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_buffer_is_referenced(r300, vbuf->buffer)) {
 156                 /* It's a very bad idea to map it... */
 157                 return FALSE;
 158             }
 159             checked[vbi] = TRUE;
 160         }
 161     }
 162     return TRUE;
 163 }
 164
 165 static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 166                                             unsigned mode,
 167                                             unsigned start,
 168                                             unsigned count)
 169 {
 170     struct pipe_vertex_element* velem;
 171     struct pipe_vertex_buffer* vbuf;
 172     unsigned vertex_element_count = r300->velems->count;
 173     unsigned i, v, vbi, dw, elem_offset, dwords;
 174
 175     /* Size of the vertex, in dwords. */
 176     unsigned vertex_size = 0;
 177
 178     /* Offsets of the attribute, in dwords, from the start of the vertex. */
 179     unsigned offset[PIPE_MAX_ATTRIBS];
 180
 181     /* Size of the vertex element, in dwords. */
 182     unsigned size[PIPE_MAX_ATTRIBS];
 183
 184     /* Stride to the same attrib in the next vertex in the vertex buffer,
 185      * in dwords. */
 186     unsigned stride[PIPE_MAX_ATTRIBS] = {0};
 187
 188     /* Mapped vertex buffers. */
 189     uint32_t* map[PIPE_MAX_ATTRIBS] = {0};
 190
 191     CS_LOCALS(r300);
 192
 193     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
 194     for (i = 0; i < vertex_element_count; i++) {
 195         velem = &r300->velems->velem[i];
 196         offset[i] = velem->src_offset / 4;
 197         size[i] = util_format_get_blocksize(velem->src_format) / 4;
 198         vertex_size += size[i];
 199         vbi = velem->vertex_buffer_index;
 200
 201         /* Map the buffer. */
 202         if (!map[vbi]) {
 203             vbuf = &r300->vertex_buffer[vbi];
 204             map[vbi] = (uint32_t*)pipe_buffer_map(r300->context.screen,
 205                                                   vbuf->buffer,
 206                                                   PIPE_BUFFER_USAGE_CPU_READ);
 207             map[vbi] += vbuf->buffer_offset / 4;
 208             stride[vbi] = vbuf->stride / 4;
 209         }
 210     }
 211
 212     dwords = 9 + count * vertex_size;
 213
 214     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 215     r300_emit_buffer_validate(r300, FALSE, NULL);
 216     r300_emit_dirty_state(r300);
 217
 218     BEGIN_CS(dwords);
 219     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 220             r300_provoking_vertex_fixes(r300, mode));
 221     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
 222     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 223     OUT_CS(count - 1);
 224     OUT_CS(0);
 225     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
 226     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
 227             r300_translate_primitive(mode));
 228
 229     /* Emit vertices. */
 230     for (v = 0; v < count; v++) {
 231         for (i = 0; i < vertex_element_count; i++) {
 232             velem = &r300->velems->velem[i];
 233             vbi = velem->vertex_buffer_index;
 234             elem_offset = offset[i] + stride[vbi] * (v + start);
 235
 236             for (dw = 0; dw < size[i]; dw++) {
 237                 OUT_CS(map[vbi][elem_offset + dw]);
 238             }
 239         }
 240     }
 241     END_CS;
 242
 243     /* Unmap buffers. */
 244     for (i = 0; i < vertex_element_count; i++) {
 245         vbi = r300->velems->velem[i].vertex_buffer_index;
 246
 247         if (map[vbi]) {
 248             vbuf = &r300->vertex_buffer[vbi];
 249             pipe_buffer_unmap(r300->context.screen, vbuf->buffer);
 250             map[vbi] = NULL;
 251         }
 252     }
 253 }
 254
 255 static void r300_emit_draw_arrays(struct r300_context *r300,
 256                                   unsigned mode,
 257                                   unsigned count)
 258 {
 259 #if defined(ENABLE_ALT_NUM_VERTS)
 260     boolean alt_num_verts = count > 65535;
 261 #else
 262     boolean alt_num_verts = FALSE;
 263 #endif
 264     CS_LOCALS(r300);
 265
 266     if (alt_num_verts) {
 267         assert(count < (1 << 24));
 268         BEGIN_CS(9);
 269         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 270     } else {
 271         BEGIN_CS(7);
 272     }
 273     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 274             r300_provoking_vertex_fixes(r300, mode));
 275     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 276     OUT_CS(count - 1);
 277     OUT_CS(0);
 278     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 279     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 280            r300_translate_primitive(mode) |
