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 static 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 || DBG_ON(r300, DBG_NO_IMMD)) {
 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->context,
 156                                           vbuf->buffer,
 157                                           R300_REF_CS | R300_REF_HW)) {
 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                              int indexBias,
 302                              unsigned minIndex,
 303                              unsigned maxIndex,
 304                              unsigned mode,
 305                              unsigned start,
 306                              unsigned count)
 307 {
 308     uint32_t count_dwords;
 309     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 310 #if defined(ENABLE_ALT_NUM_VERTS)
 311     boolean alt_num_verts = count > 65535;
 312 #else
 313     boolean alt_num_verts = FALSE;
 314 #endif
 315     CS_LOCALS(r300);
 316
 317     if (count >= (1 << 24)) {
 318         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 319                 "refusing to render.\n", count);
 320         return;
 321     }
 322
 323     assert(indexBias == 0);
 324
 325     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 326
 327     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 328         count, minIndex, maxIndex);
 329
 330     if (alt_num_verts) {
 331         BEGIN_CS(15);
 332         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 333     } else {
 334         BEGIN_CS(13);
 335     }
 336     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 337             r300_provoking_vertex_fixes(r300, mode));
 338     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 339     OUT_CS(maxIndex);
 340     OUT_CS(minIndex);
 341     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 342     if (indexSize == 4) {
 343         count_dwords = count;
 344         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 345                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 346                r300_translate_primitive(mode) |
 347                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 348     } else {
 349         count_dwords = (count + 1) / 2;
 350         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 351                r300_translate_primitive(mode) |
 352                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 353     }
 354
 355     /* INDX_BUFFER is a truly special packet3.
 356      * Unlike most other packet3, where the offset is after the count,
 357      * the order is reversed, so the relocation ends up carrying the
 358      * size of the indexbuf instead of the offset.
 359      */
 360     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 361     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 362            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 363     OUT_CS(offset_dwords << 2);
 364     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 365                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 366
 367     END_CS;
 368 }
 369
 370 /*****************************************************************************
 371  * The emission of draw packets for r300 which take care of the two-sided    *
 372  * stencil ref fallback and call r500's functions.                           *
 373  ****************************************************************************/
 374
 375 /* Set drawing for front faces. */
 376 static void r300_begin_stencil_ref_fallback(struct r300_context *r300)
 377 {
 378     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 379     CS_LOCALS(r300);
 380
 381     BEGIN_CS(2);
 382     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_BACK);
 383     END_CS;
 384 }
 385
 386 /* Set drawing for back faces. */
 387 static void r300_switch_stencil_ref_side(struct r300_context *r300)
 388 {
 389     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 390     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 391     CS_LOCALS(r300);
 392
 393     BEGIN_CS(4);
 394     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_FRONT);
 395     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 396                dsa->stencil_ref_bf | r300->stencil_ref.ref_value[1]);
 397     END_CS;
 398 }
 399
 400 /* Restore the original state. */
 401 static void r300_end_stencil_ref_fallback(struct r300_context *r300)
 402 {
 403     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 404     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 405     CS_LOCALS(r300);
 406
 407     BEGIN_CS(4);
 408     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode);
