nv50,nvc0: handle user vertex buffers

And restructure VBO validation a little in the process.

14 files changed, 423 insertions, 308 deletions

diff --git a/src/gallium/drivers/nouveau/nouveau_buffer.c b/src/gallium/drivers/nouveau/nouveau_buffer.c
index 936e2bf246a..c396e3c3986 100644
--- a/src/gallium/drivers/nouveau/nouveau_buffer.c
+++ b/src/gallium/drivers/nouveau/nouveau_buffer.c
@@ -604,30 +604,28 @@ nouveau_scratch_more(struct nouveau_context *nv, unsigned min_size)
    return ret;
 }
 
-/* Upload data to scratch memory and update buffer address.
- * Returns the bo the data resides in, if successful.
- */
-struct nouveau_bo *
+
+/* Copy data to a scratch buffer and return address & bo the data resides in. */
+uint64_t
 nouveau_scratch_data(struct nouveau_context *nv,
-                     struct nv04_resource *buf, unsigned base, unsigned size)
+                     const void *data, unsigned base, unsigned size,
+                     struct nouveau_bo **bo)
 {
-   struct nouveau_bo *bo;
    unsigned bgn = MAX2(base, nv->scratch.offset);
    unsigned end = bgn + size;
 
    if (end >= nv->scratch.end) {
       end = base + size;
       if (!nouveau_scratch_more(nv, end))
-         return NULL;
+         return 0;
       bgn = base;
    }
    nv->scratch.offset = align(end, 4);
 
-   memcpy(nv->scratch.map + bgn, buf->data + base, size);
+   memcpy(nv->scratch.map + bgn, (const uint8_t *)data + base, size);
 
-   bo = nv->scratch.current;
-   buf->address = bo->offset + (bgn - base);
-   return bo;
+   *bo = nv->scratch.current;
+   return (*bo)->offset + (bgn - base);
 }
 
 void *
diff --git a/src/gallium/drivers/nouveau/nouveau_buffer.h b/src/gallium/drivers/nouveau/nouveau_buffer.h
index 19255a3c0fd..3b8ee72e72a 100644
--- a/src/gallium/drivers/nouveau/nouveau_buffer.h
+++ b/src/gallium/drivers/nouveau/nouveau_buffer.h
@@ -89,11 +89,12 @@ boolean
 nouveau_user_buffer_upload(struct nouveau_context *, struct nv04_resource *,
                            unsigned base, unsigned size);
 
-/* Copy data to a scratch buffer, update buffer address.
- * Returns the bo the data resides in, or NULL on failure.
+/* Copy data to a scratch buffer and return address & bo the data resides in.
+ * Returns 0 on failure.
  */
-struct nouveau_bo *
+uint64_t
 nouveau_scratch_data(struct nouveau_context *,
-                     struct nv04_resource *, unsigned base, unsigned size);
+                     const void *data, unsigned base, unsigned size,
+                     struct nouveau_bo **);
 
 #endif
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index 8f5363b3684..5da0473f0e6 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -123,8 +123,11 @@ struct nv50_context {
    struct pipe_index_buffer idxbuf;
    uint32_t vbo_fifo; /* bitmask of vertex elements to be pushed to FIFO */
    uint32_t vbo_user; /* bitmask of vertex buffers pointing to user memory */
+   uint32_t vbo_constant; /* bitmask of user buffers with stride 0 */
    uint32_t vb_elt_first; /* from pipe_draw_info, for vertex upload */
    uint32_t vb_elt_limit; /* max - min element (count - 1) */
+   uint32_t instance_off; /* base vertex for instanced arrays */
+   uint32_t instance_max; /* max instance for current draw call */
 
    struct pipe_sampler_view *textures[3][PIPE_MAX_SAMPLERS];
    unsigned num_textures[3];
diff --git a/src/gallium/drivers/nv50/nv50_push.c b/src/gallium/drivers/nv50/nv50_push.c
index 1f7bc7679f0..eb6bfbccfa7 100644
--- a/src/gallium/drivers/nv50/nv50_push.c
+++ b/src/gallium/drivers/nv50/nv50_push.c
@@ -220,15 +220,17 @@ nv50_push_vbo(struct nv50_context *nv50, const struct pipe_draw_info *info)
    ctx.vertex_words = nv50->vertex->vertex_size;
 
    for (i = 0; i < nv50->num_vtxbufs; ++i) {
-      uint8_t *data;
-      struct pipe_vertex_buffer *vb = &nv50->vtxbuf[i];
-      struct nv04_resource *res = nv04_resource(vb->buffer);
+      const struct pipe_vertex_buffer *vb = &nv50->vtxbuf[i];
+      const uint8_t *data;
 
-      data = nouveau_resource_map_offset(&nv50->base, res,
-                                         vb->buffer_offset, NOUVEAU_BO_RD);
+      if (unlikely(vb->buffer))
+         data = nouveau_resource_map_offset(&nv50->base,
+            nv04_resource(vb->buffer), vb->buffer_offset, NOUVEAU_BO_RD);
+      else
+         data = vb->user_buffer;
 
       if (apply_bias && likely(!(nv50->vertex->instance_bufs & (1 << i))))
-         data += info->index_bias * vb->stride;
+         data += (ptrdiff_t)info->index_bias * vb->stride;
 
       ctx.translate->set_buffer(ctx.translate, i, data, vb->stride, ~0);
    }
@@ -304,10 +306,4 @@ nv50_push_vbo(struct nv50_context *nv50, const struct pipe_draw_info *info)
       ctx.instance_id++;
       ctx.prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
    }
-
-   if (info->indexed)
-      nouveau_resource_unmap(nv04_resource(nv50->idxbuf.buffer));
-
-   for (i = 0; i < nv50->num_vtxbufs; ++i)
-      nouveau_resource_unmap(nv04_resource(nv50->vtxbuf[i].buffer));
 }
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index 1874f3ea81a..d7efa353381 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -151,10 +151,10 @@ nv50_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
       return 1;
    case PIPE_CAP_TGSI_CAN_COMPACT_VARYINGS:
    case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
-   case PIPE_CAP_USER_VERTEX_BUFFERS:
       return 0; /* state trackers will know better */
    case PIPE_CAP_USER_CONSTANT_BUFFERS:
    case PIPE_CAP_USER_INDEX_BUFFERS:
+   case PIPE_CAP_USER_VERTEX_BUFFERS:
       return 1;
    case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
       return 256;
diff --git a/src/gallium/drivers/nv50/nv50_state.c b/src/gallium/drivers/nv50/nv50_state.c
index 1e7d17a7197..81c3fa24ab0 100644
--- a/src/gallium/drivers/nv50/nv50_state.c
+++ b/src/gallium/drivers/nv50/nv50_state.c
@@ -889,12 +889,23 @@ nv50_set_vertex_buffers(struct pipe_context *pipe,
    struct nv50_context *nv50 = nv50_context(pipe);
    unsigned i;
 
