMD_VERTEX_ELEMENT << 16) | 1);
-      OUT_BATCH((0 << BRW_VE0_INDEX_SHIFT) |
-		BRW_VE0_VALID |
-		(BRW_SURFACEFORMAT_R32G32B32A32_FLOAT << BRW_VE0_FORMAT_SHIFT) |
-		(0 << BRW_VE0_SRC_OFFSET_SHIFT));
-      OUT_BATCH((BRW_VE1_COMPONENT_STORE_0 << BRW_VE1_COMPONENT_0_SHIFT) |
-		(BRW_VE1_COMPONENT_STORE_0 << BRW_VE1_COMPONENT_1_SHIFT) |
-		(BRW_VE1_COMPONENT_STORE_0 << BRW_VE1_COMPONENT_2_SHIFT) |
-		(BRW_VE1_COMPONENT_STORE_1_FLT << BRW_VE1_COMPONENT_3_SHIFT));
-      ADVANCE_BATCH();
-      return 0;
-   }
-
-   /* Now emit vertex element (VEP) state packets.
-    *
-    */
-   BEGIN_BATCH(1 + nr * 2, IGNORE_CLIPRECTS);
-   OUT_BATCH((CMD_VERTEX_ELEMENT << 16) | ((1 + nr * 2) - 2));
-   for (i = 0; i < nr; i++) {
-      const struct pipe_vertex_element *input = &brw->curr.vertex_element[i];
-      uint32_t format = brw_translate_surface_format( input->src_format );
-      uint32_t comp0 = BRW_VE1_COMPONENT_STORE_SRC;
-      uint32_t comp1 = BRW_VE1_COMPONENT_STORE_SRC;
-      uint32_t comp2 = BRW_VE1_COMPONENT_STORE_SRC;
-      uint32_t comp3 = BRW_VE1_COMPONENT_STORE_SRC;
-
-      switch (input->nr_components) {
-      case 0: comp0 = BRW_VE1_COMPONENT_STORE_0; /* fallthrough */
-      case 1: comp1 = BRW_VE1_COMPONENT_STORE_0; /* fallthrough */
-      case 2: comp2 = BRW_VE1_COMPONENT_STORE_0; /* fallthrough */
-      case 3: comp3 = BRW_VE1_COMPONENT_STORE_1_FLT;
-	 break;
-      }
-
-      OUT_BATCH((input->vertex_buffer_index << BRW_VE0_INDEX_SHIFT) |
-		BRW_VE0_VALID |
-		(format << BRW_VE0_FORMAT_SHIFT) |
-		(input->src_offset << BRW_VE0_SRC_OFFSET_SHIFT));
+   brw_batchbuffer_data(brw->batch, brw_velems, size * 4, IGNORE_CLIPRECTS);
 
-      if (BRW_IS_IGDNG(brw))
-          OUT_BATCH((comp0 << BRW_VE1_COMPONENT_0_SHIFT) |
-                    (comp1 << BRW_VE1_COMPONENT_1_SHIFT) |
-                    (comp2 << BRW_VE1_COMPONENT_2_SHIFT) |
-                    (comp3 << BRW_VE1_COMPONENT_3_SHIFT));
-      else
-          OUT_BATCH((comp0 << BRW_VE1_COMPONENT_0_SHIFT) |
-                    (comp1 << BRW_VE1_COMPONENT_1_SHIFT) |
-                    (comp2 << BRW_VE1_COMPONENT_2_SHIFT) |
-                    (comp3 << BRW_VE1_COMPONENT_3_SHIFT) |
-                    ((i * 4) << BRW_VE1_DST_OFFSET_SHIFT));
-   }
-   ADVANCE_BATCH();
    return 0;
 }
 
@@ -396,10 +202,11 @@ static int brw_emit_vertices( struct brw_context *brw )
    if (ret)
       return ret;
 
+   /* XXX should separate this? */
    ret = brw_emit_vertex_elements( brw );
    if (ret)
       return ret;
-   
+
    return 0;
 }
 
@@ -407,7 +214,8 @@ static int brw_emit_vertices( struct brw_context *brw )
 const struct brw_tracked_state brw_vertices = {
    .dirty = {
       .mesa = (PIPE_NEW_INDEX_RANGE |
-               PIPE_NEW_VERTEX_BUFFER),
+               PIPE_NEW_VERTEX_BUFFER |
+               PIPE_NEW_VERTEX_ELEMENT),
       .brw = BRW_NEW_BATCH,
       .cache = 0,
    },
diff --git a/src/gallium/drivers/i965/brw_pipe_vertex.c b/src/gallium/drivers/i965/brw_pipe_vertex.c
index e3c48e3149..d6a840857e 100644
--- a/src/gallium/drivers/i965/brw_pipe_vertex.c
+++ b/src/gallium/drivers/i965/brw_pipe_vertex.c
@@ -1,22 +1,251 @@
 #include "brw_context.h"
+#include "brw_defines.h"
+#include "brw_structs.h"
 
+#include "util/u_memory.h"
+#include "util/u_format.h"
 
-static void brw_set_vertex_elements( struct pipe_context *pipe,
-				     unsigned count,
-				     const struct pipe_vertex_element *elements )
+
+static unsigned brw_translate_surface_format( unsigned id )
+{
+   switch (id) {
+   case PIPE_FORMAT_R64_FLOAT:
+      return BRW_SURFACEFORMAT_R64_FLOAT;
+   case PIPE_FORMAT_R64G64_FLOAT:
+      return BRW_SURFACEFORMAT_R64G64_FLOAT;
+   case PIPE_FORMAT_R64G64B64_FLOAT:
+      return BRW_SURFACEFORMAT_R64G64B64_FLOAT;
+   case PIPE_FORMAT_R64G64B64A64_FLOAT:
+      return BRW_SURFACEFORMAT_R64G64B64A64_FLOAT;
+
+   case PIPE_FORMAT_R32_FLOAT:
+      return BRW_SURFACEFORMAT_R32_FLOAT;
+   case PIPE_FORMAT_R32G32_FLOAT:
+      return BRW_SURFACEFORMAT_R32G32_FLOAT;
+   case PIPE_FORMAT_R32G32B32_FLOAT:
+      return BRW_SURFACEFORMAT_R32G32B32_FLOAT;
+   case PIPE_FORMAT_R32G32B32A32_FLOAT:
+      return BRW_SURFACEFORMAT_R32G32B32A32_FLOAT;
+
+   case PIPE_FORMAT_R32_UNORM:
+      return BRW_SURFACEFORMAT_R32_UNORM;
+   case PIPE_FORMAT_R32G32_UNORM:
+      return BRW_SURFACEFORMAT_R32G32_UNORM;
+   case PIPE_FORMAT_R32G32B32_UNORM:
+      return BRW_SURFACEFORMAT_R32G32B32_UNORM;
+   case PIPE_FORMAT_R32G32B32A32_UNORM:
+      return BRW_SURFACEFORMAT_R32G32B32A32_UNORM;
+
+   case PIPE_FORMAT_R32_USCALED:
+      return BRW_SURFACEFORMAT_R32_USCALED;
+   case PIPE_FORMAT_R32G32_USCALED:
+      return BRW_SURFACEFORMAT_R32G32_USCALED;
+   case PIPE_FORMAT_R32G32B32_USCALED:
+      return BRW_SURFACEFORMAT_R32G32B32_USCALED;
+   case PIPE_FORMAT_R32G32B32A32_USCALED:
+      return BRW_SURFACEFORMAT_R32G32B32A32_USCALED;
+
+   case PIPE_FORMAT_R32_SNORM:
+      return BRW_SURFACEFORMAT_R32_SNORM;
+   case PIPE_FORMAT_R32G32_SNORM:
+      return BRW_SURFACEFORMAT_R32G32_SNORM;
+   case PIPE_FORMAT_R32G32B32_SNORM:
+      return BRW_SURFACEFORMAT_R32G32B32_SNORM;
+   case PIPE_FORMAT_R32G32B32A32_SNORM:
+      return BRW_SURFACEFORMAT_R32G32B32A32_SNORM;
+
+   case PIPE_FORMAT_R32_SSCALED:
+      return BRW_SURFACEFORMAT_R32_SSCALED;
+   case PIPE_FORMAT_R32G32_SSCALED:
+      return BRW_SURFACEFORMAT_R32G32_SSCALED;
+   case PIPE_FORMAT_R32G32B32_SSCALED:
+      return BRW_SURFACEFORMAT_R32G32B32_SSCALED;
+   case PIPE_FORMAT_R32G32B32A32_SSCALED:
+      return BRW_SURFACEFORMAT_R32G32B32A32_SSCALED;
+
+   case PIPE_FORMAT_R16_UNORM:
+      return BRW_SURFACEFORMAT_R16_UNORM;
+   case PIPE_FORMAT_R16G16_UNORM:
+      return BRW_SURFACEFORMAT_R16G16_UNORM;
+   case PIPE_FORMAT_R16G16B16_UNORM:
+      return BRW_SURFACEFORMAT_R16G16B16_UNORM;
+   case PIPE_FORMAT_R16G16B16A16_UNORM:
+      return BRW_SURFACEFORMAT_R16G16B16A16_UNORM;
+
+   case PIPE_FORMAT_R16_USCALED:
+      return BRW_SURFACEFORMAT_R16_USCALED;
+   case PIPE_FORMAT_R16G16_USCALED:
+      return BRW_SURFACEFORMAT_R16G16_USCALED;
+   case PIPE_FORMAT_R16G16B16_USCALED:
+      return BRW_SURFACEFORMAT_R16G16B16_USCALED;
+   case PIPE_FORMAT_R16G16B16A16_USCALED:
+      return BRW_SURFACEFORMAT_R16G16B16A16_USCALED;
+
+   case PIPE_FORMAT_R16_SNORM:
+      return BRW_SURFACEFORMAT_R16_SNORM;
+   case PIPE_FORMAT_R16G16_SNORM:
+      return BRW_SURFACEFORMAT_R16G16_SNORM;
+   case PIPE_FORMAT_R16G16B16_SNORM:
+      return BRW_SURFACEFORMAT_R16G16B16_SNORM;
+   case PIPE_FORMAT_R16G16B16A16_SNORM:
+      return BRW_SURFACEFORMAT_R16G16B16A16_SNORM;
+
+   case PIPE_FORMAT_R16_SSCALED:
+      return BRW_SURFACEFORMAT_R16_SSCALED;
+   case PIPE_FORMAT_R16G16_SSCALED:
+      return BRW_SURFACEFORMAT_R16G16_SSCALED;
+   case PIPE_FORMAT_R16G16B16_SSCALED:
+      return BRW_SURFACEFORMAT_R16G16B16_SSCALED;
+   case PIPE_FORMAT_R16G16B16A16_SSCALED:
+      return BRW_SURFACEFORMAT_R16G16B16A16_SSCALED;
+
+   case PIPE_FORMAT_R8_UNORM:
+      return BRW_SURFACEFORMAT_R8_UNORM;
+   case PIPE_FORMAT_R8G8_UNORM:
+      return BRW_SURFACEFORMAT_R8G8_UNORM;
+   case PIPE_FORMAT_R8G8B8_UNORM:
+      return BRW_SURFACEFORMAT_R8G8B8_UNORM;
+   case PIPE_FORMAT_R8G8B8A8_UNORM:
+      return BRW_SURFACEFORMAT_R8G8B8A8_UNORM;
+
+   case PIPE_FORMAT_R8_USCALED:
+      return BRW_SURFACEFORMAT_R8_USCALED;
+   case PIPE_FORMAT_R8G8_USCALED:
+      return BRW_SURFACEFORMAT_R8G8_USCALED;
+   case PIPE_FORMAT_R8G8B8_USCALED:
+      return BRW_SURFACEFORMAT_R8G8B8_USCALED;
+   case PIPE_FORMAT_R8G8B8A8_USCALED:
+      return BRW_SURFACEFORMAT_R8G8B8A8_USCALED;
+
+   case PIPE_FORMAT_R8_SNORM:
+      return BRW_SURFACEFORMAT_R8_SNORM;
+   case PIPE_FORMAT_R8G8_SNORM:
+      return BRW_SURFACEFORMAT_R8G8_SNORM;
+   case PIPE_FORMAT_R8G8B8_SNORM:
+      return BRW_SURFACEFORMAT_R8G8B8_SNORM;
+   case PIPE_FORMAT_R8G8B8A8_SNORM:
+      return BRW_SURFACEFORMAT_R8G8B8A8_SNORM;
+
+   case PIPE_FORMAT_R8_SSCALED:
+      return BRW_SURFACEFORMAT_R8_SSCALED;
+   case PIPE_FORMAT_R8G8_SSCALED:
+      return BRW_SURFACEFORMAT_R8G8_SSCALED;
+   case PIPE_FORMAT_R8G8B8_SSCALED:
+      return BRW_SURFACEFORMAT_R8G8B8_SSCALED;
+   case PIPE_FORMAT_R8G8B8A8_SSCALED:
+      return BRW_SURFACEFORMAT_R8G8B8A8_SSCALED;
+
+   default:
+      assert(0);
+      return 0;
+   }
+}
+
+static void brw_translate_vertex_elements(struct brw_context *brw,
+                                          struct brw_vertex_element_packet *brw_velems,
+                                          const struct pipe_vertex_element *attribs,
+                                          unsigned count)
+{
+   unsigned i;
+
+   /* If the VS doesn't read any inputs (calculating vertex position from
+    * a state variable for some reason, for example), emit a single pad
+    * VERTEX_ELEMENT struct and bail.
+    *
+    * The stale VB state stays in place, but they don't do anything unless
+    * a VE loads from them.
+    */
+   brw_velems->header.opcode = CMD_VERTEX_ELEMENT;
+
+   if (count == 0) {
+      brw_velems->header.length = 1;
+      brw_velems->ve[0].ve0.src_offset = 0;
+      brw_velems->ve[0].ve0.src_format = BRW_SURFACEFORMAT_R32G32B32A32_FLOAT;
+      brw_velems->ve[0].ve0.valid = 1;
+      brw_velems->ve[0].ve0.vertex_buffer_index = 0;
+      brw_velems->ve[0].ve1.dst_offset = 0;
+      brw_velems->ve[0].ve1.vfcomponent0 = BRW_VE1_COMPONENT_STORE_0;
+      brw_velems->ve[0].ve1.vfcomponent1 = BRW_VE1_COMPONENT_STORE_0;
+      brw_velems->ve[0].ve1.vfcomponent2 = BRW_VE1_COMPONENT_STORE_0;
+      brw_velems->ve[0].ve1.vfcomponent3 = BRW_VE1_COMPONENT_STORE_1_FLT;
+      return;
+   }
+
+
+   /* Now emit vertex element (VEP) state packets.
+    *
+    */
+   brw_velems->header.length = (1 + count * 2) - 2;
+   for (i = 0; i < count; i++) {
+      const struct pipe_vertex_element *input = &attribs[i];
+      unsigned nr_components = util_format_get_nr_components(input->src_format);
+
+      uint32_t format = brw_translate_surface_format( input->src_format );
+      uint32_t comp0 = BRW_VE1_COMPONENT_STORE_SRC;
+      uint32_t comp1 = BRW_VE1_COMPONENT_STORE_SRC;
+      uint32_t comp2 = BRW_VE1_COMPONENT_STORE_SRC;
+      uint32_t comp3 = BRW_VE1_COMPONENT_STORE_SRC;
+
+      switch (nr_components) {
+      case 0: comp0 = BRW_VE1_COMPONENT_STORE_0; /* fallthrough */
+      case 1: comp1 = BRW_VE1_COMPONENT_STORE_0; /* fallthrough */
+      case 2: comp2 = BRW_VE1_COMPONENT_STORE_0; /* fallthrough */
+      case 3: comp3 = BRW_VE1_COMPONENT_STORE_1_FLT;
+         break;
+      }
+
+      brw_velems->ve[i].ve0.src_offset = input->src_offset;
+      brw_velems->ve[i].ve0.src_format = format;
+      brw_velems->ve[i].ve0.valid = 1;
+      brw_velems->ve[i].ve0.vertex_buffer_index = input->vertex_buffer_index;
+      brw_velems->ve[i].ve1.vfcomponent0 = comp0;
+      brw_velems->ve[i].ve1.vfcomponent1 = comp1;
+      brw_velems->ve[i].ve1.vfcomponent2 = comp2;
+      brw_velems->ve[i].ve1.vfcomponent3 = comp3;
+
+      if (BRW_IS_IGDNG(brw))
+         brw_velems->ve[i].ve1.dst_offset = 0;
+      else
+         brw_velems->ve[i].ve1.dst_offset = i * 4;
+   }
+}
+
+static void* brw_create_vertex_elements_state( struct pipe_context *pipe,
+                                               unsigned count,
+                                               const struct pipe_vertex_element *attribs )
 {
+   /* note: for the brw_swtnl.c code (if ever we need draw fallback) we'd also need
+      to store the original data */
    struct brw_context *brw = brw_context(pipe);
+   struct brw_vertex_element_packet *velems;
+   assert(count <= BRW_VEP_MAX);
+   velems = (struct brw_vertex_element_packet *) MALLOC(sizeof(struct brw_vertex_element_packet));
+   if (velems) {
+      brw_translate_vertex_elements(brw, velems, attribs, count);
+   }
+   return velems;
+}
 