 281            (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 282     END_CS;
 283 }
 284
 285 static void r300_emit_draw_elements(struct r300_context *r300,
 286                                     struct pipe_buffer* indexBuffer,
 287                                     unsigned indexSize,
 288                                     unsigned minIndex,
 289                                     unsigned maxIndex,
 290                                     unsigned mode,
 291                                     unsigned start,
 292                                     unsigned count)
 293 {
 294     uint32_t count_dwords;
 295     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 296 #if defined(ENABLE_ALT_NUM_VERTS)
 297     boolean alt_num_verts = count > 65535;
 298 #else
 299     boolean alt_num_verts = FALSE;
 300 #endif
 301     CS_LOCALS(r300);
 302
 303     assert(count < (1 << 24));
 304
 305     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 306
 307     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 308         count, minIndex, maxIndex);
 309
 310     if (alt_num_verts) {
 311         BEGIN_CS(15);
 312         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 313     } else {
 314         BEGIN_CS(13);
 315     }
 316     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 317             r300_provoking_vertex_fixes(r300, mode));
 318     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 319     OUT_CS(maxIndex);
 320     OUT_CS(minIndex);
 321     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 322     if (indexSize == 4) {
 323         count_dwords = count;
 324         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 325                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 326                r300_translate_primitive(mode) |
 327                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 328     } else {
 329         count_dwords = (count + 1) / 2;
 330         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 331                r300_translate_primitive(mode) |
 332                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 333     }
 334
 335     /* INDX_BUFFER is a truly special packet3.
 336      * Unlike most other packet3, where the offset is after the count,
 337      * the order is reversed, so the relocation ends up carrying the
 338      * size of the indexbuf instead of the offset.
 339      */
 340     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 341     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 342            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 343     OUT_CS(offset_dwords << 2);
 344     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 345                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 346
 347     END_CS;
 348 }
 349
 350 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 351                                     struct pipe_buffer** elts,
 352                                     unsigned start,
 353                                     unsigned count)
 354 {
 355     struct pipe_screen* screen = r300->context.screen;
 356     struct pipe_buffer* new_elts;
 357     unsigned char *in_map;
 358     unsigned short *out_map;
 359     unsigned i;
 360
 361     new_elts = screen->buffer_create(screen, 32,
 362                                      PIPE_BUFFER_USAGE_INDEX |
 363                                      PIPE_BUFFER_USAGE_CPU_WRITE |
 364                                      PIPE_BUFFER_USAGE_GPU_READ,
 365                                      2 * count);
 366
 367     in_map = pipe_buffer_map(screen, *elts, PIPE_BUFFER_USAGE_CPU_READ);
 368     out_map = pipe_buffer_map(screen, new_elts, PIPE_BUFFER_USAGE_CPU_WRITE);
 369
 370     in_map += start;
 371
 372     for (i = 0; i < count; i++) {
 373         *out_map = (unsigned short)*in_map;
 374         in_map++;
 375         out_map++;
 376     }
 377
 378     pipe_buffer_unmap(screen, *elts);
 379     pipe_buffer_unmap(screen, new_elts);
 380
 381     *elts = new_elts;
 382 }
 383
 384 static void r300_align_ushort_elts(struct r300_context *r300,
 385                                    struct pipe_buffer **elts,
 386                                    unsigned start, unsigned count)
 387 {
 388     struct pipe_screen* screen = r300->context.screen;
 389     struct pipe_buffer* new_elts;
 390     unsigned short *in_map;
 391     unsigned short *out_map;
 392
 393     new_elts = screen->buffer_create(screen, 32,
 394                                      PIPE_BUFFER_USAGE_INDEX |
 395                                      PIPE_BUFFER_USAGE_CPU_WRITE |