-   for (i = 0; i < count; ++i)
+   nv50->vbo_user = nv50->vbo_constant = 0;
+
+   for (i = 0; i < count; ++i) {
+      nv50->vtxbuf[i].stride = vb[i].stride;
       pipe_resource_reference(&nv50->vtxbuf[i].buffer, vb[i].buffer);
+      if (!vb[i].buffer && vb[i].user_buffer) {
+         nv50->vtxbuf[i].user_buffer = vb[i].user_buffer;
+         nv50->vbo_user |= 1 << i;
+         if (!vb[i].stride)
+            nv50->vbo_constant |= 1 << i;
+      } else {
+         nv50->vtxbuf[i].buffer_offset = vb[i].buffer_offset;
+      }
+   }
    for (; i < nv50->num_vtxbufs; ++i)
       pipe_resource_reference(&nv50->vtxbuf[i].buffer, NULL);
 
-   memcpy(nv50->vtxbuf, vb, sizeof(*vb) * count);
    nv50->num_vtxbufs = count;
 
    nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_VERTEX);
diff --git a/src/gallium/drivers/nv50/nv50_stateobj.h b/src/gallium/drivers/nv50/nv50_stateobj.h
index eec146552b4..f75608ca726 100644
--- a/src/gallium/drivers/nv50/nv50_stateobj.h
+++ b/src/gallium/drivers/nv50/nv50_stateobj.h
@@ -50,11 +50,12 @@ struct nv50_vertex_element {
 };
 
 struct nv50_vertex_stateobj {
+   uint32_t min_instance_div[PIPE_MAX_ATTRIBS];
+   uint16_t vb_access_size[PIPE_MAX_ATTRIBS];
    struct translate *translate;
    unsigned num_elements;
    uint32_t instance_elts;
    uint32_t instance_bufs;
-   uint16_t vb_access_size[PIPE_MAX_ATTRIBS];
    boolean need_conversion;
    unsigned vertex_size;
    unsigned packet_vertex_limit;
diff --git a/src/gallium/drivers/nv50/nv50_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index d21d699bebe..4cecdea3bfb 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -62,6 +62,9 @@ nv50_vertex_state_create(struct pipe_context *pipe,
 
     memset(so->vb_access_size, 0, sizeof(so->vb_access_size));
 
+    for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
+       so->min_instance_div[i] = 0xffffffff;
+
     transkey.nr_elements = 0;
     transkey.output_stride = 0;
 
@@ -109,6 +112,8 @@ nv50_vertex_state_create(struct pipe_context *pipe,
             if (unlikely(ve->instance_divisor)) {
                so->instance_elts |= 1 << i;
                so->instance_bufs |= 1 << vbi;
+               if (ve->instance_divisor < so->min_instance_div[vbi])
+                  so->min_instance_div[vbi] = ve->instance_divisor;
             }
         }
     }
@@ -130,14 +135,12 @@ static void
 nv50_emit_vtxattr(struct nv50_context *nv50, struct pipe_vertex_buffer *vb,
                   struct pipe_vertex_element *ve, unsigned attr)
 {
-   const void *data;
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
-   struct nv04_resource *res = nv04_resource(vb->buffer);
+   const void *data = (const uint8_t *)vb->user_buffer + ve->src_offset;
    float v[4];
    const unsigned nc = util_format_get_nr_components(ve->src_format);
 
-   data = nouveau_resource_map_offset(&nv50->base, res, vb->buffer_offset +
-                                      ve->src_offset, NOUVEAU_BO_RD);
+   assert(vb->user_buffer);
 
    util_format_read_4f(ve->src_format, v, 0, data, 0, 0, 0, 1, 1);
 
@@ -175,8 +178,8 @@ nv50_emit_vtxattr(struct nv50_context *nv50, struct pipe_vertex_buffer *vb,
 }
 
 static INLINE void
-nv50_vbuf_range(struct nv50_context *nv50, int vbi,
-                uint32_t *base, uint32_t *size)
+nv50_user_vbuf_range(struct nv50_context *nv50, int vbi,
+                     uint32_t *base, uint32_t *size)
 {
    if (unlikely(nv50->vertex->instance_bufs & (1 << vbi))) {
       /* TODO: use min and max instance divisor to get a proper range */
@@ -192,66 +195,43 @@ nv50_vbuf_range(struct nv50_context *nv50, int vbi,
 }
 
 static void
-nv50_prevalidate_vbufs(struct nv50_context *nv50, unsigned limits[])
+nv50_upload_user_buffers(struct nv50_context *nv50,
+                         uint64_t addrs[], uint32_t limits[])
 {
-   const uint32_t bo_flags = NOUVEAU_BO_RD | NOUVEAU_BO_GART;
-   struct nouveau_bo *bo;
-   struct pipe_vertex_buffer *vb;
-   struct nv04_resource *buf;
-   int i;
-   uint32_t base, size;
+   unsigned b;
 
-   nv50->vbo_fifo = nv50->vbo_user = 0;
+   for (b = 0; b < nv50->num_vtxbufs; ++b) {
+      struct nouveau_bo *bo;
+      const struct pipe_vertex_buffer *vb = &nv50->vtxbuf[b];
+      uint32_t base, size;
 
-   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_VERTEX);
-
-   for (i = 0; i < nv50->num_vtxbufs; ++i) {
-      vb = &nv50->vtxbuf[i];
-      limits[i] = 0;
-      if (!vb->stride)
+      if (!(nv50->vbo_user & (1 << b)) || !vb->stride)
          continue;
-      buf = nv04_resource(vb->buffer);
-
-      if (nouveau_resource_mapped_by_gpu(vb->buffer)) {
-         BCTX_REFN(nv50->bufctx_3d, VERTEX, buf, RD);
-      } else {
-         if (nv50->vbo_push_hint) {
-            nv50->vbo_fifo = ~0;
-            return;
-         }
-         nv50->base.vbo_dirty = TRUE;
-
-         if (buf->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY) {
-            assert(vb->stride > vb->buffer_offset);
-            nv50->vbo_user |= 1 << i;
-            nv50_vbuf_range(nv50, i, &base, &size);
-            limits[i] = base + size - 1;
-            bo = nouveau_scratch_data(&nv50->base, buf, base, size);
-            if (bo)
-               BCTX_REFN_bo(nv50->bufctx_3d, VERTEX_TMP, bo_flags, bo);
-         } else {
-            if (nouveau_buffer_migrate(&nv50->base, buf, NOUVEAU_BO_GART))
-               BCTX_REFN(nv50->bufctx_3d, VERTEX, buf, RD);
-         }
-      }
+      nv50_user_vbuf_range(nv50, b, &base, &size);
+
+      limits[b] = base + size - 1;
+      addrs[b] = nouveau_scratch_data(&nv50->base, vb->user_buffer, base, size,
+                                      &bo);
+      if (addrs[b])
+         BCTX_REFN_bo(nv50->bufctx_3d, VERTEX_TMP, NOUVEAU_BO_GART |
+                      NOUVEAU_BO_RD, bo);
    }
+   nv50->base.vbo_dirty = TRUE;
 }
 
