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(pipe->screen)->caps->is_r500 &&
 424                             count > 65536;
 425 #else
 426     boolean alt_num_verts = FALSE;
 427 #endif
 428     unsigned short_count;
 429
 430     if (!u_trim_pipe_prim(mode, &count)) {
 431         return;
 432     }
 433
 434     if (indexSize == 1) {
 435         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 436         indexSize = 2;
 437         start = 0;
 438     } else if (indexSize == 2 && start % 2 != 0) {
 439         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 440         start = 0;
 441     }
 442
 443     r300_update_derived_state(r300);
 444
 445     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 446
 447     /* 128 dwords for emit_aos and emit_draw_elements */
 448     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 449     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 450     r300_emit_dirty_state(r300);
 451     r300_emit_aos(r300, 0);
 452
 453     u_upload_flush(r300->upload_vb);
 454     u_upload_flush(r300->upload_ib);
 455     if (alt_num_verts || count <= 65535) {
 456         r300_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 457                                 maxIndex, mode, start, count);
 458     } else {
 459         do {
 460             short_count = MIN2(count, 65534);
 461             r300_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 462                                     maxIndex, mode, start, short_count);
 463
 464             start += short_count;
 465             count -= short_count;
 466
 467             /* 16 spare dwords are enough for emit_draw_elements. */
 468             if (count && r300_reserve_cs_space(r300, 16)) {
 469                 r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 470                 r300_emit_dirty_state(r300);
 471                 r300_emit_aos(r300, 0);
 472             }
 473         } while (count);
 474     }
 475
 476     if (indexBuffer != orgIndexBuffer) {
 477         pipe_buffer_reference( &indexBuffer, NULL );
 478     }
 479 }
 480
 481 /* Simple helpers for context setup. Should probably be moved to util. */
 482 void r300_draw_elements(struct pipe_context* pipe,
 483                         struct pipe_buffer* indexBuffer,
 484                         unsigned indexSize, unsigned mode,
 485                         unsigned start, unsigned count)
 486 {
 487     struct r300_context *r300 = r300_context(pipe);
 488
 489     pipe->draw_range_elements(pipe, indexBuffer, indexSize, 0,
 490                               r300->vertex_buffer_max_index,
 491                               mode, start, count);
 492 }
 493
 494 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 495                       unsigned start, unsigned count)
 496 {
 497     struct r300_context* r300 = r300_context(pipe);
 498 #if defined(ENABLE_ALT_NUM_VERTS)
 499     boolean alt_num_verts = r300_screen(pipe->screen)->caps->is_r500 &&
 500                             count > 65536;
 501 #else
 502     boolean alt_num_verts = FALSE;
 503 #endif
 504     unsigned short_count;
 505
 506     if (!u_trim_pipe_prim(mode, &count)) {
 507         return;
 508     }
 509
 510     r300_update_derived_state(r300);
 511
 512     if (immd_is_good_idea(r300, count)) {
 513         r300_emit_draw_arrays_immediate(r300, mode, start, count);
 514     } else {
 515         /* Make sure there are at least 128 spare dwords in the command buffer.
 516          * (most of it being consumed by emit_aos) */
 517         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 518         r300_emit_buffer_validate(r300, TRUE, NULL);
 519         r300_emit_dirty_state(r300);
 520
 521         if (alt_num_verts || count <= 65535) {
 522             r300_emit_aos(r300, start);
 523             r300_emit_draw_arrays(r300, mode, count);
 524         } else {
 525             do {
 526                 short_count = MIN2(count, 65535);
 527                 r300_emit_aos(r300, start);
 528                 r300_emit_draw_arrays(r300, mode, short_count);
 529
 530                 start += short_count;
 531                 count -= short_count;
 532
 533                 /* Again, we emit both AOS and draw_arrays so there should be
 534                  * at least 128 spare dwords. */
 535                 if (count && r300_reserve_cs_space(r300, 128)) {
 536                     r300_emit_buffer_validate(r300, TRUE, NULL);
 537                     r300_emit_dirty_state(r300);
 538                 }
 539             } while (count);
 540         }
 541         u_upload_flush(r300->upload_vb);
 542     }
 543 }
 544
 545 /****************************************************************************
 546  * The rest of this file is for SW TCL rendering only. Please be polite and *
 547  * keep these functions separated so that they are easier to locate. ~C.    *
 548  ***************************************************************************/
 549
 550 /* SW TCL arrays, using Draw. */
 551 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 552                                unsigned mode,
 553                                unsigned start,
 554                                unsigned count)
 555 {
 556     struct r300_context* r300 = r300_context(pipe);
 557     int i;
 558
