[g3dvl] move zscan into shaders
[mesa.git] / src / gallium / drivers / nv50 / nv50_vbo.c
index 911eabca1cf4b6393ff97009674eec342cf68c0f..abdb9ce2f936a01d43b07067c5283a4ac74c5980 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright 2008 Ben Skeggs
+ * Copyright 2010 Christoph Bumiller
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
 #include "pipe/p_state.h"
 #include "util/u_inlines.h"
 #include "util/u_format.h"
+#include "translate/translate.h"
 
-#include "nouveau/nouveau_util.h"
 #include "nv50_context.h"
 #include "nv50_resource.h"
 
-static INLINE uint32_t
-nv50_vbo_type_to_hw(enum pipe_format format)
-{
-       const struct util_format_description *desc;
-
-       desc = util_format_description(format);
-       assert(desc);
-
-       switch (desc->channel[0].type) {
-       case UTIL_FORMAT_TYPE_FLOAT:
-               return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_FLOAT;
-       case UTIL_FORMAT_TYPE_UNSIGNED:
-               if (desc->channel[0].normalized) {
-                       return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_UNORM;
-               }
-               return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_USCALED;
-       case UTIL_FORMAT_TYPE_SIGNED:
-               if (desc->channel[0].normalized) {
-                       return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_SNORM;
-               }
-               return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_SSCALED;
-       /*
-       case PIPE_FORMAT_TYPE_UINT:
-               return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_UINT;
-       case PIPE_FORMAT_TYPE_SINT:
-               return NV50TCL_VERTEX_ARRAY_ATTRIB_TYPE_SINT; */
-       default:
-               return 0;
-       }
-}
+#include "nv50_3d.xml.h"
 
-static INLINE uint32_t
-nv50_vbo_size_to_hw(unsigned size, unsigned nr_c)
+void
+nv50_vertex_state_delete(struct pipe_context *pipe,
+                         void *hwcso)
 {
-       static const uint32_t hw_values[] = {
-               0, 0, 0, 0,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_8,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_8_8,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_8_8_8,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_8_8_8_8,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_16,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_16_16,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_16_16_16,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_16_16_16_16,
-               0, 0, 0, 0,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_32,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_32_32,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_32_32_32,
-               NV50TCL_VERTEX_ARRAY_ATTRIB_FORMAT_32_32_32_32 };
-
-       /* we'd also have R11G11B10 and R10G10B10A2 */
-
-       assert(nr_c > 0 && nr_c <= 4);
-
-       if (size > 32)
-               return 0;
-       size >>= (3 - 2);
-
-       return hw_values[size + (nr_c - 1)];
+   struct nv50_vertex_stateobj *so = hwcso;
+
+   if (so->translate)
+      so->translate->release(so->translate);
+   FREE(hwcso);
 }
 
-static INLINE uint32_t
-nv50_vbo_vtxelt_to_hw(struct pipe_vertex_element *ve)
+void *
+nv50_vertex_state_create(struct pipe_context *pipe,
+                         unsigned num_elements,
+                         const struct pipe_vertex_element *elements)
 {
-       uint32_t hw_type, hw_size;
-       enum pipe_format pf = ve->src_format;
-       const struct util_format_description *desc;
-       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, nr_components);
-
-       if (!hw_type || !hw_size) {
-               NOUVEAU_ERR("unsupported vbo format: %s\n", util_format_name(pf));
-               abort();
-               return 0x24e80000;
-       }
+    struct nv50_vertex_stateobj *so;
+    struct translate_key transkey;
+    unsigned i;
+
+    so = MALLOC(sizeof(*so) +
+                num_elements * sizeof(struct nv50_vertex_element));
+    if (!so)
+        return NULL;
+    so->num_elements = num_elements;
+    so->instance_elts = 0;
+    so->instance_bufs = 0;
+    so->need_conversion = FALSE;
+
+    transkey.nr_elements = 0;
+    transkey.output_stride = 0;
+
+    for (i = 0; i < num_elements; ++i) {
+        const struct pipe_vertex_element *ve = &elements[i];
+        const unsigned vbi = ve->vertex_buffer_index;
+        enum pipe_format fmt = ve->src_format;
+
+        so->element[i].pipe = elements[i];
+        so->element[i].state = nv50_format_table[fmt].vtx;
+
+        if (!so->element[i].state) {
+            switch (util_format_get_nr_components(fmt)) {
+            case 1: fmt = PIPE_FORMAT_R32_FLOAT; break;
+            case 2: fmt = PIPE_FORMAT_R32G32_FLOAT; break;
+            case 3: fmt = PIPE_FORMAT_R32G32B32_FLOAT; break;
+            case 4: fmt = PIPE_FORMAT_R32G32B32A32_FLOAT; break;
+            default:
+                assert(0);
+                return NULL;
+            }
+            so->element[i].state = nv50_format_table[fmt].vtx;
+            so->need_conversion = TRUE;
+        }
+        so->element[i].state |= i;
+
+        if (1) {
+            unsigned j = transkey.nr_elements++;
+
+            transkey.element[j].type = TRANSLATE_ELEMENT_NORMAL;
+            transkey.element[j].input_format = ve->src_format;
+            transkey.element[j].input_buffer = vbi;
+            transkey.element[j].input_offset = ve->src_offset;
+            transkey.element[j].instance_divisor = ve->instance_divisor;
+
+            transkey.element[j].output_format = fmt;
+            transkey.element[j].output_offset = transkey.output_stride;
+            transkey.output_stride += (util_format_get_stride(fmt, 1) + 3) & ~3;
+
+            if (unlikely(ve->instance_divisor)) {
+               so->instance_elts |= 1 << i;
+               so->instance_bufs |= 1 << vbi;
+            }
+        }
+    }
+
+    so->translate = translate_create(&transkey);
+    so->vertex_size = transkey.output_stride / 4;
+    so->packet_vertex_limit = NV04_PFIFO_MAX_PACKET_LEN /
+       MAX2(so->vertex_size, 1);
+
+    return so;
+}
 