 static void
 nv50_update_user_vbufs(struct nv50_context *nv50)
 {
-   const uint32_t bo_flags = NOUVEAU_BO_RD | NOUVEAU_BO_GART;
-   struct nouveau_bo *bo;
+   uint64_t address[PIPE_MAX_ATTRIBS];
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
-   uint32_t base, offset, size;
-   int i;
+   unsigned i;
    uint32_t written = 0;
 
    for (i = 0; i < nv50->vertex->num_elements; ++i) {
       struct pipe_vertex_element *ve = &nv50->vertex->element[i].pipe;
-      const int b = ve->vertex_buffer_index;
+      const unsigned b = ve->vertex_buffer_index;
       struct pipe_vertex_buffer *vb = &nv50->vtxbuf[b];
-      struct nv04_resource *buf = nv04_resource(vb->buffer);
+      uint32_t base, size;
 
       if (!(nv50->vbo_user & (1 << b)))
          continue;
@@ -260,22 +240,24 @@ nv50_update_user_vbufs(struct nv50_context *nv50)
          nv50_emit_vtxattr(nv50, vb, ve, i);
          continue;
       }
-      nv50_vbuf_range(nv50, b, &base, &size);
+      nv50_user_vbuf_range(nv50, b, &base, &size);
 
       if (!(written & (1 << b))) {
+         struct nouveau_bo *bo;
+         const uint32_t bo_flags = NOUVEAU_BO_GART | NOUVEAU_BO_RD;
          written |= 1 << b;
-         bo = nouveau_scratch_data(&nv50->base, buf, base, size);
-         if (bo)
+         address[b] = nouveau_scratch_data(&nv50->base, vb->user_buffer,
+                                           base, size, &bo);
+         if (address[b])
             BCTX_REFN_bo(nv50->bufctx_3d, VERTEX_TMP, bo_flags, bo);
       }
-      offset = vb->buffer_offset + ve->src_offset;
 
       BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
-      PUSH_DATAh(push, buf->address + base + size - 1);
-      PUSH_DATA (push, buf->address + base + size - 1);
+      PUSH_DATAh(push, address[b] + base + size - 1);
+      PUSH_DATA (push, address[b] + base + size - 1);
       BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
-      PUSH_DATAh(push, buf->address + offset);
-      PUSH_DATA (push, buf->address + offset);
+      PUSH_DATAh(push, address[b] + ve->src_offset);
+      PUSH_DATA (push, address[b] + ve->src_offset);
    }
    nv50->base.vbo_dirty = TRUE;
 }
@@ -292,90 +274,112 @@ nv50_release_user_vbufs(struct nv50_context *nv50)
 void
 nv50_vertex_arrays_validate(struct nv50_context *nv50)
 {
+   uint64_t addrs[PIPE_MAX_ATTRIBS];
+   uint32_t limits[PIPE_MAX_ATTRIBS];
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_vertex_stateobj *vertex = nv50->vertex;
    struct pipe_vertex_buffer *vb;
    struct nv50_vertex_element *ve;
+   uint32_t mask;
+   uint32_t refd = 0;
    unsigned i;
-   unsigned limits[PIPE_MAX_ATTRIBS]; /* user vertex buffer limits */
+   const unsigned n = MAX2(vertex->num_elements, nv50->state.num_vtxelts);
 
-   if (unlikely(vertex->need_conversion)) {
+   if (unlikely(vertex->need_conversion))
       nv50->vbo_fifo = ~0;
-      nv50->vbo_user = 0;
-   } else {
-      nv50_prevalidate_vbufs(nv50, limits);
+   else
+   if (nv50->vbo_user & ~nv50->vbo_constant)
+      nv50->vbo_fifo = nv50->vbo_push_hint ? ~0 : 0;
+   else
+      nv50->vbo_fifo = 0;
+
+   /* update vertex format state */
+   BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_ATTRIB(0)), n);
+   if (nv50->vbo_fifo) {
+      nv50->state.num_vtxelts = vertex->num_elements;
+      for (i = 0; i < vertex->num_elements; ++i)
+         PUSH_DATA (push, vertex->element[i].state);
+      for (; i < n; ++i)
+         PUSH_DATA (push, NV50_3D_VERTEX_ATTRIB_INACTIVE);
+      for (i = 0; i < n; ++i) {
+         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
+         PUSH_DATA (push, 0);
+      }
+      return;
    }
-
-   BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_ATTRIB(0)), vertex->num_elements);
    for (i = 0; i < vertex->num_elements; ++i) {
+      const unsigned b = vertex->element[i].pipe.vertex_buffer_index;
       ve = &vertex->element[i];
-      vb = &nv50->vtxbuf[ve->pipe.vertex_buffer_index];
+      vb = &nv50->vtxbuf[b];
 
-      if (likely(vb->stride) || nv50->vbo_fifo) {
+      if (likely(vb->stride) || !(nv50->vbo_user & (1 << b)))
          PUSH_DATA(push, ve->state);
-      } else {
+      else
          PUSH_DATA(push, ve->state | NV50_3D_VERTEX_ARRAY_ATTRIB_CONST);
-         nv50->vbo_fifo &= ~(1 << i);
-      }
    }
+   for (; i < n; ++i)
+      PUSH_DATA(push, NV50_3D_VERTEX_ATTRIB_INACTIVE);
+
+   /* update per-instance enables */
+   mask = vertex->instance_elts ^ nv50->state.instance_elts;
+   while (mask) {
+      const int i = ffs(mask) - 1;
+      mask &= ~(1 << i);
+      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
+      PUSH_DATA (push, (vertex->instance_elts >> i) & 1);
+   }
+   nv50->state.instance_elts = vertex->instance_elts;
+
+   if (nv50->vbo_user & ~nv50->vbo_constant)
+      nv50_upload_user_buffers(nv50, addrs, limits);
 
+   /* update buffers and set constant attributes */
    for (i = 0; i < vertex->num_elements; ++i) {
-      struct nv04_resource *res;
-      unsigned limit, offset;
-      
+      uint64_t address, limit;
+      const unsigned b = vertex->element[i].pipe.vertex_buffer_index;
       ve = &vertex->element[i];
-      vb = &nv50->vtxbuf[ve->pipe.vertex_buffer_index];
+      vb = &nv50->vtxbuf[b];
 
