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 /*****************************************************************************
 166  * The emission of draw packets for r500. Older GPUs may use these functions *
 167  * after resolving fallback issues (e.g. stencil ref two-sided).             *
 168  ****************************************************************************/
 169
 170 void r500_emit_draw_arrays_immediate(struct r300_context *r300,
 171                                      unsigned mode,
 172                                      unsigned start,
 173                                      unsigned count)
 174 {
 175     struct pipe_vertex_element* velem;
 176     struct pipe_vertex_buffer* vbuf;
 177     unsigned vertex_element_count = r300->velems->count;
 178     unsigned i, v, vbi, dw, elem_offset, dwords;
 179
 180     /* Size of the vertex, in dwords. */
 181     unsigned vertex_size = 0;
 182
 183     /* Offsets of the attribute, in dwords, from the start of the vertex. */
 184     unsigned offset[PIPE_MAX_ATTRIBS];
 185
 186     /* Size of the vertex element, in dwords. */
 187     unsigned size[PIPE_MAX_ATTRIBS];
 188
 189     /* Stride to the same attrib in the next vertex in the vertex buffer,
 190      * in dwords. */
 191     unsigned stride[PIPE_MAX_ATTRIBS] = {0};
 192
 193     /* Mapped vertex buffers. */
 194     uint32_t* map[PIPE_MAX_ATTRIBS] = {0};
 195
 196     CS_LOCALS(r300);
 197
 198     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
 199     for (i = 0; i < vertex_element_count; i++) {
 200         velem = &r300->velems->velem[i];
 201         offset[i] = velem->src_offset / 4;
 202         size[i] = util_format_get_blocksize(velem->src_format) / 4;
 203         vertex_size += size[i];
 204         vbi = velem->vertex_buffer_index;
 205
 206         /* Map the buffer. */
 207         if (!map[vbi]) {
 208             vbuf = &r300->vertex_buffer[vbi];
 209             map[vbi] = (uint32_t*)pipe_buffer_map(r300->context.screen,
 210                                                   vbuf->buffer,
 211                                                   PIPE_BUFFER_USAGE_CPU_READ);
 212             map[vbi] += vbuf->buffer_offset / 4;
 213             stride[vbi] = vbuf->stride / 4;
 214         }
 215     }
 216
 217     dwords = 9 + count * vertex_size;
 218
 219     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 220     r300_emit_buffer_validate(r300, FALSE, NULL);
 221     r300_emit_dirty_state(r300);
 222
 223     BEGIN_CS(dwords);
 224     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 225             r300_provoking_vertex_fixes(r300, mode));
 226     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
 227     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 228     OUT_CS(count - 1);
 229     OUT_CS(0);
 230     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
 231     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
 232             r300_translate_primitive(mode));
 233
 234     /* Emit vertices. */
 235     for (v = 0; v < count; v++) {
 236         for (i = 0; i < vertex_element_count; i++) {
 237             velem = &r300->velems->velem[i];
 238             vbi = velem->vertex_buffer_index;
 239             elem_offset = offset[i] + stride[vbi] * (v + start);
 240
 241             for (dw = 0; dw < size[i]; dw++) {
 242                 OUT_CS(map[vbi][elem_offset + dw]);
 243             }
 244         }
 245     }
 246     END_CS;
 247
 248     /* Unmap buffers. */
 249     for (i = 0; i < vertex_element_count; i++) {
 250         vbi = r300->velems->velem[i].vertex_buffer_index;
 251
 252         if (map[vbi]) {
 253             vbuf = &r300->vertex_buffer[vbi];
 254             pipe_buffer_unmap(r300->context.screen, vbuf->buffer);
 255             map[vbi] = NULL;
 256         }
 257     }
 258 }
 259
 260 void r500_emit_draw_arrays(struct r300_context *r300,
 261                            unsigned mode,
 262                            unsigned count)
 263 {
 264 #if defined(ENABLE_ALT_NUM_VERTS)
 265     boolean alt_num_verts = count > 65535;
 266 #else
 267     boolean alt_num_verts = FALSE;
 268 #endif
 269     CS_LOCALS(r300);
 270
 271     if (alt_num_verts) {
 272         assert(count < (1 << 24));
 273         BEGIN_CS(9);
 274         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 275     } else {
 276         BEGIN_CS(7);
 277     }
 278     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 279             r300_provoking_vertex_fixes(r300, mode));
 280     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 281     OUT_CS(count - 1);
 282     OUT_CS(0);
 283     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 284     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 285            r300_translate_primitive(mode) |
 286            (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 287     END_CS;
 288 }
 289
 290 void r500_emit_draw_elements(struct r300_context *r300,