 396                                      PIPE_BUFFER_USAGE_GPU_READ,
 397                                      2 * count);
 398
 399     in_map = pipe_buffer_map(screen, *elts, PIPE_BUFFER_USAGE_CPU_READ);
 400     out_map = pipe_buffer_map(screen, new_elts, PIPE_BUFFER_USAGE_CPU_WRITE);
 401
 402     memcpy(out_map, in_map+start, 2 * count);
 403
 404     pipe_buffer_unmap(screen, *elts);
 405     pipe_buffer_unmap(screen, new_elts);
 406
 407     *elts = new_elts;
 408 }
 409
 410 /* This is the fast-path drawing & emission for HW TCL. */
 411 void r300_draw_range_elements(struct pipe_context* pipe,
 412                               struct pipe_buffer* indexBuffer,
 413                               unsigned indexSize,
 414                               unsigned minIndex,
 415                               unsigned maxIndex,
 416                               unsigned mode,
 417                               unsigned start,
 418                               unsigned count)
 419 {
 420     struct r300_context* r300 = r300_context(pipe);
 421     struct pipe_buffer* orgIndexBuffer = indexBuffer;
 422 #if defined(ENABLE_ALT_NUM_VERTS)
 423     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 424                             count > 65536;
 425 #else
 426     boolean alt_num_verts = FALSE;
 427 #endif
 428     unsigned short_count;
 429
 430     if (r300->skip_rendering) {
 431         return;
 432     }
 433
 434     if (!u_trim_pipe_prim(mode, &count)) {
 435         return;
 436     }
 437
 438     if (indexSize == 1) {
 439         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 440         indexSize = 2;
 441         start = 0;
 442     } else if (indexSize == 2 && start % 2 != 0) {
 443         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 444         start = 0;
 445     }
 446
 447     r300_update_derived_state(r300);
 448
 449     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 450
 451     /* 128 dwords for emit_aos and emit_draw_elements */
 452     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 453     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 454     r300_emit_dirty_state(r300);
 455     r300_emit_aos(r300, 0);
 456
 457     u_upload_flush(r300->upload_vb);
 458     u_upload_flush(r300->upload_ib);
 459     if (alt_num_verts || count <= 65535) {
 460         r300_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 461                                 maxIndex, mode, start, count);
 462     } else {
 463         do {
 464             short_count = MIN2(count, 65534);
 465             r300_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 466                                     maxIndex, mode, start, short_count);
 467
 468             start += short_count;
 469             count -= short_count;
 470
 471             /* 16 spare dwords are enough for emit_draw_elements. */
 472             if (count && r300_reserve_cs_space(r300, 16)) {
 473                 r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 474                 r300_emit_dirty_state(r300);
 475                 r300_emit_aos(r300, 0);
 476             }
 477         } while (count);
 478     }
 479
 480     if (indexBuffer != orgIndexBuffer) {
 481         pipe_buffer_reference( &indexBuffer, NULL );
 482     }
 483 }
 484
 485 /* Simple helpers for context setup. Should probably be moved to util. */
 486 void r300_draw_elements(struct pipe_context* pipe,
 487                         struct pipe_buffer* indexBuffer,
 488                         unsigned indexSize, unsigned mode,
 489                         unsigned start, unsigned count)
 490 {
 491     struct r300_context *r300 = r300_context(pipe);
 492
 493     pipe->draw_range_elements(pipe, indexBuffer, indexSize, 0,
 494                               r300->vertex_buffer_max_index,
 495                               mode, start, count);
 496 }
 497
 498 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 499                       unsigned start, unsigned count)
 500 {
 501     struct r300_context* r300 = r300_context(pipe);
 502 #if defined(ENABLE_ALT_NUM_VERTS)
 503     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 504                             count > 65536;
 505 #else
 506     boolean alt_num_verts = FALSE;
 507 #endif
 508     unsigned short_count;
 509
 510     if (r300->skip_rendering) {
 511         return;
 512     }
 513
 514     if (!u_trim_pipe_prim(mode, &count)) {
 515         return;
 516     }
 517
 518     r300_update_derived_state(r300);
 519
 520     if (immd_is_good_idea(r300, count)) {
 521         r300_emit_draw_arrays_immediate(r300, mode, start, count);
 522     } else {
 523         /* Make sure there are at least 128 spare dwords in the command buffer.