-      if (unlikely(ve->pipe.instance_divisor)) {
-         if (!(nv50->state.instance_elts & (1 << i))) {
-            BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
-            PUSH_DATA (push, 1);
-         }
-         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_DIVISOR(i)), 1);
-         PUSH_DATA (push, ve->pipe.instance_divisor);
-      } else
-      if (unlikely(nv50->state.instance_elts & (1 << i))) {
-         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
-         PUSH_DATA (push, 0);
-      }
-
-      res = nv04_resource(vb->buffer);
-
-      if (nv50->vbo_fifo || unlikely(vb->stride == 0)) {
-         if (!nv50->vbo_fifo)
-            nv50_emit_vtxattr(nv50, vb, &ve->pipe, i);
+      if (unlikely(nv50->vbo_constant & (1 << b))) {
          BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
          PUSH_DATA (push, 0);
+         nv50_emit_vtxattr(nv50, vb, &ve->pipe, i);
          continue;
+      } else
+      if (nv50->vbo_user & (1 << b)) {
+         address = addrs[b] + ve->pipe.src_offset;
+         limit = addrs[b] + limits[b];
+      } else {
+         struct nv04_resource *buf = nv04_resource(vb->buffer);
+         if (!(refd & (1 << b))) {
+            refd |= 1 << b;
+            BCTX_REFN(nv50->bufctx_3d, VERTEX, buf, RD);
+         }
+         address = buf->address + vb->buffer_offset + ve->pipe.src_offset;
+         limit = buf->address + buf->base.width0 - 1;
       }
 
-      offset = ve->pipe.src_offset + vb->buffer_offset;
-      limit = limits[ve->pipe.vertex_buffer_index];
-      if (!limit)
-         limit = vb->buffer->width0 - 1;
-
-      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
-      PUSH_DATA (push, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
+      if (unlikely(ve->pipe.instance_divisor)) {
+         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 4);
+         PUSH_DATA (push, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
+         PUSH_DATAh(push, address);
+         PUSH_DATA (push, address);
+         PUSH_DATA (push, ve->pipe.instance_divisor);
+      } else {
+         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 3);
+         PUSH_DATA (push, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
+         PUSH_DATAh(push, address);
+         PUSH_DATA (push, address);
+      }
       BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
-      PUSH_DATAh(push, res->address + limit);
-      PUSH_DATA (push, res->address + limit);
-      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
-      PUSH_DATAh(push, res->address + offset);
-      PUSH_DATA (push, res->address + offset);
+      PUSH_DATAh(push, limit);
+      PUSH_DATA (push, limit);
    }
    for (; i < nv50->state.num_vtxelts; ++i) {
-      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_ATTRIB(i)), 1);
-      PUSH_DATA (push, NV50_3D_VERTEX_ATTRIB_INACTIVE);
-      if (unlikely(nv50->state.instance_elts & (1 << i))) {
-         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
-         PUSH_DATA (push, 0);
-      }
       BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
       PUSH_DATA (push, 0);
    }
-
    nv50->state.num_vtxelts = vertex->num_elements;
-   nv50->state.instance_elts = vertex->instance_elts;
 }
 
 #define NV50_PRIM_GL_CASE(n) \
@@ -703,6 +707,8 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    /* NOTE: caller must ensure that (min_index + index_bias) is >= 0 */
    nv50->vb_elt_first = info->min_index + info->index_bias;
    nv50->vb_elt_limit = info->max_index - info->min_index;
+   nv50->instance_off = info->start_instance;
+   nv50->instance_max = info->instance_count - 1;
 
    /* For picking only a few vertices from a large user buffer, push is better,
     * if index count is larger and we expect repeated vertices, suggest upload.
@@ -710,11 +716,13 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    nv50->vbo_push_hint = /* the 64 is heuristic */
       !(info->indexed && ((nv50->vb_elt_limit + 64) < info->count));
 
-   if (nv50->vbo_push_hint != !!nv50->vbo_fifo)
-      nv50->dirty |= NV50_NEW_ARRAYS;
-
-   if (nv50->vbo_user && !(nv50->dirty & (NV50_NEW_VERTEX | NV50_NEW_ARRAYS)))
-      nv50_update_user_vbufs(nv50);
+   if (nv50->vbo_user && !(nv50->dirty & (NV50_NEW_ARRAYS | NV50_NEW_VERTEX))) {
+      if (!!nv50->vbo_fifo != nv50->vbo_push_hint)
+         nv50->dirty |= NV50_NEW_ARRAYS;
+      else
+      if (!nv50->vbo_fifo)
+         nv50_update_user_vbufs(nv50);
+   }
 
    if (unlikely(nv50->num_so_targets && !nv50->gmtyprog))
       nv50->state.prim_size = nv50_pipe_prim_to_prim_size[info->mode];
diff --git a/src/gallium/drivers/nvc0/nvc0_context.h b/src/gallium/drivers/nvc0/nvc0_context.h
index 261cfb11007..ef792227b34 100644
--- a/src/gallium/drivers/nvc0/nvc0_context.h
+++ b/src/gallium/drivers/nvc0/nvc0_context.h
@@ -131,6 +131,8 @@ struct nvc0_context {
    uint32_t vbo_user; /* bitmask of vertex buffers pointing to user memory */
    uint32_t vb_elt_first; /* from pipe_draw_info, for vertex upload */
    uint32_t vb_elt_limit; /* max - min element (count - 1) */
+   uint32_t instance_off; /* current base vertex for instanced arrays */
+   uint32_t instance_max; /* last instance for current draw call */
 
    struct pipe_sampler_view *textures[5][PIPE_MAX_SAMPLERS];
    unsigned num_textures[5];
diff --git a/src/gallium/drivers/nvc0/nvc0_screen.c b/src/gallium/drivers/nvc0/nvc0_screen.c
index 3698e71924f..0717ac8f725 100644
--- a/src/gallium/drivers/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nvc0/nvc0_screen.c
@@ -139,10 +139,10 @@ nvc0_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
       return 1;
    case PIPE_CAP_TGSI_CAN_COMPACT_VARYINGS:
    case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
-   case PIPE_CAP_USER_VERTEX_BUFFERS:
       return 0; /* state trackers will know better */
    case PIPE_CAP_USER_CONSTANT_BUFFERS:
    case PIPE_CAP_USER_INDEX_BUFFERS:
+   case PIPE_CAP_USER_VERTEX_BUFFERS:
       return 1;
    case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
       return 256;
diff --git a/src/gallium/drivers/nvc0/nvc0_state.c b/src/gallium/drivers/nvc0/nvc0_state.c
index 5eee9d4bf88..01bd5cef484 100644
--- a/src/gallium/drivers/nvc0/nvc0_state.c
+++ b/src/gallium/drivers/nvc0/nvc0_state.c
@@ -756,13 +756,20 @@ nvc0_set_vertex_buffers(struct pipe_context *pipe,
     uint32_t constant_vbos = 0;
     unsigned i;
 