 291                              struct pipe_buffer* indexBuffer,
 292                              unsigned indexSize,
 293                              unsigned minIndex,
 294                              unsigned maxIndex,
 295                              unsigned mode,
 296                              unsigned start,
 297                              unsigned count)
 298 {
 299     uint32_t count_dwords;
 300     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 301 #if defined(ENABLE_ALT_NUM_VERTS)
 302     boolean alt_num_verts = count > 65535;
 303 #else
 304     boolean alt_num_verts = FALSE;
 305 #endif
 306     CS_LOCALS(r300);
 307
 308     assert(count < (1 << 24));
 309
 310     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 311
 312     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 313         count, minIndex, maxIndex);
 314
 315     if (alt_num_verts) {
 316         BEGIN_CS(15);
 317         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 318     } else {
 319         BEGIN_CS(13);
 320     }
 321     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 322             r300_provoking_vertex_fixes(r300, mode));
 323     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 324     OUT_CS(maxIndex);
 325     OUT_CS(minIndex);
 326     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 327     if (indexSize == 4) {
 328         count_dwords = count;
 329         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 330                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 331                r300_translate_primitive(mode) |
 332                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 333     } else {
 334         count_dwords = (count + 1) / 2;
 335         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 336                r300_translate_primitive(mode) |
 337                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 338     }
 339
 340     /* INDX_BUFFER is a truly special packet3.
 341      * Unlike most other packet3, where the offset is after the count,
 342      * the order is reversed, so the relocation ends up carrying the
 343      * size of the indexbuf instead of the offset.
 344      */
 345     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 346     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 347            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 348     OUT_CS(offset_dwords << 2);
 349     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 350                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 351
 352     END_CS;
 353 }
 354
 355 /*****************************************************************************
 356  * The emission of draw packets for r300 which take care of the two-sided    *
 357  * stencil ref fallback and call r500's functions.                           *
 358  ****************************************************************************/
 359
 360 /* Set drawing for front faces. */
 361 static void r300_begin_stencil_ref_fallback(struct r300_context *r300)
 362 {
 363     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 364     CS_LOCALS(r300);
 365
 366     BEGIN_CS(2);
 367     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_BACK);
 368     END_CS;
 369 }
 370
 371 /* Set drawing for back faces. */
 372 static void r300_switch_stencil_ref_side(struct r300_context *r300)
 373 {
 374     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 375     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 376     CS_LOCALS(r300);
 377
 378     BEGIN_CS(4);
 379     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_FRONT);
 380     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 381                dsa->stencil_ref_bf | r300->stencil_ref.ref_value[1]);
 382     END_CS;
 383 }
 384
 385 /* Restore the original state. */
 386 static void r300_end_stencil_ref_fallback(struct r300_context *r300)
 387 {
 388     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 389     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 390     CS_LOCALS(r300);
 391
 392     BEGIN_CS(4);
 393     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode);
 394     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 395                dsa->stencil_ref_mask | r300->stencil_ref.ref_value[0]);
 396     END_CS;
 397 }
 398
 399 void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 400                                      unsigned mode,
 401                                      unsigned start,
 402                                      unsigned count)
 403 {
 404     if (!r300->stencil_ref_bf_fallback) {
 405         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 406     } else {
 407         r300_begin_stencil_ref_fallback(r300);
 408         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 409         r300_switch_stencil_ref_side(r300);
 410         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 411         r300_end_stencil_ref_fallback(r300);
