src/gallium/drivers/r300/r300_render.c

   1 /*
   2  * Copyright 2009 Corbin Simpson <MostAwesomeDude@gmail.com>
   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  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  22
  23 /* r300_render: Vertex and index buffer primitive emission. Contains both
  24  * HW TCL fastpath rendering, and SW TCL Draw-assisted rendering. */
  25
  26 #include "draw/draw_context.h"
  27 #include "draw/draw_vbuf.h"
  28
  29 #include "util/u_inlines.h"
  30
  31 #include "util/u_format.h"
  32 #include "util/u_memory.h"
  33 #include "util/u_upload_mgr.h"
  34 #include "util/u_prim.h"
  35
  36 #include "r300_cs.h"
  37 #include "r300_context.h"
  38 #include "r300_screen_buffer.h"
  39 #include "r300_emit.h"
  40 #include "r300_reg.h"
  41 #include "r300_render.h"
  42 #include "r300_state_derived.h"
  43
  44 /* XXX The DRM rejects VAP_ALT_NUM_VERTICES.. */
  45 //#define ENABLE_ALT_NUM_VERTS
  46
  47 uint32_t r300_translate_primitive(unsigned prim)
  48 {
  49     switch (prim) {
  50         case PIPE_PRIM_POINTS:
  51             return R300_VAP_VF_CNTL__PRIM_POINTS;
  52         case PIPE_PRIM_LINES:
  53             return R300_VAP_VF_CNTL__PRIM_LINES;
  54         case PIPE_PRIM_LINE_LOOP:
  55             return R300_VAP_VF_CNTL__PRIM_LINE_LOOP;
  56         case PIPE_PRIM_LINE_STRIP:
  57             return R300_VAP_VF_CNTL__PRIM_LINE_STRIP;
  58         case PIPE_PRIM_TRIANGLES:
  59             return R300_VAP_VF_CNTL__PRIM_TRIANGLES;
  60         case PIPE_PRIM_TRIANGLE_STRIP:
  61             return R300_VAP_VF_CNTL__PRIM_TRIANGLE_STRIP;
  62         case PIPE_PRIM_TRIANGLE_FAN:
  63             return R300_VAP_VF_CNTL__PRIM_TRIANGLE_FAN;
  64         case PIPE_PRIM_QUADS:
  65             return R300_VAP_VF_CNTL__PRIM_QUADS;
  66         case PIPE_PRIM_QUAD_STRIP:
  67             return R300_VAP_VF_CNTL__PRIM_QUAD_STRIP;
  68         case PIPE_PRIM_POLYGON:
  69             return R300_VAP_VF_CNTL__PRIM_POLYGON;
  70         default:
  71             return 0;
  72     }
  73 }
  74
  75 static uint32_t r300_provoking_vertex_fixes(struct r300_context *r300,
  76                                             unsigned mode)
  77 {
  78     struct r300_rs_state* rs = (struct r300_rs_state*)r300->rs_state.state;
  79     uint32_t color_control = rs->color_control;
  80
  81     /* By default (see r300_state.c:r300_create_rs_state) color_control is
  82      * initialized to provoking the first vertex.
  83      *
  84      * Triangle fans must be reduced to the second vertex, not the first, in
  85      * Gallium flatshade-first mode, as per the GL spec.
  86      * (http://www.opengl.org/registry/specs/ARB/provoking_vertex.txt)
  87      *
  88      * Quads never provoke correctly in flatshade-first mode. The first
  89      * vertex is never considered as provoking, so only the second, third,
  90      * and fourth vertices can be selected, and both "third" and "last" modes
  91      * select the fourth vertex. This is probably due to D3D lacking quads.
  92      *
  93      * Similarly, polygons reduce to the first, not the last, vertex, when in
  94      * "last" mode, and all other modes start from the second vertex.
  95      *
  96      * ~ C.