 409     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 410                dsa->stencil_ref_mask | r300->stencil_ref.ref_value[0]);
 411     END_CS;
 412 }
 413
 414 void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 415                                      unsigned mode,
 416                                      unsigned start,
 417                                      unsigned count)
 418 {
 419     if (!r300->stencil_ref_bf_fallback) {
 420         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 421     } else {
 422         r300_begin_stencil_ref_fallback(r300);
 423         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 424         r300_switch_stencil_ref_side(r300);
 425         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 426         r300_end_stencil_ref_fallback(r300);
 427     }
 428 }
 429
 430 void r300_emit_draw_arrays(struct r300_context *r300,
 431                            unsigned mode,
 432                            unsigned count)
 433 {
 434     if (!r300->stencil_ref_bf_fallback) {
 435         r500_emit_draw_arrays(r300, mode, count);
 436     } else {
 437         r300_begin_stencil_ref_fallback(r300);
 438         r500_emit_draw_arrays(r300, mode, count);
 439         r300_switch_stencil_ref_side(r300);
 440         r500_emit_draw_arrays(r300, mode, count);
 441         r300_end_stencil_ref_fallback(r300);
 442     }
 443 }
 444
 445 void r300_emit_draw_elements(struct r300_context *r300,
 446                              struct pipe_resource* indexBuffer,
 447                              unsigned indexSize,
 448                              int indexBias,
 449                              unsigned minIndex,
 450                              unsigned maxIndex,
 451                              unsigned mode,
 452                              unsigned start,
 453                              unsigned count)
 454 {
 455     if (!r300->stencil_ref_bf_fallback) {
 456         r500_emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 457                                 minIndex, maxIndex, mode, start, count);
 458     } else {
 459         r300_begin_stencil_ref_fallback(r300);
 460         r500_emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 461                                 minIndex, maxIndex, mode, start, count);
 462         r300_switch_stencil_ref_side(r300);
 463         r500_emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 464                                 minIndex, maxIndex, mode, start, count);
 465         r300_end_stencil_ref_fallback(r300);
 466     }
 467 }
 468
 469 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 470                                     struct pipe_resource** elts,
 471                                     unsigned start,
 472                                     unsigned count)
 473 {
 474     struct pipe_context* context = &r300->context;
 475     struct pipe_screen* screen = r300->context.screen;
 476     struct pipe_resource* new_elts;
 477     unsigned char *in_map;
 478     unsigned short *out_map;
 479     struct pipe_transfer *src_transfer, *dst_transfer;
 480     unsigned i;
 481
 482     new_elts = pipe_buffer_create(screen,
 483                                   PIPE_BIND_INDEX_BUFFER,
 484                                   2 * count);
 485
 486     in_map = pipe_buffer_map(context, *elts, PIPE_TRANSFER_READ, &src_transfer);
 487     out_map = pipe_buffer_map(context, new_elts, PIPE_TRANSFER_WRITE, &dst_transfer);
 488
 489     in_map += start;
 490
 491     for (i = 0; i < count; i++) {
 492         *out_map = (unsigned short)*in_map;
 493         in_map++;
 494         out_map++;
 495     }
 496
 497     pipe_buffer_unmap(context, *elts, src_transfer);
 498     pipe_buffer_unmap(context, new_elts, dst_transfer);
 499
 500     *elts = new_elts;
 501 }
 502
 503 static void r300_align_ushort_elts(struct r300_context *r300,
 504                                    struct pipe_resource **elts,
 505                                    unsigned start, unsigned count)
 506 {
 507     struct pipe_context* context = &r300->context;
 508     struct pipe_transfer *in_transfer = NULL;
 509     struct pipe_transfer *out_transfer = NULL;
 510     struct pipe_resource* new_elts;
 511     unsigned short *in_map;
 512     unsigned short *out_map;
 513
 514     new_elts = pipe_buffer_create(context->screen,
 515                                   PIPE_BIND_INDEX_BUFFER,
 516                                   2 * count);
 517
 518     in_map = pipe_buffer_map(context, *elts,
 519                              PIPE_TRANSFER_READ, &in_transfer);
 520     out_map = pipe_buffer_map(context, new_elts,