 524          * (most of it being consumed by emit_aos) */
 525         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 526         r300_emit_buffer_validate(r300, TRUE, NULL);
 527         r300_emit_dirty_state(r300);
 528
 529         if (alt_num_verts || count <= 65535) {
 530             r300_emit_aos(r300, start);
 531             r300_emit_draw_arrays(r300, mode, count);
 532         } else {
 533             do {
 534                 short_count = MIN2(count, 65535);
 535                 r300_emit_aos(r300, start);
 536                 r300_emit_draw_arrays(r300, mode, short_count);
 537
 538                 start += short_count;
 539                 count -= short_count;
 540
 541                 /* Again, we emit both AOS and draw_arrays so there should be
 542                  * at least 128 spare dwords. */
 543                 if (count && r300_reserve_cs_space(r300, 128)) {
 544                     r300_emit_buffer_validate(r300, TRUE, NULL);
 545                     r300_emit_dirty_state(r300);
 546                 }
 547             } while (count);
 548         }
 549         u_upload_flush(r300->upload_vb);
 550     }
 551 }
 552
 553 /****************************************************************************
 554  * The rest of this file is for SW TCL rendering only. Please be polite and *
 555  * keep these functions separated so that they are easier to locate. ~C.    *
 556  ***************************************************************************/
 557
 558 /* SW TCL arrays, using Draw. */
 559 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 560                                unsigned mode,
 561                                unsigned start,
 562                                unsigned count)
 563 {
 564     struct r300_context* r300 = r300_context(pipe);
 565     int i;
 566
 567     if (r300->skip_rendering) {
 568         return;
 569     }
 570
 571     if (!u_trim_pipe_prim(mode, &count)) {
 572         return;
 573     }
 574
 575     for (i = 0; i < r300->vertex_buffer_count; i++) {
 576         void* buf = pipe_buffer_map(pipe->screen,
 577                                     r300->vertex_buffer[i].buffer,
 578                                     PIPE_BUFFER_USAGE_CPU_READ);
 579         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 580     }
 581
 582     draw_set_mapped_element_buffer(r300->draw, 0, NULL);
 583
 584     draw_arrays(r300->draw, mode, start, count);
 585
 586     for (i = 0; i < r300->vertex_buffer_count; i++) {
 587         pipe_buffer_unmap(pipe->screen, r300->vertex_buffer[i].buffer);
 588         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 589     }
 590 }
 591
 592 /* SW TCL elements, using Draw. */
 593 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 594                                        struct pipe_buffer* indexBuffer,
 595                                        unsigned indexSize,
 596                                        unsigned minIndex,
 597                                        unsigned maxIndex,
 598                                        unsigned mode,
 599                                        unsigned start,
 600                                        unsigned count)
 601 {
 602     struct r300_context* r300 = r300_context(pipe);
 603     int i;
 604     void* indices;
 605
 606     if (r300->skip_rendering) {
 607         return;
 608     }
 609
 610     if (!u_trim_pipe_prim(mode, &count)) {
 611         return;
 612     }
 613
 614     for (i = 0; i < r300->vertex_buffer_count; i++) {
 615         void* buf = pipe_buffer_map(pipe->screen,
 616                                     r300->vertex_buffer[i].buffer,
 617                                     PIPE_BUFFER_USAGE_CPU_READ);
 618         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 619     }
 620
 621     indices = pipe_buffer_map(pipe->screen, indexBuffer,
 622                               PIPE_BUFFER_USAGE_CPU_READ);