 412     }
 413 }
 414
 415 void r300_emit_draw_arrays(struct r300_context *r300,
 416                            unsigned mode,
 417                            unsigned count)
 418 {
 419     if (!r300->stencil_ref_bf_fallback) {
 420         r500_emit_draw_arrays(r300, mode, count);
 421     } else {
 422         r300_begin_stencil_ref_fallback(r300);
 423         r500_emit_draw_arrays(r300, mode, count);
 424         r300_switch_stencil_ref_side(r300);
 425         r500_emit_draw_arrays(r300, mode, count);
 426         r300_end_stencil_ref_fallback(r300);
 427     }
 428 }
 429
 430 void r300_emit_draw_elements(struct r300_context *r300,
 431                              struct pipe_buffer* indexBuffer,
 432                              unsigned indexSize,
 433                              unsigned minIndex,
 434                              unsigned maxIndex,
 435                              unsigned mode,
 436                              unsigned start,
 437                              unsigned count)
 438 {
 439     if (!r300->stencil_ref_bf_fallback) {
 440         r500_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 441                                 maxIndex, mode, start, count);
 442     } else {
 443         r300_begin_stencil_ref_fallback(r300);
 444         r500_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 445                                 maxIndex, mode, start, count);
 446         r300_switch_stencil_ref_side(r300);
 447         r500_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 448                                 maxIndex, mode, start, count);
 449         r300_end_stencil_ref_fallback(r300);
 450     }
 451 }
 452
 453 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 454                                     struct pipe_buffer** elts,
 455                                     unsigned start,
 456                                     unsigned count)
 457 {
 458     struct pipe_screen* screen = r300->context.screen;
 459     struct pipe_buffer* new_elts;
 460     unsigned char *in_map;
 461     unsigned short *out_map;
 462     unsigned i;
 463
 464     new_elts = screen->buffer_create(screen, 32,
 465                                      PIPE_BUFFER_USAGE_INDEX |
 466                                      PIPE_BUFFER_USAGE_CPU_WRITE |
 467                                      PIPE_BUFFER_USAGE_GPU_READ,
 468                                      2 * count);
 469
 470     in_map = pipe_buffer_map(screen, *elts, PIPE_BUFFER_USAGE_CPU_READ);
 471     out_map = pipe_buffer_map(screen, new_elts, PIPE_BUFFER_USAGE_CPU_WRITE);
 472
 473     in_map += start;
 474
 475     for (i = 0; i < count; i++) {
 476         *out_map = (unsigned short)*in_map;
 477         in_map++;
 478         out_map++;
 479     }
 480
 481     pipe_buffer_unmap(screen, *elts);
 482     pipe_buffer_unmap(screen, new_elts);
 483
 484     *elts = new_elts;
 485 }
 486
 487 static void r300_align_ushort_elts(struct r300_context *r300,
 488                                    struct pipe_buffer **elts,
 489                                    unsigned start, unsigned count)
 490 {
 491     struct pipe_screen* screen = r300->context.screen;
 492     struct pipe_buffer* new_elts;
 493     unsigned short *in_map;
 494     unsigned short *out_map;
 495
 496     new_elts = screen->buffer_create(screen, 32,
 497                                      PIPE_BUFFER_USAGE_INDEX |
 498                                      PIPE_BUFFER_USAGE_CPU_WRITE |
 499                                      PIPE_BUFFER_USAGE_GPU_READ,
 500                                      2 * count);
 501
 502     in_map = pipe_buffer_map(screen, *elts, PIPE_BUFFER_USAGE_CPU_READ);
 503     out_map = pipe_buffer_map(screen, new_elts, PIPE_BUFFER_USAGE_CPU_WRITE);
 504
 505     memcpy(out_map, in_map+start, 2 * count);
 506
 507     pipe_buffer_unmap(screen, *elts);
 508     pipe_buffer_unmap(screen, new_elts);
 509
 510     *elts = new_elts;
 511 }
 512
 513 /* This is the fast-path drawing & emission for HW TCL. */
 514 void r300_draw_range_elements(struct pipe_context* pipe,
 515                               struct pipe_buffer* indexBuffer,
 516                               unsigned indexSize,
 517                               unsigned minIndex,
 518                               unsigned maxIndex,
 519                               unsigned mode,
 520                               unsigned start,
 521                               unsigned count)
 522 {
 523     struct r300_context* r300 = r300_context(pipe);
 524     struct pipe_buffer* orgIndexBuffer = indexBuffer;
 525 #if defined(ENABLE_ALT_NUM_VERTS)
 526     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 527                             count > 65536;
 528 #else
 529     boolean alt_num_verts = FALSE;
 530 #endif
 531     unsigned short_count;
 532
 533     if (r300->skip_rendering) {
 534         return;
 535     }
 536