 521                               PIPE_TRANSFER_WRITE, &out_transfer);
 522
 523     memcpy(out_map, in_map+start, 2 * count);
 524
 525     pipe_buffer_unmap(context, *elts, in_transfer);
 526     pipe_buffer_unmap(context, new_elts, out_transfer);
 527
 528     *elts = new_elts;
 529 }
 530
 531 /* This is the fast-path drawing & emission for HW TCL. */
 532 void r300_draw_range_elements(struct pipe_context* pipe,
 533                               struct pipe_resource* indexBuffer,
 534                               unsigned indexSize,
 535                               int indexBias,
 536                               unsigned minIndex,
 537                               unsigned maxIndex,
 538                               unsigned mode,
 539                               unsigned start,
 540                               unsigned count)
 541 {
 542     struct r300_context* r300 = r300_context(pipe);
 543     struct pipe_resource* orgIndexBuffer = indexBuffer;
 544 #if defined(ENABLE_ALT_NUM_VERTS)
 545     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 546                             count > 65536;
 547 #else
 548     boolean alt_num_verts = FALSE;
 549 #endif
 550     unsigned short_count;
 551
 552     if (r300->skip_rendering) {
 553         return;
 554     }
 555
 556     if (!u_trim_pipe_prim(mode, &count)) {
 557         return;
 558     }
 559
 560     if (indexSize == 1) {
 561         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 562         indexSize = 2;
 563         start = 0;
 564     } else if (indexSize == 2 && start % 2 != 0) {
 565         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 566         start = 0;
 567     }
 568
 569     r300_update_derived_state(r300);
 570
 571     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 572
 573     /* 128 dwords for emit_aos and emit_draw_elements */
 574     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 575     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 576     r300_emit_dirty_state(r300);
 577     r300_emit_aos(r300, 0);
 578
 579     u_upload_flush(r300->upload_vb);
 580     u_upload_flush(r300->upload_ib);
 581     if (alt_num_verts || count <= 65535) {
 582         r300->emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 583                                  minIndex, maxIndex, mode, start, count);
 584     } else {
 585         do {
 586             short_count = MIN2(count, 65534);
 587             r300->emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 588                                      minIndex, maxIndex,
 589                                      mode, start, short_count);
 590
 591             start += short_count;
 592             count -= short_count;
 593
 594             /* 16 spare dwords are enough for emit_draw_elements. */
 595             if (count && r300_reserve_cs_space(r300, 16)) {
 596                 r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 597                 r300_emit_dirty_state(r300);
 598                 r300_emit_aos(r300, 0);
 599             }
 600         } while (count);
 601     }
 602
 603     if (indexBuffer != orgIndexBuffer) {
 604         pipe_resource_reference( &indexBuffer, NULL );
 605     }
 606 }
 607
 608 /* Simple helpers for context setup. Should probably be moved to util. */
 609 void r300_draw_elements(struct pipe_context* pipe,
 610                         struct pipe_resource* indexBuffer,
 611                         unsigned indexSize, int indexBias, unsigned mode,
 612                         unsigned start, unsigned count)
 613 {
 614     struct r300_context *r300 = r300_context(pipe);
 615
 616     pipe->draw_range_elements(pipe, indexBuffer, indexSize, indexBias,
 617                               0, r300->vertex_buffer_max_index,
 618                               mode, start, count);
 619 }
 620
 621 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 622                       unsigned start, unsigned count)
 623 {
 624     struct r300_context* r300 = r300_context(pipe);
 625 #if defined(ENABLE_ALT_NUM_VERTS)
 626     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 627                             count > 65536;
 628 #else
 629     boolean alt_num_verts = FALSE;
 630 #endif
 631     unsigned short_count;
 632
 633     if (r300->skip_rendering) {
 634         return;
 635     }
 636
 637     if (!u_trim_pipe_prim(mode, &count)) {
 638         return;
 639     }
 640
 641     r300_update_derived_state(r300);
 642
 643     if (immd_is_good_idea(r300, count)) {
 644         r300->emit_draw_arrays_immediate(r300, mode, start, count);
 645     } else {
 646         /* Make sure there are at least 128 spare dwords in the command buffer.