-       if (desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_Z) /* BGRA */
-               hw_size |= (1 << 31); /* no real swizzle bits :-( */
+#define NV50_3D_VERTEX_ATTRIB_INACTIVE              \
+   NV50_3D_VERTEX_ARRAY_ATTRIB_TYPE_FLOAT |         \
+   NV50_3D_VERTEX_ARRAY_ATTRIB_FORMAT_32_32_32_32 | \
+   NV50_3D_VERTEX_ARRAY_ATTRIB_CONST
 
-       return (hw_type | hw_size);
+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_channel *chan = nv50->screen->base.channel;
+   struct nv04_resource *res = nv04_resource(vb->buffer);
+   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);
+
+   util_format_read_4f(ve->src_format, v, 0, data, 0, 0, 0, 1, 1);
+
+   switch (nc) {
+   case 4:
+      BEGIN_RING(chan, RING_3D(VTX_ATTR_4F_X(attr)), 4);
+      OUT_RINGf (chan, v[0]);
+      OUT_RINGf (chan, v[1]);
+      OUT_RINGf (chan, v[2]);
+      OUT_RINGf (chan, v[3]);
+      break;
+   case 3:
+      BEGIN_RING(chan, RING_3D(VTX_ATTR_3F_X(attr)), 3);
+      OUT_RINGf (chan, v[0]);
+      OUT_RINGf (chan, v[1]);
+      OUT_RINGf (chan, v[2]);
+      break;
+   case 2:
+      BEGIN_RING(chan, RING_3D(VTX_ATTR_2F_X(attr)), 2);
+      OUT_RINGf (chan, v[0]);
+      OUT_RINGf (chan, v[1]);
+      break;
+   case 1:
+      if (attr == nv50->vertprog->vp.edgeflag) {
+         BEGIN_RING(chan, RING_3D(EDGEFLAG_ENABLE), 1);
+         OUT_RING  (chan, v[0] ? 1 : 0);
+      }
+      BEGIN_RING(chan, RING_3D(VTX_ATTR_1F(attr)), 1);
+      OUT_RINGf (chan, v[0]);
+      break;
+   default:
+      assert(0);
+      break;
+   }
 }
 
-struct instance {
-       struct nouveau_bo *bo;
-       unsigned delta;
-       unsigned stride;
-       unsigned step;
-       unsigned divisor;
-};
+static INLINE void
+nv50_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 */
+      *base = 0;
+      *size = nv50->vtxbuf[vbi].buffer->width0;
+   } else {
+      assert(nv50->vbo_max_index != ~0);
+      *base = nv50->vbo_min_index * nv50->vtxbuf[vbi].stride;
+      *size = (nv50->vbo_max_index -
+               nv50->vbo_min_index + 1) * nv50->vtxbuf[vbi].stride;
+   }
+}
 
 static void
-instance_init(struct nv50_context *nv50, struct instance *a, unsigned first)
+nv50_prevalidate_vbufs(struct nv50_context *nv50)
 {
-       int i;
-
-       for (i = 0; i < nv50->vtxelt->num_elements; i++) {
-               struct pipe_vertex_element *ve = &nv50->vtxelt->pipe[i];
-               struct pipe_vertex_buffer *vb;
-
-               a[i].divisor = ve->instance_divisor;
-               if (a[i].divisor) {
-                       vb = &nv50->vtxbuf[ve->vertex_buffer_index];
-
-                       a[i].bo = nv50_resource(vb->buffer)->bo;
-                       a[i].stride = vb->stride;
-                       a[i].step = first % a[i].divisor;
-                       a[i].delta = vb->buffer_offset + ve->src_offset +
-                                    (first * a[i].stride);
-               }
-       }
+   struct pipe_vertex_buffer *vb;
+   struct nv04_resource *buf;
+   int i;
+   uint32_t base, size;
+
+   nv50->vbo_fifo = nv50->vbo_user = 0;
+
+   nv50_bufctx_reset(nv50, NV50_BUFCTX_VERTEX);
+
+   for (i = 0; i < nv50->num_vtxbufs; ++i) {
+      vb = &nv50->vtxbuf[i];
+      if (!vb->stride)
+         continue;
+      buf = nv04_resource(vb->buffer);
+
+      /* NOTE: user buffers with temporary storage count as mapped by GPU */
+      if (!nouveau_resource_mapped_by_gpu(vb->buffer)) {
+         if (nv50->vbo_push_hint) {
+            nv50->vbo_fifo = ~0;
+            continue;
+         } else {
+            if (buf->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY) {
+               nv50->vbo_user |= 1 << i;
+               assert(vb->stride > vb->buffer_offset);
+               nv50_vbuf_range(nv50, i, &base, &size);
+               nouveau_user_buffer_upload(buf, base, size);
+            } else {
+               nouveau_buffer_migrate(&nv50->base, buf, NOUVEAU_BO_GART);
+            }
+            nv50->base.vbo_dirty = TRUE;
+         }
+      }
+      nv50_bufctx_add_resident(nv50, NV50_BUFCTX_VERTEX, buf, NOUVEAU_BO_RD);
+      nouveau_buffer_adjust_score(&nv50->base, buf, 1);
+   }
 }
 
