src/gallium/drivers/nv50/nv50_vbo.c

   1 /*
   2  * Copyright 2010 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #include "pipe/p_context.h"
  24 #include "pipe/p_state.h"
  25 #include "util/u_inlines.h"
  26 #include "util/u_format.h"
  27 #include "translate/translate.h"
  28
  29 #include "nv50_context.h"
  30 #include "nv50_resource.h"
  31
  32 #include "nv50_3d.xml.h"
  33
  34 void
  35 nv50_vertex_state_delete(struct pipe_context *pipe,
  36                          void *hwcso)
  37 {
  38    struct nv50_vertex_stateobj *so = hwcso;
  39
  40    if (so->translate)
  41       so->translate->release(so->translate);
  42    FREE(hwcso);
  43 }
  44
  45 void *
  46 nv50_vertex_state_create(struct pipe_context *pipe,
  47                          unsigned num_elements,
  48                          const struct pipe_vertex_element *elements)
  49 {
  50     struct nv50_vertex_stateobj *so;
  51     struct translate_key transkey;
  52     unsigned i;
  53
  54     assert(num_elements);
  55
  56     so = MALLOC(sizeof(*so) +
  57                 (num_elements - 1) * sizeof(struct nv50_vertex_element));
  58     if (!so)
  59         return NULL;
  60     so->num_elements = num_elements;
  61     so->instance_elts = 0;
  62     so->instance_bufs = 0;
  63     so->need_conversion = FALSE;
  64
  65     transkey.nr_elements = 0;
  66     transkey.output_stride = 0;
  67
  68     for (i = 0; i < num_elements; ++i) {
  69         const struct pipe_vertex_element *ve = &elements[i];
  70         const unsigned vbi = ve->vertex_buffer_index;
  71         enum pipe_format fmt = ve->src_format;
  72
  73         so->element[i].pipe = elements[i];
  74         so->element[i].state = nv50_format_table[fmt].vtx;
  75
  76         if (!so->element[i].state) {
  77             switch (util_format_get_nr_components(fmt)) {
  78             case 1: fmt = PIPE_FORMAT_R32_FLOAT; break;
  79             case 2: fmt = PIPE_FORMAT_R32G32_FLOAT; break;
  80             case 3: fmt = PIPE_FORMAT_R32G32B32_FLOAT; break;
  81             case 4: fmt = PIPE_FORMAT_R32G32B32A32_FLOAT; break;
  82             default:
  83                 assert(0);
  84                 return NULL;
  85             }
  86             so->element[i].state = nv50_format_table[fmt].vtx;
  87             so->need_conversion = TRUE;
  88         }
  89         so->element[i].state |= i;
  90
  91         if (1) {
  92             unsigned j = transkey.nr_elements++;
  93
  94             transkey.element[j].type = TRANSLATE_ELEMENT_NORMAL;
  95             transkey.element[j].input_format = ve->src_format;
  96             transkey.element[j].input_buffer = vbi;
  97             transkey.element[j].input_offset = ve->src_offset;
  98             transkey.element[j].instance_divisor = ve->instance_divisor;
  99
 100             transkey.element[j].output_format = fmt;
 101             transkey.element[j].output_offset = transkey.output_stride;
 102             transkey.output_stride += (util_format_get_stride(fmt, 1) + 3) & ~3;
 103
 104             if (unlikely(ve->instance_divisor)) {
 105                so->instance_elts |= 1 << i;
 106                so->instance_bufs |= 1 << vbi;
 107             }
 108         }
 109     }
 110
 111     so->translate = translate_create(&transkey);
 112     so->vertex_size = transkey.output_stride / 4;
 113     so->packet_vertex_limit = NV04_PFIFO_MAX_PACKET_LEN /
 114        MAX2(so->vertex_size, 1);
 115
 116     return so;
 117 }
 118
 119 #define NV50_3D_VERTEX_ATTRIB_INACTIVE              \
 120    NV50_3D_VERTEX_ARRAY_ATTRIB_TYPE_FLOAT |         \
 121    NV50_3D_VERTEX_ARRAY_ATTRIB_FORMAT_32_32_32_32 | \
 122    NV50_3D_VERTEX_ARRAY_ATTRIB_CONST
 123
 124 static void