 623     draw_set_mapped_element_buffer_range(r300->draw, indexSize,
 624                                          minIndex, maxIndex, indices);
 625
 626     draw_arrays(r300->draw, mode, start, count);
 627
 628     for (i = 0; i < r300->vertex_buffer_count; i++) {
 629         pipe_buffer_unmap(pipe->screen, r300->vertex_buffer[i].buffer);
 630         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 631     }
 632
 633     pipe_buffer_unmap(pipe->screen, indexBuffer);
 634     draw_set_mapped_element_buffer_range(r300->draw, 0, start,
 635                                          start + count - 1, NULL);
 636 }
 637
 638 /* Object for rendering using Draw. */
 639 struct r300_render {
 640     /* Parent class */
 641     struct vbuf_render base;
 642
 643     /* Pipe context */
 644     struct r300_context* r300;
 645
 646     /* Vertex information */
 647     size_t vertex_size;
 648     unsigned prim;
 649     unsigned hwprim;
 650
 651     /* VBO */
 652     struct pipe_buffer* vbo;
 653     size_t vbo_size;
 654     size_t vbo_offset;
 655     size_t vbo_max_used;
 656     void * vbo_ptr;
 657 };
 658
 659 static INLINE struct r300_render*
 660 r300_render(struct vbuf_render* render)
 661 {
 662     return (struct r300_render*)render;
 663 }
 664
 665 static const struct vertex_info*
 666 r300_render_get_vertex_info(struct vbuf_render* render)
 667 {
 668     struct r300_render* r300render = r300_render(render);
 669     struct r300_context* r300 = r300render->r300;
 670
 671     r300_update_derived_state(r300);
 672
 673     return &r300->vertex_info;
 674 }
 675
 676 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 677                                                    ushort vertex_size,
 678                                                    ushort count)
 679 {
 680     struct r300_render* r300render = r300_render(render);
 681     struct r300_context* r300 = r300render->r300;
 682     struct pipe_screen* screen = r300->context.screen;
 683     size_t size = (size_t)vertex_size * (size_t)count;
 684
 685     if (size + r300render->vbo_offset > r300render->vbo_size)
 686     {
 687         pipe_buffer_reference(&r300->vbo, NULL);
 688         r300render->vbo = pipe_buffer_create(screen,
 689                                              64,
 690                                              PIPE_BUFFER_USAGE_VERTEX,
 691                                              R300_MAX_DRAW_VBO_SIZE);
 692         r300render->vbo_offset = 0;
 693         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 694     }
 695
 696     r300render->vertex_size = vertex_size;
 697     r300->vbo = r300render->vbo;
 698     r300->vbo_offset = r300render->vbo_offset;
 699
 700     return (r300render->vbo) ? TRUE : FALSE;
 701 }
 702
 703 static void* r300_render_map_vertices(struct vbuf_render* render)
 704 {
 705     struct r300_render* r300render = r300_render(render);
 706     struct pipe_screen* screen = r300render->r300->context.screen;
 707
 708     r300render->vbo_ptr = pipe_buffer_map(screen, r300render->vbo,
 709                                           PIPE_BUFFER_USAGE_CPU_WRITE);
 710
 711     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 712 }
 713
 714 static void r300_render_unmap_vertices(struct vbuf_render* render,
 715                                              ushort min,
 716                                              ushort max)
 717 {
 718     struct r300_render* r300render = r300_render(render);
 719     struct pipe_screen* screen = r300render->r300->context.screen;
 720     CS_LOCALS(r300render->r300);
 721     BEGIN_CS(2);
 722     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 723     END_CS;
 724
 725     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 726                                     r300render->vertex_size * (max + 1));
 727     pipe_buffer_unmap(screen, r300render->vbo);
 728 }
 729
 730 static void r300_render_release_vertices(struct vbuf_render* render)
 731 {