 static void
-instance_step(struct nv50_context *nv50, struct instance *a)
+nv50_update_user_vbufs(struct nv50_context *nv50)
 {
-       struct nouveau_channel *chan = nv50->screen->tesla->channel;
-       struct nouveau_grobj *tesla = nv50->screen->tesla;
-       int i;
-
-       for (i = 0; i < nv50->vtxelt->num_elements; i++) {
-               if (!a[i].divisor)
-                       continue;
-
-               BEGIN_RING(chan, tesla,
-                          NV50TCL_VERTEX_ARRAY_START_HIGH(i), 2);
-               OUT_RELOCh(chan, a[i].bo, a[i].delta, NOUVEAU_BO_RD |
-                          NOUVEAU_BO_VRAM | NOUVEAU_BO_GART);
-               OUT_RELOCl(chan, a[i].bo, a[i].delta, NOUVEAU_BO_RD |
-                          NOUVEAU_BO_VRAM | NOUVEAU_BO_GART);
-               if (++a[i].step == a[i].divisor) {
-                       a[i].step = 0;
-                       a[i].delta += a[i].stride;
-               }
-       }
+   struct nouveau_channel *chan = nv50->screen->base.channel;
+   uint32_t base, offset, size;
+   int 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;
+      struct pipe_vertex_buffer *vb = &nv50->vtxbuf[b];
+      struct nv04_resource *buf = nv04_resource(vb->buffer);
+
+      if (!(nv50->vbo_user & (1 << b)))
+         continue;
+
+      if (!vb->stride) {
+         nv50_emit_vtxattr(nv50, vb, ve, i);
+         continue;
+      }
+      nv50_vbuf_range(nv50, b, &base, &size);
+
+      if (!(written & (1 << b))) {
+         written |= 1 << b;
+         nouveau_user_buffer_upload(buf, base, size);
+      }
+      offset = vb->buffer_offset + ve->src_offset;
+
+      MARK_RING (chan, 6, 4);
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
+      OUT_RESRCh(chan, buf, base + size - 1, NOUVEAU_BO_RD);
+      OUT_RESRCl(chan, buf, base + size - 1, NOUVEAU_BO_RD);
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
+      OUT_RESRCh(chan, buf, offset, NOUVEAU_BO_RD);
+      OUT_RESRCl(chan, buf, offset, NOUVEAU_BO_RD);
+   }
+   nv50->base.vbo_dirty = TRUE;
 }
 
-void
-nv50_draw_arrays_instanced(struct pipe_context *pipe,
-                          unsigned mode, unsigned start, unsigned count,
-                          unsigned startInstance, unsigned instanceCount)
+static INLINE void
+nv50_release_user_vbufs(struct nv50_context *nv50)
 {
-       struct nv50_context *nv50 = nv50_context(pipe);
-       struct nouveau_channel *chan = nv50->screen->tesla->channel;
-       struct nouveau_grobj *tesla = nv50->screen->tesla;
-       struct instance a[16];
-       unsigned prim = nv50_prim(mode);
-
-       instance_init(nv50, a, startInstance);
-       if (!nv50_state_validate(nv50, 10 + 16*3))
-               return;
-
-       if (nv50->vbo_fifo) {
-               nv50_push_elements_instanced(pipe, NULL, 0, 0, mode, start,
-                                            count, startInstance,
-                                            instanceCount);
-               return;
-       }
-
-       BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
-       OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
-       OUT_RING  (chan, startInstance);
-       while (instanceCount--) {
-               if (AVAIL_RING(chan) < (7 + 16*3)) {
-                       FIRE_RING(chan);
-                       if (!nv50_state_validate(nv50, 7 + 16*3)) {
-                               assert(0);
-                               return;
-                       }
-               }
-               instance_step(nv50, a);
-
-               BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-               OUT_RING  (chan, prim);
-               BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
-               OUT_RING  (chan, start);
-               OUT_RING  (chan, count);
-               BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-               OUT_RING  (chan, 0);
-
-               prim |= (1 << 28);
-       }
+   uint32_t vbo_user = nv50->vbo_user;
+
+   while (vbo_user) {
+      int i = ffs(vbo_user) - 1;
+      vbo_user &= ~(1 << i);
+
+      nouveau_buffer_release_gpu_storage(nv04_resource(nv50->vtxbuf[i].buffer));
+   }
 }
 