 537     if (!u_trim_pipe_prim(mode, &count)) {
 538         return;
 539     }
 540
 541     if (indexSize == 1) {
 542         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 543         indexSize = 2;
 544         start = 0;
 545     } else if (indexSize == 2 && start % 2 != 0) {
 546         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 547         start = 0;
 548     }
 549
 550     r300_update_derived_state(r300);
 551
 552     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 553
 554     /* 128 dwords for emit_aos and emit_draw_elements */
 555     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 556     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 557     r300_emit_dirty_state(r300);
 558     r300_emit_aos(r300, 0);
 559
 560     u_upload_flush(r300->upload_vb);
 561     u_upload_flush(r300->upload_ib);
 562     if (alt_num_verts || count <= 65535) {
 563         r300->emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 564                                  maxIndex, mode, start, count);
 565     } else {
 566         do {
 567             short_count = MIN2(count, 65534);
 568             r300->emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
 569                                       maxIndex, mode, start, short_count);
 570
 571             start += short_count;
 572             count -= short_count;
 573
 574             /* 16 spare dwords are enough for emit_draw_elements. */
 575             if (count && r300_reserve_cs_space(r300, 16)) {
 576                 r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 577                 r300_emit_dirty_state(r300);
 578                 r300_emit_aos(r300, 0);
 579             }
 580         } while (count);
 581     }
 582
 583     if (indexBuffer != orgIndexBuffer) {
 584         pipe_buffer_reference( &indexBuffer, NULL );
 585     }
 586 }
 587
 588 /* Simple helpers for context setup. Should probably be moved to util. */
 589 void r300_draw_elements(struct pipe_context* pipe,
 590                         struct pipe_buffer* indexBuffer,
 591                         unsigned indexSize, unsigned mode,
 592                         unsigned start, unsigned count)
 593 {
 594     struct r300_context *r300 = r300_context(pipe);
 595
 596     pipe->draw_range_elements(pipe, indexBuffer, indexSize, 0,
 597                               r300->vertex_buffer_max_index,
 598                               mode, start, count);
 599 }
 600
 601 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 602                       unsigned start, unsigned count)
 603 {
 604     struct r300_context* r300 = r300_context(pipe);
 605 #if defined(ENABLE_ALT_NUM_VERTS)
 606     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 607                             count > 65536;
 608 #else
 609     boolean alt_num_verts = FALSE;
 610 #endif
 611     unsigned short_count;
 612
 613     if (r300->skip_rendering) {
 614         return;
 615     }
 616
 617     if (!u_trim_pipe_prim(mode, &count)) {
 618         return;
 619     }
 620
 621     r300_update_derived_state(r300);
 622
 623     if (immd_is_good_idea(r300, count)) {
 624         r300->emit_draw_arrays_immediate(r300, mode, start, count);
 625     } else {
 626         /* Make sure there are at least 128 spare dwords in the command buffer.
 627          * (most of it being consumed by emit_aos) */
 628         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 629         r300_emit_buffer_validate(r300, TRUE, NULL);
 630         r300_emit_dirty_state(r300);
 631
 632         if (alt_num_verts || count <= 65535) {
 633             r300_emit_aos(r300, start);
 634             r300->emit_draw_arrays(r300, mode, count);
 635         } else {
 636             do {
 637                 short_count = MIN2(count, 65535);
 638                 r300_emit_aos(r300, start);
 639                 r300->emit_draw_arrays(r300, mode, short_count);
 640
 641                 start += short_count;
 642                 count -= short_count;
 643
 644                 /* Again, we emit both AOS and draw_arrays so there should be
 645                  * at least 128 spare dwords. */
 646                 if (count && r300_reserve_cs_space(r300, 128)) {
 647                     r300_emit_buffer_validate(r300, TRUE, NULL);
 648                     r300_emit_dirty_state(r300);
 649                 }
 650             } while (count);
 651         }
 652         u_upload_flush(r300->upload_vb);
 653     }
 654 }
 655
 656 /****************************************************************************
 657  * The rest of this file is for SW TCL rendering only. Please be polite and *
 658  * keep these functions separated so that they are easier to locate. ~C.    *
 659  ***************************************************************************/
 660
 661 /* SW TCL arrays, using Draw. */
 662 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 663                                unsigned mode,