+    nvc0->vbo_user = 0;
+
     if (count != nvc0->num_vtxbufs) {
        for (i = 0; i < count; ++i) {
           pipe_resource_reference(&nvc0->vtxbuf[i].buffer, vb[i].buffer);
-          nvc0->vtxbuf[i].buffer_offset = vb[i].buffer_offset;
+          if (vb[i].user_buffer) {
+             nvc0->vbo_user |= 1 << i;
+             nvc0->vtxbuf[i].user_buffer = vb[i].user_buffer;
+             if (!vb[i].stride)
+                constant_vbos |= 1 << i;
+          } else {
+             nvc0->vtxbuf[i].buffer_offset = vb[i].buffer_offset;
+          }
           nvc0->vtxbuf[i].stride = vb[i].stride;
-          if (!vb[i].stride)
-             constant_vbos |= 1 << i;
        }
        for (; i < nvc0->num_vtxbufs; ++i)
           pipe_resource_reference(&nvc0->vtxbuf[i].buffer, NULL);
@@ -771,6 +778,13 @@ nvc0_set_vertex_buffers(struct pipe_context *pipe,
        nvc0->dirty |= NVC0_NEW_ARRAYS;
     } else {
        for (i = 0; i < count; ++i) {
+          if (vb[i].user_buffer) {
+             nvc0->vtxbuf[i].user_buffer = vb[i].user_buffer;
+             nvc0->vbo_user |= 1 << i;
+             if (!vb[i].stride)
+                constant_vbos |= 1 << i;
+             assert(!vb[i].buffer);
+          }
           if (nvc0->vtxbuf[i].buffer == vb[i].buffer &&
               nvc0->vtxbuf[i].buffer_offset == vb[i].buffer_offset &&
               nvc0->vtxbuf[i].stride == vb[i].stride)
@@ -778,10 +792,7 @@ nvc0_set_vertex_buffers(struct pipe_context *pipe,
           pipe_resource_reference(&nvc0->vtxbuf[i].buffer, vb[i].buffer);
           nvc0->vtxbuf[i].buffer_offset = vb[i].buffer_offset;
           nvc0->vtxbuf[i].stride = vb[i].stride;
-          if (likely(vb[i].stride))
-             nvc0->dirty |= NVC0_NEW_ARRAYS;
-          else
-             constant_vbos |= 1 << i;
+          nvc0->dirty |= NVC0_NEW_ARRAYS;
        }
     }
     if (constant_vbos != nvc0->constant_vbos) {
diff --git a/src/gallium/drivers/nvc0/nvc0_stateobj.h b/src/gallium/drivers/nvc0/nvc0_stateobj.h
index 5afbffb7d22..edab60bea5b 100644
--- a/src/gallium/drivers/nvc0/nvc0_stateobj.h
+++ b/src/gallium/drivers/nvc0/nvc0_stateobj.h
@@ -49,11 +49,13 @@ struct nvc0_vertex_element {
 };
 
 struct nvc0_vertex_stateobj {
+   uint32_t min_instance_div[PIPE_MAX_ATTRIBS];
+   uint16_t vb_access_size[PIPE_MAX_ATTRIBS];
    struct translate *translate;
    unsigned num_elements;
    uint32_t instance_elts;
    uint32_t instance_bufs;
-   uint16_t vb_access_size[PIPE_MAX_ATTRIBS];
+   boolean shared_slots;
    boolean need_conversion; /* e.g. VFETCH cannot convert f64 to f32 */
    unsigned size; /* size of vertex in bytes (when packed) */
    struct nvc0_vertex_element element[0];
diff --git a/src/gallium/drivers/nvc0/nvc0_vbo.c b/src/gallium/drivers/nvc0/nvc0_vbo.c
index c1c9050209a..61a6341b852 100644
--- a/src/gallium/drivers/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nvc0/nvc0_vbo.c
@@ -52,6 +52,7 @@ nvc0_vertex_state_create(struct pipe_context *pipe,
     struct nvc0_vertex_stateobj *so;
     struct translate_key transkey;
     unsigned i;
+    unsigned src_offset_max = 0;
 
     so = MALLOC(sizeof(*so) +
                 num_elements * sizeof(struct nvc0_vertex_element));
@@ -60,10 +61,14 @@ nvc0_vertex_state_create(struct pipe_context *pipe,
     so->num_elements = num_elements;
     so->instance_elts = 0;
     so->instance_bufs = 0;
+    so->shared_slots = FALSE;
     so->need_conversion = FALSE;
 
     memset(so->vb_access_size, 0, sizeof(so->vb_access_size));
 
+    for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
+       so->min_instance_div[i] = 0xffffffff;
+
     transkey.nr_elements = 0;
     transkey.output_stride = 0;
 
@@ -91,12 +96,16 @@ nvc0_vertex_state_create(struct pipe_context *pipe,
         }
         size = util_format_get_blocksize(fmt);
 
+        src_offset_max = MAX2(src_offset_max, ve->src_offset);
+
         if (so->vb_access_size[vbi] < (ve->src_offset + size))
            so->vb_access_size[vbi] = ve->src_offset + size;
 
         if (unlikely(ve->instance_divisor)) {
            so->instance_elts |= 1 << i;
            so->instance_bufs |= 1 << vbi;
+           if (ve->instance_divisor < so->min_instance_div[vbi])
+              so->min_instance_div[vbi] = ve->instance_divisor;
         }
 
         if (1) {
@@ -129,6 +138,17 @@ nvc0_vertex_state_create(struct pipe_context *pipe,
     so->size = transkey.output_stride;
     so->translate = translate_create(&transkey);
 
+    if (so->instance_elts || src_offset_max >= (1 << 14))
+       return so;
+    so->shared_slots = TRUE;
+
+    for (i = 0; i < num_elements; ++i) {
+       const unsigned b = elements[i].vertex_buffer_index;
+       const unsigned s = elements[i].src_offset;
+       so->element[i].state &= ~NVC0_3D_VERTEX_ATTRIB_FORMAT_BUFFER__MASK;
+       so->element[i].state |= b << NVC0_3D_VERTEX_ATTRIB_FORMAT_BUFFER__SHIFT;
+       so->element[i].state |= s << NVC0_3D_VERTEX_ATTRIB_FORMAT_OFFSET__SHIFT;
+    }
     return so;
 }
 