  97      */
  98
  99     if (rs->rs.flatshade_first) {
 100         switch (mode) {
 101             case PIPE_PRIM_TRIANGLE_FAN:
 102                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_SECOND;
 103                 break;
 104             case PIPE_PRIM_QUADS:
 105             case PIPE_PRIM_QUAD_STRIP:
 106             case PIPE_PRIM_POLYGON:
 107                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_LAST;
 108                 break;
 109             default:
 110                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_FIRST;
 111                 break;
 112         }
 113     } else {
 114         color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_LAST;
 115     }
 116
 117     return color_control;
 118 }
 119
 120 /* Check if the requested number of dwords is available in the CS and
 121  * if not, flush. Return TRUE if the flush occured. */
 122 static boolean r300_reserve_cs_space(struct r300_context *r300,
 123                                      unsigned dwords)
 124 {
 125     if (!r300->rws->check_cs(r300->rws, dwords)) {
 126         r300->context.flush(&r300->context, 0, NULL);
 127         return TRUE;
 128     }
 129     return FALSE;
 130 }
 131
 132 static boolean immd_is_good_idea(struct r300_context *r300,
 133                                  unsigned count)
 134 {
 135     struct pipe_vertex_element* velem;
 136     struct pipe_vertex_buffer* vbuf;
 137     boolean checked[PIPE_MAX_ATTRIBS] = {0};
 138     unsigned vertex_element_count = r300->velems->count;
 139     unsigned i, vbi;
 140
 141     if (count > 10) {
 142         return FALSE;
 143     }
 144
 145     /* We shouldn't map buffers referenced by CS, busy buffers,
 146      * and ones placed in VRAM. */
 147     /* XXX Check for VRAM buffers. */
 148     for (i = 0; i < vertex_element_count; i++) {
 149         velem = &r300->velems->velem[i];
 150         vbi = velem->vertex_buffer_index;
 151
 152         if (!checked[vbi]) {
 153             vbuf = &r300->vertex_buffer[vbi];
 154
 155             if (r300->context.is_resource_referenced(&r300->context,
 156                                                      vbuf->buffer,
 157                                                      0, 0)) {
 158                 /* It's a very bad idea to map it... */
 159                 return FALSE;
 160             }
 161             checked[vbi] = TRUE;
 162         }
 163     }
 164     return TRUE;
 165 }
 166
 167 /*****************************************************************************
 168  * The emission of draw packets for r500. Older GPUs may use these functions *
 169  * after resolving fallback issues (e.g. stencil ref two-sided).             *
 170  ****************************************************************************/
 171
 172 void r500_emit_draw_arrays_immediate(struct r300_context *r300,
 173                                      unsigned mode,
 174                                      unsigned start,
 175                                      unsigned count)
 176 {
 177     struct pipe_vertex_element* velem;
 178     struct pipe_vertex_buffer* vbuf;
 179     unsigned vertex_element_count = r300->velems->count;
 180     unsigned i, v, vbi, dw, elem_offset, dwords;
 181
 182     /* Size of the vertex, in dwords. */
 183     unsigned vertex_size = 0;
 184
 185     /* Offsets of the attribute, in dwords, from the start of the vertex. */
 186     unsigned offset[PIPE_MAX_ATTRIBS];
 187
 188     /* Size of the vertex element, in dwords. */
 189     unsigned size[PIPE_MAX_ATTRIBS];
 190
 191     /* Stride to the same attrib in the next vertex in the vertex buffer,
 192      * in dwords. */
 193     unsigned stride[PIPE_MAX_ATTRIBS] = {0};
 194
 195     /* Mapped vertex buffers. */
 196     uint32_t* map[PIPE_MAX_ATTRIBS] = {0};
 197     struct pipe_transfer* transfer[PIPE_MAX_ATTRIBS] = {NULL};
 198
 199     CS_LOCALS(r300);
 200
 201     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
 202     for (i = 0; i < vertex_element_count; i++) {
 203         velem = &r300->velems->velem[i];
 204         offset[i] = velem->src_offset / 4;
 205         size[i] = util_format_get_blocksize(velem->src_format) / 4;
 206         vertex_size += size[i];
 207         vbi = velem->vertex_buffer_index;
 208
 209         /* Map the buffer. */
 210         if (!map[vbi]) {
 211             vbuf = &r300->vertex_buffer[vbi];
 212             map[vbi] = (uint32_t*)pipe_buffer_map(&r300->context,
 213                                                   vbuf->buffer,
 214                                                   PIPE_TRANSFER_READ,
 215                                                   &transfer[vbi]);
 216             map[vbi] += vbuf->buffer_offset / 4;
 217             stride[vbi] = vbuf->stride / 4;
 218         }
 219     }
 220
 221     dwords = 9 + count * vertex_size;
 222
 223     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 224     r300_emit_buffer_validate(r300, FALSE, NULL);
 225     r300_emit_dirty_state(r300);