 559     if (!u_trim_pipe_prim(mode, &count)) {
 560         return;
 561     }
 562
 563     for (i = 0; i < r300->vertex_buffer_count; i++) {
 564         void* buf = pipe_buffer_map(pipe->screen,
 565                                     r300->vertex_buffer[i].buffer,
 566                                     PIPE_BUFFER_USAGE_CPU_READ);
 567         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 568     }
 569
 570     draw_set_mapped_element_buffer(r300->draw, 0, NULL);
 571
 572     draw_set_mapped_constant_buffer(r300->draw,
 573                                     PIPE_SHADER_VERTEX,
 574                                     0,
 575                                     r300->shader_constants[PIPE_SHADER_VERTEX].constants,
 576                                     r300->shader_constants[PIPE_SHADER_VERTEX].count *
 577                 (sizeof(float) * 4));
 578
 579     draw_arrays(r300->draw, mode, start, count);
 580
 581     for (i = 0; i < r300->vertex_buffer_count; i++) {
 582         pipe_buffer_unmap(pipe->screen, r300->vertex_buffer[i].buffer);
 583         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 584     }
 585 }
 586
 587 /* SW TCL elements, using Draw. */
 588 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 589                                        struct pipe_buffer* indexBuffer,
 590                                        unsigned indexSize,
 591                                        unsigned minIndex,
 592                                        unsigned maxIndex,
 593                                        unsigned mode,
 594                                        unsigned start,
 595                                        unsigned count)
 596 {
 597     struct r300_context* r300 = r300_context(pipe);
 598     int i;
 599     void* indices;
 600
 601     if (!u_trim_pipe_prim(mode, &count)) {
 602         return;
 603     }
 604
 605     for (i = 0; i < r300->vertex_buffer_count; i++) {
 606         void* buf = pipe_buffer_map(pipe->screen,
 607                                     r300->vertex_buffer[i].buffer,
 608                                     PIPE_BUFFER_USAGE_CPU_READ);
 609         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 610     }
 611
 612     indices = pipe_buffer_map(pipe->screen, indexBuffer,
 613                               PIPE_BUFFER_USAGE_CPU_READ);
 614     draw_set_mapped_element_buffer_range(r300->draw, indexSize,
 615                                          minIndex, maxIndex, indices);
 616
 617     draw_set_mapped_constant_buffer(r300->draw,
 618                                     PIPE_SHADER_VERTEX,
 619                                     0,
 620             r300->shader_constants[PIPE_SHADER_VERTEX].constants,
 621             r300->shader_constants[PIPE_SHADER_VERTEX].count *
 622                 (sizeof(float) * 4));
 623
 624     draw_arrays(r300->draw, mode, start, count);
 625
 626     for (i = 0; i < r300->vertex_buffer_count; i++) {
 627         pipe_buffer_unmap(pipe->screen, r300->vertex_buffer[i].buffer);
 628         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 629     }
 630
 631     pipe_buffer_unmap(pipe->screen, indexBuffer);
 632     draw_set_mapped_element_buffer_range(r300->draw, 0, start,
 633                                          start + count - 1, NULL);
 634 }
 635
 636 /* Object for rendering using Draw. */
 637 struct r300_render {
 638     /* Parent class */
 639     struct vbuf_render base;
 640
 641     /* Pipe context */
 642     struct r300_context* r300;
 643
 644     /* Vertex information */
 645     size_t vertex_size;
 646     unsigned prim;
 647     unsigned hwprim;
 648
 649     /* VBO */
 650     struct pipe_buffer* vbo;
 651     size_t vbo_size;
 652     size_t vbo_offset;
 653     size_t vbo_max_used;
 654     void * vbo_ptr;
 655 };
 656
 657 static INLINE struct r300_render*
 658 r300_render(struct vbuf_render* render)
 659 {
 660     return (struct r300_render*)render;
 661 }
 662
 663 static const struct vertex_info*
 664 r300_render_get_vertex_info(struct vbuf_render* render)
 665 {
 666     struct r300_render* r300render = r300_render(render);
 667     struct r300_context* r300 = r300render->r300;
 668
 669     r300_update_derived_state(r300);
 670
 671     return &r300->vertex_info;
 672 }
 673
 674 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 675                                                    ushort vertex_size,
 676                                                    ushort count)
 677 {
 678     struct r300_render* r300render = r300_render(render);
 679     struct r300_context* r300 = r300render->r300;
 680     struct pipe_screen* screen = r300->context.screen;
 681     size_t size = (size_t)vertex_size * (size_t)count;
 682
 683     if (size + r300render->vbo_offset > r300render->vbo_size)
 684     {
 685         pipe_buffer_reference(&r300->vbo, NULL);
 686         r300render->vbo = pipe_buffer_create(screen,
 687                                              64,
 688                                              PIPE_BUFFER_USAGE_VERTEX,
 689                                              R300_MAX_DRAW_VBO_SIZE);
 690         r300render->vbo_offset = 0;
 691         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 692     }
 693
 694     r300render->vertex_size = vertex_size;
 695     r300->vbo = r300render->vbo;