 void
-nv50_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
-                unsigned count)
+nv50_vertex_arrays_validate(struct nv50_context *nv50)
 {
-       nv50_draw_arrays_instanced(pipe, mode, start, count, 0, 1);
+   struct nouveau_channel *chan = nv50->screen->base.channel;
+   struct nv50_vertex_stateobj *vertex = nv50->vertex;
+   struct pipe_vertex_buffer *vb;
+   struct nv50_vertex_element *ve;
+   unsigned i;
+
+   if (unlikely(vertex->need_conversion)) {
+      nv50->vbo_fifo = ~0;
+      nv50->vbo_user = 0;
+   } else {
+      nv50_prevalidate_vbufs(nv50);
+   }
+
+   BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_ATTRIB(0)), vertex->num_elements);
+   for (i = 0; i < vertex->num_elements; ++i) {
+      ve = &vertex->element[i];
+      vb = &nv50->vtxbuf[ve->pipe.vertex_buffer_index];
+
+      if (likely(vb->stride) || nv50->vbo_fifo) {
+         OUT_RING(chan, ve->state);
+      } else {
+         OUT_RING(chan, ve->state | NV50_3D_VERTEX_ARRAY_ATTRIB_CONST);
+         nv50->vbo_fifo &= ~(1 << i);
+      }
+   }
+
+   for (i = 0; i < vertex->num_elements; ++i) {
+      struct nv04_resource *res;
+      unsigned size, offset;
+      
+      ve = &vertex->element[i];
+      vb = &nv50->vtxbuf[ve->pipe.vertex_buffer_index];
+
+      if (unlikely(ve->pipe.instance_divisor)) {
+         if (!(nv50->state.instance_elts & (1 << i))) {
+            BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
+            OUT_RING  (chan, 1);
+         }
+         BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_DIVISOR(i)), 1);
+         OUT_RING  (chan, ve->pipe.instance_divisor);
+      } else
+      if (unlikely(nv50->state.instance_elts & (1 << i))) {
+         BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
+         OUT_RING  (chan, 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);
+         BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FETCH(i)), 1);
+         OUT_RING  (chan, 0);
+         continue;
+      }
+
+      size = vb->buffer->width0;
+      offset = ve->pipe.src_offset + vb->buffer_offset;
+
+      MARK_RING (chan, 8, 4);
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FETCH(i)), 1);
+      OUT_RING  (chan, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
+      OUT_RESRCh(chan, res, size - 1, NOUVEAU_BO_RD);
+      OUT_RESRCl(chan, res, size - 1, NOUVEAU_BO_RD);
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
+      OUT_RESRCh(chan, res, offset, NOUVEAU_BO_RD);
+      OUT_RESRCl(chan, res, offset, NOUVEAU_BO_RD);
+   }
+   for (; i < nv50->state.num_vtxelts; ++i) {
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_ATTRIB(i)), 1);
+      OUT_RING  (chan, NV50_3D_VERTEX_ATTRIB_INACTIVE);
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FETCH(i)), 1);
+      OUT_RING  (chan, 0);
+   }
+
+   nv50->state.num_vtxelts = vertex->num_elements;
+   nv50->state.instance_elts = vertex->instance_elts;
 }
 
-struct inline_ctx {
-       struct nv50_context *nv50;
-       void *map;
-};
+#define NV50_PRIM_GL_CASE(n) \
+   case PIPE_PRIM_##n: return NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_##n
 
-static void
-inline_elt08(void *priv, unsigned start, unsigned count)
+static INLINE unsigned
+nv50_prim_gl(unsigned prim)
 {
-       struct inline_ctx *ctx = priv;
-       struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
-       struct nouveau_channel *chan = tesla->channel;
-       uint8_t *map = (uint8_t *)ctx->map + start;
-
-       if (count & 1) {
-               BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32, 1);
-               OUT_RING  (chan, map[0]);
-               map++;
-               count &= ~1;
-       }
-
-       count >>= 1;
-       if (!count)
-               return;
-
-       BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U16, count);
-       while (count--) {
-               OUT_RING(chan, (map[1] << 16) | map[0]);
-               map += 2;
-       }
+   switch (prim) {
+   NV50_PRIM_GL_CASE(POINTS);
+   NV50_PRIM_GL_CASE(LINES);
+   NV50_PRIM_GL_CASE(LINE_LOOP);
+   NV50_PRIM_GL_CASE(LINE_STRIP);
+   NV50_PRIM_GL_CASE(TRIANGLES);
+   NV50_PRIM_GL_CASE(TRIANGLE_STRIP);
+   NV50_PRIM_GL_CASE(TRIANGLE_FAN);
+   NV50_PRIM_GL_CASE(QUADS);
+   NV50_PRIM_GL_CASE(QUAD_STRIP);
+   NV50_PRIM_GL_CASE(POLYGON);
+   NV50_PRIM_GL_CASE(LINES_ADJACENCY);
+   NV50_PRIM_GL_CASE(LINE_STRIP_ADJACENCY);
+   NV50_PRIM_GL_CASE(TRIANGLES_ADJACENCY);
+   NV50_PRIM_GL_CASE(TRIANGLE_STRIP_ADJACENCY);
+   default:
+      return NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_POINTS;
+      break;
+   }
 }
 
 static void
-inline_elt16(void *priv, unsigned start, unsigned count)
+nv50_draw_vbo_flush_notify(struct nouveau_channel *chan)
 {
-       struct inline_ctx *ctx = priv;
-       struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
-       struct nouveau_channel *chan = tesla->channel;
-       uint16_t *map = (uint16_t *)ctx->map + start;
-
-       if (count & 1) {
-               BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32, 1);
-               OUT_RING  (chan, map[0]);
-               count &= ~1;
-               map++;
-       }
-
-       count >>= 1;
-       if (!count)
-               return;
-
-       BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U16, count);
-       while (count--) {
-               OUT_RING(chan, (map[1] << 16) | map[0]);
-               map += 2;
-       }
-}
+   struct nv50_context *nv50 = chan->user_private;
 
-static void
-inline_elt32(void *priv, unsigned start, unsigned count)
-{
-       struct inline_ctx *ctx = priv;
-       struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
-       struct nouveau_channel *chan = tesla->channel;
+   nouveau_fence_update(&nv50->screen->base, TRUE);
 
-       BEGIN_RING_NI(chan, tesla, NV50TCL_VB_ELEMENT_U32, count);
-       OUT_RINGp    (chan, (uint32_t *)ctx->map + start, count);
+   nv50_bufctx_emit_relocs(nv50);
 }
 
 static void
-inline_edgeflag(void *priv, boolean enabled)
+nv50_draw_arrays(struct nv50_context *nv50,
+                 unsigned mode, unsigned start, unsigned count,
+                 unsigned instance_count)
 {
-       struct inline_ctx *ctx = priv;
-       struct nouveau_grobj *tesla = ctx->nv50->screen->tesla;
-       struct nouveau_channel *chan = tesla->channel;
+   struct nouveau_channel *chan = nv50->screen->base.channel;
+   unsigned prim;
 