 226
 227     BEGIN_CS(dwords);
 228     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 229             r300_provoking_vertex_fixes(r300, mode));
 230     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
 231     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 232     OUT_CS(count - 1);
 233     OUT_CS(0);
 234     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
 235     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
 236             r300_translate_primitive(mode));
 237
 238     /* Emit vertices. */
 239     for (v = 0; v < count; v++) {
 240         for (i = 0; i < vertex_element_count; i++) {
 241             velem = &r300->velems->velem[i];
 242             vbi = velem->vertex_buffer_index;
 243             elem_offset = offset[i] + stride[vbi] * (v + start);
 244
 245             for (dw = 0; dw < size[i]; dw++) {
 246                 OUT_CS(map[vbi][elem_offset + dw]);
 247             }
 248         }
 249     }
 250     END_CS;
 251
 252     /* Unmap buffers. */
 253     for (i = 0; i < vertex_element_count; i++) {
 254         vbi = r300->velems->velem[i].vertex_buffer_index;
 255
 256         if (map[vbi]) {
 257             vbuf = &r300->vertex_buffer[vbi];
 258             pipe_buffer_unmap(&r300->context, vbuf->buffer, transfer[vbi]);
 259             map[vbi] = NULL;
 260         }
 261     }
 262 }
 263
 264 void r500_emit_draw_arrays(struct r300_context *r300,
 265                            unsigned mode,
 266                            unsigned count)
 267 {
 268 #if defined(ENABLE_ALT_NUM_VERTS)
 269     boolean alt_num_verts = count > 65535;
 270 #else
 271     boolean alt_num_verts = FALSE;
 272 #endif
 273     CS_LOCALS(r300);
 274
 275     if (alt_num_verts) {
 276         if (count >= (1 << 24)) {
 277             fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 278                     "refusing to render.\n", count);
 279             return;
 280         }
 281         BEGIN_CS(9);
 282         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 283     } else {
 284         BEGIN_CS(7);
 285     }
 286     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 287             r300_provoking_vertex_fixes(r300, mode));
 288     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 289     OUT_CS(count - 1);
 290     OUT_CS(0);
 291     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 292     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 293            r300_translate_primitive(mode) |
 294            (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 295     END_CS;
 296 }
 297
 298 void r500_emit_draw_elements(struct r300_context *r300,
 299                              struct pipe_resource* indexBuffer,
 300                              unsigned indexSize,
 301                              unsigned minIndex,
 302                              unsigned maxIndex,
 303                              unsigned mode,
 304                              unsigned start,
 305                              unsigned count)
 306 {
 307     uint32_t count_dwords;
 308     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 309 #if defined(ENABLE_ALT_NUM_VERTS)
 310     boolean alt_num_verts = count > 65535;
 311 #else
 312     boolean alt_num_verts = FALSE;
 313 #endif
 314     CS_LOCALS(r300);
 315
 316     if (count >= (1 << 24)) {
 317         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 318                 "refusing to render.\n", count);
 319         return;
 320     }
 321
 322     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 323
 324     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 325         count, minIndex, maxIndex);
 326
 327     if (alt_num_verts) {
 328         BEGIN_CS(15);
 329         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 330     } else {
 331         BEGIN_CS(13);
 332     }
 333     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 334             r300_provoking_vertex_fixes(r300, mode));
 335     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 336     OUT_CS(maxIndex);
 337     OUT_CS(minIndex);
 338     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 339     if (indexSize == 4) {
 340         count_dwords = count;
 341         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 342                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 343                r300_translate_primitive(mode) |
 344                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 345     } else {
 346         count_dwords = (count + 1) / 2;
 347         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 348                r300_translate_primitive(mode) |
 349                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 350     }
 351
 352     /* INDX_BUFFER is a truly special packet3.
 353      * Unlike most other packet3, where the offset is after the count,
 354      * the order is reversed, so the relocation ends up carrying the
 355      * size of the indexbuf instead of the offset.