-   memcpy(brw->curr.vertex_element, elements, count * sizeof(elements[0]));
-   brw->curr.num_vertex_elements = count;
+static void brw_bind_vertex_elements_state(struct pipe_context *pipe,
+                                           void *velems)
+{
+   struct brw_context *brw = brw_context(pipe);
+   struct brw_vertex_element_packet *brw_velems = (struct brw_vertex_element_packet *) velems;
+
+   brw->curr.velems = brw_velems;
 
    brw->state.dirty.mesa |= PIPE_NEW_VERTEX_ELEMENT;
 }
 
+static void brw_delete_vertex_elements_state(struct pipe_context *pipe, void *velems)
+{
+   FREE( velems );
+}
+
 
 static void brw_set_vertex_buffers(struct pipe_context *pipe,
-				   unsigned count,
-				   const struct pipe_vertex_buffer *buffers)
+                                   unsigned count,
+                                   const struct pipe_vertex_buffer *buffers)
 {
    struct brw_context *brw = brw_context(pipe);
    unsigned i;
@@ -49,7 +278,9 @@ void
 brw_pipe_vertex_init( struct brw_context *brw )
 {
    brw->base.set_vertex_buffers = brw_set_vertex_buffers;
-   brw->base.set_vertex_elements = brw_set_vertex_elements;
+   brw->base.create_vertex_elements_state = brw_create_vertex_elements_state;
+   brw->base.bind_vertex_elements_state = brw_bind_vertex_elements_state;
+   brw->base.delete_vertex_elements_state = brw_delete_vertex_elements_state;
 }
 
 
diff --git a/src/gallium/drivers/i965/brw_structs.h b/src/gallium/drivers/i965/brw_structs.h
index bf10bc04de..e97ddeb5e1 100644
--- a/src/gallium/drivers/i965/brw_structs.h
+++ b/src/gallium/drivers/i965/brw_structs.h
@@ -28,7 +28,7 @@
   * Authors:
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
-        
+
 
 #ifndef BRW_STRUCTS_H
 #define BRW_STRUCTS_H
@@ -1149,7 +1149,7 @@ struct brw_vertex_element_state
       GLuint valid:1; 
       GLuint vertex_buffer_index:5; 
    } ve0;
-   
+
    struct
    {
       GLuint dst_offset:8; 
diff --git a/src/gallium/drivers/identity/id_context.c b/src/gallium/drivers/identity/id_context.c
index 8248b2a413..05a9dfab77 100644
--- a/src/gallium/drivers/identity/id_context.c
+++ b/src/gallium/drivers/identity/id_context.c
@@ -377,6 +377,42 @@ identity_delete_vs_state(struct pipe_context *_pipe,
                          vs);
 }
 
+
+static void
+identity_create_vertex_elements_state(struct pipe_context *_pipe,
+                                      unsigned num_elements,
+                                      const struct pipe_vertex_element *vertex_elements)
+{
+   struct identity_context *id_pipe = identity_context(_pipe);
+   struct pipe_context *pipe = id_pipe->pipe;
+
+   pipe->create_vertex_elements_state(pipe,
+                                      num_elements,
+                                      vertex_elements);
+}
+
+static void
+identity_bind_vertex_elements_state(struct pipe_context *_pipe,
+                                    void *velems)
+{
+   struct identity_context *id_pipe = identity_context(_pipe);
+   struct pipe_context *pipe = id_pipe->pipe;
+
+   pipe->bind_vertex_elements_state(pipe,
+                                    velems);
+}
+
+static void
+identity_delete_vertex_elements_state(struct pipe_context *_pipe,
+                                      void *velems)
+{
+   struct identity_context *id_pipe = identity_context(_pipe);
+   struct pipe_context *pipe = id_pipe->pipe;
+
+   pipe->delete_vertex_elements_state(pipe,
+                                      velems);
+}
+
 static void
 identity_set_blend_color(struct pipe_context *_pipe,
                          const struct pipe_blend_color *blend_color)
@@ -563,20 +599,6 @@ identity_set_vertex_buffers(struct pipe_context *_pipe,
                             num_buffers,
                             buffers);
 }
-
-static void
-identity_set_vertex_elements(struct pipe_context *_pipe,
-                             unsigned num_elements,
-                             const struct pipe_vertex_element *vertex_elements)
-{
-   struct identity_context *id_pipe = identity_context(_pipe);
-   struct pipe_context *pipe = id_pipe->pipe;
-
-   pipe->set_vertex_elements(pipe,
-                             num_elements,
-                             vertex_elements);
-}
-
 static void
 identity_surface_copy(struct pipe_context *_pipe,
                       struct pipe_surface *_dst,
@@ -733,6 +755,9 @@ identity_context_create(struct pipe_screen *_screen, struct pipe_context *pipe)
    id_pipe->base.create_vs_state = identity_create_vs_state;
    id_pipe->base.bind_vs_state = identity_bind_vs_state;
    id_pipe->base.delete_vs_state = identity_delete_vs_state;
+   id_pipe->base.create_vertex_elements_state = identity_create_vertex_elements_state;
+   id_pipe->base.bind_vertex_elements_state = identity_bind_vertex_elements_state;
+   id_pipe->base.delete_vertex_elements_state = identity_delete_vertex_elements_state;
    id_pipe->base.set_blend_color = identity_set_blend_color;
    id_pipe->base.set_stencil_ref = identity_set_stencil_ref;
    id_pipe->base.set_clip_state = identity_set_clip_state;
@@ -744,7 +769,6 @@ identity_context_create(struct pipe_screen *_screen, struct pipe_context *pipe)
    id_pipe->base.set_fragment_sampler_textures = identity_set_fragment_sampler_textures;
    id_pipe->base.set_vertex_sampler_textures = identity_set_vertex_sampler_textures;
    id_pipe->base.set_vertex_buffers = identity_set_vertex_buffers;
-   id_pipe->base.set_vertex_elements = identity_set_vertex_elements;
    id_pipe->base.surface_copy = identity_surface_copy;
    id_pipe->base.surface_fill = identity_surface_fill;
    id_pipe->base.clear = identity_clear;
diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c
index e31ae6a3fc..d94efec16a 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.c
+++ b/src/gallium/drivers/llvmpipe/lp_context.c
@@ -145,6 +145,10 @@ llvmpipe_create_context( struct pipe_screen *screen, void *priv )
    llvmpipe->pipe.bind_vs_state   = llvmpipe_bind_vs_state;
    llvmpipe->pipe.delete_vs_state = llvmpipe_delete_vs_state;
 
+   llvmpipe->pipe.create_vertex_elements_state = llvmpipe_create_vertex_elements_state;
+   llvmpipe->pipe.bind_vertex_elements_state = llvmpipe_bind_vertex_elements_state;
+   llvmpipe->pipe.delete_vertex_elements_state = llvmpipe_delete_vertex_elements_state;
+
    llvmpipe->pipe.set_blend_color = llvmpipe_set_blend_color;
    llvmpipe->pipe.set_stencil_ref = llvmpipe_set_stencil_ref;
    llvmpipe->pipe.set_clip_state = llvmpipe_set_clip_state;
@@ -157,7 +161,6 @@ llvmpipe_create_context( struct pipe_screen *screen, void *priv )
    llvmpipe->pipe.set_viewport_state = llvmpipe_set_viewport_state;
 
    llvmpipe->pipe.set_vertex_buffers = llvmpipe_set_vertex_buffers;
-   llvmpipe->pipe.set_vertex_elements = llvmpipe_set_vertex_elements;
 
    llvmpipe->pipe.draw_arrays = llvmpipe_draw_arrays;
    llvmpipe->pipe.draw_elements = llvmpipe_draw_elements;
diff --git a/src/gallium/drivers/llvmpipe/lp_context.h b/src/gallium/drivers/llvmpipe/lp_context.h
index 955c7eb8e0..217ec59b68 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_context.h
@@ -46,6 +46,7 @@ struct lp_fragment_shader;
 struct lp_vertex_shader;
 struct lp_blend_state;
 struct setup_context;
+struct lp_velems_state;
 
 struct llvmpipe_context {
    struct pipe_context pipe;  /**< base class */
@@ -58,6 +59,7 @@ struct llvmpipe_context {
    const struct pipe_rasterizer_state *rasterizer;
    struct lp_fragment_shader *fs;
    const struct lp_vertex_shader *vs;
+   const struct lp_velems_state *velems;
 