 125 nv50_emit_vtxattr(struct nv50_context *nv50, struct pipe_vertex_buffer *vb,
 126                   struct pipe_vertex_element *ve, unsigned attr)
 127 {
 128    const void *data;
 129    struct nouveau_channel *chan = nv50->screen->base.channel;
 130    struct nv04_resource *res = nv04_resource(vb->buffer);
 131    float v[4];
 132    const unsigned nc = util_format_get_nr_components(ve->src_format);
 133
 134    data = nouveau_resource_map_offset(&nv50->pipe, res, vb->buffer_offset +
 135                                       ve->src_offset, NOUVEAU_BO_RD);
 136
 137    util_format_read_4f(ve->src_format, v, 0, data, 0, 0, 0, 1, 1);
 138
 139    switch (nc) {
 140    case 4:
 141       BEGIN_RING(chan, RING_3D(VTX_ATTR_4F_X(attr)), 4);
 142       OUT_RINGf (chan, v[0]);
 143       OUT_RINGf (chan, v[1]);
 144       OUT_RINGf (chan, v[2]);
 145       OUT_RINGf (chan, v[3]);
 146       break;
 147    case 3:
 148       BEGIN_RING(chan, RING_3D(VTX_ATTR_3F_X(attr)), 3);
 149       OUT_RINGf (chan, v[0]);
 150       OUT_RINGf (chan, v[1]);
 151       OUT_RINGf (chan, v[2]);
 152       break;
 153    case 2:
 154       BEGIN_RING(chan, RING_3D(VTX_ATTR_2F_X(attr)), 2);
 155       OUT_RINGf (chan, v[0]);
 156       OUT_RINGf (chan, v[1]);
 157       break;
 158    case 1:
 159       if (attr == nv50->vertprog->vp.edgeflag) {
 160          BEGIN_RING(chan, RING_3D(EDGEFLAG_ENABLE), 1);
 161          OUT_RING  (chan, v[0] ? 1 : 0);
 162       }
 163       BEGIN_RING(chan, RING_3D(VTX_ATTR_1F(attr)), 1);
 164       OUT_RINGf (chan, v[0]);
 165       break;
 166    default:
 167       assert(0);
 168       break;
 169    }
 170 }
 171
 172 static INLINE void
 173 nv50_vbuf_range(struct nv50_context *nv50, int vbi,
 174                 uint32_t *base, uint32_t *size)
 175 {
 176    if (unlikely(nv50->vertex->instance_bufs & (1 << vbi))) {
 177       /* TODO: use min and max instance divisor to get a proper range */
 178       *base = 0;
 179       *size = nv50->vtxbuf[vbi].buffer->width0;
 180    } else {
 181       assert(nv50->vbo_max_index != ~0);
 182       *base = nv50->vbo_min_index * nv50->vtxbuf[vbi].stride;
 183       *size = (nv50->vbo_max_index -
 184                nv50->vbo_min_index + 1) * nv50->vtxbuf[vbi].stride;
 185    }
 186 }
 187
 188 static void
 189 nv50_prevalidate_vbufs(struct nv50_context *nv50)
 190 {
 191    struct pipe_vertex_buffer *vb;
 192    struct nv04_resource *buf;
 193    int i;
 194    uint32_t base, size;
 195
 196    nv50->vbo_fifo = nv50->vbo_user = 0;
 197
 198    nv50_bufctx_reset(nv50, NV50_BUFCTX_VERTEX);
 199
 200    for (i = 0; i < nv50->num_vtxbufs; ++i) {
 201       vb = &nv50->vtxbuf[i];
 202       if (!vb->stride)
 203          continue;
 204       buf = nv04_resource(vb->buffer);
 205
 206       /* NOTE: user buffers with temporary storage count as mapped by GPU */
 207       if (!nouveau_resource_mapped_by_gpu(vb->buffer)) {
 208          if (nv50->vbo_push_hint) {
 209             nv50->vbo_fifo = ~0;
 210             continue;
 211          } else {
 212             if (buf->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY) {
 213                nv50->vbo_user |= 1 << i;
 214                assert(vb->stride > vb->buffer_offset);
 215                nv50_vbuf_range(nv50, i, &base, &size);
 216                nouveau_user_buffer_upload(buf, base, size);
 217             } else {
 218                nouveau_buffer_migrate(&nv50->pipe, buf, NOUVEAU_BO_GART);
 219             }
 220             nv50->vbo_dirty = TRUE;
 221          }
 222       }
 223       nv50_bufctx_add_resident(nv50, NV50_BUFCTX_VERTEX, buf, NOUVEAU_BO_RD);
 224       nouveau_buffer_adjust_score(&nv50->pipe, buf, 1);
 225    }
 226 }
 227
 228 static void
 229 nv50_update_user_vbufs(struct nv50_context *nv50)