 356      */
 357     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 358     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 359            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 360     OUT_CS(offset_dwords << 2);
 361     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 362                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 363
 364     END_CS;
 365 }
 366
 367 /*****************************************************************************
 368  * The emission of draw packets for r300 which take care of the two-sided    *
 369  * stencil ref fallback and call r500's functions.                           *
 370  ****************************************************************************/
 371
 372 /* Set drawing for front faces. */
 373 static void r300_begin_stencil_ref_fallback(struct r300_context *r300)
 374 {
 375     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 376     CS_LOCALS(r300);
 377
 378     BEGIN_CS(2);
 379     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_BACK);
 380     END_CS;
 381 }
 382
 383 /* Set drawing for back faces. */
 384 static void r300_switch_stencil_ref_side(struct r300_context *r300)
 385 {
 386     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 387     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 388     CS_LOCALS(r300);
 389
 390     BEGIN_CS(4);
 391     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_FRONT);
 392     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 393                dsa->stencil_ref_bf | r300->stencil_ref.ref_value[1]);
 394     END_CS;
 395 }
 396
 397 /* Restore the original state. */
 398 static void r300_end_stencil_ref_fallback(struct r300_context *r300)
 399 {
 400     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 401     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 402     CS_LOCALS(r300);
 403
 404     BEGIN_CS(4);
 405     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode);
 406     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 407                dsa->stencil_ref_mask | r300->stencil_ref.ref_value[0]);
 408     END_CS;
 409 }
 410
 411 void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 412                                      unsigned mode,
 413                                      unsigned start,
 414                                      unsigned count)
 415 {
 416     if (!r300->stencil_ref_bf_fallback) {
 417         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 418     } else {
 419         r300_begin_stencil_ref_fallback(r300);
 420         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 421         r300_switch_stencil_ref_side(r300);
 422         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 423         r300_end_stencil_ref_fallback(r300);
 424     }
 425 }
 426
 427 void r300_emit_draw_arrays(struct r300_context *r300,
 428                            unsigned mode,
 429                            unsigned count)
 430 {
 431     if (!r300->stencil_ref_bf_fallback) {
 432         r500_emit_draw_arrays(r300, mode, count);
 433     } else {
 434         r300_begin_stencil_ref_fallback(r300);
 435         r500_emit_draw_arrays(r300, mode, count);
 436         r300_switch_stencil_ref_side(r300);
 437         r500_emit_draw_arrays(r300, mode, count);
 438         r300_end_stencil_ref_fallback(r300);
 439     }
 440 }
 441
 442 void r300_emit_draw_elements(struct r300_context *r300,
 443                              struct pipe_resource* indexBuffer,
 444                              unsigned indexSize,
 445                              unsigned minIndex,
 446                              unsigned maxIndex,
 447                              unsigned mode,
 448                              unsigned start,
 449                              unsigned count)
 450 {
 451     if (!r300->stencil_ref_bf_fallback) {
 452         r500_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 453                                 maxIndex, mode, start, count);
 454     } else {
 455         r300_begin_stencil_ref_fallback(r300);
 456         r500_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 457                                 maxIndex, mode, start, count);
 458         r300_switch_stencil_ref_side(r300);
 459         r500_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 460                                 maxIndex, mode, start, count);
 461         r300_end_stencil_ref_fallback(r300);
 462     }
 463 }
 464
 465 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 466                                     struct pipe_resource** elts,
 467                                     unsigned start,
 468                                     unsigned count)
 469 {
 470     struct pipe_context* context = &r300->context;
 471     struct pipe_screen* screen = r300->context.screen;
 472     struct pipe_resource* new_elts;
 473     unsigned char *in_map;
 474     unsigned short *out_map;
 475     struct pipe_transfer *src_transfer, *dst_transfer;
 476     unsigned i;
 477
 478     new_elts = pipe_buffer_create(screen,
 479                                   PIPE_BIND_INDEX_BUFFER,
 480                                   2 * count);
 481
 482     in_map = pipe_buffer_map(context, *elts, PIPE_TRANSFER_READ, &src_transfer);
 483     out_map = pipe_buffer_map(context, new_elts, PIPE_TRANSFER_WRITE, &dst_transfer);
 484
 485     in_map += start;
 486
 487     for (i = 0; i < count; i++) {
 488         *out_map = (unsigned short)*in_map;
 489         in_map++;
 490         out_map++;
 491     }
 492