    /** Other rendering state */
    struct pipe_blend_color blend_color;
@@ -71,13 +73,11 @@ struct llvmpipe_context {
    struct pipe_texture *vertex_textures[PIPE_MAX_VERTEX_SAMPLERS];
    struct pipe_viewport_state viewport;
    struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS];
-   struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
 
    unsigned num_samplers;
    unsigned num_textures;
    unsigned num_vertex_samplers;
    unsigned num_vertex_textures;
-   unsigned num_vertex_elements;
    unsigned num_vertex_buffers;
 
    unsigned dirty; /**< Mask of LP_NEW_x flags */
diff --git a/src/gallium/drivers/llvmpipe/lp_state.h b/src/gallium/drivers/llvmpipe/lp_state.h
index 9beba32271..6dbdc195bf 100644
--- a/src/gallium/drivers/llvmpipe/lp_state.h
+++ b/src/gallium/drivers/llvmpipe/lp_state.h
@@ -119,6 +119,10 @@ struct lp_vertex_shader {
    struct draw_vertex_shader *draw_data;
 };
 
+struct lp_velems_state {
+   unsigned count;
+   struct pipe_vertex_element velem[PIPE_MAX_ATTRIBS];
+};
 
 
 void *
@@ -176,8 +180,14 @@ void *llvmpipe_create_vs_state(struct pipe_context *,
 void llvmpipe_bind_vs_state(struct pipe_context *, void *);
 void llvmpipe_delete_vs_state(struct pipe_context *, void *);
 
+void *llvmpipe_create_vertex_elements_state(struct pipe_context *,
+                                            unsigned count,
+                                            const struct pipe_vertex_element *);
+void llvmpipe_bind_vertex_elements_state(struct pipe_context *, void *);
+void llvmpipe_delete_vertex_elements_state(struct pipe_context *, void *);
+
 void llvmpipe_set_polygon_stipple( struct pipe_context *,
-				  const struct pipe_poly_stipple * );
+                                   const struct pipe_poly_stipple * );
 
 void llvmpipe_set_scissor_state( struct pipe_context *,
                                  const struct pipe_scissor_state * );
@@ -194,10 +204,6 @@ llvmpipe_set_vertex_sampler_textures(struct pipe_context *,
 void llvmpipe_set_viewport_state( struct pipe_context *,
                                   const struct pipe_viewport_state * );
 
-void llvmpipe_set_vertex_elements(struct pipe_context *,
-                                  unsigned count,
-                                  const struct pipe_vertex_element *);
-
 void llvmpipe_set_vertex_buffers(struct pipe_context *,
                                  unsigned count,
                                  const struct pipe_vertex_buffer *);
diff --git a/src/gallium/drivers/llvmpipe/lp_state_vertex.c b/src/gallium/drivers/llvmpipe/lp_state_vertex.c
index 57ac25ea0c..2ddd110a5f 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_vertex.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_vertex.c
@@ -35,24 +35,40 @@
 #include "draw/draw_context.h"
 
 
+void *
+llvmpipe_create_vertex_elements_state(struct pipe_context *pipe,
+                                      unsigned count,
+                                      const struct pipe_vertex_element *attribs)
+{
+   struct lp_velems_state *velems;
+   assert(count <= PIPE_MAX_ATTRIBS);
+   velems = (struct lp_velems_state *) MALLOC(sizeof(struct lp_velems_state));
+   if (velems) {
+      velems->count = count;
+      memcpy(velems->velem, attribs, sizeof(*attribs) * count);
+   }
+   return velems;
+}
+
 void
-llvmpipe_set_vertex_elements(struct pipe_context *pipe,
-                             unsigned count,
-                             const struct pipe_vertex_element *attribs)
+llvmpipe_bind_vertex_elements_state(struct pipe_context *pipe,
+                                    void *velems)
 {
    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+   struct lp_velems_state *lp_velems = (struct lp_velems_state *) velems;
 
-   assert(count <= PIPE_MAX_ATTRIBS);
-
-   memcpy(llvmpipe->vertex_element, attribs,
-          count * sizeof(struct pipe_vertex_element));
-   llvmpipe->num_vertex_elements = count;
+   llvmpipe->velems = lp_velems;
 
    llvmpipe->dirty |= LP_NEW_VERTEX;
 
-   draw_set_vertex_elements(llvmpipe->draw, count, attribs);
+   draw_set_vertex_elements(llvmpipe->draw, lp_velems->count, lp_velems->velem);
 }
 
+void
+llvmpipe_delete_vertex_elements_state(struct pipe_context *pipe, void *velems)
+{
+   FREE( velems );
+}
 
 void
 llvmpipe_set_vertex_buffers(struct pipe_context *pipe,
diff --git a/src/gallium/drivers/nv30/nv30_context.h b/src/gallium/drivers/nv30/nv30_context.h
index ea259aadf3..1786460aec 100644
--- a/src/gallium/drivers/nv30/nv30_context.h
+++ b/src/gallium/drivers/nv30/nv30_context.h
@@ -107,6 +107,11 @@ struct nv30_state {
 	struct nouveau_stateobj *hw[NV30_STATE_MAX];
 };
 
+struct nv30_vtxelt_state {
+	struct pipe_vertex_element pipe[16];
+	unsigned num_elements;
+};
+
 struct nv30_context {
 	struct pipe_context pipe;
 
@@ -142,8 +147,7 @@ struct nv30_context {
 	unsigned dirty_samplers;
 	struct pipe_vertex_buffer vtxbuf[PIPE_MAX_ATTRIBS];
 	unsigned vtxbuf_nr;
-	struct pipe_vertex_element vtxelt[PIPE_MAX_ATTRIBS];
-	unsigned vtxelt_nr;
+	struct nv30_vtxelt_state *vtxelt;
 };
 
 static INLINE struct nv30_context *
diff --git a/src/gallium/drivers/nv30/nv30_state.c b/src/gallium/drivers/nv30/nv30_state.c
index d911c80707..24b15a63ac 100644
--- a/src/gallium/drivers/nv30/nv30_state.c
+++ b/src/gallium/drivers/nv30/nv30_state.c
@@ -669,15 +669,34 @@ nv30_set_vertex_buffers(struct pipe_context *pipe, unsigned count,
 	/*nv30->draw_dirty |= NV30_NEW_ARRAYS;*/
 }
 
+static void *
+nv30_vtxelts_state_create(struct pipe_context *pipe,
+			  unsigned num_elements,
+			  const struct pipe_vertex_element *elements)
+{
+	struct nv30_vtxelt_state *cso = CALLOC_STRUCT(nv30_vtxelt_state);
+
+	assert(num_elements < 16); /* not doing fallbacks yet */
+	cso->num_elements = num_elements;
+	memcpy(cso->pipe, elements, num_elements * sizeof(*elements));
+
+/*	nv30_vtxelt_construct(cso);*/
+
+	return (void *)cso;
+}
+
 static void
-nv30_set_vertex_elements(struct pipe_context *pipe, unsigned count,
-			 const struct pipe_vertex_element *ve)
+nv30_vtxelts_state_delete(struct pipe_context *pipe, void *hwcso)
 {
-	struct nv30_context *nv30 = nv30_context(pipe);
+	FREE(hwcso);
+}
 
-	memcpy(nv30->vtxelt, ve, sizeof(*ve) * count);
-	nv30->vtxelt_nr = count;
+static void
+nv30_vtxelts_state_bind(struct pipe_context *pipe, void *hwcso)
+{
+	struct nv30_context *nv30 = nv30_context(pipe);
 
+	nv30->vtxelt = hwcso;
 	nv30->dirty |= NV30_NEW_ARRAYS;
 	/*nv30->draw_dirty |= NV30_NEW_ARRAYS;*/
 }
@@ -722,7 +741,10 @@ nv30_init_state_functions(struct nv30_context *nv30)
 	nv30->pipe.set_scissor_state = nv30_set_scissor_state;
 	nv30->pipe.set_viewport_state = nv30_set_viewport_state;
 
+	nv30->pipe.create_vertex_elements_state = nv30_vtxelts_state_create;
+	nv30->pipe.delete_vertex_elements_state = nv30_vtxelts_state_delete;
+	nv30->pipe.bind_vertex_elements_state = nv30_vtxelts_state_bind;
+
 	nv30->pipe.set_vertex_buffers = nv30_set_vertex_buffers;
-	nv30->pipe.set_vertex_elements = nv30_set_vertex_elements;
 }
 
diff --git a/src/gallium/drivers/nv30/nv30_vbo.c b/src/gallium/drivers/nv30/nv30_vbo.c
index e48823a913..f3856bb5a5 100644
--- a/src/gallium/drivers/nv30/nv30_vbo.c
+++ b/src/gallium/drivers/nv30/nv30_vbo.c
@@ -492,16 +492,16 @@ nv30_vbo_validate(struct nv30_context *nv30)
 	int hw;
 
 	vtxbuf = so_new(3, 17, 18);
-	so_method(vtxbuf, rankine, NV34TCL_VTXBUF_ADDRESS(0), nv30->vtxelt_nr);
+	so_method(vtxbuf, rankine, NV34TCL_VTXBUF_ADDRESS(0), nv30->vtxelt->num_elements);
 	vtxfmt = so_new(1, 16, 0);
-	so_method(vtxfmt, rankine, NV34TCL_VTXFMT(0), nv30->vtxelt_nr);
+	so_method(vtxfmt, rankine, NV34TCL_VTXFMT(0), nv30->vtxelt->num_elements);
 
-	for (hw = 0; hw < nv30->vtxelt_nr; hw++) {
+	for (hw = 0; hw < nv30->vtxelt->num_elements; hw++) {
 		struct pipe_vertex_element *ve;
 		struct pipe_vertex_buffer *vb;
 		unsigned type, ncomp;
 
-		ve = &nv30->vtxelt[hw];
+		ve = &nv30->vtxelt->pipe[hw];
 		vb = &nv30->vtxbuf[ve->vertex_buffer_index];
 
 		if (!vb->stride) {
diff --git a/src/gallium/drivers/nv40/nv40_context.h b/src/gallium/drivers/nv40/nv40_context.h
index 97fb6a2ef9..2550ec654b 100644
--- a/src/gallium/drivers/nv40/nv40_context.h
+++ b/src/gallium/drivers/nv40/nv40_context.h
@@ -107,6 +107,12 @@ struct nv40_state {
 	struct nouveau_stateobj *hw[NV40_STATE_MAX];
 };
 
+
+struct nv40_vtxelt_state {
+	struct pipe_vertex_element pipe[16];
+	unsigned num_elements;
+};
+
 struct nv40_context {
 	struct pipe_context pipe;
 
@@ -157,8 +163,7 @@ struct nv40_context {
 	unsigned dirty_samplers;
 	struct pipe_vertex_buffer vtxbuf[PIPE_MAX_ATTRIBS];
 	unsigned vtxbuf_nr;
-	struct pipe_vertex_element vtxelt[PIPE_MAX_ATTRIBS];
-	unsigned vtxelt_nr;
+	struct nv40_vtxelt_state *vtxelt;
 };
 
 static INLINE struct nv40_context *
diff --git a/src/gallium/drivers/nv40/nv40_state.c b/src/gallium/drivers/nv40/nv40_state.c
index 4f28675e64..ee471e6ce4 100644
--- a/src/gallium/drivers/nv40/nv40_state.c
+++ b/src/gallium/drivers/nv40/nv40_state.c
@@ -684,15 +684,34 @@ nv40_set_vertex_buffers(struct pipe_context *pipe, unsigned count,
 	nv40->draw_dirty |= NV40_NEW_ARRAYS;
 }
 
+static void *
+nv40_vtxelts_state_create(struct pipe_context *pipe,
+			  unsigned num_elements,
+			  const struct pipe_vertex_element *elements)
+{
+	struct nv40_vtxelt_state *cso = CALLOC_STRUCT(nv40_vtxelt_state);
+
+	assert(num_elements < 16); /* not doing fallbacks yet */
+	cso->num_elements = num_elements;
+	memcpy(cso->pipe, elements, num_elements * sizeof(*elements));
+
+/*	nv40_vtxelt_construct(cso);*/
+
+	return (void *)cso;
+}
+
 static void
-nv40_set_vertex_elements(struct pipe_context *pipe, unsigned count,
-			 const struct pipe_vertex_element *ve)
+nv40_vtxelts_state_delete(struct pipe_context *pipe, void *hwcso)
 {
-	struct nv40_context *nv40 = nv40_context(pipe);
+	FREE(hwcso);
+}
 