 647          * (most of it being consumed by emit_aos) */
 648         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 649         r300_emit_buffer_validate(r300, TRUE, NULL);
 650         r300_emit_dirty_state(r300);
 651
 652         if (alt_num_verts || count <= 65535) {
 653             r300_emit_aos(r300, start);
 654             r300->emit_draw_arrays(r300, mode, count);
 655         } else {
 656             do {
 657                 short_count = MIN2(count, 65535);
 658                 r300_emit_aos(r300, start);
 659                 r300->emit_draw_arrays(r300, mode, short_count);
 660
 661                 start += short_count;
 662                 count -= short_count;
 663
 664                 /* Again, we emit both AOS and draw_arrays so there should be
 665                  * at least 128 spare dwords. */
 666                 if (count && r300_reserve_cs_space(r300, 128)) {
 667                     r300_emit_buffer_validate(r300, TRUE, NULL);
 668                     r300_emit_dirty_state(r300);
 669                 }
 670             } while (count);
 671         }
 672         u_upload_flush(r300->upload_vb);
 673     }
 674 }
 675
 676 /****************************************************************************
 677  * The rest of this file is for SW TCL rendering only. Please be polite and *
 678  * keep these functions separated so that they are easier to locate. ~C.    *
 679  ***************************************************************************/
 680
 681 /* SW TCL arrays, using Draw. */
 682 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 683                                unsigned mode,
 684                                unsigned start,
 685                                unsigned count)
 686 {
 687     struct r300_context* r300 = r300_context(pipe);
 688     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 689     int i;
 690
 691     if (r300->skip_rendering) {
 692         return;
 693     }
 694
 695     if (!u_trim_pipe_prim(mode, &count)) {
 696         return;
 697     }
 698
 699     for (i = 0; i < r300->vertex_buffer_count; i++) {
 700         void* buf = pipe_buffer_map(pipe,
 701                                     r300->vertex_buffer[i].buffer,
 702                                     PIPE_TRANSFER_READ,
 703                                     &vb_transfer[i]);
 704         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 705     }
 706
 707     draw_set_mapped_element_buffer(r300->draw, 0, 0, NULL);
 708
 709     draw_arrays(r300->draw, mode, start, count);
 710
 711     for (i = 0; i < r300->vertex_buffer_count; i++) {
 712         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 713                           vb_transfer[i]);
 714         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 715     }
 716 }
 717
 718 /* SW TCL elements, using Draw. */
 719 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 720                                        struct pipe_resource* indexBuffer,
 721                                        unsigned indexSize,
 722                                        int indexBias,
 723                                        unsigned minIndex,
 724                                        unsigned maxIndex,
 725                                        unsigned mode,
 726                                        unsigned start,
 727                                        unsigned count)
 728 {
 729     struct r300_context* r300 = r300_context(pipe);
 730     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 731     struct pipe_transfer *ib_transfer;
 732     int i;
 733     void* indices;
 734
 735     if (r300->skip_rendering) {
 736         return;
 737     }
 738
 739     if (!u_trim_pipe_prim(mode, &count)) {
 740         return;
 741     }
 742
 743     for (i = 0; i < r300->vertex_buffer_count; i++) {
 744         void* buf = pipe_buffer_map(pipe,
 745                                     r300->vertex_buffer[i].buffer,
 746                                     PIPE_TRANSFER_READ,
 747                                     &vb_transfer[i]);
 748         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 749     }
 750
 751     indices = pipe_buffer_map(pipe, indexBuffer,
 752                               PIPE_TRANSFER_READ, &ib_transfer);
 753     draw_set_mapped_element_buffer_range(r300->draw, indexSize, indexBias,
 754                                          minIndex, maxIndex, indices);
 755
 756     draw_arrays(r300->draw, mode, start, count);
 757
 758     for (i = 0; i < r300->vertex_buffer_count; i++) {
 759         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 760                           vb_transfer[i]);