 732     struct r300_render* r300render = r300_render(render);
 733
 734     r300render->vbo_offset += r300render->vbo_max_used;
 735     r300render->vbo_max_used = 0;
 736 }
 737
 738 static boolean r300_render_set_primitive(struct vbuf_render* render,
 739                                                unsigned prim)
 740 {
 741     struct r300_render* r300render = r300_render(render);
 742
 743     r300render->prim = prim;
 744     r300render->hwprim = r300_translate_primitive(prim);
 745
 746     return TRUE;
 747 }
 748
 749 static void r300_render_draw_arrays(struct vbuf_render* render,
 750                                           unsigned start,
 751                                           unsigned count)
 752 {
 753     struct r300_render* r300render = r300_render(render);
 754     struct r300_context* r300 = r300render->r300;
 755
 756     CS_LOCALS(r300);
 757
 758     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
 759     r300_emit_buffer_validate(r300, FALSE, NULL);
 760     r300_emit_dirty_state(r300);
 761
 762     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 763
 764     BEGIN_CS(2);
 765     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 766     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 767            r300render->hwprim);
 768     END_CS;
 769 }
 770
 771 static void r300_render_draw(struct vbuf_render* render,
 772                                    const ushort* indices,
 773                                    uint count)
 774 {
 775     struct r300_render* r300render = r300_render(render);
 776     struct r300_context* r300 = r300render->r300;
 777     int i;
 778     unsigned dwords = 2 + (count+1)/2;
 779
 780     CS_LOCALS(r300);
 781
 782     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 783     r300_emit_buffer_validate(r300, FALSE, NULL);
 784     r300_emit_dirty_state(r300);
 785
 786     BEGIN_CS(dwords);
 787     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 788     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 789            r300render->hwprim);
 790     for (i = 0; i < count-1; i += 2) {
 791         OUT_CS(indices[i+1] << 16 | indices[i]);
 792     }
 793     if (count % 2) {
 794         OUT_CS(indices[count-1]);
 795     }
 796     END_CS;
 797 }
 798
 799 static void r300_render_destroy(struct vbuf_render* render)
 800 {
 801     FREE(render);
 802 }
 803
 804 static struct vbuf_render* r300_render_create(struct r300_context* r300)
 805 {
 806     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
 807
 808     r300render->r300 = r300;
 809
 810     /* XXX find real numbers plz */
 811     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
 812     r300render->base.max_indices = 16 * 1024;
 813
 814     r300render->base.get_vertex_info = r300_render_get_vertex_info;
 815     r300render->base.allocate_vertices = r300_render_allocate_vertices;
 816     r300render->base.map_vertices = r300_render_map_vertices;
 817     r300render->base.unmap_vertices = r300_render_unmap_vertices;
 818     r300render->base.set_primitive = r300_render_set_primitive;
 819     r300render->base.draw = r300_render_draw;
 820     r300render->base.draw_arrays = r300_render_draw_arrays;
 821     r300render->base.release_vertices = r300_render_release_vertices;
 822     r300render->base.destroy = r300_render_destroy;
 823
 824     r300render->vbo = NULL;
 825     r300render->vbo_size = 0;
 826     r300render->vbo_offset = 0;
 827
 828     return &r300render->base;
 829 }
 830
 831 struct draw_stage* r300_draw_stage(struct r300_context* r300)
 832 {
 833     struct vbuf_render* render;
 834     struct draw_stage* stage;
 835
 836     render = r300_render_create(r300);
 837
 838     if (!render) {
 839         return NULL;
 840     }
 841
 842     stage = draw_vbuf_stage(r300->draw, render);
 843
 844     if (!stage) {
 845         render->destroy(render);
 846         return NULL;
 847     }
 848
 849     draw_set_render(r300->draw, render);
 850
 851     return stage;
 852 }