-	memcpy(nv40->vtxelt, ve, sizeof(*ve) * count);
-	nv40->vtxelt_nr = count;
+static void
+nv40_vtxelts_state_bind(struct pipe_context *pipe, void *hwcso)
+{
+	struct nv40_context *nv40 = nv40_context(pipe);
 
+	nv40->vtxelt = hwcso;
 	nv40->dirty |= NV40_NEW_ARRAYS;
 	nv40->draw_dirty |= NV40_NEW_ARRAYS;
 }
@@ -737,7 +756,10 @@ nv40_init_state_functions(struct nv40_context *nv40)
 	nv40->pipe.set_scissor_state = nv40_set_scissor_state;
 	nv40->pipe.set_viewport_state = nv40_set_viewport_state;
 
+	nv40->pipe.create_vertex_elements_state = nv40_vtxelts_state_create;
+	nv40->pipe.delete_vertex_elements_state = nv40_vtxelts_state_delete;
+	nv40->pipe.bind_vertex_elements_state = nv40_vtxelts_state_bind;
+
 	nv40->pipe.set_vertex_buffers = nv40_set_vertex_buffers;
-	nv40->pipe.set_vertex_elements = nv40_set_vertex_elements;
 }
 
diff --git a/src/gallium/drivers/nv40/nv40_state_emit.c b/src/gallium/drivers/nv40/nv40_state_emit.c
index 8990f303ce..297d71f4fa 100644
--- a/src/gallium/drivers/nv40/nv40_state_emit.c
+++ b/src/gallium/drivers/nv40/nv40_state_emit.c
@@ -174,7 +174,7 @@ nv40_state_validate_swtnl(struct nv40_context *nv40)
 
 	if (nv40->draw_dirty & NV40_NEW_ARRAYS) {
 		draw_set_vertex_buffers(draw, nv40->vtxbuf_nr, nv40->vtxbuf);
-		draw_set_vertex_elements(draw, nv40->vtxelt_nr, nv40->vtxelt);	
+		draw_set_vertex_elements(draw, nv40->vtxelt->num_elements, nv40->vtxelt->pipe);	
 	}
 
 	nv40_state_do_validate(nv40, swtnl_states);
diff --git a/src/gallium/drivers/nv40/nv40_vbo.c b/src/gallium/drivers/nv40/nv40_vbo.c
index 7812460d2e..fabdf4bf23 100644
--- a/src/gallium/drivers/nv40/nv40_vbo.c
+++ b/src/gallium/drivers/nv40/nv40_vbo.c
@@ -493,16 +493,16 @@ nv40_vbo_validate(struct nv40_context *nv40)
 	int hw;
 
 	vtxbuf = so_new(3, 17, 18);
-	so_method(vtxbuf, curie, NV40TCL_VTXBUF_ADDRESS(0), nv40->vtxelt_nr);
+	so_method(vtxbuf, curie, NV40TCL_VTXBUF_ADDRESS(0), nv40->vtxelt->num_elements);
 	vtxfmt = so_new(1, 16, 0);
-	so_method(vtxfmt, curie, NV40TCL_VTXFMT(0), nv40->vtxelt_nr);
+	so_method(vtxfmt, curie, NV40TCL_VTXFMT(0), nv40->vtxelt->num_elements);
 
-	for (hw = 0; hw < nv40->vtxelt_nr; hw++) {
+	for (hw = 0; hw < nv40->vtxelt->num_elements; hw++) {
 		struct pipe_vertex_element *ve;
 		struct pipe_vertex_buffer *vb;
 		unsigned type, ncomp;
 
-		ve = &nv40->vtxelt[hw];
+		ve = &nv40->vtxelt->pipe[hw];
 		vb = &nv40->vtxbuf[ve->vertex_buffer_index];
 
 		if (!vb->stride) {
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index 044437e75f..e91a61b27d 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)
 {
@@ -169,8 +175,7 @@ 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];
@@ -218,6 +223,7 @@ extern void nv50_draw_elements_instanced(struct pipe_context *pipe,
 					 unsigned count,
 					 unsigned startInstance,
 					 unsigned instanceCount);
+extern void nv50_vtxelt_construct(struct nv50_vtxelt_stateobj *cso);
 extern void nv50_vbo_validate(struct nv50_context *nv50);
 
 /* nv50_clear.c */
diff --git a/src/gallium/drivers/nv50/nv50_state.c b/src/gallium/drivers/nv50/nv50_state.c
index 7d304907b6..ffbf3473a1 100644
--- a/src/gallium/drivers/nv50/nv50_state.c
+++ b/src/gallium/drivers/nv50/nv50_state.c
@@ -720,15 +720,34 @@ 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 +797,10 @@ 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_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index 1c8ee0b9ad..c1dcb93b48 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -139,15 +139,16 @@ 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));
@@ -222,11 +223,12 @@ 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];
+	enum pipe_format pf = nv50->vtxelt->pipe[i].src_format;
+	unsigned nr_components = util_format_get_nr_components(pf);
 
-	util_format_read_4f(nv50->vtxelt[i].src_format,
-			    v, 0, data, 0, 0, 0, 1, 1);
+	util_format_read_4f(pf, v, 0, data, 0, 0, 0, 1, 1);
 
-	switch (nv50->vtxelt[i].nr_components) {
+	switch (nr_components) {
 	case 4:
 		BEGIN_RING(chan, tesla, NV50TCL_VTX_ATTR_4F_X(i), 4);
 		OUT_RINGf (chan, v[0]);
@@ -263,16 +265,17 @@ init_per_instance_arrays_immd(struct nv50_context *nv50,
 	struct nouveau_bo *bo;
 	unsigned i, b, count = 0;
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
+	for (i = 0; i < nv50->vtxelt->num_elements; ++i) {
+		if (!nv50->vtxelt->pipe[i].instance_divisor)
 			continue;
 		++count;
-		b = nv50->vtxelt[i].vertex_buffer_index;
+		b = nv50->vtxelt->pipe[i].vertex_buffer_index;
 
-		pos[i] = nv50->vtxelt[i].src_offset +
+		pos[i] = nv50->vtxelt->pipe[i].src_offset +
 			nv50->vtxbuf[b].buffer_offset +
 			startInstance * nv50->vtxbuf[b].stride;
-		step[i] = startInstance % nv50->vtxelt[i].instance_divisor;
+		step[i] = startInstance %
+			nv50->vtxelt->pipe[i].instance_divisor;
 
 		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
 		if (!bo->map)
@@ -293,22 +296,22 @@ init_per_instance_arrays(struct nv50_context *nv50,
 	struct nouveau_channel *chan = tesla->channel;
 	struct nouveau_bo *bo;
 	struct nouveau_stateobj *so;
-	unsigned i, b, count = 0;
+	unsigned i, b, count = 0, num_elements = nv50->vtxelt->num_elements;
 	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);
+	so = so_new(num_elements, num_elements * 2, num_elements * 2);
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
+	for (i = 0; i < nv50->vtxelt->num_elements; ++i) {
+		if (!nv50->vtxelt->pipe[i].instance_divisor)
 			continue;
 		++count;
-		b = nv50->vtxelt[i].vertex_buffer_index;
+		b = nv50->vtxelt->pipe[i].vertex_buffer_index;
 
-		pos[i] = nv50->vtxelt[i].src_offset +
+		pos[i] = nv50->vtxelt->pipe[i].src_offset +
 			nv50->vtxbuf[b].buffer_offset +
 			startInstance * nv50->vtxbuf[b].stride;
 
@@ -316,7 +319,8 @@ init_per_instance_arrays(struct nv50_context *nv50,
 			step[i] = 0;
 			continue;
 		}
-		step[i] = startInstance % nv50->vtxelt[i].instance_divisor;
+		step[i] = startInstance %
+			nv50->vtxelt->pipe[i].instance_divisor;
 
 		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
 
@@ -341,12 +345,12 @@ step_per_instance_arrays_immd(struct nv50_context *nv50,
 	struct nouveau_bo *bo;
 	unsigned i, b;
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
+	for (i = 0; i < nv50->vtxelt->num_elements; ++i) {
+		if (!nv50->vtxelt->pipe[i].instance_divisor)
 			continue;
-		if (++step[i] != nv50->vtxelt[i].instance_divisor)
+		if (++step[i] != nv50->vtxelt->pipe[i].instance_divisor)
 			continue;
-		b = nv50->vtxelt[i].vertex_buffer_index;
+		b = nv50->vtxelt->pipe[i].vertex_buffer_index;
 		bo = nouveau_bo(nv50->vtxbuf[b].buffer);
 
 		step[i] = 0;
@@ -364,7 +368,7 @@ step_per_instance_arrays(struct nv50_context *nv50,
 	struct nouveau_channel *chan = tesla->channel;
 	struct nouveau_bo *bo;
 	struct nouveau_stateobj *so;
-	unsigned i, b;
+	unsigned i, b, num_elements = nv50->vtxelt->num_elements;
 	const uint32_t rl = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
 
 	if (nv50->vbo_fifo) {
@@ -372,14 +376,14 @@ step_per_instance_arrays(struct nv50_context *nv50,
 		return;
 	}
 
-	so = so_new(nv50->vtxelt_nr, nv50->vtxelt_nr * 2, nv50->vtxelt_nr * 2);
+	so = so_new(num_elements, num_elements * 2, num_elements * 2);
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
-		if (!nv50->vtxelt[i].instance_divisor)
+	for (i = 0; i < nv50->vtxelt->num_elements; ++i) {
+		if (!nv50->vtxelt->pipe[i].instance_divisor)
 			continue;
-		b = nv50->vtxelt[i].vertex_buffer_index;
+		b = nv50->vtxelt->pipe[i].vertex_buffer_index;
 
-		if (++step[i] == nv50->vtxelt[i].instance_divisor) {
+		if (++step[i] == nv50->vtxelt->pipe[i].instance_divisor) {
 			step[i] = 0;
 			pos[i] += nv50->vtxbuf[b].stride;
 		}
@@ -726,6 +730,7 @@ 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 +741,10 @@ 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 +781,18 @@ 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);
+	}
+}
+
+void
 nv50_vbo_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -794,19 +812,19 @@ nv50_vbo_validate(struct nv50_context *nv50)
 	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)) {
@@ -855,7 +873,7 @@ 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);
@@ -1016,13 +1034,13 @@ emit_prepare(struct nv50_context *nv50, struct nv50_vbo_emitctx *emit,
 	emit->nr_ve = 0;
 	emit->vtx_dwords = 0;
 
-	for (i = 0; i < nv50->vtxelt_nr; ++i) {
+	for (i = 0; i < nv50->vtxelt->num_elements; ++i) {
 		struct pipe_vertex_element *ve;
 		struct pipe_vertex_buffer *vb;
-		unsigned n, size;
+		unsigned n, size, nr_components;
 		const struct util_format_description *desc;
 
-		ve = &nv50->vtxelt[i];
+		ve = &nv50->vtxelt->pipe[i];
 		vb = &nv50->vtxbuf[ve->vertex_buffer_index];
 		if (!(nv50->vbo_fifo & (1 << i)) || ve->instance_divisor)
 			continue;
@@ -1038,8 +1056,9 @@ emit_prepare(struct nv50_context *nv50, struct nv50_vbo_emitctx *emit,
 
 		size = util_format_get_component_bits(
 			ve->src_format, UTIL_FORMAT_COLORSPACE_RGB, 0);
+		nr_components = util_format_get_nr_components(ve->src_format);
 
-		assert(ve->nr_components > 0 && ve->nr_components <= 4);
+		assert(nr_components > 0 && nr_components <= 4);
 
 		/* It shouldn't be necessary to push the implicit 1s
 		 * for case 3 and size 8 cases 1, 2, 3.
@@ -1049,25 +1068,25 @@ emit_prepare(struct nv50_context *nv50, struct nv50_vbo_emitctx *emit,
 			NOUVEAU_ERR("unsupported vtxelt size: %u\n", size);
 			return FALSE;
 		case 32:
-			switch (ve->nr_components) {
+			switch (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;
+			emit->vtx_dwords += nr_components;
 			break;
 		case 16:
-			switch (ve->nr_components) {
+			switch (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;
+			emit->vtx_dwords += (nr_components + 1) >> 1;
 			break;
 		case 8:
-			switch (ve->nr_components) {
+			switch (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;
diff --git a/src/gallium/drivers/r300/r300_blit.c b/src/gallium/drivers/r300/r300_blit.c
index 513cc0f5d4..b7ad6b2020 100644
--- a/src/gallium/drivers/r300/r300_blit.c
+++ b/src/gallium/drivers/r300/r300_blit.c
@@ -36,6 +36,7 @@ static void r300_blitter_save_states(struct r300_context* r300)
     util_blitter_save_vertex_shader(r300->blitter, r300->vs_state.state);
     util_blitter_save_viewport(r300->blitter, &r300->viewport);
     util_blitter_save_clip(r300->blitter, &r300->clip);
+    util_blitter_save_vertex_elements(r300->blitter, r300->velems);
 }
 