 493     pipe_buffer_unmap(context, *elts, src_transfer);
 494     pipe_buffer_unmap(context, new_elts, dst_transfer);
 495
 496     *elts = new_elts;
 497 }
 498
 499 static void r300_align_ushort_elts(struct r300_context *r300,
 500                                    struct pipe_resource **elts,
 501                                    unsigned start, unsigned count)
 502 {
 503     struct pipe_context* context = &r300->context;
 504     struct pipe_transfer *in_transfer = NULL;
 505     struct pipe_transfer *out_transfer = NULL;
 506     struct pipe_resource* new_elts;
 507     unsigned short *in_map;
 508     unsigned short *out_map;
 509
 510     new_elts = pipe_buffer_create(context->screen,
 511                                   PIPE_BIND_INDEX_BUFFER,
 512                                   2 * count);
 513
 514     in_map = pipe_buffer_map(context, *elts,
 515                              PIPE_TRANSFER_READ, &in_transfer);
 516     out_map = pipe_buffer_map(context, new_elts,
 517                               PIPE_TRANSFER_WRITE, &out_transfer);
 518
 519     memcpy(out_map, in_map+start, 2 * count);
 520
 521     pipe_buffer_unmap(context, *elts, in_transfer);
 522     pipe_buffer_unmap(context, new_elts, out_transfer);
 523
 524     *elts = new_elts;
 525 }
 526
 527 /* This is the fast-path drawing & emission for HW TCL. */
 528 void r300_draw_range_elements(struct pipe_context* pipe,
 529                               struct pipe_resource* indexBuffer,
 530                               unsigned indexSize,
 531                               unsigned minIndex,
 532                               unsigned maxIndex,
 533                               unsigned mode,
 534                               unsigned start,
 535                               unsigned count)
 536 {
 537     struct r300_context* r300 = r300_context(pipe);
 538     struct pipe_resource* orgIndexBuffer = indexBuffer;
 539 #if defined(ENABLE_ALT_NUM_VERTS)
 540     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 541                             count > 65536;
 542 #else
 543     boolean alt_num_verts = FALSE;
 544 #endif
 545     unsigned short_count;
 546
 547     if (r300->skip_rendering) {
 548         return;
 549     }
 550
 551     if (!u_trim_pipe_prim(mode, &count)) {
 552         return;
 553     }
 554
 555     if (indexSize == 1) {
 556         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 557         indexSize = 2;
 558         start = 0;
 559     } else if (indexSize == 2 && start % 2 != 0) {
 560         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 561         start = 0;
 562     }
 563
 564     r300_update_derived_state(r300);
 565
 566     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 567
 568     /* 128 dwords for emit_aos and emit_draw_elements */
 569     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 570     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 571     r300_emit_dirty_state(r300);
 572     r300_emit_aos(r300, 0);
 573
 574     u_upload_flush(r300->upload_vb);
 575     u_upload_flush(r300->upload_ib);
 576     if (alt_num_verts || count <= 65535) {
 577         r300->emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 578                                  maxIndex, mode, start, count);
 579     } else {
 580         do {
 581             short_count = MIN2(count, 65534);
 582             r300->emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 583                                       maxIndex, mode, start, short_count);
 584
 585             start += short_count;
 586             count -= short_count;
 587
 588             /* 16 spare dwords are enough for emit_draw_elements. */
 589             if (count && r300_reserve_cs_space(r300, 16)) {
 590                 r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 591                 r300_emit_dirty_state(r300);
 592                 r300_emit_aos(r300, 0);
 593             }
 594         } while (count);
 595     }
 596
 597     if (indexBuffer != orgIndexBuffer) {
 598         pipe_resource_reference( &indexBuffer, NULL );
 599     }
 600 }
 601
 602 /* Simple helpers for context setup. Should probably be moved to util. */
 603 void r300_draw_elements(struct pipe_context* pipe,
 604                         struct pipe_resource* indexBuffer,
 605                         unsigned indexSize, unsigned mode,
 606                         unsigned start, unsigned count)
 607 {
 608     struct r300_context *r300 = r300_context(pipe);
 609
 610     pipe->draw_range_elements(pipe, indexBuffer, indexSize, 0,
 611                               r300->vertex_buffer_max_index,
 612                               mode, start, count);
 613 }
 614
 615 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 616                       unsigned start, unsigned count)
 617 {
 618     struct r300_context* r300 = r300_context(pipe);
 619 #if defined(ENABLE_ALT_NUM_VERTS)
 620     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 621                             count > 65536;
 622 #else
 623     boolean alt_num_verts = FALSE;
 624 #endif
 625     unsigned short_count;
 626
 627     if (r300->skip_rendering) {
 628         return;
 629     }
 630
 631     if (!u_trim_pipe_prim(mode, &count)) {
 632         return;
 633     }
 634
 635     r300_update_derived_state(r300);
 636
 637     if (immd_is_good_idea(r300, count)) {
 638         r300->emit_draw_arrays_immediate(r300, mode, start, count);
 639     } else {
 640         /* Make sure there are at least 128 spare dwords in the command buffer.