-       BEGIN_RING(chan, tesla, NV50TCL_EDGEFLAG_ENABLE, 1);
-       OUT_RING  (chan, enabled ? 1 : 0);
+   chan->flush_notify = nv50_draw_vbo_flush_notify;
+   chan->user_private = nv50;
+
+   prim = nv50_prim_gl(mode);
+
+   while (instance_count--) {
+      BEGIN_RING(chan, RING_3D(VERTEX_BEGIN_GL), 1);
+      OUT_RING  (chan, prim);
+      BEGIN_RING(chan, RING_3D(VERTEX_BUFFER_FIRST), 2);
+      OUT_RING  (chan, start);
+      OUT_RING  (chan, count);
+      BEGIN_RING(chan, RING_3D(VERTEX_END_GL), 1);
+      OUT_RING  (chan, 0);
+
+      prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
+   }
+
+   chan->flush_notify = nv50_default_flush_notify;
 }
 
 static void
-nv50_draw_elements_inline(struct pipe_context *pipe,
-                         struct pipe_resource *indexBuffer, unsigned indexSize,
-                         unsigned mode, unsigned start, unsigned count,
-                         unsigned startInstance, unsigned instanceCount)
+nv50_draw_elements_inline_u08(struct nouveau_channel *chan, uint8_t *map,
+                              unsigned start, unsigned count)
 {
-       struct nv50_context *nv50 = nv50_context(pipe);
-       struct nouveau_channel *chan = nv50->screen->tesla->channel;
-       struct nouveau_grobj *tesla = nv50->screen->tesla;
-       struct pipe_transfer *transfer;
-       struct instance a[16];
-       struct inline_ctx ctx;
-       struct u_split_prim s;
-       boolean nzi = FALSE;
-       unsigned overhead;
-
-       overhead = 16*3; /* potential instance adjustments */
-       overhead += 4; /* Begin()/End() */
-       overhead += 4; /* potential edgeflag disable/reenable */
-       overhead += 3; /* potentially 3 VTX_ELT_U16/U32 packet headers */
-
-       s.priv = &ctx;
-       if (indexSize == 1)
-               s.emit = inline_elt08;
-       else
-       if (indexSize == 2)
-               s.emit = inline_elt16;
-       else
-               s.emit = inline_elt32;
-       s.edge = inline_edgeflag;
-
-       ctx.nv50 = nv50;
-       ctx.map = pipe_buffer_map(pipe, indexBuffer, PIPE_TRANSFER_READ, &transfer);
-       assert(ctx.map);
-       if (!ctx.map)
-               return;
-
-       instance_init(nv50, a, startInstance);
-       if (!nv50_state_validate(nv50, overhead + 6 + 3))
-               return;
-
-       BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
-       OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
-       OUT_RING  (chan, startInstance);
-       while (instanceCount--) {
-               unsigned max_verts;
-               boolean done;
-
-               u_split_prim_init(&s, mode, start, count);
-               do {
-                       if (AVAIL_RING(chan) < (overhead + 6)) {
-                               FIRE_RING(chan);
-                               if (!nv50_state_validate(nv50, (overhead + 6))) {
-                                       assert(0);
-                                       return;
-                               }
-                       }
-
-                       max_verts = AVAIL_RING(chan) - overhead;
-                       if (max_verts > 2047)
-                               max_verts = 2047;
-                       if (indexSize != 4)
-                               max_verts <<= 1;
-                       instance_step(nv50, a);
-
-                       BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-                       OUT_RING  (chan, nv50_prim(s.mode) | (nzi ? (1<<28) : 0));
-                       done = u_split_prim_next(&s, max_verts);
-                       BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-                       OUT_RING  (chan, 0);
-               } while (!done);
-
-               nzi = TRUE;
-       }
-
-       pipe_buffer_unmap(pipe, indexBuffer, transfer);
+   map += start;
+
+   if (count & 3) {
+      unsigned i;
+      BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U32), count & 3);
+      for (i = 0; i < (count & 3); ++i)
+         OUT_RING(chan, *map++);
+      count &= ~3;
+   }
+   while (count) {
+      unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 4) / 4;
+
+      BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U8), nr);
+      for (i = 0; i < nr; ++i) {
+         OUT_RING(chan,
+                  (map[3] << 24) | (map[2] << 16) | (map[1] << 8) | map[0]);
+         map += 4;
+      }
+      count -= nr * 4;
+   }
 }
 