 /* Clear currently bound buffers. */
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index c2825d5f26..80e69924f0 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -278,6 +278,21 @@ struct r300_texture {
     enum r300_buffer_tiling microtile, macrotile;
 };
 
+struct r300_vertex_info {
+    /* Parent class */
+    struct vertex_info vinfo;
+
+    /* R300_VAP_PROG_STREAK_CNTL_[0-7] */
+    uint32_t vap_prog_stream_cntl[8];
+    /* R300_VAP_PROG_STREAK_CNTL_EXT_[0-7] */
+    uint32_t vap_prog_stream_cntl_ext[8];
+};
+
+struct r300_velems_state {
+    unsigned count;
+    struct pipe_vertex_element velem[PIPE_MAX_ATTRIBS];
+};
+
 extern struct pipe_viewport_state r300_viewport_identity;
 
 struct r300_context {
@@ -350,8 +365,7 @@ struct r300_context {
     int vertex_buffer_count;
     int vertex_buffer_max_index;
     /* Vertex elements for Gallium. */
-    struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
-    int vertex_element_count;
+    struct r300_velems_state *velems;
 
     /* Vertex info for Draw. */
     struct vertex_info vertex_info;
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index 51fc590e5d..65377b3865 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -757,9 +757,9 @@ void r300_emit_textures_state(struct r300_context *r300,
 void r300_emit_aos(struct r300_context* r300, unsigned offset)
 {
     struct pipe_vertex_buffer *vb1, *vb2, *vbuf = r300->vertex_buffer;
-    struct pipe_vertex_element *velem = r300->vertex_element;
+    struct pipe_vertex_element *velem = r300->velems->velem;
     int i;
-    unsigned size1, size2, aos_count = r300->vertex_element_count;
+    unsigned size1, size2, aos_count = r300->velems->count;
     unsigned packet_size = (aos_count * 3 + 1) / 2;
     CS_LOCALS(r300);
 
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index c8420bcdd5..d0e06ade40 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -143,7 +143,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 {
     struct pipe_vertex_element* velem;
     struct pipe_vertex_buffer* vbuf;
-    unsigned vertex_element_count = r300->vertex_element_count;
+    unsigned vertex_element_count = r300->velems->count;
     unsigned i, v, vbi, dw, elem_offset, dwords;
 
     /* Size of the vertex, in dwords. */
@@ -166,7 +166,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 
     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
     for (i = 0; i < vertex_element_count; i++) {
-        velem = &r300->vertex_element[i];
+        velem = &r300->velems->velem[i];
         offset[i] = velem->src_offset / 4;
         size[i] = util_format_get_blocksize(velem->src_format) / 4;
         vertex_size += size[i];
@@ -202,7 +202,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
     /* Emit vertices. */
     for (v = 0; v < count; v++) {
         for (i = 0; i < vertex_element_count; i++) {
-            velem = &r300->vertex_element[i];
+            velem = &r300->velems->velem[i];
             vbi = velem->vertex_buffer_index;
             elem_offset = offset[i] + stride[vbi] * (v + start);
 
@@ -215,7 +215,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 
     /* Unmap buffers. */
     for (i = 0; i < vertex_element_count; i++) {
-        vbi = r300->vertex_element[i].vertex_buffer_index;
+        vbi = r300->velems->velem[i].vertex_buffer_index;
 
         if (map[vbi]) {
             vbuf = &r300->vertex_buffer[vbi];
@@ -319,6 +319,31 @@ static void r300_emit_draw_elements(struct r300_context *r300,
     END_CS;
 }
 
+static boolean r300_setup_vertex_buffers(struct r300_context *r300)
+{
+    struct pipe_vertex_buffer *vbuf = r300->vertex_buffer;
+    struct pipe_vertex_element *velem = r300->velems->velem;
+    struct pipe_buffer *pbuf;
+
+validate:
+    for (int i = 0; i < r300->velems->count; i++) {
+        pbuf = vbuf[velem[i].vertex_buffer_index].buffer;
+
+        if (!r300->winsys->add_buffer(r300->winsys, pbuf,
+                                      RADEON_GEM_DOMAIN_GTT, 0)) {
+            r300->context.flush(&r300->context, 0, NULL);
+            goto validate;
+        }
+    }
+
+    if (!r300->winsys->validate(r300->winsys)) {
+        r300->context.flush(&r300->context, 0, NULL);
+        return r300->winsys->validate(r300->winsys);
+    }
+
+    return TRUE;
+}
+
 static void r300_shorten_ubyte_elts(struct r300_context* r300,
                                     struct pipe_buffer** elts,
                                     unsigned count)
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index 8c9f604622..7683a423f2 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -1054,11 +1054,11 @@ static void r300_set_vertex_buffers(struct pipe_context* pipe,
 static boolean r300_validate_aos(struct r300_context *r300)
 {
     struct pipe_vertex_buffer *vbuf = r300->vertex_buffer;
-    struct pipe_vertex_element *velem = r300->vertex_element;
+    struct pipe_vertex_element *velem = r300->velems->velem;
     int i;
 
     /* Check if formats and strides are aligned to the size of DWORD. */
-    for (i = 0; i < r300->vertex_element_count; i++) {
+    for (i = 0; i < r300->velems->count; i++) {
         if (vbuf[velem[i].vertex_buffer_index].stride % 4 != 0 ||
             util_format_get_blocksize(velem[i].src_format) % 4 != 0) {
             return FALSE;
@@ -1067,20 +1067,37 @@ static boolean r300_validate_aos(struct r300_context *r300)
     return TRUE;
 }
 
-static void r300_set_vertex_elements(struct pipe_context* pipe,
-                                    unsigned count,
-                                    const struct pipe_vertex_element* elements)
+static void* r300_create_vertex_elements_state(struct pipe_context* pipe,
+                                               unsigned count,
+                                               const struct pipe_vertex_element* attribs)
 {
-    struct r300_context* r300 = r300_context(pipe);
+    struct r300_velems_state *velems;
+
+    /*XXX should precalculate state here instead of later */
+    assert(count <= PIPE_MAX_ATTRIBS);
+    velems = CALLOC_STRUCT(r300_velems_state);
+    if (velems) {
+        velems->count = count;
+        memcpy(velems->velem, attribs, sizeof(struct pipe_vertex_element) * count);
+    }
+    return velems;
+}
+
+static void r300_bind_vertex_elements_state(struct pipe_context *pipe,
+                                            void *velems)
+{
+    struct r300_context *r300 = r300_context(pipe);
+    struct r300_velems_state *r300_velems = (struct r300_velems_state *) velems;
 
-    memcpy(r300->vertex_element,
-           elements,
-           sizeof(struct pipe_vertex_element) * count);
-    r300->vertex_element_count = count;
+    if (velems == NULL) {
+        return;
+    }
+
+    r300->velems = r300_velems;
 
     if (r300->draw) {
         draw_flush(r300->draw);
-        draw_set_vertex_elements(r300->draw, count, elements);
+        draw_set_vertex_elements(r300->draw, r300_velems->count, r300_velems->velem);
     }
 
     if (!r300_validate_aos(r300)) {
@@ -1090,6 +1107,11 @@ static void r300_set_vertex_elements(struct pipe_context* pipe,
     }
 }
 
+static void r300_delete_vertex_elements_state(struct pipe_context *pipe, void *velems)
+{
+   FREE(velems);
+}
+
 static void* r300_create_vs_state(struct pipe_context* pipe,
                                   const struct pipe_shader_state* shader)
 {
@@ -1262,7 +1284,10 @@ void r300_init_state_functions(struct r300_context* r300)
     r300->context.set_viewport_state = r300_set_viewport_state;
 
     r300->context.set_vertex_buffers = r300_set_vertex_buffers;
-    r300->context.set_vertex_elements = r300_set_vertex_elements;
+
+    r300->context.create_vertex_elements_state = r300_create_vertex_elements_state;
+    r300->context.bind_vertex_elements_state = r300_bind_vertex_elements_state;
+    r300->context.delete_vertex_elements_state = r300_delete_vertex_elements_state;
 
     r300->context.create_vs_state = r300_create_vs_state;
     r300->context.bind_vs_state = r300_bind_vs_state;
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index e9e40747ef..4422581b51 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -127,10 +127,11 @@ static void r300_vertex_psc(struct r300_context* r300)
     DBG(r300, DBG_DRAW, "r300: vs expects %d attribs, routing %d elements"
             " in psc\n",
             vs->info.num_inputs,
+            r300->velems->count);
             r300->vertex_element_count);
 
-    for (i = 0; i < r300->vertex_element_count; i++) {
-        format = r300->vertex_element[i].src_format;
+    for (i = 0; i < r300->velems->count; i++) {
+        format = r300->velems->velem[i].src_format;
 
         type = r300_translate_vertex_data_type(format) |
             (stream_tab[i] << R300_DST_VEC_LOC_SHIFT);
diff --git a/src/gallium/drivers/softpipe/sp_context.c b/src/gallium/drivers/softpipe/sp_context.c
index ddc35bcd62..e8f2318e48 100644
--- a/src/gallium/drivers/softpipe/sp_context.c
+++ b/src/gallium/drivers/softpipe/sp_context.c
@@ -245,6 +245,10 @@ softpipe_create_context( struct pipe_screen *screen,
    softpipe->pipe.bind_gs_state   = softpipe_bind_gs_state;
    softpipe->pipe.delete_gs_state = softpipe_delete_gs_state;
 
+   softpipe->pipe.create_vertex_elements_state = softpipe_create_vertex_elements_state;
+   softpipe->pipe.bind_vertex_elements_state = softpipe_bind_vertex_elements_state;
+   softpipe->pipe.delete_vertex_elements_state = softpipe_delete_vertex_elements_state;
+
    softpipe->pipe.set_blend_color = softpipe_set_blend_color;
    softpipe->pipe.set_stencil_ref = softpipe_set_stencil_ref;
    softpipe->pipe.set_clip_state = softpipe_set_clip_state;
@@ -257,7 +261,6 @@ softpipe_create_context( struct pipe_screen *screen,
    softpipe->pipe.set_viewport_state = softpipe_set_viewport_state;
 
    softpipe->pipe.set_vertex_buffers = softpipe_set_vertex_buffers;
-   softpipe->pipe.set_vertex_elements = softpipe_set_vertex_elements;
 
    softpipe->pipe.draw_arrays = softpipe_draw_arrays;
    softpipe->pipe.draw_elements = softpipe_draw_elements;
diff --git a/src/gallium/drivers/softpipe/sp_context.h b/src/gallium/drivers/softpipe/sp_context.h
index 95def72c54..396d46d4ba 100644
--- a/src/gallium/drivers/softpipe/sp_context.h
+++ b/src/gallium/drivers/softpipe/sp_context.h
@@ -45,6 +45,7 @@ struct softpipe_tile_cache;
 struct softpipe_tex_tile_cache;
 struct sp_fragment_shader;
 struct sp_vertex_shader;
+struct sp_velems_state;
 
 
 struct softpipe_context {
@@ -59,6 +60,7 @@ struct softpipe_context {
    struct sp_fragment_shader *fs;
    struct sp_vertex_shader *vs;
    struct sp_geometry_shader *gs;
+   struct sp_velems_state *velems;
 
    /** Other rendering state */
    struct pipe_blend_color blend_color;
@@ -72,13 +74,11 @@ struct softpipe_context {
    struct pipe_texture *vertex_textures[PIPE_MAX_VERTEX_SAMPLERS];
    struct pipe_viewport_state viewport;
    struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS];
-   struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
 
    unsigned num_samplers;
    unsigned num_textures;
    unsigned num_vertex_samplers;
    unsigned num_vertex_textures;
-   unsigned num_vertex_elements;
    unsigned num_vertex_buffers;
 
    unsigned dirty; /**< Mask of SP_NEW_x flags */
diff --git a/src/gallium/drivers/softpipe/sp_state.h b/src/gallium/drivers/softpipe/sp_state.h
index 4370bbeaee..6b01c0f4d7 100644
--- a/src/gallium/drivers/softpipe/sp_state.h
+++ b/src/gallium/drivers/softpipe/sp_state.h
@@ -100,6 +100,11 @@ struct sp_geometry_shader {
    struct draw_geometry_shader *draw_data;
 };
 