 230 {
 231    struct nouveau_channel *chan = nv50->screen->base.channel;
 232    uint32_t base, offset, size;
 233    int i;
 234    uint32_t written = 0;
 235
 236    for (i = 0; i < nv50->vertex->num_elements; ++i) {
 237       struct pipe_vertex_element *ve = &nv50->vertex->element[i].pipe;
 238       const int b = ve->vertex_buffer_index;
 239       struct pipe_vertex_buffer *vb = &nv50->vtxbuf[b];
 240       struct nv04_resource *buf = nv04_resource(vb->buffer);
 241
 242       if (!(nv50->vbo_user & (1 << b)))
 243          continue;
 244
 245       if (!vb->stride) {
 246          nv50_emit_vtxattr(nv50, vb, ve, i);
 247          continue;
 248       }
 249       nv50_vbuf_range(nv50, b, &base, &size);
 250
 251       if (!(written & (1 << b))) {
 252          written |= 1 << b;
 253          nouveau_user_buffer_upload(buf, base, size);
 254       }
 255       offset = vb->buffer_offset + ve->src_offset;
 256
 257       MARK_RING (chan, 6, 4);
 258       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
 259       OUT_RESRCh(chan, buf, base + size - 1, NOUVEAU_BO_RD);
 260       OUT_RESRCl(chan, buf, base + size - 1, NOUVEAU_BO_RD);
 261       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
 262       OUT_RESRCh(chan, buf, offset, NOUVEAU_BO_RD);
 263       OUT_RESRCl(chan, buf, offset, NOUVEAU_BO_RD);
 264    }
 265    nv50->vbo_dirty = TRUE;
 266 }
 267
 268 static INLINE void
 269 nv50_release_user_vbufs(struct nv50_context *nv50)
 270 {
 271    uint32_t vbo_user = nv50->vbo_user;
 272
 273    while (vbo_user) {
 274       int i = ffs(vbo_user) - 1;
 275       vbo_user &= ~(1 << i);
 276
 277       nouveau_buffer_release_gpu_storage(nv04_resource(nv50->vtxbuf[i].buffer));
 278    }
 279 }
 280
 281 void
 282 nv50_vertex_arrays_validate(struct nv50_context *nv50)
 283 {
 284    struct nouveau_channel *chan = nv50->screen->base.channel;
 285    struct nv50_vertex_stateobj *vertex = nv50->vertex;
 286    struct pipe_vertex_buffer *vb;
 287    struct nv50_vertex_element *ve;
 288    unsigned i;
 289
 290    if (unlikely(vertex->need_conversion)) {
 291       nv50->vbo_fifo = ~0;
 292       nv50->vbo_user = 0;
 293    } else {
 294       nv50_prevalidate_vbufs(nv50);
 295    }
 296
 297    BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_ATTRIB(0)), vertex->num_elements);
 298    for (i = 0; i < vertex->num_elements; ++i) {
 299       ve = &vertex->element[i];
 300       vb = &nv50->vtxbuf[ve->pipe.vertex_buffer_index];
 301
 302       if (likely(vb->stride) || nv50->vbo_fifo) {
 303          OUT_RING(chan, ve->state);
 304       } else {
 305          OUT_RING(chan, ve->state | NV50_3D_VERTEX_ARRAY_ATTRIB_CONST);
 306          nv50->vbo_fifo &= ~(1 << i);
 307       }
 308    }
 309
 310    for (i = 0; i < vertex->num_elements; ++i) {
 311       struct nv04_resource *res;
 312       unsigned size, offset;
 313
 314       ve = &vertex->element[i];
 315       vb = &nv50->vtxbuf[ve->pipe.vertex_buffer_index];
 316
 317       if (unlikely(ve->pipe.instance_divisor)) {
 318          if (!(nv50->state.instance_elts & (1 << i))) {
 319             BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
 320             OUT_RING  (chan, 1);
 321          }
 322          BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_DIVISOR(i)), 1);
 323          OUT_RING  (chan, ve->pipe.instance_divisor);
 324       } else
 325       if (unlikely(nv50->state.instance_elts & (1 << i))) {
 326          BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
 327          OUT_RING  (chan, 0);
 328       }
 329
 330       res = nv04_resource(vb->buffer);
 331
 332       if (nv50->vbo_fifo || unlikely(vb->stride == 0)) {
 333          if (!nv50->vbo_fifo)