 696     r300->vbo_offset = r300render->vbo_offset;
 697
 698     return (r300render->vbo) ? TRUE : FALSE;
 699 }
 700
 701 static void* r300_render_map_vertices(struct vbuf_render* render)
 702 {
 703     struct r300_render* r300render = r300_render(render);
 704     struct pipe_screen* screen = r300render->r300->context.screen;
 705
 706     r300render->vbo_ptr = pipe_buffer_map(screen, r300render->vbo,
 707                                           PIPE_BUFFER_USAGE_CPU_WRITE);
 708
 709     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 710 }
 711
 712 static void r300_render_unmap_vertices(struct vbuf_render* render,
 713                                              ushort min,
 714                                              ushort max)
 715 {
 716     struct r300_render* r300render = r300_render(render);
 717     struct pipe_screen* screen = r300render->r300->context.screen;
 718     CS_LOCALS(r300render->r300);
 719     BEGIN_CS(2);
 720     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 721     END_CS;
 722
 723     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 724                                     r300render->vertex_size * (max + 1));
 725     pipe_buffer_unmap(screen, r300render->vbo);
 726 }
 727
 728 static void r300_render_release_vertices(struct vbuf_render* render)
 729 {
 730     struct r300_render* r300render = r300_render(render);
 731
 732     r300render->vbo_offset += r300render->vbo_max_used;
 733     r300render->vbo_max_used = 0;
 734 }
 735
 736 static boolean r300_render_set_primitive(struct vbuf_render* render,
 737                                                unsigned prim)
 738 {
 739     struct r300_render* r300render = r300_render(render);
 740
 741     r300render->prim = prim;
 742     r300render->hwprim = r300_translate_primitive(prim);
 743
 744     return TRUE;
 745 }
 746
 747 static void r300_render_draw_arrays(struct vbuf_render* render,
 748                                           unsigned start,
 749                                           unsigned count)
 750 {
 751     struct r300_render* r300render = r300_render(render);
 752     struct r300_context* r300 = r300render->r300;
 753
 754     CS_LOCALS(r300);
 755
 756     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
 757     r300_emit_buffer_validate(r300, FALSE, NULL);
 758     r300_emit_dirty_state(r300);
 759
 760     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 761
 762     BEGIN_CS(2);
 763     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 764     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 765            r300render->hwprim);
 766     END_CS;
 767 }
 768
 769 static void r300_render_draw(struct vbuf_render* render,
 770                                    const ushort* indices,
 771                                    uint count)
 772 {
 773     struct r300_render* r300render = r300_render(render);
 774     struct r300_context* r300 = r300render->r300;
 775     int i;
 776     unsigned dwords = 2 + (count+1)/2;
 777
 778     CS_LOCALS(r300);
 779
 780     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 781     r300_emit_buffer_validate(r300, FALSE, NULL);
 782     r300_emit_dirty_state(r300);
 783
 784     BEGIN_CS(dwords);
 785     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 786     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 787            r300render->hwprim);
 788     for (i = 0; i < count-1; i += 2) {
 789         OUT_CS(indices[i+1] << 16 | indices[i]);
 790     }
 791     if (count % 2) {
 792         OUT_CS(indices[count-1]);
 793     }
 794     END_CS;
 795 }
 796
 797 static void r300_render_destroy(struct vbuf_render* render)
 798 {
 799     FREE(render);
 800 }
 801
 802 static struct vbuf_render* r300_render_create(struct r300_context* r300)
 803 {
 804     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
 805
 806     r300render->r300 = r300;
 807
 808     /* XXX find real numbers plz */
 809     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
 810     r300render->base.max_indices = 16 * 1024;
 811
 812     r300render->base.get_vertex_info = r300_render_get_vertex_info;
 813     r300render->base.allocate_vertices = r300_render_allocate_vertices;
 814     r300render->base.map_vertices = r300_render_map_vertices;
 815     r300render->base.unmap_vertices = r300_render_unmap_vertices;
 816     r300render->base.set_primitive = r300_render_set_primitive;
 817     r300render->base.draw = r300_render_draw;
 818     r300render->base.draw_arrays = r300_render_draw_arrays;
 819     r300render->base.release_vertices = r300_render_release_vertices;
 820     r300render->base.destroy = r300_render_destroy;
 821
 822     r300render->vbo = NULL;
 823     r300render->vbo_size = 0;
 824     r300render->vbo_offset = 0;
 825
 826     return &r300render->base;
 827 }
 828
 829 struct draw_stage* r300_draw_stage(struct r300_context* r300)
 830 {
 831     struct vbuf_render* render;
 832     struct draw_stage* stage;
 833
 834     render = r300_render_create(r300);
 835
 836     if (!render) {
 837         return NULL;
 838     }
 839
 840     stage = draw_vbuf_stage(r300->draw, render);
 841
 842     if (!stage) {
 843         render->destroy(render);
 844         return NULL;
 845     }
 846
 847     draw_set_render(r300->draw, render);
 848
 849     return stage;
 850 }