@@ -143,54 +163,47 @@ nvc0_vertex_state_create(struct pipe_context *pipe,
     ((c) << NVC0_3D_VTX_ATTR_DEFINE_COMP__SHIFT))
 
 static void
-nvc0_update_constant_vertex_attribs(struct nvc0_context *nvc0)
+nvc0_set_constant_vertex_attrib(struct nvc0_context *nvc0, const unsigned a)
 {
-   uint32_t mask = nvc0->state.constant_elts;
-
-   while (unlikely(mask)) {
-      const int i = ffs(mask) - 1;
-      uint32_t mode;
-      struct nouveau_pushbuf *push = nvc0->base.pushbuf;
-      struct pipe_vertex_element *ve = &nvc0->vertex->element[i].pipe;
-      struct pipe_vertex_buffer *vb = &nvc0->vtxbuf[ve->vertex_buffer_index];
-      const struct util_format_description *desc;
-      void *dst;
-      const void *src = nouveau_resource_map_offset(&nvc0->base,
-         nv04_resource(vb->buffer),
-         vb->buffer_offset + ve->src_offset, NOUVEAU_BO_RD);
-
-      mask &= ~(1 << i);
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   struct pipe_vertex_element *ve = &nvc0->vertex->element[a].pipe;
+   struct pipe_vertex_buffer *vb = &nvc0->vtxbuf[ve->vertex_buffer_index];
+   uint32_t mode;
+   const struct util_format_description *desc;
+   void *dst;
+   const void *src = (const uint8_t *)vb->user_buffer + ve->src_offset;
+   assert(!vb->buffer);
 
-      desc = util_format_description(ve->src_format);
+   desc = util_format_description(ve->src_format);
 
-      PUSH_SPACE(push, 6);
-      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 5);
-      dst = push->cur + 1;
-      if (desc->channel[0].pure_integer) {
-         if (desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
-            mode = VTX_ATTR(i, 4, SINT, 32);
-            desc->unpack_rgba_sint(dst, 0, src, 0, 1, 1);
-         } else {
-            mode = VTX_ATTR(i, 4, UINT, 32);
-            desc->unpack_rgba_uint(dst, 0, src, 0, 1, 1);
-         }
+   PUSH_SPACE(push, 6);
+   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 5);
+   dst = &push->cur[1];
+   if (desc->channel[0].pure_integer) {
+      if (desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
+         mode = VTX_ATTR(a, 4, SINT, 32);
+         desc->unpack_rgba_sint(dst, 0, src, 0, 1, 1);
       } else {
-         mode = VTX_ATTR(i, 4, FLOAT, 32);
-         desc->unpack_rgba_float(dst, 0, src, 0, 1, 1);
+         mode = VTX_ATTR(a, 4, UINT, 32);
+         desc->unpack_rgba_uint(dst, 0, src, 0, 1, 1);
       }
-      *push->cur = mode;
-      push->cur += 5;
+   } else {
+      mode = VTX_ATTR(a, 4, FLOAT, 32);
+      desc->unpack_rgba_float(dst, 0, src, 0, 1, 1);
    }
+   push->cur[0] = mode;
+   push->cur += 5;
 }
 
 static INLINE void
-nvc0_vbuf_range(struct nvc0_context *nvc0, int vbi,
-                uint32_t *base, uint32_t *size)
+nvc0_user_vbuf_range(struct nvc0_context *nvc0, int vbi,
+                     uint32_t *base, uint32_t *size)
 {
    if (unlikely(nvc0->vertex->instance_bufs & (1 << vbi))) {
-      /* TODO: use min and max instance divisor to get a proper range */
-      *base = 0;
-      *size = nvc0->vtxbuf[vbi].buffer->width0;
+      const uint32_t div = nvc0->vertex->min_instance_div[vbi];
+      *base = nvc0->instance_off * nvc0->vtxbuf[vbi].stride;
+      *size = (nvc0->instance_max / div) * nvc0->vtxbuf[vbi].stride +
+         nvc0->vertex->vb_access_size[vbi];
    } else {
       /* NOTE: if there are user buffers, we *must* have index bounds */
       assert(nvc0->vb_elt_limit != ~0);
@@ -200,100 +213,196 @@ nvc0_vbuf_range(struct nvc0_context *nvc0, int vbi,
    }
 }
 
-/* Return whether to use alternative vertex submission mode (translate),
- * and validate vertex buffers and upload user arrays (if normal mode).
- */
-static uint8_t
-nvc0_prevalidate_vbufs(struct nvc0_context *nvc0, unsigned limits[])
+static INLINE void
+nvc0_release_user_vbufs(struct nvc0_context *nvc0)
 {
-   const uint32_t bo_flags = NOUVEAU_BO_RD | NOUVEAU_BO_GART;
-   struct nouveau_bo *bo;
-   struct pipe_vertex_buffer *vb;
-   struct nv04_resource *buf;
-   int i;
-   uint32_t base, size;
-
-   nvc0->vbo_user = 0;
-
-   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_VTX);
-
-   for (i = 0; i < nvc0->num_vtxbufs; ++i) {
-      vb = &nvc0->vtxbuf[i];
-      limits[i] = 0;
-      if (!vb->stride)
-         continue;
-      buf = nv04_resource(vb->buffer);
-
-      if (!nouveau_resource_mapped_by_gpu(vb->buffer)) {
-         if (nvc0->vbo_push_hint)
-            return 1;
-         nvc0->base.vbo_dirty = TRUE;
-
-         if (buf->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY) {
-            assert(vb->stride > vb->buffer_offset);
-            nvc0->vbo_user |= 1 << i;
-            nvc0_vbuf_range(nvc0, i, &base, &size);
-            limits[i] = base + size - 1;
-            bo = nouveau_scratch_data(&nvc0->base, buf, base, size);
-            if (bo)
-               BCTX_REFN_bo(nvc0->bufctx_3d, VTX_TMP, bo_flags, bo);
-            continue;
-         } else {
-            nouveau_buffer_migrate(&nvc0->base, buf, NOUVEAU_BO_GART);
-         }
-      }
-      BCTX_REFN(nvc0->bufctx_3d, VTX, buf, RD);
+   if (nvc0->vbo_user) {
+      nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_VTX_TMP);
+      nouveau_scratch_done(&nvc0->base);
    }
-   return 0;
 }
 
 static void
 nvc0_update_user_vbufs(struct nvc0_context *nvc0)
 {
-   const uint32_t bo_flags = NOUVEAU_BO_RD | NOUVEAU_BO_GART;
-   struct nouveau_bo *bo;
+   uint64_t address[PIPE_MAX_ATTRIBS];
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
-   uint32_t base, offset, size;
    int i;
    uint32_t written = 0;
 
    PUSH_SPACE(push, nvc0->vertex->num_elements * 8);
-
    for (i = 0; i < nvc0->vertex->num_elements; ++i) {
       struct pipe_vertex_element *ve = &nvc0->vertex->element[i].pipe;
-      const int b = ve->vertex_buffer_index;
+      const unsigned b = ve->vertex_buffer_index;
       struct pipe_vertex_buffer *vb = &nvc0->vtxbuf[b];
-      struct nv04_resource *buf = nv04_resource(vb->buffer);
+      uint32_t base, size;
 