 641          * (most of it being consumed by emit_aos) */
 642         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 643         r300_emit_buffer_validate(r300, TRUE, NULL);
 644         r300_emit_dirty_state(r300);
 645
 646         if (alt_num_verts || count <= 65535) {
 647             r300_emit_aos(r300, start);
 648             r300->emit_draw_arrays(r300, mode, count);
 649         } else {
 650             do {
 651                 short_count = MIN2(count, 65535);
 652                 r300_emit_aos(r300, start);
 653                 r300->emit_draw_arrays(r300, mode, short_count);
 654
 655                 start += short_count;
 656                 count -= short_count;
 657
 658                 /* Again, we emit both AOS and draw_arrays so there should be
 659                  * at least 128 spare dwords. */
 660                 if (count && r300_reserve_cs_space(r300, 128)) {
 661                     r300_emit_buffer_validate(r300, TRUE, NULL);
 662                     r300_emit_dirty_state(r300);
 663                 }
 664             } while (count);
 665         }
 666         u_upload_flush(r300->upload_vb);
 667     }
 668 }
 669
 670 /****************************************************************************
 671  * The rest of this file is for SW TCL rendering only. Please be polite and *
 672  * keep these functions separated so that they are easier to locate. ~C.    *
 673  ***************************************************************************/
 674
 675 /* SW TCL arrays, using Draw. */
 676 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 677                                unsigned mode,
 678                                unsigned start,
 679                                unsigned count)
 680 {
 681     struct r300_context* r300 = r300_context(pipe);
 682     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 683     int i;
 684
 685     if (r300->skip_rendering) {
 686         return;
 687     }
 688
 689     if (!u_trim_pipe_prim(mode, &count)) {
 690         return;
 691     }
 692
 693     for (i = 0; i < r300->vertex_buffer_count; i++) {
 694         void* buf = pipe_buffer_map(pipe,
 695                                     r300->vertex_buffer[i].buffer,
 696                                     PIPE_TRANSFER_READ,
 697                                     &vb_transfer[i]);
 698         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 699     }
 700
 701     draw_set_mapped_element_buffer(r300->draw, 0, NULL);
 702
 703     draw_arrays(r300->draw, mode, start, count);
 704
 705     for (i = 0; i < r300->vertex_buffer_count; i++) {
 706         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 707                           vb_transfer[i]);
 708         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 709     }
 710 }
 711
 712 /* SW TCL elements, using Draw. */
 713 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 714                                        struct pipe_resource* indexBuffer,
 715                                        unsigned indexSize,
 716                                        unsigned minIndex,
 717                                        unsigned maxIndex,
 718                                        unsigned mode,
 719                                        unsigned start,
 720                                        unsigned count)
 721 {
 722     struct r300_context* r300 = r300_context(pipe);
 723     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 724     struct pipe_transfer *ib_transfer;
 725     int i;
 726     void* indices;
 727
 728     if (r300->skip_rendering) {
 729         return;
 730     }
 731
 732     if (!u_trim_pipe_prim(mode, &count)) {
 733         return;
 734     }
 735
 736     for (i = 0; i < r300->vertex_buffer_count; i++) {
 737         void* buf = pipe_buffer_map(pipe,
 738                                     r300->vertex_buffer[i].buffer,
 739                                     PIPE_TRANSFER_READ,
 740                                     &vb_transfer[i]);
 741         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 742     }
 743
 744     indices = pipe_buffer_map(pipe, indexBuffer,
 745                               PIPE_TRANSFER_READ, &ib_transfer);
 746     draw_set_mapped_element_buffer_range(r300->draw, indexSize,
 747                                          minIndex, maxIndex, indices);
 748
 749     draw_arrays(r300->draw, mode, start, count);
 750
 751     for (i = 0; i < r300->vertex_buffer_count; i++) {
 752         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 753                           vb_transfer[i]);