 334             nv50_emit_vtxattr(nv50, vb, &ve->pipe, i);
 335          BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FETCH(i)), 1);
 336          OUT_RING  (chan, 0);
 337          continue;
 338       }
 339
 340       size = vb->buffer->width0;
 341       offset = ve->pipe.src_offset + vb->buffer_offset;
 342
 343       MARK_RING (chan, 8, 4);
 344       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FETCH(i)), 1);
 345       OUT_RING  (chan, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
 346       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
 347       OUT_RESRCh(chan, res, size - 1, NOUVEAU_BO_RD);
 348       OUT_RESRCl(chan, res, size - 1, NOUVEAU_BO_RD);
 349       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
 350       OUT_RESRCh(chan, res, offset, NOUVEAU_BO_RD);
 351       OUT_RESRCl(chan, res, offset, NOUVEAU_BO_RD);
 352    }
 353    for (; i < nv50->state.num_vtxelts; ++i) {
 354       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_ATTRIB(i)), 1);
 355       OUT_RING  (chan, NV50_3D_VERTEX_ATTRIB_INACTIVE);
 356       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FETCH(i)), 1);
 357       OUT_RING  (chan, 0);
 358    }
 359
 360    nv50->state.num_vtxelts = vertex->num_elements;
 361    nv50->state.instance_elts = vertex->instance_elts;
 362 }
 363
 364 #define NV50_PRIM_GL_CASE(n) \
 365    case PIPE_PRIM_##n: return NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_##n
 366
 367 static INLINE unsigned
 368 nv50_prim_gl(unsigned prim)
 369 {
 370    switch (prim) {
 371    NV50_PRIM_GL_CASE(POINTS);
 372    NV50_PRIM_GL_CASE(LINES);
 373    NV50_PRIM_GL_CASE(LINE_LOOP);
 374    NV50_PRIM_GL_CASE(LINE_STRIP);
 375    NV50_PRIM_GL_CASE(TRIANGLES);
 376    NV50_PRIM_GL_CASE(TRIANGLE_STRIP);
 377    NV50_PRIM_GL_CASE(TRIANGLE_FAN);
 378    NV50_PRIM_GL_CASE(QUADS);
 379    NV50_PRIM_GL_CASE(QUAD_STRIP);
 380    NV50_PRIM_GL_CASE(POLYGON);
 381    NV50_PRIM_GL_CASE(LINES_ADJACENCY);
 382    NV50_PRIM_GL_CASE(LINE_STRIP_ADJACENCY);
 383    NV50_PRIM_GL_CASE(TRIANGLES_ADJACENCY);
 384    NV50_PRIM_GL_CASE(TRIANGLE_STRIP_ADJACENCY);
 385    default:
 386       return NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_POINTS;
 387       break;
 388    }
 389 }
 390
 391 static void
 392 nv50_draw_vbo_flush_notify(struct nouveau_channel *chan)
 393 {
 394    struct nv50_context *nv50 = chan->user_private;
 395
 396    nouveau_fence_update(&nv50->screen->base, TRUE);
 397
 398    nv50_bufctx_emit_relocs(nv50);
 399 }
 400
 401 static void
 402 nv50_draw_arrays(struct nv50_context *nv50,
 403                  unsigned mode, unsigned start, unsigned count,
 404                  unsigned instance_count)
 405 {
 406    struct nouveau_channel *chan = nv50->screen->base.channel;
 407    unsigned prim;
 408
 409    chan->flush_notify = nv50_draw_vbo_flush_notify;
 410    chan->user_private = nv50;
 411
 412    prim = nv50_prim_gl(mode);
 413
 414    while (instance_count--) {
 415       BEGIN_RING(chan, RING_3D(VERTEX_BEGIN_GL), 1);
 416       OUT_RING  (chan, prim);
 417       BEGIN_RING(chan, RING_3D(VERTEX_BUFFER_FIRST), 2);
 418       OUT_RING  (chan, start);
 419       OUT_RING  (chan, count);
 420       BEGIN_RING(chan, RING_3D(VERTEX_END_GL), 1);
 421       OUT_RING  (chan, 0);
 422
 423       prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
 424    }
 425
 426    chan->flush_notify = nv50_default_flush_notify;
 427 }
 428
 429 static void
 430 nv50_draw_elements_inline_u08(struct nouveau_channel *chan, uint8_t *map,
 431                               unsigned start, unsigned count)
 432 {
 433    map += start;
 434
 435    if (count & 3) {
 436       unsigned i;
 437       BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U32), count & 3);