-      if (!(nvc0->vbo_user & (1 << b)) || !vb->stride)
+      if (!(nvc0->vbo_user & (1 << b)))
+         continue;
+      if (!vb->stride) {
+         nvc0_set_constant_vertex_attrib(nvc0, i);
          continue;
-      nvc0_vbuf_range(nvc0, b, &base, &size);
+      }
+      nvc0_user_vbuf_range(nvc0, b, &base, &size);
 
       if (!(written & (1 << b))) {
+         struct nouveau_bo *bo;
+         const uint32_t bo_flags = NOUVEAU_BO_RD | NOUVEAU_BO_GART;
          written |= 1 << b;
-         bo = nouveau_scratch_data(&nvc0->base, buf, base, size);
+         address[b] = nouveau_scratch_data(&nvc0->base, vb->user_buffer,
+                                           base, size, &bo);
          if (bo)
             BCTX_REFN_bo(nvc0->bufctx_3d, VTX_TMP, bo_flags, bo);
       }
-      offset = vb->buffer_offset + ve->src_offset;
 
       BEGIN_1IC0(push, NVC0_3D(MACRO_VERTEX_ARRAY_SELECT), 5);
       PUSH_DATA (push, i);
-      PUSH_DATAh(push, buf->address + base + size - 1);
-      PUSH_DATA (push, buf->address + base + size - 1);
-      PUSH_DATAh(push, buf->address + offset);
-      PUSH_DATA (push, buf->address + offset);
+      PUSH_DATAh(push, address[b] + base + size - 1);
+      PUSH_DATA (push, address[b] + base + size - 1);
+      PUSH_DATAh(push, address[b] + ve->src_offset);
+      PUSH_DATA (push, address[b] + ve->src_offset);
    }
    nvc0->base.vbo_dirty = TRUE;
 }
 
-static INLINE void
-nvc0_release_user_vbufs(struct nvc0_context *nvc0)
+static void
+nvc0_update_user_vbufs_shared(struct nvc0_context *nvc0)
 {
-   if (nvc0->vbo_user) {
-      nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_VTX_TMP);
-      nouveau_scratch_done(&nvc0->base);
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   uint32_t mask = nvc0->vbo_user & ~nvc0->constant_vbos;
+
+   PUSH_SPACE(push, nvc0->num_vtxbufs * 8);
+   while (mask) {
+      struct nouveau_bo *bo;
+      const uint32_t bo_flags = NOUVEAU_BO_RD | NOUVEAU_BO_GART;
+      uint64_t address;
+      uint32_t base, size;
+      const int b = ffs(mask) - 1;
+      mask &= ~(1 << b);
+
+      nvc0_user_vbuf_range(nvc0, b, &base, &size);
+
+      address = nouveau_scratch_data(&nvc0->base, nvc0->vtxbuf[b].user_buffer,
+                                     base, size, &bo);
+      if (bo)
+         BCTX_REFN_bo(nvc0->bufctx_3d, VTX_TMP, bo_flags, bo);
+
+      BEGIN_1IC0(push, NVC0_3D(MACRO_VERTEX_ARRAY_SELECT), 5);
+      PUSH_DATA (push, b);
+      PUSH_DATAh(push, address + base + size - 1);
+      PUSH_DATA (push, address + base + size - 1);
+      PUSH_DATAh(push, address);
+      PUSH_DATA (push, address);
+   }
+
+   mask = nvc0->state.constant_elts;
+   while (mask) {
+      int i = ffs(mask) - 1;
+      mask &= ~(1 << i);
+      nvc0_set_constant_vertex_attrib(nvc0, i);
+   }
+}
+
+static void
+nvc0_validate_vertex_buffers(struct nvc0_context *nvc0)
+{
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   const struct nvc0_vertex_stateobj *vertex = nvc0->vertex;
+   uint32_t refd = 0;
+   unsigned i;
+
+   PUSH_SPACE(push, vertex->num_elements * 8);
+   for (i = 0; i < vertex->num_elements; ++i) {
+      const struct nvc0_vertex_element *ve;
+      const struct pipe_vertex_buffer *vb;
+      struct nv04_resource *res;
+      unsigned b;
+      unsigned limit, offset;
+
+      if (nvc0->state.constant_elts & (1 << i))
+         continue;
+      ve = &vertex->element[i];
+      b = ve->pipe.vertex_buffer_index;
+      vb = &nvc0->vtxbuf[b];
+
+      if (!vb->buffer) {
+         if (vb->stride) {
+            if (ve->pipe.instance_divisor) {
+               BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_DIVISOR(i)), 1);
+               PUSH_DATA (push, ve->pipe.instance_divisor);
+            }
+            BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(i)), 1);
+            PUSH_DATA (push, (1 << 12) | vb->stride);
+         }
+         /* address/value set in nvc0_update_user_vbufs */
+         continue;
+      }
+      res = nv04_resource(vb->buffer);
+      offset = ve->pipe.src_offset + vb->buffer_offset;
+      limit = vb->buffer->width0 - 1;
+
+      if (unlikely(ve->pipe.instance_divisor)) {
+         BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(i)), 4);
+         PUSH_DATA (push, (1 << 12) | vb->stride);
+         PUSH_DATAh(push, res->address + offset);
+         PUSH_DATA (push, res->address + offset);
+         PUSH_DATA (push, ve->pipe.instance_divisor);
+      } else {
+         BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(i)), 3);
+         PUSH_DATA (push, (1 << 12) | vb->stride);
+         PUSH_DATAh(push, res->address + offset);
+         PUSH_DATA (push, res->address + offset);
+      }
+      BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
+      PUSH_DATAh(push, res->address + limit);
+      PUSH_DATA (push, res->address + limit);
+
+      if (!(refd & (1 << b))) {
+         refd |= 1 << b;
+         BCTX_REFN(nvc0->bufctx_3d, VTX, res, RD);
+      }
+   }
+   if (nvc0->vbo_user)
+      nvc0_update_user_vbufs(nvc0);
+}
+
+static void
+nvc0_validate_vertex_buffers_shared(struct nvc0_context *nvc0)
+{
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   unsigned b;
+   const uint32_t mask = nvc0->vbo_user;
+
+   PUSH_SPACE(push, nvc0->num_vtxbufs * 8);
+   for (b = 0; b < nvc0->num_vtxbufs; ++b) {
+      struct pipe_vertex_buffer *vb = &nvc0->vtxbuf[b];
+      struct nv04_resource *buf;
+      uint32_t offset, limit;
+
+      if (mask & (1 << b)) {
+         if (vb->stride) {
+            BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(b)), 1);
+            PUSH_DATA (push, NVC0_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
+         }
+         /* address/value set in nvc0_update_user_vbufs_shared */
+         continue;
+      }
+      buf = nv04_resource(vb->buffer);
+      offset = vb->buffer_offset;
+      limit = buf->base.width0 - 1;
+
+      BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(b)), 3);
+      PUSH_DATA (push, NVC0_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
+      PUSH_DATAh(push, buf->address + offset);
+      PUSH_DATA (push, buf->address + offset);
+      BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_LIMIT_HIGH(b)), 2);
+      PUSH_DATAh(push, buf->address + limit);
+      PUSH_DATA (push, buf->address + limit);
+
+      BCTX_REFN(nvc0->bufctx_3d, VTX, buf, RD);
    }
+   if (nvc0->vbo_user)
+      nvc0_update_user_vbufs_shared(nvc0);
 }
 
 void
@@ -301,20 +410,19 @@ nvc0_vertex_arrays_validate(struct nvc0_context *nvc0)
 {
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
    struct nvc0_vertex_stateobj *vertex = nvc0->vertex;
-   struct pipe_vertex_buffer *vb;
    struct nvc0_vertex_element *ve;
    uint32_t const_vbos;
    unsigned i;
-   unsigned limits[PIPE_MAX_ATTRIBS];
    uint8_t vbo_mode;
    boolean update_vertex;
 