 754         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 755     }
 756
 757     pipe_buffer_unmap(pipe, indexBuffer,
 758                       ib_transfer);
 759     draw_set_mapped_element_buffer_range(r300->draw, 0, start,
 760                                          start + count - 1, NULL);
 761 }
 762
 763 /* Object for rendering using Draw. */
 764 struct r300_render {
 765     /* Parent class */
 766     struct vbuf_render base;
 767
 768     /* Pipe context */
 769     struct r300_context* r300;
 770
 771     /* Vertex information */
 772     size_t vertex_size;
 773     unsigned prim;
 774     unsigned hwprim;
 775
 776     /* VBO */
 777     struct pipe_resource* vbo;
 778     size_t vbo_size;
 779     size_t vbo_offset;
 780     size_t vbo_max_used;
 781     void * vbo_ptr;
 782
 783    struct pipe_transfer *vbo_transfer;
 784 };
 785
 786 static INLINE struct r300_render*
 787 r300_render(struct vbuf_render* render)
 788 {
 789     return (struct r300_render*)render;
 790 }
 791
 792 static const struct vertex_info*
 793 r300_render_get_vertex_info(struct vbuf_render* render)
 794 {
 795     struct r300_render* r300render = r300_render(render);
 796     struct r300_context* r300 = r300render->r300;
 797
 798     r300_update_derived_state(r300);
 799
 800     return &r300->vertex_info;
 801 }
 802
 803 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 804                                                    ushort vertex_size,
 805                                                    ushort count)
 806 {
 807     struct r300_render* r300render = r300_render(render);
 808     struct r300_context* r300 = r300render->r300;
 809     struct pipe_screen* screen = r300->context.screen;
 810     size_t size = (size_t)vertex_size * (size_t)count;
 811
 812     if (size + r300render->vbo_offset > r300render->vbo_size)
 813     {
 814         pipe_resource_reference(&r300->vbo, NULL);
 815         r300render->vbo = pipe_buffer_create(screen,
 816                                              PIPE_BIND_VERTEX_BUFFER,
 817                                              R300_MAX_DRAW_VBO_SIZE);
 818         r300render->vbo_offset = 0;
 819         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 820     }
 821
 822     r300render->vertex_size = vertex_size;
 823     r300->vbo = r300render->vbo;
 824     r300->vbo_offset = r300render->vbo_offset;
 825
 826     return (r300render->vbo) ? TRUE : FALSE;
 827 }
 828
 829 static void* r300_render_map_vertices(struct vbuf_render* render)
 830 {
 831     struct r300_render* r300render = r300_render(render);
 832
 833     r300render->vbo_ptr = pipe_buffer_map(&r300render->r300->context,
 834                                           r300render->vbo,
 835                                           PIPE_TRANSFER_WRITE,
 836                                           &r300render->vbo_transfer);
 837
 838     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 839 }
 840
 841 static void r300_render_unmap_vertices(struct vbuf_render* render,
 842                                              ushort min,
 843                                              ushort max)
 844 {
 845     struct r300_render* r300render = r300_render(render);
 846     struct pipe_context* context = &r300render->r300->context;
 847     CS_LOCALS(r300render->r300);
 848     BEGIN_CS(2);
 849     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 850     END_CS;
 851
 852     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 853                                     r300render->vertex_size * (max + 1));
 854     pipe_buffer_unmap(context, r300render->vbo, r300render->vbo_transfer);
 855 }
 856
 857 static void r300_render_release_vertices(struct vbuf_render* render)
 858 {
 859     struct r300_render* r300render = r300_render(render);
 860
 861     r300render->vbo_offset += r300render->vbo_max_used;
 862     r300render->vbo_max_used = 0;
 863 }
 864
 865 static boolean r300_render_set_primitive(struct vbuf_render* render,
 866                                                unsigned prim)
 867 {
 868     struct r300_render* r300render = r300_render(render);
 869
 870     r300render->prim = prim;
 871     r300render->hwprim = r300_translate_primitive(prim);
 872
 873     return TRUE;
 874 }
 875
 876 static void r500_render_draw_arrays(struct vbuf_render* render,
 877                                     unsigned start,
 878                                     unsigned count)
 879 {
 880     struct r300_render* r300render = r300_render(render);
 881     struct r300_context* r300 = r300render->r300;
 882
 883     CS_LOCALS(r300);
 884