+struct sp_velems_state {
+   unsigned count;
+   struct pipe_vertex_element velem[PIPE_MAX_ATTRIBS];
+};
+
 
 void *
 softpipe_create_blend_state(struct pipe_context *,
@@ -160,8 +165,14 @@ void *softpipe_create_gs_state(struct pipe_context *,
 void softpipe_bind_gs_state(struct pipe_context *, void *);
 void softpipe_delete_gs_state(struct pipe_context *, void *);
 
+void *softpipe_create_vertex_elements_state(struct pipe_context *,
+                                            unsigned count,
+                                            const struct pipe_vertex_element *);
+void softpipe_bind_vertex_elements_state(struct pipe_context *, void *);
+void softpipe_delete_vertex_elements_state(struct pipe_context *, void *);
+
 void softpipe_set_polygon_stipple( struct pipe_context *,
-				  const struct pipe_poly_stipple * );
+                                   const struct pipe_poly_stipple * );
 
 void softpipe_set_scissor_state( struct pipe_context *,
                                  const struct pipe_scissor_state * );
@@ -178,10 +189,6 @@ softpipe_set_vertex_sampler_textures(struct pipe_context *,
 void softpipe_set_viewport_state( struct pipe_context *,
                                   const struct pipe_viewport_state * );
 
-void softpipe_set_vertex_elements(struct pipe_context *,
-                                  unsigned count,
-                                  const struct pipe_vertex_element *);
-
 void softpipe_set_vertex_buffers(struct pipe_context *,
                                  unsigned count,
                                  const struct pipe_vertex_buffer *);
diff --git a/src/gallium/drivers/softpipe/sp_state_vertex.c b/src/gallium/drivers/softpipe/sp_state_vertex.c
index b491d92ed1..a151758ddc 100644
--- a/src/gallium/drivers/softpipe/sp_state_vertex.c
+++ b/src/gallium/drivers/softpipe/sp_state_vertex.c
@@ -32,27 +32,44 @@
 #include "sp_context.h"
 #include "sp_state.h"
 
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 
 
+void *
+softpipe_create_vertex_elements_state(struct pipe_context *pipe,
+                                      unsigned count,
+                                      const struct pipe_vertex_element *attribs)
+{
+   struct sp_velems_state *velems;
+   assert(count <= PIPE_MAX_ATTRIBS);
+   velems = (struct sp_velems_state *) MALLOC(sizeof(struct sp_velems_state));
+   if (velems) {
+      velems->count = count;
+      memcpy(velems->velem, attribs, sizeof(*attribs) * count);
+   }
+   return velems;
+}
+
 void
-softpipe_set_vertex_elements(struct pipe_context *pipe,
-                             unsigned count,
-                             const struct pipe_vertex_element *attribs)
+softpipe_bind_vertex_elements_state(struct pipe_context *pipe,
+                                    void *velems)
 {
    struct softpipe_context *softpipe = softpipe_context(pipe);
+   struct sp_velems_state *sp_velems = (struct sp_velems_state *) velems;
 
-   assert(count <= PIPE_MAX_ATTRIBS);
-
-   memcpy(softpipe->vertex_element, attribs,
-          count * sizeof(struct pipe_vertex_element));
-   softpipe->num_vertex_elements = count;
+   softpipe->velems = sp_velems;
 
    softpipe->dirty |= SP_NEW_VERTEX;
 
-   draw_set_vertex_elements(softpipe->draw, count, attribs);
+   draw_set_vertex_elements(softpipe->draw, sp_velems->count, sp_velems->velem);
 }
 
+void
+softpipe_delete_vertex_elements_state(struct pipe_context *pipe, void *velems)
+{
+   FREE( velems );
+}
 
 void
 softpipe_set_vertex_buffers(struct pipe_context *pipe,
diff --git a/src/gallium/drivers/svga/svga_context.h b/src/gallium/drivers/svga/svga_context.h
index 03302e2a6e..791d30edc0 100644
--- a/src/gallium/drivers/svga/svga_context.h
+++ b/src/gallium/drivers/svga/svga_context.h
@@ -169,6 +169,11 @@ struct svga_sampler_state {
    unsigned view_max_lod;
 };
 
+struct svga_velems_state {
+   unsigned count;
+   struct pipe_vertex_element velem[PIPE_MAX_ATTRIBS];
+};
+
 /* Use to calculate differences between state emitted to hardware and
  * current driver-calculated state.  
  */
@@ -178,13 +183,13 @@ struct svga_state
    const struct svga_depth_stencil_state *depth;
    const struct svga_rasterizer_state *rast;
    const struct svga_sampler_state *sampler[PIPE_MAX_SAMPLERS];
+   const struct svga_velems_state *velems;
 
    struct pipe_texture *texture[PIPE_MAX_SAMPLERS]; /* or texture ID's? */
    struct svga_fragment_shader *fs;
    struct svga_vertex_shader *vs;
 
    struct pipe_vertex_buffer vb[PIPE_MAX_ATTRIBS];
-   struct pipe_vertex_element ve[PIPE_MAX_ATTRIBS];
    struct pipe_buffer *cb[PIPE_SHADER_TYPES];
 
    struct pipe_framebuffer_state framebuffer;
@@ -204,7 +209,6 @@ struct svga_state
 
    unsigned num_samplers;
    unsigned num_textures;
-   unsigned num_vertex_elements;
    unsigned num_vertex_buffers;
    unsigned reduced_prim;
 
diff --git a/src/gallium/drivers/svga/svga_pipe_vertex.c b/src/gallium/drivers/svga/svga_pipe_vertex.c
index 836b8441da..1715a47fc6 100644
--- a/src/gallium/drivers/svga/svga_pipe_vertex.c
+++ b/src/gallium/drivers/svga/svga_pipe_vertex.c
@@ -26,6 +26,7 @@
 #include "util/u_inlines.h"
 #include "pipe/p_defines.h"
 #include "util/u_math.h"
+#include "util/u_memory.h"
 #include "tgsi/tgsi_parse.h"
 
 #include "svga_screen.h"
@@ -64,20 +65,37 @@ static void svga_set_vertex_buffers(struct pipe_context *pipe,
    svga->dirty |= SVGA_NEW_VBUFFER;
 }
 
-static void svga_set_vertex_elements(struct pipe_context *pipe,
-                                     unsigned count,
-                                     const struct pipe_vertex_element *elements)
+
+static void *
+svga_create_vertex_elements_state(struct pipe_context *pipe,
+                                  unsigned count,
+                                  const struct pipe_vertex_element *attribs)
 {
-   struct svga_context *svga = svga_context(pipe);
-   unsigned i;
+   struct svga_velems_state *velems;
+   assert(count <= PIPE_MAX_ATTRIBS);
+   velems = (struct svga_velems_state *) MALLOC(sizeof(struct svga_velems_state));
+   if (velems) {
+      velems->count = count;
+      memcpy(velems->velem, attribs, sizeof(*attribs) * count);
+   }
+   return velems;
+}
 
-   for (i = 0; i < count; i++)
-      svga->curr.ve[i] = elements[i];
+static void svga_bind_vertex_elements_state(struct pipe_context *pipe,
+                                            void *velems)
+{
+   struct svga_context *svga = svga_context(pipe);
+   struct svga_velems_state *svga_velems = (struct svga_velems_state *) velems;
 
-   svga->curr.num_vertex_elements = count;
+   svga->curr.velems = svga_velems;
    svga->dirty |= SVGA_NEW_VELEMENT;
 }
 
+static void svga_delete_vertex_elements_state(struct pipe_context *pipe,
+                                              void *velems)
+{
+   FREE(velems);
+}
 
 void svga_cleanup_vertex_state( struct svga_context *svga )
 {
@@ -91,7 +109,9 @@ void svga_cleanup_vertex_state( struct svga_context *svga )
 void svga_init_vertex_functions( struct svga_context *svga )
 {
    svga->pipe.set_vertex_buffers = svga_set_vertex_buffers;
-   svga->pipe.set_vertex_elements = svga_set_vertex_elements;
+   svga->pipe.create_vertex_elements_state = svga_create_vertex_elements_state;
+   svga->pipe.bind_vertex_elements_state = svga_bind_vertex_elements_state;
+   svga->pipe.delete_vertex_elements_state = svga_delete_vertex_elements_state;
 }
 
 
diff --git a/src/gallium/drivers/svga/svga_state_need_swtnl.c b/src/gallium/drivers/svga/svga_state_need_swtnl.c
index d774e3e504..10d473584d 100644
--- a/src/gallium/drivers/svga/svga_state_need_swtnl.c
+++ b/src/gallium/drivers/svga/svga_state_need_swtnl.c
@@ -76,8 +76,8 @@ static int update_need_swvfetch( struct svga_context *svga,
    unsigned i;
    boolean need_swvfetch = FALSE;
 
-   for (i = 0; i < svga->curr.num_vertex_elements; i++) {
-      svga->state.sw.ve_format[i] = svga_translate_vertex_format(svga->curr.ve[i].src_format);
+   for (i = 0; i < svga->curr.velems->count; i++) {
+      svga->state.sw.ve_format[i] = svga_translate_vertex_format(svga->curr.velems->velem[i].src_format);
       if (svga->state.sw.ve_format[i] == SVGA3D_DECLTYPE_MAX) {
          need_swvfetch = TRUE;
          break;
diff --git a/src/gallium/drivers/svga/svga_state_vdecl.c b/src/gallium/drivers/svga/svga_state_vdecl.c
index ded903170b..f531e22304 100644
--- a/src/gallium/drivers/svga/svga_state_vdecl.c
+++ b/src/gallium/drivers/svga/svga_state_vdecl.c
@@ -95,17 +95,17 @@ upload_user_buffers( struct svga_context *svga )
 static int emit_hw_vs_vdecl( struct svga_context *svga,
                              unsigned dirty )
 {
-   const struct pipe_vertex_element *ve = svga->curr.ve;
+   const struct pipe_vertex_element *ve = svga->curr.velems->velem;
    SVGA3dVertexDecl decl;
    unsigned i;
 
-   assert(svga->curr.num_vertex_elements >=
+   assert(svga->curr.velems->count >=
           svga->curr.vs->base.info.file_count[TGSI_FILE_INPUT]);
 
    svga_hwtnl_reset_vdecl( svga->hwtnl, 
-                           svga->curr.num_vertex_elements );
+                           svga->curr.velems->count );
 
-   for (i = 0; i < svga->curr.num_vertex_elements; i++) {
+   for (i = 0; i < svga->curr.velems->count; i++) {
       const struct pipe_vertex_buffer *vb = &svga->curr.vb[ve[i].vertex_buffer_index];
       unsigned usage, index;
 
diff --git a/src/gallium/drivers/svga/svga_state_vs.c b/src/gallium/drivers/svga/svga_state_vs.c
index d7999fe53d..781f7bf533 100644
--- a/src/gallium/drivers/svga/svga_state_vs.c
+++ b/src/gallium/drivers/svga/svga_state_vs.c
@@ -186,8 +186,8 @@ static int update_zero_stride( struct svga_context *svga,
    svga->curr.zero_stride_vertex_elements = 0;
    svga->curr.num_zero_stride_vertex_elements = 0;
 
-   for (i = 0; i < svga->curr.num_vertex_elements; i++) {
-      const struct pipe_vertex_element *vel = &svga->curr.ve[i];
+   for (i = 0; i < svga->curr.velems->count; i++) {
+      const struct pipe_vertex_element *vel = &svga->curr.velems->velem[i];
       const struct pipe_vertex_buffer *vbuffer = &svga->curr.vb[
          vel->vertex_buffer_index];
       if (vbuffer->stride == 0) {
diff --git a/src/gallium/drivers/svga/svga_swtnl_state.c b/src/gallium/drivers/svga/svga_swtnl_state.c
index 35f36a828f..246d34e649 100644
--- a/src/gallium/drivers/svga/svga_swtnl_state.c
+++ b/src/gallium/drivers/svga/svga_swtnl_state.c
@@ -99,8 +99,8 @@ static int update_swtnl_draw( struct svga_context *svga,
 
    if (dirty & SVGA_NEW_VELEMENT)
       draw_set_vertex_elements(svga->swtnl.draw, 
-                               svga->curr.num_vertex_elements, 
-                               svga->curr.ve );
+                               svga->curr.velems->count, 
+                               svga->curr.velems->velem );
 
    if (dirty & SVGA_NEW_CLIP)
       draw_set_clip_state(svga->swtnl.draw, 
diff --git a/src/gallium/drivers/trace/tr_context.c b/src/gallium/drivers/trace/tr_context.c
index df40fbade6..133521f45e 100644
--- a/src/gallium/drivers/trace/tr_context.c
+++ b/src/gallium/drivers/trace/tr_context.c
@@ -773,6 +773,70 @@ trace_context_delete_vs_state(struct pipe_context *_pipe,
 }
 