-void
-nv50_draw_elements_instanced(struct pipe_context *pipe,
-                            struct pipe_resource *indexBuffer,
-                            unsigned indexSize, int indexBias,
-                            unsigned mode, unsigned start, unsigned count,
-                            unsigned startInstance, unsigned instanceCount)
+static void
+nv50_draw_elements_inline_u16(struct nouveau_channel *chan, uint16_t *map,
+                              unsigned start, unsigned count)
 {
-       struct nv50_context *nv50 = nv50_context(pipe);
-       struct nouveau_channel *chan = nv50->screen->tesla->channel;
-       struct nouveau_grobj *tesla = nv50->screen->tesla;
-       struct instance a[16];
-       unsigned prim = nv50_prim(mode);
-
-       instance_init(nv50, a, startInstance);
-       if (!nv50_state_validate(nv50, 13 + 16*3))
-               return;
-
-       assert(indexBias == 0);
-
-       if (nv50->vbo_fifo) {
-               nv50_push_elements_instanced(pipe, indexBuffer, indexSize,
-                                            indexBias, mode, start, count,
-                                            startInstance, instanceCount);
-               return;
-       } else
-       if (!(indexBuffer->bind & PIPE_BIND_INDEX_BUFFER) || indexSize == 1) {
-               nv50_draw_elements_inline(pipe, indexBuffer, indexSize,
-                                         mode, start, count, startInstance,
-                                         instanceCount);
-               return;
-       }
-
-       BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 2);
-       OUT_RING  (chan, NV50_CB_AUX | (24 << 8));
-       OUT_RING  (chan, startInstance);
-       while (instanceCount--) {
-               if (AVAIL_RING(chan) < (7 + 16*3)) {
-                       FIRE_RING(chan);
-                       if (!nv50_state_validate(nv50, 10 + 16*3)) {
-                               assert(0);
-                               return;
-                       }
-               }
-               instance_step(nv50, a);
-
-               BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
-               OUT_RING  (chan, prim);
-               if (indexSize == 4) {
-                       BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U32 | 0x30000, 0);
-                       OUT_RING  (chan, count);
-                       nouveau_pushbuf_submit(chan, 
-                                              nv50_resource(indexBuffer)->bo,
-                                              start << 2, count << 2);
-               } else
-               if (indexSize == 2) {
-                       unsigned vb_start = (start & ~1);
-                       unsigned vb_end = (start + count + 1) & ~1;
-                       unsigned dwords = (vb_end - vb_start) >> 1;
-
-                       BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16_SETUP, 1);
-                       OUT_RING  (chan, ((start & 1) << 31) | count);
-                       BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16 | 0x30000, 0);
-                       OUT_RING  (chan, dwords);
-                       nouveau_pushbuf_submit(chan,
-                                              nv50_resource(indexBuffer)->bo,
-                                              vb_start << 1, dwords << 2);
-                       BEGIN_RING(chan, tesla, NV50TCL_VB_ELEMENT_U16_SETUP, 1);
-                       OUT_RING  (chan, 0);
-               }
-               BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
-               OUT_RING  (chan, 0);
-
-               prim |= (1 << 28);
-       }
+   map += start;
+
+   if (count & 1) {
+      count &= ~1;
+      BEGIN_RING(chan, RING_3D(VB_ELEMENT_U32), 1);
+      OUT_RING  (chan, *map++);
+   }
+   while (count) {
+      unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 2) / 2;
+
+      BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U16), nr);
+      for (i = 0; i < nr; ++i) {
+         OUT_RING(chan, (map[1] << 16) | map[0]);
+         map += 2;
+      }
+      count -= nr * 2;
+   }
 }
 
-void
-nv50_draw_elements(struct pipe_context *pipe,
-                  struct pipe_resource *indexBuffer,
-                  unsigned indexSize, int indexBias,
-                  unsigned mode, unsigned start, unsigned count)
+static void
+nv50_draw_elements_inline_u32(struct nouveau_channel *chan, uint32_t *map,
+                              unsigned start, unsigned count)
 {
-       nv50_draw_elements_instanced(pipe, indexBuffer, indexSize, indexBias,
-                                    mode, start, count, 0, 1);
-}
+   map += start;
 
-static INLINE boolean
-nv50_vbo_static_attrib(struct nv50_context *nv50, unsigned attrib,
-                      struct nouveau_stateobj **pso,
-                      struct pipe_vertex_element *ve,
-                      struct pipe_vertex_buffer *vb)
+   while (count) {
+      const unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
 
-{
-       struct nouveau_stateobj *so;
-       struct nouveau_grobj *tesla = nv50->screen->tesla;
-       struct nouveau_bo *bo = nv50_resource(vb->buffer)->bo;
-       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)
-               return FALSE;
-
-       util_format_read_4f(ve->src_format, v, 0, (uint8_t *)bo->map +
-                           (vb->buffer_offset + ve->src_offset), 0,
-                           0, 0, 1, 1);
-       so = *pso;
-       if (!so)
-               *pso = so = so_new(nv50->vtxelt->num_elements,
-                                  nv50->vtxelt->num_elements * 4, 0);
-
-       switch (nr_components) {
-       case 4:
-               so_method(so, tesla, NV50TCL_VTX_ATTR_4F_X(attrib), 4);
-               so_data  (so, fui(v[0]));
-               so_data  (so, fui(v[1]));
-               so_data  (so, fui(v[2]));
-               so_data  (so, fui(v[3]));
-               break;
-       case 3:
-               so_method(so, tesla, NV50TCL_VTX_ATTR_3F_X(attrib), 3);
-               so_data  (so, fui(v[0]));
-               so_data  (so, fui(v[1]));
-               so_data  (so, fui(v[2]));
-               break;
-       case 2:
-               so_method(so, tesla, NV50TCL_VTX_ATTR_2F_X(attrib), 2);
-               so_data  (so, fui(v[0]));
-               so_data  (so, fui(v[1]));
-               break;
-       case 1:
-               if (attrib == nv50->vertprog->cfg.edgeflag_in) {
-                       so_method(so, tesla, NV50TCL_EDGEFLAG_ENABLE, 1);
-                       so_data  (so, v[0] ? 1 : 0);
-               }
-               so_method(so, tesla, NV50TCL_VTX_ATTR_1F(attrib), 1);
-               so_data  (so, fui(v[0]));
-               break;
-       default:
-               nouveau_bo_unmap(bo);
-               return FALSE;
-       }
-
-       nouveau_bo_unmap(bo);
-       return TRUE;
+      BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U32), nr);
+      OUT_RINGp    (chan, map, nr);
+
+      map += nr;
+      count -= nr;
+   }
 }
 