 885     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
 886     r300_emit_buffer_validate(r300, FALSE, NULL);
 887     r300_emit_dirty_state(r300);
 888
 889     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 890
 891     BEGIN_CS(2);
 892     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 893     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 894            r300render->hwprim);
 895     END_CS;
 896 }
 897
 898 static void r500_render_draw(struct vbuf_render* render,
 899                              const ushort* indices,
 900                              uint count)
 901 {
 902     struct r300_render* r300render = r300_render(render);
 903     struct r300_context* r300 = r300render->r300;
 904     int i;
 905     unsigned dwords = 2 + (count+1)/2;
 906
 907     CS_LOCALS(r300);
 908
 909     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 910     r300_emit_buffer_validate(r300, FALSE, NULL);
 911     r300_emit_dirty_state(r300);
 912
 913     BEGIN_CS(dwords);
 914     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 915     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 916            r300render->hwprim);
 917     for (i = 0; i < count-1; i += 2) {
 918         OUT_CS(indices[i+1] << 16 | indices[i]);
 919     }
 920     if (count % 2) {
 921         OUT_CS(indices[count-1]);
 922     }
 923     END_CS;
 924 }
 925
 926 static void r300_render_draw_arrays(struct vbuf_render* render,
 927                                     unsigned start,
 928                                     unsigned count)
 929 {
 930     struct r300_context* r300 = r300_render(render)->r300;
 931
 932     if (!r300->stencil_ref_bf_fallback) {
 933         r500_render_draw_arrays(render, start, count);
 934     } else {
 935         r300_begin_stencil_ref_fallback(r300);
 936         r500_render_draw_arrays(render, start, count);
 937         r300_switch_stencil_ref_side(r300);
 938         r500_render_draw_arrays(render, start, count);
 939         r300_end_stencil_ref_fallback(r300);
 940     }
 941 }
 942
 943 static void r300_render_draw(struct vbuf_render* render,
 944                              const ushort* indices,
 945                              uint count)
 946 {
 947     struct r300_context* r300 = r300_render(render)->r300;
 948
 949     if (!r300->stencil_ref_bf_fallback) {
 950         r500_render_draw(render, indices, count);
 951     } else {
 952         r300_begin_stencil_ref_fallback(r300);
 953         r500_render_draw(render, indices, count);
 954         r300_switch_stencil_ref_side(r300);
 955         r500_render_draw(render, indices, count);
 956         r300_end_stencil_ref_fallback(r300);
 957     }
 958 }
 959
 960 static void r300_render_destroy(struct vbuf_render* render)
 961 {
 962     FREE(render);
 963 }
 964
 965 static struct vbuf_render* r300_render_create(struct r300_context* r300)
 966 {
 967     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
 968
 969     r300render->r300 = r300;
 970
 971     /* XXX find real numbers plz */
 972     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
 973     r300render->base.max_indices = 16 * 1024;
 974
 975     r300render->base.get_vertex_info = r300_render_get_vertex_info;
 976     r300render->base.allocate_vertices = r300_render_allocate_vertices;
 977     r300render->base.map_vertices = r300_render_map_vertices;
 978     r300render->base.unmap_vertices = r300_render_unmap_vertices;
 979     r300render->base.set_primitive = r300_render_set_primitive;
 980     if (r300->screen->caps.is_r500) {
 981         r300render->base.draw = r500_render_draw;
 982         r300render->base.draw_arrays = r500_render_draw_arrays;
 983     } else {
 984         r300render->base.draw = r300_render_draw;
 985         r300render->base.draw_arrays = r300_render_draw_arrays;
 986     }
 987     r300render->base.release_vertices = r300_render_release_vertices;
 988     r300render->base.destroy = r300_render_destroy;
 989
 990     r300render->vbo = NULL;
 991     r300render->vbo_size = 0;
 992     r300render->vbo_offset = 0;
 993
 994     return &r300render->base;
 995 }
 996
 997 struct draw_stage* r300_draw_stage(struct r300_context* r300)
 998 {
 999     struct vbuf_render* render;
1000     struct draw_stage* stage;
1001
1002     render = r300_render_create(r300);
1003
1004     if (!render) {
1005         return NULL;
1006     }
1007
1008     stage = draw_vbuf_stage(r300->draw, render);
1009
1010     if (!stage) {
1011         render->destroy(render);
1012         return NULL;
1013     }
1014
1015     draw_set_render(r300->draw, render);
1016
1017     return stage;
1018 }