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_VTX);
+
    if (unlikely(vertex->need_conversion) ||
        unlikely(nvc0->vertprog->vp.edgeflag < PIPE_MAX_ATTRIBS)) {
-      nvc0->vbo_user = 0;
       vbo_mode = 3;
    } else {
-      vbo_mode = nvc0_prevalidate_vbufs(nvc0, limits);
+      vbo_mode = (nvc0->vbo_user && nvc0->vbo_push_hint) ? 1 : 0;
    }
    const_vbos = vbo_mode ? 0 : nvc0->constant_vbos;
 
@@ -386,38 +494,10 @@ nvc0_vertex_arrays_validate(struct nvc0_context *nvc0)
    if (nvc0->state.vbo_mode) /* using translate, don't set up arrays here */
       return;
 
-   PUSH_SPACE(push, vertex->num_elements * 8);
-   for (i = 0; i < vertex->num_elements; ++i) {
-      struct nv04_resource *res;
-      unsigned limit, offset;
-
-      if (nvc0->state.constant_elts & (1 << i))
-         continue;
-      ve = &vertex->element[i];
-      vb = &nvc0->vtxbuf[ve->pipe.vertex_buffer_index];
-
-      res = nv04_resource(vb->buffer);
-      offset = ve->pipe.src_offset + vb->buffer_offset;
-      limit = limits[ve->pipe.vertex_buffer_index];
-      if (!limit)
-         limit = vb->buffer->width0 - 1;
-
-      if (unlikely(ve->pipe.instance_divisor)) {
-         BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(i)), 4);
-         PUSH_DATA (push, (1 << 12) | vb->stride);
-         PUSH_DATAh(push, res->address + offset);
-         PUSH_DATA (push, res->address + offset);
-         PUSH_DATA (push, ve->pipe.instance_divisor);
-      } else {
-         BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_FETCH(i)), 3);
-         PUSH_DATA (push, (1 << 12) | vb->stride);
-         PUSH_DATAh(push, res->address + offset);
-         PUSH_DATA (push, res->address + offset);
-      }
-      BEGIN_NVC0(push, NVC0_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
-      PUSH_DATAh(push, res->address + limit);
-      PUSH_DATA (push, res->address + limit);
-   }
+   if (vertex->shared_slots)
+      nvc0_validate_vertex_buffers_shared(nvc0);
+   else
+      nvc0_validate_vertex_buffers(nvc0);
 }
 
 void
@@ -710,6 +790,8 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    /* NOTE: caller must ensure that (min_index + index_bias) is >= 0 */
    nvc0->vb_elt_first = info->min_index + info->index_bias;
    nvc0->vb_elt_limit = info->max_index - info->min_index;
+   nvc0->instance_off = info->start_instance;
+   nvc0->instance_max = info->instance_count - 1;
 
    /* For picking only a few vertices from a large user buffer, push is better,
     * if index count is larger and we expect repeated vertices, suggest upload.
@@ -717,20 +799,18 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    nvc0->vbo_push_hint =
       info->indexed && (nvc0->vb_elt_limit >= (info->count * 2));
 
-   /* Check whether we want to switch vertex-submission mode,
-    * and if not, update user vbufs.
-    */
-   if (!(nvc0->dirty & NVC0_NEW_ARRAYS)) {
-      if (nvc0->vbo_push_hint) {
-         if (nvc0->vbo_user)
-            nvc0->dirty |= NVC0_NEW_ARRAYS; /* switch to translate mode */
-      } else
-      if (nvc0->state.vbo_mode == 1) {
-         nvc0->dirty |= NVC0_NEW_ARRAYS; /* back to normal mode */
+   /* Check whether we want to switch vertex-submission mode. */
+   if (nvc0->vbo_user && !(nvc0->dirty & (NVC0_NEW_ARRAYS | NVC0_NEW_VERTEX))) {
+      if (nvc0->vbo_push_hint != !!nvc0->state.vbo_mode)
+         if (nvc0->state.vbo_mode != 3)
+            nvc0->dirty |= NVC0_NEW_ARRAYS;
+
+      if (!(nvc0->dirty & NVC0_NEW_ARRAYS) && nvc0->state.vbo_mode == 0) {
+         if (nvc0->vertex->shared_slots)
+            nvc0_update_user_vbufs_shared(nvc0);
+         else
+            nvc0_update_user_vbufs(nvc0);
       }
-      if (nvc0->vbo_user &&
-          !(nvc0->dirty & (NVC0_NEW_VERTEX | NVC0_NEW_ARRAYS)))
-         nvc0_update_user_vbufs(nvc0);
    }
 
    /* 8 as minimum to avoid immediate double validation of new buffers */
@@ -743,7 +823,6 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
       push->kick_notify = nvc0_default_kick_notify;
       return;
    }
-   nvc0_update_constant_vertex_attribs(nvc0);
 
    /* space for base instance, flush, and prim restart */
    PUSH_SPACE(push, 8);
diff --git a/src/gallium/drivers/nvc0/nvc0_vbo_translate.c b/src/gallium/drivers/nvc0/nvc0_vbo_translate.c
index 6317c21a8d7..42d9e055c71 100644
--- a/src/gallium/drivers/nvc0/nvc0_vbo_translate.c
+++ b/src/gallium/drivers/nvc0/nvc0_vbo_translate.c
@@ -65,8 +65,11 @@ nvc0_vertex_configure_translate(struct nvc0_context *nvc0, int32_t index_bias)
       const uint8_t *map;
       const struct pipe_vertex_buffer *vb = &nvc0->vtxbuf[i];
 
-      map = nouveau_resource_map_offset(&nvc0->base,
-               nv04_resource(vb->buffer), vb->buffer_offset, NOUVEAU_BO_RD);
+      if (likely(!vb->buffer))
+         map = (const uint8_t *)vb->user_buffer;
+      else
+         map = nouveau_resource_map_offset(&nvc0->base,
+            nv04_resource(vb->buffer), vb->buffer_offset, NOUVEAU_BO_RD);
 
       if (index_bias && !unlikely(nvc0->vertex->instance_bufs & (1 << i)))
          map += (intptr_t)index_bias * vb->stride;