 438       for (i = 0; i < (count & 3); ++i)
 439          OUT_RING(chan, *map++);
 440       count &= ~3;
 441    }
 442    while (count) {
 443       unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 4) / 4;
 444
 445       BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U8), nr);
 446       for (i = 0; i < nr; ++i) {
 447          OUT_RING(chan,
 448                   (map[3] << 24) | (map[2] << 16) | (map[1] << 8) | map[0]);
 449          map += 4;
 450       }
 451       count -= nr * 4;
 452    }
 453 }
 454
 455 static void
 456 nv50_draw_elements_inline_u16(struct nouveau_channel *chan, uint16_t *map,
 457                               unsigned start, unsigned count)
 458 {
 459    map += start;
 460
 461    if (count & 1) {
 462       count &= ~1;
 463       BEGIN_RING(chan, RING_3D(VB_ELEMENT_U32), 1);
 464       OUT_RING  (chan, *map++);
 465    }
 466    while (count) {
 467       unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 2) / 2;
 468
 469       BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U16), nr);
 470       for (i = 0; i < nr; ++i) {
 471          OUT_RING(chan, (map[1] << 16) | map[0]);
 472          map += 2;
 473       }
 474       count -= nr * 2;
 475    }
 476 }
 477
 478 static void
 479 nv50_draw_elements_inline_u32(struct nouveau_channel *chan, uint32_t *map,
 480                               unsigned start, unsigned count)
 481 {
 482    map += start;
 483
 484    while (count) {
 485       const unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
 486
 487       BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U32), nr);
 488       OUT_RINGp    (chan, map, nr);
 489
 490       map += nr;
 491       count -= nr;
 492    }
 493 }
 494
 495 static void
 496 nv50_draw_elements_inline_u32_short(struct nouveau_channel *chan, uint32_t *map,
 497                                     unsigned start, unsigned count)
 498 {
 499    map += start;
 500
 501    if (count & 1) {
 502       count--;
 503       BEGIN_RING(chan, RING_3D(VB_ELEMENT_U32), 1);
 504       OUT_RING  (chan, *map++);
 505    }
 506    while (count) {
 507       unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 2) / 2;
 508
 509       BEGIN_RING_NI(chan, RING_3D(VB_ELEMENT_U16), nr);
 510       for (i = 0; i < nr; ++i) {
 511          OUT_RING(chan, (map[1] << 16) | map[0]);
 512          map += 2;
 513       }
 514       count -= nr * 2;
 515    }
 516 }
 517
 518 static void
 519 nv50_draw_elements(struct nv50_context *nv50, boolean shorten,
 520                    unsigned mode, unsigned start, unsigned count,
 521                    unsigned instance_count, int32_t index_bias)
 522 {
 523    struct nouveau_channel *chan = nv50->screen->base.channel;
 524    void *data;
 525    unsigned prim;
 526    const unsigned index_size = nv50->idxbuf.index_size;
 527
 528    chan->flush_notify = nv50_draw_vbo_flush_notify;
 529    chan->user_private = nv50;
 530
 531    prim = nv50_prim_gl(mode);
 532
 533    if (index_bias != nv50->state.index_bias) {
 534       BEGIN_RING(chan, RING_3D(VB_ELEMENT_BASE), 1);
 535       OUT_RING  (chan, index_bias);
 536       nv50->state.index_bias = index_bias;
 537    }
 538
 539    if (nouveau_resource_mapped_by_gpu(nv50->idxbuf.buffer) && 0) {
 540       struct nv04_resource *res = nv04_resource(nv50->idxbuf.buffer);
 541       unsigned offset = res->offset + nv50->idxbuf.offset;
 542
 543       nouveau_buffer_adjust_score(&nv50->pipe, res, 1);
 544
 545       while (instance_count--) {
 546          BEGIN_RING(chan, RING_3D(VERTEX_BEGIN_GL), 1);
 547          OUT_RING  (chan, mode);
 548
 549          switch (index_size) {
 550          case 4:
 551          {
 552             WAIT_RING (chan, 2);