 
+static INLINE void *
+trace_context_create_vertex_elements_state(struct pipe_context *_pipe,
+                                           unsigned num_elements,
+                                           const struct  pipe_vertex_element *elements)
+{
+   struct trace_context *tr_ctx = trace_context(_pipe);
+   struct pipe_context *pipe = tr_ctx->pipe;
+   void * result;
+
+   trace_dump_call_begin("pipe_context", "create_vertex_elements_state");
+
+   trace_dump_arg(ptr, pipe);
+   trace_dump_arg(uint, num_elements);
+
+   trace_dump_arg_begin("elements");
+   trace_dump_struct_array(vertex_element, elements, num_elements);
+   trace_dump_arg_end();
+
+   result = pipe->create_vertex_elements_state(pipe, num_elements, elements);
+
+   trace_dump_ret(ptr, result);
+
+   trace_dump_call_end();
+
+   return result;
+}
+
+
+static INLINE void
+trace_context_bind_vertex_elements_state(struct pipe_context *_pipe,
+                                         void *state)
+{
+   struct trace_context *tr_ctx = trace_context(_pipe);
+   struct pipe_context *pipe = tr_ctx->pipe;
+
+   trace_dump_call_begin("pipe_context", "bind_vertex_elements_state");
+
+   trace_dump_arg(ptr, pipe);
+   trace_dump_arg(ptr, state);
+
+   pipe->bind_vertex_elements_state(pipe, state);
+
+   trace_dump_call_end();
+}
+
+
+static INLINE void
+trace_context_delete_vertex_elements_state(struct pipe_context *_pipe,
+                                           void *state)
+{
+   struct trace_context *tr_ctx = trace_context(_pipe);
+   struct pipe_context *pipe = tr_ctx->pipe;
+
+   trace_dump_call_begin("pipe_context", "delete_verte_elements_state");
+
+   trace_dump_arg(ptr, pipe);
+   trace_dump_arg(ptr, state);
+
+   pipe->delete_vertex_elements_state(pipe, state);
+
+   trace_dump_call_end();
+}
+
+
 static INLINE void
 trace_context_set_blend_color(struct pipe_context *_pipe,
                               const struct pipe_blend_color *state)
@@ -1048,29 +1112,6 @@ trace_context_set_vertex_buffers(struct pipe_context *_pipe,
 
 
 static INLINE void
-trace_context_set_vertex_elements(struct pipe_context *_pipe,
-                                  unsigned num_elements,
-                                  const struct pipe_vertex_element *elements)
-{
-   struct trace_context *tr_ctx = trace_context(_pipe);
-   struct pipe_context *pipe = tr_ctx->pipe;
-
-   trace_dump_call_begin("pipe_context", "set_vertex_elements");
-
-   trace_dump_arg(ptr, pipe);
-   trace_dump_arg(uint, num_elements);
-
-   trace_dump_arg_begin("elements");
-   trace_dump_struct_array(vertex_element, elements, num_elements);
-   trace_dump_arg_end();
-
-   pipe->set_vertex_elements(pipe, num_elements, elements);
-
-   trace_dump_call_end();
-}
-
-
-static INLINE void
 trace_context_surface_copy(struct pipe_context *_pipe,
                            struct pipe_surface *dest,
                            unsigned destx, unsigned desty,
@@ -1303,6 +1344,9 @@ trace_context_create(struct trace_screen *tr_scr,
    tr_ctx->base.create_vs_state = trace_context_create_vs_state;
    tr_ctx->base.bind_vs_state = trace_context_bind_vs_state;
    tr_ctx->base.delete_vs_state = trace_context_delete_vs_state;
+   tr_ctx->base.create_vertex_elements_state = trace_context_create_vertex_elements_state;
+   tr_ctx->base.bind_vertex_elements_state = trace_context_bind_vertex_elements_state;
+   tr_ctx->base.delete_vertex_elements_state = trace_context_delete_vertex_elements_state;
    tr_ctx->base.set_blend_color = trace_context_set_blend_color;
    tr_ctx->base.set_stencil_ref = trace_context_set_stencil_ref;
    tr_ctx->base.set_clip_state = trace_context_set_clip_state;
@@ -1314,7 +1358,6 @@ trace_context_create(struct trace_screen *tr_scr,
    tr_ctx->base.set_fragment_sampler_textures = trace_context_set_fragment_sampler_textures;
    tr_ctx->base.set_vertex_sampler_textures = trace_context_set_vertex_sampler_textures;
    tr_ctx->base.set_vertex_buffers = trace_context_set_vertex_buffers;
-   tr_ctx->base.set_vertex_elements = trace_context_set_vertex_elements;
    if (pipe->surface_copy)
       tr_ctx->base.surface_copy = trace_context_surface_copy;
    if (pipe->surface_fill)
diff --git a/src/gallium/drivers/trace/tr_dump_state.c b/src/gallium/drivers/trace/tr_dump_state.c
index f97d963dba..f82dd01c69 100644
--- a/src/gallium/drivers/trace/tr_dump_state.c
+++ b/src/gallium/drivers/trace/tr_dump_state.c
@@ -479,7 +479,6 @@ void trace_dump_vertex_element(const struct pipe_vertex_element *state)
    trace_dump_member(uint, state, src_offset);
 
    trace_dump_member(uint, state, vertex_buffer_index);
-   trace_dump_member(uint, state, nr_components);
 
    trace_dump_member(format, state, src_format);
 
diff --git a/src/gallium/include/pipe/p_context.h b/src/gallium/include/pipe/p_context.h
index f82b77903e..376b01aa69 100644
--- a/src/gallium/include/pipe/p_context.h
+++ b/src/gallium/include/pipe/p_context.h
@@ -177,6 +177,12 @@ struct pipe_context {
    void   (*bind_gs_state)(struct pipe_context *, void *);
    void   (*delete_gs_state)(struct pipe_context *, void *);
 
+   void * (*create_vertex_elements_state)(struct pipe_context *,
+                                          unsigned num_elements,
+                                          const struct pipe_vertex_element *);
+   void   (*bind_vertex_elements_state)(struct pipe_context *, void *);
+   void   (*delete_vertex_elements_state)(struct pipe_context *, void *);
+
    /*@}*/
 
    /**
@@ -220,9 +226,6 @@ struct pipe_context {
                                unsigned num_buffers,
                                const struct pipe_vertex_buffer * );
 
-   void (*set_vertex_elements)( struct pipe_context *,
-                                unsigned num_elements,
-                                const struct pipe_vertex_element * );
    /*@}*/
 
 
diff --git a/src/gallium/include/pipe/p_state.h b/src/gallium/include/pipe/p_state.h
index 72635d1031..3a97d888ce 100644
--- a/src/gallium/include/pipe/p_state.h
+++ b/src/gallium/include/pipe/p_state.h
@@ -373,7 +373,6 @@ struct pipe_vertex_element
     * this attribute live in?
     */
    unsigned vertex_buffer_index:8;
-   unsigned nr_components:8;
  
    enum pipe_format src_format;
 };
diff --git a/src/gallium/state_trackers/python/p_context.i b/src/gallium/state_trackers/python/p_context.i
index 3f36ccb621..5c44462e80 100644
--- a/src/gallium/state_trackers/python/p_context.i
+++ b/src/gallium/state_trackers/python/p_context.i
@@ -51,7 +51,7 @@ struct st_context {
    void set_blend( const struct pipe_blend_state *state ) {
       cso_set_blend($self->cso, state);
    }
-   
+
    void set_fragment_sampler( unsigned index, const struct pipe_sampler_state *state ) {
       cso_single_sampler($self->cso, index, state);
       cso_single_sampler_done($self->cso);
@@ -222,9 +222,9 @@ struct st_context {
    void set_vertex_elements(unsigned num) 
    {
       $self->num_vertex_elements = num;
-      $self->pipe->set_vertex_elements($self->pipe, 
-                                       $self->num_vertex_elements, 
-                                       $self->vertex_elements);
+      cso_set_vertex_elements($self->cso,
+                              $self->num_vertex_elements, 
+                              $self->vertex_elements);
    }
 
    /*
diff --git a/src/gallium/state_trackers/vega/api_masks.c b/src/gallium/state_trackers/vega/api_masks.c
index 9c123a4cf9..7eb5ea1f07 100644
--- a/src/gallium/state_trackers/vega/api_masks.c
+++ b/src/gallium/state_trackers/vega/api_masks.c
@@ -86,6 +86,8 @@ draw_clear_quad(struct vg_context *st,
 
    /* draw */
    if (buf) {
+      cso_set_vertex_elements(st->cso_context, 2, st->velems);
+
       util_draw_vertex_buffer(pipe, buf, 0,
                               PIPE_PRIM_TRIANGLE_FAN,
                               4,  /* verts */
diff --git a/src/gallium/state_trackers/vega/polygon.c b/src/gallium/state_trackers/vega/polygon.c
index c06dbf5206..eef2c1eb87 100644
--- a/src/gallium/state_trackers/vega/polygon.c
+++ b/src/gallium/state_trackers/vega/polygon.c
@@ -292,12 +292,12 @@ static void draw_polygon(struct vg_context *ctx,
    pipe->set_vertex_buffers(pipe, 1, &vbuffer);
 
    /* tell pipe about the vertex attributes */
+   memset(&velement, 0, sizeof(velement));
    velement.src_offset = 0;
    velement.instance_divisor = 0;
    velement.vertex_buffer_index = 0;
    velement.src_format = PIPE_FORMAT_R32G32_FLOAT;
-   velement.nr_components = COMPONENTS;
-   pipe->set_vertex_elements(pipe, 1, &velement);
+   cso_set_vertex_elements(ctx->cso_context, 1, &velement);
 
    /* draw */
    pipe->draw_arrays(pipe, PIPE_PRIM_TRIANGLE_FAN, 
diff --git a/src/gallium/state_trackers/vega/renderer.c b/src/gallium/state_trackers/vega/renderer.c
index 05620efa9c..47e8b470a1 100644
--- a/src/gallium/state_trackers/vega/renderer.c
+++ b/src/gallium/state_trackers/vega/renderer.c
@@ -210,6 +210,7 @@ void renderer_draw_quad(struct renderer *r,
    buf = setup_vertex_data(r, x1, y1, x2, y2, depth);
 
    if (buf) {
+      cso_set_vertex_elements(r->cso, 2, r->owner->velems);
       util_draw_vertex_buffer(r->pipe, buf, 0,
                               PIPE_PRIM_TRIANGLE_FAN,
                               4,  /* verts */
@@ -248,6 +249,7 @@ void renderer_draw_texture(struct renderer *r,
                                s0, t0, s1, t1, 0.0f);
 
    if (buf) {
+      cso_set_vertex_elements(r->cso, 2, r->owner->velems);
       util_draw_vertex_buffer(pipe, buf, 0,
                            PIPE_PRIM_TRIANGLE_FAN,
                            4,  /* verts */
@@ -370,6 +372,7 @@ void renderer_copy_texture(struct renderer *ctx,
                          0.0f);
 
    if (buf) {
+      cso_set_vertex_elements(ctx->cso, 2, ctx->owner->velems);
       util_draw_vertex_buffer(ctx->pipe, buf, 0,
                               PIPE_PRIM_TRIANGLE_FAN,
                               4,  /* verts */
@@ -535,6 +538,7 @@ void renderer_copy_surface(struct renderer *ctx,
                            (float) dstX1, (float) dstY1, z);
 
    if (buf) {
+      cso_set_vertex_elements(ctx->cso, 2, ctx->owner->velems);
       util_draw_vertex_buffer(ctx->pipe, buf, 0,
                               PIPE_PRIM_TRIANGLE_FAN,
                               4,  /* verts */
@@ -587,6 +591,7 @@ void renderer_texture_quad(struct renderer *r,
                           s0, t0, s1, t1, 0.0f);
 
    if (buf) {
+      cso_set_vertex_elements(r->cso, 2, r->owner->velems);
       util_draw_vertex_buffer(pipe, buf, 0,
                               PIPE_PRIM_TRIANGLE_FAN,
                               4,  /* verts */
diff --git a/src/gallium/state_trackers/vega/vg_context.c b/src/gallium/state_trackers/vega/vg_context.c
index 426bf9bc62..170391ec03 100644
--- a/src/gallium/state_trackers/vega/vg_context.c
+++ b/src/gallium/state_trackers/vega/vg_context.c
@@ -72,6 +72,7 @@ struct vg_context * vg_create_context(struct pipe_context *pipe,
                                       struct vg_context *share)
 {
    struct vg_context *ctx;
+   unsigned i;
 
    ctx = CALLOC_STRUCT(vg_context);
 