-void
-nv50_vtxelt_construct(struct nv50_vtxelt_stateobj *cso)
+static void
+nv50_draw_elements_inline_u32_short(struct nouveau_channel *chan, uint32_t *map,
+                                    unsigned start, unsigned count)
 {
-       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);
-       }
+   map += start;
+
+   if (count & 1) {
+      count--;
+      BEGIN_RING(chan, RING_3D(VB_ELEMENT_U32), 1);
+      OUT_RING  (chan, *map++);
+   }
+   while (count) {
+      unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 2) / 2;
+
+      BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U16), nr);
+      for (i = 0; i < nr; ++i) {
+         OUT_RING(chan, (map[1] << 16) | map[0]);
+         map += 2;
+      }
+      count -= nr * 2;
+   }
 }
 
-struct nouveau_stateobj *
-nv50_vbo_validate(struct nv50_context *nv50)
+static void
+nv50_draw_elements(struct nv50_context *nv50, boolean shorten,
+                   unsigned mode, unsigned start, unsigned count,
+                   unsigned instance_count, int32_t index_bias)
 {
-       struct nouveau_grobj *tesla = nv50->screen->tesla;
-       struct nouveau_stateobj *vtxbuf, *vtxfmt, *vtxattr;
-       unsigned i, n_ve;
-
-       /* don't validate if Gallium took away our buffers */
-       if (nv50->vtxbuf_nr == 0)
-               return NULL;
-
-       nv50->vbo_fifo = 0;
-       if (nv50->screen->force_push ||
-           nv50->vertprog->cfg.edgeflag_in < 16)
-               nv50->vbo_fifo = 0xffff;
-
-       for (i = 0; i < nv50->vtxbuf_nr; i++) {
-               if (nv50->vtxbuf[i].stride &&
-                   !(nv50->vtxbuf[i].buffer->bind & PIPE_BIND_VERTEX_BUFFER))
-                       nv50->vbo_fifo = 0xffff;
-       }
-
-       n_ve = MAX2(nv50->vtxelt->num_elements, nv50->state.vtxelt_nr);
-
-       vtxattr = NULL;
-       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->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 = nv50_resource(vb->buffer)->bo;
-               uint32_t hw = nv50->vtxelt->hw[i];
-
-               if (!vb->stride &&
-                   nv50_vbo_static_attrib(nv50, i, &vtxattr, ve, vb)) {
-                       so_data(vtxfmt, hw | (1 << 4));
-
-                       so_method(vtxbuf, tesla,
-                                 NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
-                       so_data  (vtxbuf, 0);
-
-                       nv50->vbo_fifo &= ~(1 << i);
-                       continue;
-               }
-
-               if (nv50->vbo_fifo) {
-                       so_data  (vtxfmt, hw | (ve->instance_divisor ? (1 << 4) : i));
-                       so_method(vtxbuf, tesla,
-                                 NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
-                       so_data  (vtxbuf, 0);
-                       continue;
-               }
-
-               so_data(vtxfmt, hw | i);
-
-               so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_FORMAT(i), 3);
-               so_data  (vtxbuf, 0x20000000 |
-                         (ve->instance_divisor ? 0 : vb->stride));
-               so_reloc (vtxbuf, bo, vb->buffer_offset +
-                         ve->src_offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
-                         NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
-               so_reloc (vtxbuf, bo, vb->buffer_offset +
-                         ve->src_offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
-                         NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
-
-               /* vertex array limits */
-               so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_LIMIT_HIGH(i), 2);
-               so_reloc (vtxbuf, bo, vb->buffer->width0 - 1,
-                         NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD |
-                         NOUVEAU_BO_HIGH, 0, 0);
-               so_reloc (vtxbuf, bo, vb->buffer->width0 - 1,
-                         NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD |
-                         NOUVEAU_BO_LOW, 0, 0);
-       }
-       for (; i < n_ve; ++i) {
-               so_data  (vtxfmt, 0x7e080010);
-
-               so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
-               so_data  (vtxbuf, 0);
-       }
-       nv50->state.vtxelt_nr = nv50->vtxelt->num_elements;
-
-       so_ref (vtxbuf, &nv50->state.vtxbuf);
-       so_ref (vtxattr, &nv50->state.vtxattr);
-       so_ref (NULL, &vtxbuf);
-       so_ref (NULL, &vtxattr);
-       return vtxfmt;
+   struct nouveau_channel *chan = nv50->screen->base.channel;
+   void *data;
+   unsigned prim;
+   const unsigned index_size = nv50->idxbuf.index_size;
+
+   chan->flush_notify = nv50_draw_vbo_flush_notify;
+   chan->user_private = nv50;
+
+   prim = nv50_prim_gl(mode);
+
+   if (index_bias != nv50->state.index_bias) {
+      BEGIN_RING(chan, RING_3D(VB_ELEMENT_BASE), 1);
+      OUT_RING  (chan, index_bias);
+      nv50->state.index_bias = index_bias;
+   }
+
+   if (nouveau_resource_mapped_by_gpu(nv50->idxbuf.buffer)) {
+      struct nv04_resource *res = nv04_resource(nv50->idxbuf.buffer);
+
+      start += nv50->idxbuf.offset >> (index_size >> 1);
+
+      nouveau_buffer_adjust_score(&nv50->base, res, 1);
+
+      while (instance_count--) {
+         BEGIN_RING(chan, RING_3D(VERTEX_BEGIN_GL), 1);
+         OUT_RING  (chan, mode);
+
+         switch (index_size) {
+         case 4:
+         {
+            WAIT_RING (chan, 2);
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U32) | 0x30000, 0);
+            OUT_RING  (chan, count);
+            nouveau_pushbuf_submit(chan, res->bo, res->offset + start * 4,
+                                   count * 4);
+         }
+            break;
+         case 2:
+         {
+            unsigned pb_start = (start & ~1);
+            unsigned pb_words = (((start + count + 1) & ~1) - pb_start) >> 1;
+
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U16_SETUP), 1);
+            OUT_RING  (chan, (start << 31) | count);
+            WAIT_RING (chan, 2);
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U16) | 0x30000, 0);
+            OUT_RING  (chan, pb_words);
+            nouveau_pushbuf_submit(chan, res->bo, res->offset + pb_start * 2,
+                                   pb_words * 4);
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U16_SETUP), 1);
+            OUT_RING  (chan, 0);
+            break;
+         }
+         case 1:
+         {
+            unsigned pb_start = (start & ~3);
+            unsigned pb_words = (((start + count + 3) & ~3) - pb_start) >> 1;
+
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U8_SETUP), 1);
+            OUT_RING  (chan, (start << 30) | count);
+            WAIT_RING (chan, 2);
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U8) | 0x30000, 0);
+            OUT_RING  (chan, pb_words);
+            nouveau_pushbuf_submit(chan, res->bo, res->offset + pb_start,
+                                   pb_words * 4);
+            BEGIN_RING(chan, RING_3D(VB_ELEMENT_U8_SETUP), 1);
+            OUT_RING  (chan, 0);
+            break;
+         }
+         default:
+            assert(0);
+            return;
+         }
+         BEGIN_RING(chan, RING_3D(VERTEX_END_GL), 1);
+         OUT_RING  (chan, 0);
+
+         nv50_resource_fence(res, NOUVEAU_BO_RD);
+
+         mode |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
+      }
+   } else {
+      data = nouveau_resource_map_offset(&nv50->base,
+                                         nv04_resource(nv50->idxbuf.buffer),
+                                         nv50->idxbuf.offset, NOUVEAU_BO_RD);
+      if (!data)
+         return;
+
+      while (instance_count--) {
+         BEGIN_RING(chan, RING_3D(VERTEX_BEGIN_GL), 1);
+         OUT_RING  (chan, prim);
+         switch (index_size) {
+         case 1:
+            nv50_draw_elements_inline_u08(chan, data, start, count);
+            break;
+         case 2:
+            nv50_draw_elements_inline_u16(chan, data, start, count);
+            break;
+         case 4:
+            if (shorten)
+               nv50_draw_elements_inline_u32_short(chan, data, start, count);
+            else
+               nv50_draw_elements_inline_u32(chan, data, start, count);
+            break;
+         default:
+            assert(0);
+            return;
+         }
+         BEGIN_RING(chan, RING_3D(VERTEX_END_GL), 1);
+         OUT_RING  (chan, 0);
+
+         prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
+      }
+   }
+
+   chan->flush_notify = nv50_default_flush_notify;
 }
 