 553             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U32) | 0x30000, 0);
 554             OUT_RING  (chan, count);
 555             nouveau_pushbuf_submit(chan, res->bo,
 556                                    (start << 2) + offset,
 557                                    (count << 2));
 558          }
 559             break;
 560          case 2:
 561          {
 562             unsigned pb_start = (start & ~1);
 563             unsigned pb_words = (((start + count + 1) & ~1) - pb_start) >> 1;
 564
 565             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U16_SETUP), 1);
 566             OUT_RING  (chan, (start << 31) | count);
 567             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U16) | 0x30000, 0);
 568             OUT_RING  (chan, pb_words);
 569             nouveau_pushbuf_submit(chan, res->bo,
 570                                    (pb_start << 1) + offset, pb_words << 2);
 571             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U16_SETUP), 1);
 572             OUT_RING  (chan, 0);
 573             break;
 574          }
 575          case 1:
 576          {
 577             unsigned pb_start = (start & ~3);
 578             unsigned pb_words = (((start + count + 3) & ~3) - pb_start) >> 1;
 579
 580             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U8_SETUP), 1);
 581             OUT_RING  (chan, (start << 30) | count);
 582             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U8) | 0x30000, 0);
 583             OUT_RING  (chan, pb_words);
 584             nouveau_pushbuf_submit(chan, res->bo,
 585                                    pb_start + offset, pb_words << 2);
 586             BEGIN_RING(chan, RING_3D(VB_ELEMENT_U8_SETUP), 1);
 587             OUT_RING  (chan, 0);
 588             break;
 589          }
 590          default:
 591             assert(0);
 592             return;
 593          }
 594
 595          nv50_resource_fence(res, NOUVEAU_BO_RD);
 596
 597          mode |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
 598       }
 599    } else {
 600       data = nouveau_resource_map_offset(&nv50->pipe,
 601                                          nv04_resource(nv50->idxbuf.buffer),
 602                                          nv50->idxbuf.offset, NOUVEAU_BO_RD);
 603       if (!data)
 604          return;
 605
 606       while (instance_count--) {
 607          BEGIN_RING(chan, RING_3D(VERTEX_BEGIN_GL), 1);
 608          OUT_RING  (chan, prim);
 609          switch (index_size) {
 610          case 1:
 611             nv50_draw_elements_inline_u08(chan, data, start, count);
 612             break;
 613          case 2:
 614             nv50_draw_elements_inline_u16(chan, data, start, count);
 615             break;
 616          case 4:
 617             if (shorten)
 618                nv50_draw_elements_inline_u32_short(chan, data, start, count);
 619             else
 620                nv50_draw_elements_inline_u32(chan, data, start, count);
 621             break;
 622          default:
 623             assert(0);
 624             return;
 625          }
 626          BEGIN_RING(chan, RING_3D(VERTEX_END_GL), 1);
 627          OUT_RING  (chan, 0);
 628
 629          prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
 630       }
 631    }
 632
 633    chan->flush_notify = nv50_default_flush_notify;
 634 }
 635
 636 void
 637 nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
 638 {
 639    struct nv50_context *nv50 = nv50_context(pipe);
 640    struct nouveau_channel *chan = nv50->screen->base.channel;
 641
 642    /* For picking only a few vertices from a large user buffer, push is better,
 643     * if index count is larger and we expect repeated vertices, suggest upload.