 761         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 762     }
 763
 764     pipe_buffer_unmap(pipe, indexBuffer,
 765                       ib_transfer);
 766     draw_set_mapped_element_buffer_range(r300->draw, 0, 0,
 767                                          start, start + count - 1,
 768                                          NULL);
 769 }
 770
 771 /* Object for rendering using Draw. */
 772 struct r300_render {
 773     /* Parent class */
 774     struct vbuf_render base;
 775
 776     /* Pipe context */
 777     struct r300_context* r300;
 778
 779     /* Vertex information */
 780     size_t vertex_size;
 781     unsigned prim;
 782     unsigned hwprim;
 783
 784     /* VBO */
 785     struct pipe_resource* vbo;
 786     size_t vbo_size;
 787     size_t vbo_offset;
 788     size_t vbo_max_used;
 789     void * vbo_ptr;
 790
 791    struct pipe_transfer *vbo_transfer;
 792 };
 793
 794 static INLINE struct r300_render*
 795 r300_render(struct vbuf_render* render)
 796 {
 797     return (struct r300_render*)render;
 798 }
 799
 800 static const struct vertex_info*
 801 r300_render_get_vertex_info(struct vbuf_render* render)
 802 {
 803     struct r300_render* r300render = r300_render(render);
 804     struct r300_context* r300 = r300render->r300;
 805
 806     r300_update_derived_state(r300);
 807
 808     return &r300->vertex_info;
 809 }
 810
 811 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 812                                                    ushort vertex_size,
 813                                                    ushort count)
 814 {
 815     struct r300_render* r300render = r300_render(render);
 816     struct r300_context* r300 = r300render->r300;
 817     struct pipe_screen* screen = r300->context.screen;
 818     size_t size = (size_t)vertex_size * (size_t)count;
 819
 820     if (size + r300render->vbo_offset > r300render->vbo_size)
 821     {
 822         pipe_resource_reference(&r300->vbo, NULL);
 823         r300render->vbo = pipe_buffer_create(screen,
 824                                              PIPE_BIND_VERTEX_BUFFER,
 825                                              R300_MAX_DRAW_VBO_SIZE);
 826         r300render->vbo_offset = 0;
 827         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 828     }
 829
 830     r300render->vertex_size = vertex_size;
 831     r300->vbo = r300render->vbo;
 832     r300->vbo_offset = r300render->vbo_offset;
 833
 834     return (r300render->vbo) ? TRUE : FALSE;
 835 }
 836
 837 static void* r300_render_map_vertices(struct vbuf_render* render)
 838 {
 839     struct r300_render* r300render = r300_render(render);
 840
 841     r300render->vbo_ptr = pipe_buffer_map(&r300render->r300->context,
 842                                           r300render->vbo,
 843                                           PIPE_TRANSFER_WRITE,
 844                                           &r300render->vbo_transfer);
 845
 846     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 847 }
 848
 849 static void r300_render_unmap_vertices(struct vbuf_render* render,
 850                                              ushort min,
 851                                              ushort max)
 852 {
 853     struct r300_render* r300render = r300_render(render);
 854     struct pipe_context* context = &r300render->r300->context;
 855     CS_LOCALS(r300render->r300);
 856     BEGIN_CS(2);
 857     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 858     END_CS;
 859
 860     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 861                                     r300render->vertex_size * (max + 1));
 862     pipe_buffer_unmap(context, r300render->vbo, r300render->vbo_transfer);
 863 }
 864
 865 static void r300_render_release_vertices(struct vbuf_render* render)
 866 {
 867     struct r300_render* r300render = r300_render(render);
 868
 869     r300render->vbo_offset += r300render->vbo_max_used;
 870     r300render->vbo_max_used = 0;
 871 }
 872
 873 static boolean r300_render_set_primitive(struct vbuf_render* render,
 874                                                unsigned prim)
 875 {
 876     struct r300_render* r300render = r300_render(render);
 877
 878     r300render->prim = prim;
 879     r300render->hwprim = r300_translate_primitive(prim);
 880
 881     return TRUE;
 882 }
 883
 884 static void r500_render_draw_arrays(struct vbuf_render* render,
 885                                     unsigned start,
 886                                     unsigned count)
 887 {
 888     struct r300_render* r300render = r300_render(render);
 889     struct r300_context* r300 = r300render->r300;