-
+void
+nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
+{
+   struct nv50_context *nv50 = nv50_context(pipe);
+   struct nouveau_channel *chan = nv50->screen->base.channel;
+
+   /* 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.
+    */
+   nv50->vbo_push_hint = /* the 64 is heuristic */
+      !(info->indexed &&
+        ((info->max_index - info->min_index + 64) < info->count));
+
+   nv50->vbo_min_index = info->min_index;
+   nv50->vbo_max_index = info->max_index;
+
+   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);
+
+   nv50_state_validate(nv50);
+
+   if (nv50->vbo_fifo) {
+      nv50_push_vbo(nv50, info);
+      return;
+   }
+
+   if (nv50->state.instance_base != info->start_instance) {
+      nv50->state.instance_base = info->start_instance;
+      /* NOTE: this does not affect the shader input, should it ? */
+      BEGIN_RING(chan, RING_3D(VB_INSTANCE_BASE), 1);
+      OUT_RING  (chan, info->start_instance);
+   }
+
+   if (nv50->base.vbo_dirty) {
+      BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FLUSH), 1);
+      OUT_RING  (chan, 0);
+      nv50->base.vbo_dirty = FALSE;
+   }
+
+   if (!info->indexed) {
+      nv50_draw_arrays(nv50,
+                       info->mode, info->start, info->count,
+                       info->instance_count);
+   } else {
+      boolean shorten = info->max_index <= 65535;
+
+      assert(nv50->idxbuf.buffer);
+
+      if (info->primitive_restart != nv50->state.prim_restart) {
+         if (info->primitive_restart) {
+            BEGIN_RING(chan, RING_3D(PRIM_RESTART_ENABLE), 2);
+            OUT_RING  (chan, 1);
+            OUT_RING  (chan, info->restart_index);
+
+            if (info->restart_index > 65535)
+               shorten = FALSE;
+         } else {
+            BEGIN_RING(chan, RING_3D(PRIM_RESTART_ENABLE), 1);
+            OUT_RING  (chan, 0);
+         }
+         nv50->state.prim_restart = info->primitive_restart;
+      } else
+      if (info->primitive_restart) {
+         BEGIN_RING(chan, RING_3D(PRIM_RESTART_INDEX), 1);
+         OUT_RING  (chan, info->restart_index);
+
+         if (info->restart_index > 65535)
+            shorten = FALSE;
+      }
+
+      nv50_draw_elements(nv50, shorten,
+                         info->mode, info->start, info->count,
+                         info->instance_count, info->index_bias);
+   }
+
+   nv50_release_user_vbufs(nv50);
+}