 644     */
 645    nv50->vbo_push_hint = /* the 64 is heuristic */
 646       !(info->indexed &&
 647         ((info->max_index - info->min_index + 64) < info->count));
 648
 649    nv50->vbo_min_index = info->min_index;
 650    nv50->vbo_max_index = info->max_index;
 651
 652    if (nv50->vbo_push_hint != !!nv50->vbo_fifo)
 653       nv50->dirty |= NV50_NEW_ARRAYS;
 654
 655    if (nv50->vbo_user && !(nv50->dirty & (NV50_NEW_VERTEX | NV50_NEW_ARRAYS)))
 656       nv50_update_user_vbufs(nv50);
 657
 658    nv50_state_validate(nv50);
 659
 660    if (nv50->vbo_fifo) {
 661       nv50_push_vbo(nv50, info);
 662       return;
 663    }
 664
 665    if (nv50->state.instance_base != info->start_instance) {
 666       nv50->state.instance_base = info->start_instance;
 667       /* NOTE: this does not affect the shader input, should it ? */
 668       BEGIN_RING(chan, RING_3D(VB_INSTANCE_BASE), 1);
 669       OUT_RING  (chan, info->start_instance);
 670    }
 671
 672    if (nv50->vbo_dirty) {
 673       BEGIN_RING(chan, RING_3D(VERTEX_ARRAY_FLUSH), 1);
 674       OUT_RING  (chan, 0);
 675       nv50->vbo_dirty = FALSE;
 676    }
 677
 678    if (!info->indexed) {
 679       nv50_draw_arrays(nv50,
 680                        info->mode, info->start, info->count,
 681                        info->instance_count);
 682    } else {
 683       boolean shorten = info->max_index <= 65535;
 684
 685       assert(nv50->idxbuf.buffer);
 686
 687       if (info->primitive_restart != nv50->state.prim_restart) {
 688          if (info->primitive_restart) {
 689             BEGIN_RING(chan, RING_3D(PRIM_RESTART_ENABLE), 2);
 690             OUT_RING  (chan, 1);
 691             OUT_RING  (chan, info->restart_index);
 692
 693             if (info->restart_index > 65535)
 694                shorten = FALSE;
 695          } else {
 696             BEGIN_RING(chan, RING_3D(PRIM_RESTART_ENABLE), 1);
 697             OUT_RING  (chan, 0);
 698          }
 699          nv50->state.prim_restart = info->primitive_restart;
 700       } else
 701       if (info->primitive_restart) {
 702          BEGIN_RING(chan, RING_3D(PRIM_RESTART_INDEX), 1);
 703          OUT_RING  (chan, info->restart_index);
 704
 705          if (info->restart_index > 65535)
 706             shorten = FALSE;
 707       }
 708
 709       nv50_draw_elements(nv50, shorten,
 710                          info->mode, info->start, info->count,
 711                          info->instance_count, info->index_bias);
 712    }
 713
 714    nv50_release_user_vbufs(nv50);
 715 }