 664                                unsigned start,
 665                                unsigned count)
 666 {
 667     struct r300_context* r300 = r300_context(pipe);
 668     int i;
 669
 670     if (r300->skip_rendering) {
 671         return;
 672     }
 673
 674     if (!u_trim_pipe_prim(mode, &count)) {
 675         return;
 676     }
 677
 678     for (i = 0; i < r300->vertex_buffer_count; i++) {
 679         void* buf = pipe_buffer_map(pipe->screen,
 680                                     r300->vertex_buffer[i].buffer,
 681                                     PIPE_BUFFER_USAGE_CPU_READ);
 682         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 683     }
 684
 685     draw_set_mapped_element_buffer(r300->draw, 0, NULL);
 686
 687     draw_arrays(r300->draw, mode, start, count);
 688
 689     for (i = 0; i < r300->vertex_buffer_count; i++) {
 690         pipe_buffer_unmap(pipe->screen, r300->vertex_buffer[i].buffer);
 691         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 692     }
 693 }
 694
 695 /* SW TCL elements, using Draw. */
 696 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 697                                        struct pipe_buffer* indexBuffer,
 698                                        unsigned indexSize,
 699                                        unsigned minIndex,
 700                                        unsigned maxIndex,
 701                                        unsigned mode,
 702                                        unsigned start,
 703                                        unsigned count)
 704 {
 705     struct r300_context* r300 = r300_context(pipe);
 706     int i;
 707     void* indices;
 708
 709     if (r300->skip_rendering) {
 710         return;
 711     }
 712
 713     if (!u_trim_pipe_prim(mode, &count)) {
 714         return;
 715     }
 716
 717     for (i = 0; i < r300->vertex_buffer_count; i++) {
 718         void* buf = pipe_buffer_map(pipe->screen,
 719                                     r300->vertex_buffer[i].buffer,
 720                                     PIPE_BUFFER_USAGE_CPU_READ);
 721         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 722     }
 723
 724     indices = pipe_buffer_map(pipe->screen, indexBuffer,
 725                               PIPE_BUFFER_USAGE_CPU_READ);
 726     draw_set_mapped_element_buffer_range(r300->draw, indexSize,
 727                                          minIndex, maxIndex, indices);
 728
 729     draw_arrays(r300->draw, mode, start, count);
 730
 731     for (i = 0; i < r300->vertex_buffer_count; i++) {
 732         pipe_buffer_unmap(pipe->screen, r300->vertex_buffer[i].buffer);
 733         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 734     }
 735
 736     pipe_buffer_unmap(pipe->screen, indexBuffer);
 737     draw_set_mapped_element_buffer_range(r300->draw, 0, start,
 738                                          start + count - 1, NULL);
 739 }
 740
 741 /* Object for rendering using Draw. */
 742 struct r300_render {
 743     /* Parent class */
 744     struct vbuf_render base;
 745
 746     /* Pipe context */
 747     struct r300_context* r300;
 748
 749     /* Vertex information */
 750     size_t vertex_size;
 751     unsigned prim;
 752     unsigned hwprim;
 753
 754     /* VBO */
 755     struct pipe_buffer* vbo;
 756     size_t vbo_size;
 757     size_t vbo_offset;
 758     size_t vbo_max_used;
 759     void * vbo_ptr;
 760 };
 761
 762 static INLINE struct r300_render*
 763 r300_render(struct vbuf_render* render)
 764 {
 765     return (struct r300_render*)render;
 766 }
 767
 768 static const struct vertex_info*
 769 r300_render_get_vertex_info(struct vbuf_render* render)
 770 {
 771     struct r300_render* r300render = r300_render(render);
 772     struct r300_context* r300 = r300render->r300;
 773
 774     r300_update_derived_state(r300);
 775
 776     return &r300->vertex_info;
 777 }
 778
 779 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 780                                                    ushort vertex_size,
 781                                                    ushort count)
 782 {
 783     struct r300_render* r300render = r300_render(render);
 784     struct r300_context* r300 = r300render->r300;
 785     struct pipe_screen* screen = r300->context.screen;
 786     size_t size = (size_t)vertex_size * (size_t)count;
 787
 788     if (size + r300render->vbo_offset > r300render->vbo_size)
 789     {
 790         pipe_buffer_reference(&r300->vbo, NULL);
 791         r300render->vbo = pipe_buffer_create(screen,
 792                                              64,
 793                                              PIPE_BUFFER_USAGE_VERTEX,
 794                                              R300_MAX_DRAW_VBO_SIZE);
 795         r300render->vbo_offset = 0;
 796         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 797     }
 798
 799     r300render->vertex_size = vertex_size;
 800     r300->vbo = r300render->vbo;
 801     r300->vbo_offset = r300render->vbo_offset;