 890
 891     CS_LOCALS(r300);
 892
 893     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
 894     r300_emit_buffer_validate(r300, FALSE, NULL);
 895     r300_emit_dirty_state(r300);
 896
 897     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 898
 899     BEGIN_CS(2);
 900     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 901     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 902            r300render->hwprim);
 903     END_CS;
 904 }
 905
 906 static void r500_render_draw(struct vbuf_render* render,
 907                              const ushort* indices,
 908                              uint count)
 909 {
 910     struct r300_render* r300render = r300_render(render);
 911     struct r300_context* r300 = r300render->r300;
 912     int i;
 913     unsigned dwords = 2 + (count+1)/2;
 914
 915     CS_LOCALS(r300);
 916
 917     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 918     r300_emit_buffer_validate(r300, FALSE, NULL);
 919     r300_emit_dirty_state(r300);
 920
 921     BEGIN_CS(dwords);
 922     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 923     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 924            r300render->hwprim);
 925     for (i = 0; i < count-1; i += 2) {
 926         OUT_CS(indices[i+1] << 16 | indices[i]);
 927     }
 928     if (count % 2) {
 929         OUT_CS(indices[count-1]);
 930     }
 931     END_CS;
 932 }
 933
 934 static void r300_render_draw_arrays(struct vbuf_render* render,
 935                                     unsigned start,
 936                                     unsigned count)
 937 {
 938     struct r300_context* r300 = r300_render(render)->r300;
 939
 940     if (!r300->stencil_ref_bf_fallback) {
 941         r500_render_draw_arrays(render, start, count);
 942     } else {
 943         r300_begin_stencil_ref_fallback(r300);
 944         r500_render_draw_arrays(render, start, count);
 945         r300_switch_stencil_ref_side(r300);
 946         r500_render_draw_arrays(render, start, count);
 947         r300_end_stencil_ref_fallback(r300);
 948     }
 949 }
 950
 951 static void r300_render_draw(struct vbuf_render* render,
 952                              const ushort* indices,
 953                              uint count)
 954 {
 955     struct r300_context* r300 = r300_render(render)->r300;
 956
 957     if (!r300->stencil_ref_bf_fallback) {
 958         r500_render_draw(render, indices, count);
 959     } else {
 960         r300_begin_stencil_ref_fallback(r300);
 961         r500_render_draw(render, indices, count);
 962         r300_switch_stencil_ref_side(r300);
 963         r500_render_draw(render, indices, count);
 964         r300_end_stencil_ref_fallback(r300);
 965     }
 966 }
 967
 968 static void r300_render_destroy(struct vbuf_render* render)
 969 {
 970     FREE(render);
 971 }
 972
 973 static struct vbuf_render* r300_render_create(struct r300_context* r300)
 974 {
 975     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
 976
 977     r300render->r300 = r300;
 978
 979     /* XXX find real numbers plz */
 980     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
 981     r300render->base.max_indices = 16 * 1024;
 982
 983     r300render->base.get_vertex_info = r300_render_get_vertex_info;
 984     r300render->base.allocate_vertices = r300_render_allocate_vertices;
 985     r300render->base.map_vertices = r300_render_map_vertices;
 986     r300render->base.unmap_vertices = r300_render_unmap_vertices;
 987     r300render->base.set_primitive = r300_render_set_primitive;
 988     if (r300->screen->caps.is_r500) {
 989         r300render->base.draw = r500_render_draw;
 990         r300render->base.draw_arrays = r500_render_draw_arrays;
 991     } else {
 992         r300render->base.draw = r300_render_draw;
 993         r300render->base.draw_arrays = r300_render_draw_arrays;
 994     }
 995     r300render->base.release_vertices = r300_render_release_vertices;
 996     r300render->base.destroy = r300_render_destroy;
 997
 998     r300render->vbo = NULL;
 999     r300render->vbo_size = 0;
1000     r300render->vbo_offset = 0;
1001
1002     return &r300render->base;
1003 }
1004
1005 struct draw_stage* r300_draw_stage(struct r300_context* r300)
1006 {
1007     struct vbuf_render* render;
1008     struct draw_stage* stage;
1009
1010     render = r300_render_create(r300);
1011
1012     if (!render) {
1013         return NULL;
1014     }
1015
1016     stage = draw_vbuf_stage(r300->draw, render);
1017
1018     if (!stage) {
1019         render->destroy(render);
1020         return NULL;
1021     }
1022
1023     draw_set_render(r300->draw, render);
1024
1025     return stage;
1026 }