@@ -103,6 +104,13 @@ struct vg_context * vg_create_context(struct pipe_context *pipe,
    ctx->blend_sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
    ctx->blend_sampler.normalized_coords = 0;
 
+   for (i = 0; i < 2; i++) {
+      ctx->velems[i].src_offset = i * 4 * sizeof(float);
+      ctx->velems[i].instance_divisor = 0;
+      ctx->velems[i].vertex_buffer_index = 0;
+      ctx->velems[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+   }
+
    vg_set_error(ctx, VG_NO_ERROR);
 
    ctx->owned_objects[VG_OBJECT_PAINT] = cso_hash_create();
diff --git a/src/gallium/state_trackers/vega/vg_context.h b/src/gallium/state_trackers/vega/vg_context.h
index bc88c8d139..804e9e76d7 100644
--- a/src/gallium/state_trackers/vega/vg_context.h
+++ b/src/gallium/state_trackers/vega/vg_context.h
@@ -146,6 +146,7 @@ struct vg_context
    struct vg_shader *clear_vs;
    struct vg_shader *texture_vs;
    struct pipe_buffer *vs_const_buffer;
+   struct pipe_vertex_element velems[2];
 };
 
 struct vg_object {
diff --git a/src/gallium/state_trackers/xorg/xorg_renderer.c b/src/gallium/state_trackers/xorg/xorg_renderer.c
index 83b0d31e38..1eb926360b 100644
--- a/src/gallium/state_trackers/xorg/xorg_renderer.c
+++ b/src/gallium/state_trackers/xorg/xorg_renderer.c
@@ -68,6 +68,8 @@ renderer_draw(struct xorg_renderer *r)
 
 
    if (buf) {
+      cso_set_vertex_elements(r->cso, r->attrs_per_vertex, r->velems);
+
       util_draw_vertex_buffer(pipe, buf, 0,
                               PIPE_PRIM_QUADS,
                               num_verts,  /* verts */
@@ -92,6 +94,7 @@ renderer_init_state(struct xorg_renderer *r)
 {
    struct pipe_depth_stencil_alpha_state dsa;
    struct pipe_rasterizer_state raster;
+   unsigned i;
 
    /* set common initial clip state */
    memset(&dsa, 0, sizeof(struct pipe_depth_stencil_alpha_state));
@@ -103,6 +106,14 @@ renderer_init_state(struct xorg_renderer *r)
    raster.gl_rasterization_rules = 1;
    cso_set_rasterizer(r->cso, &raster);
 
+   /* vertex elements state */
+   memset(&r->velems[0], 0, sizeof(r->velems[0]) * 3);
+   for (i = 0; i < 3; i++) {
+      r->velems[i].src_offset = i * 4 * sizeof(float);
+      r->velems[i].instance_divisor = 0;
+      r->velems[i].vertex_buffer_index = 0;
+      r->velems[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+   }
 }
 
 
@@ -600,6 +611,8 @@ void renderer_draw_yuv(struct xorg_renderer *r,
    if (buf) {
       const int num_attribs = 2; /*pos + tex coord*/
 
+      cso_set_vertex_elements(r->cso, num_attribs, r->velems);
+
       util_draw_vertex_buffer(pipe, buf, 0,
                               PIPE_PRIM_QUADS,
                               4,  /* verts */
diff --git a/src/gallium/state_trackers/xorg/xorg_renderer.h b/src/gallium/state_trackers/xorg/xorg_renderer.h
index af6aa0567d..3d00628719 100644
--- a/src/gallium/state_trackers/xorg/xorg_renderer.h
+++ b/src/gallium/state_trackers/xorg/xorg_renderer.h
@@ -28,6 +28,7 @@ struct xorg_renderer {
 
    float buffer[BUF_SIZE];
    int buffer_size;
+   struct pipe_vertex_element velems[3];
 
    /* number of attributes per vertex for the current
     * draw operation */
diff --git a/src/mesa/state_tracker/st_cb_bitmap.c b/src/mesa/state_tracker/st_cb_bitmap.c
index 0332d4dbdf..f326601c3b 100644
--- a/src/mesa/state_tracker/st_cb_bitmap.c
+++ b/src/mesa/state_tracker/st_cb_bitmap.c
@@ -440,6 +440,7 @@ draw_bitmap_quad(GLcontext *ctx, GLint x, GLint y, GLfloat z,
    cso_save_viewport(cso);
    cso_save_fragment_shader(cso);
    cso_save_vertex_shader(cso);
+   cso_save_vertex_elements(cso);
 
    /* rasterizer state: just scissor */
    st->bitmap.rasterizer.scissor = ctx->Scissor.Enabled;
@@ -490,6 +491,8 @@ draw_bitmap_quad(GLcontext *ctx, GLint x, GLint y, GLfloat z,
       cso_set_viewport(cso, &vp);
    }
 
+   cso_set_vertex_elements(cso, 3, st->velems_util_draw);
+
    /* convert Z from [0,1] to [-1,-1] to match viewport Z scale/bias */
    z = z * 2.0 - 1.0;
 
@@ -509,6 +512,7 @@ draw_bitmap_quad(GLcontext *ctx, GLint x, GLint y, GLfloat z,
    cso_restore_viewport(cso);
    cso_restore_fragment_shader(cso);
    cso_restore_vertex_shader(cso);
+   cso_restore_vertex_elements(cso);
 }
 
 
diff --git a/src/mesa/state_tracker/st_cb_clear.c b/src/mesa/state_tracker/st_cb_clear.c
index 9e66eed363..de86062fc4 100644
--- a/src/mesa/state_tracker/st_cb_clear.c
+++ b/src/mesa/state_tracker/st_cb_clear.c
@@ -213,6 +213,7 @@ clear_with_quad(GLcontext *ctx,
    cso_save_clip(st->cso_context);
    cso_save_fragment_shader(st->cso_context);
    cso_save_vertex_shader(st->cso_context);
+   cso_save_vertex_elements(st->cso_context);
 
    /* blend state: RGBA masking */
    {
@@ -264,6 +265,8 @@ clear_with_quad(GLcontext *ctx,
       cso_set_depth_stencil_alpha(st->cso_context, &depth_stencil);
    }
 
+   cso_set_vertex_elements(st->cso_context, 2, st->velems_util_draw);
+
    cso_set_rasterizer(st->cso_context, &st->clear.raster);
 
    /* viewport state: viewport matching window dims */
@@ -297,6 +300,8 @@ clear_with_quad(GLcontext *ctx,
    cso_restore_clip(st->cso_context);
    cso_restore_fragment_shader(st->cso_context);
    cso_restore_vertex_shader(st->cso_context);
+   cso_restore_vertex_elements(st->cso_context);
+
 }
 
 
diff --git a/src/mesa/state_tracker/st_cb_drawpixels.c b/src/mesa/state_tracker/st_cb_drawpixels.c
index 98453321f8..c609435a15 100644
--- a/src/mesa/state_tracker/st_cb_drawpixels.c
+++ b/src/mesa/state_tracker/st_cb_drawpixels.c
@@ -529,6 +529,7 @@ draw_textured_quad(GLcontext *ctx, GLint x, GLint y, GLfloat z,
    cso_save_sampler_textures(cso);
    cso_save_fragment_shader(cso);
    cso_save_vertex_shader(cso);
+   cso_save_vertex_elements(cso);
 
    /* rasterizer state: just scissor */
    {
@@ -581,6 +582,8 @@ draw_textured_quad(GLcontext *ctx, GLint x, GLint y, GLfloat z,
       cso_set_viewport(cso, &vp);
    }
 
+   cso_set_vertex_elements(cso, 3, st->velems_util_draw);
+
    /* texture state: */
    if (st->pixel_xfer.pixelmap_enabled) {
       struct pipe_texture *textures[2];
@@ -615,6 +618,7 @@ draw_textured_quad(GLcontext *ctx, GLint x, GLint y, GLfloat z,
    cso_restore_sampler_textures(cso);
    cso_restore_fragment_shader(cso);
    cso_restore_vertex_shader(cso);
+   cso_restore_vertex_elements(cso);
 }
 
 
diff --git a/src/mesa/state_tracker/st_context.c b/src/mesa/state_tracker/st_context.c
index de8beaf5e2..0358a70726 100644
--- a/src/mesa/state_tracker/st_context.c
+++ b/src/mesa/state_tracker/st_context.c
@@ -141,6 +141,14 @@ st_create_context_priv( GLcontext *ctx, struct pipe_context *pipe )
    for (i = 0; i < PIPE_MAX_SAMPLERS; i++)
       st->state.sampler_list[i] = &st->state.samplers[i];
 
+   for (i = 0; i < 3; i++) {
+      memset(&st->velems_util_draw[i], 0, sizeof(struct pipe_vertex_element));
+      st->velems_util_draw[i].src_offset = i * 4 * sizeof(float);
+      st->velems_util_draw[i].instance_divisor = 0;
+      st->velems_util_draw[i].vertex_buffer_index = 0;
+      st->velems_util_draw[i].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+   }
+
    /* we want all vertex data to be placed in buffer objects */
    vbo_use_buffer_objects(ctx);
 
diff --git a/src/mesa/state_tracker/st_context.h b/src/mesa/state_tracker/st_context.h
index 045c029c30..6622361a7e 100644
--- a/src/mesa/state_tracker/st_context.h
+++ b/src/mesa/state_tracker/st_context.h
@@ -174,6 +174,9 @@ struct st_context
       unsigned vbuf_slot;
    } clear;
 
+   /** used for anything using util_draw_vertex_buffer */
+   struct pipe_vertex_element velems_util_draw[3];
+
    void *passthrough_fs;  /**< simple pass-through frag shader */
 
    struct gen_mipmap_state *gen_mipmap;
diff --git a/src/mesa/state_tracker/st_draw.c b/src/mesa/state_tracker/st_draw.c
index 32b9a473cf..4d2e39108d 100644
--- a/src/mesa/state_tracker/st_draw.c
+++ b/src/mesa/state_tracker/st_draw.c
@@ -57,6 +57,7 @@
 #include "pipe/p_defines.h"
 #include "util/u_inlines.h"
 #include "util/u_format.h"
+#include "cso_cache/cso_context.h"
 
 
 static GLuint double_types[4] = {
@@ -368,7 +369,6 @@ setup_interleaved_attribs(GLcontext *ctx,
          (unsigned) (arrays[mesaAttr]->Ptr - offset0);
       velements[attr].instance_divisor = 0;
       velements[attr].vertex_buffer_index = 0;
-      velements[attr].nr_components = arrays[mesaAttr]->Size;
       velements[attr].src_format =
          st_pipe_vertex_format(arrays[mesaAttr]->Type,
                                arrays[mesaAttr]->Size,
@@ -458,7 +458,6 @@ setup_non_interleaved_attribs(GLcontext *ctx,
       vbuffer[attr].max_index = max_index;
       velements[attr].instance_divisor = 0;
       velements[attr].vertex_buffer_index = attr;
-      velements[attr].nr_components = arrays[mesaAttr]->Size;
       velements[attr].src_format
          = st_pipe_vertex_format(arrays[mesaAttr]->Type,
                                  arrays[mesaAttr]->Size,
@@ -596,14 +595,13 @@ st_draw_vbo(GLcontext *ctx,
       for (i = 0; i < num_velements; i++) {
          printf("vlements[%d].vbuffer_index = %u\n", i, velements[i].vertex_buffer_index);
          printf("vlements[%d].src_offset = %u\n", i, velements[i].src_offset);
-         printf("vlements[%d].nr_comps = %u\n", i, velements[i].nr_components);
          printf("vlements[%d].format = %s\n", i, util_format_name(velements[i].src_format));
       }
    }
 #endif
 
    pipe->set_vertex_buffers(pipe, num_vbuffers, vbuffer);
-   pipe->set_vertex_elements(pipe, num_velements, velements);
+   cso_set_vertex_elements(ctx->st->cso_context, num_velements, velements);
 
    if (num_vbuffers == 0 || num_velements == 0)
       return;
diff --git a/src/mesa/state_tracker/st_draw_feedback.c b/src/mesa/state_tracker/st_draw_feedback.c
index 087f2f22bb..26a5b3fcd6 100644
--- a/src/mesa/state_tracker/st_draw_feedback.c
+++ b/src/mesa/state_tracker/st_draw_feedback.c
@@ -178,7 +178,6 @@ st_feedback_draw_vbo(GLcontext *ctx,
       vbuffers[attr].max_index = max_index;
       velements[attr].instance_divisor = 0;
       velements[attr].vertex_buffer_index = attr;
-      velements[attr].nr_components = arrays[mesaAttr]->Size;
       velements[attr].src_format = 
          st_pipe_vertex_format(arrays[mesaAttr]->Type,
                                arrays[mesaAttr]->Size,