 802
 803     return (r300render->vbo) ? TRUE : FALSE;
 804 }
 805
 806 static void* r300_render_map_vertices(struct vbuf_render* render)
 807 {
 808     struct r300_render* r300render = r300_render(render);
 809     struct pipe_screen* screen = r300render->r300->context.screen;
 810
 811     r300render->vbo_ptr = pipe_buffer_map(screen, r300render->vbo,
 812                                           PIPE_BUFFER_USAGE_CPU_WRITE);
 813
 814     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 815 }
 816
 817 static void r300_render_unmap_vertices(struct vbuf_render* render,
 818                                              ushort min,
 819                                              ushort max)
 820 {
 821     struct r300_render* r300render = r300_render(render);
 822     struct pipe_screen* screen = r300render->r300->context.screen;
 823     CS_LOCALS(r300render->r300);
 824     BEGIN_CS(2);
 825     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 826     END_CS;
 827
 828     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 829                                     r300render->vertex_size * (max + 1));
 830     pipe_buffer_unmap(screen, r300render->vbo);
 831 }
 832
 833 static void r300_render_release_vertices(struct vbuf_render* render)
 834 {
 835     struct r300_render* r300render = r300_render(render);
 836
 837     r300render->vbo_offset += r300render->vbo_max_used;
 838     r300render->vbo_max_used = 0;
 839 }
 840
 841 static boolean r300_render_set_primitive(struct vbuf_render* render,
 842                                                unsigned prim)
 843 {
 844     struct r300_render* r300render = r300_render(render);
 845
 846     r300render->prim = prim;
 847     r300render->hwprim = r300_translate_primitive(prim);
 848
 849     return TRUE;
 850 }
 851
 852 static void r500_render_draw_arrays(struct vbuf_render* render,
 853                                     unsigned start,
 854                                     unsigned count)
 855 {
 856     struct r300_render* r300render = r300_render(render);
 857     struct r300_context* r300 = r300render->r300;
 858
 859     CS_LOCALS(r300);
 860
 861     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
 862     r300_emit_buffer_validate(r300, FALSE, NULL);
 863     r300_emit_dirty_state(r300);
 864
 865     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 866
 867     BEGIN_CS(2);
 868     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 869     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 870            r300render->hwprim);
 871     END_CS;
 872 }
 873
 874 static void r500_render_draw(struct vbuf_render* render,
 875                              const ushort* indices,
 876                              uint count)
 877 {
 878     struct r300_render* r300render = r300_render(render);
 879     struct r300_context* r300 = r300render->r300;
 880     int i;
 881     unsigned dwords = 2 + (count+1)/2;
 882
 883     CS_LOCALS(r300);
 884
 885     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 886     r300_emit_buffer_validate(r300, FALSE, NULL);
 887     r300_emit_dirty_state(r300);
 888
 889     BEGIN_CS(dwords);
 890     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 891     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 892            r300render->hwprim);
 893     for (i = 0; i < count-1; i += 2) {
 894         OUT_CS(indices[i+1] << 16 | indices[i]);
 895     }
 896     if (count % 2) {
 897         OUT_CS(indices[count-1]);
 898     }
 899     END_CS;
 900 }
 901
 902 static void r300_render_draw_arrays(struct vbuf_render* render,
 903                                     unsigned start,
 904                                     unsigned count)
 905 {
 906     struct r300_context* r300 = r300_render(render)->r300;
 907
 908     if (!r300->stencil_ref_bf_fallback) {
 909         r500_render_draw_arrays(render, start, count);
 910     } else {
 911         r300_begin_stencil_ref_fallback(r300);
 912         r500_render_draw_arrays(render, start, count);
 913         r300_switch_stencil_ref_side(r300);
 914         r500_render_draw_arrays(render, start, count);
 915         r300_end_stencil_ref_fallback(r300);
 916     }
 917 }
 918
 919 static void r300_render_draw(struct vbuf_render* render,
 920                              const ushort* indices,
 921                              uint count)
 922 {
 923     struct r300_context* r300 = r300_render(render)->r300;
 924
 925     if (!r300->stencil_ref_bf_fallback) {
 926         r500_render_draw(render, indices, count);
 927     } else {
 928         r300_begin_stencil_ref_fallback(r300);
 929         r500_render_draw(render, indices, count);
 930         r300_switch_stencil_ref_side(r300);
 931         r500_render_draw(render, indices, count);
 932         r300_end_stencil_ref_fallback(r300);
 933     }
 934 }
 935
 936 static void r300_render_destroy(struct vbuf_render* render)
 937 {
 938     FREE(render);
 939 }
 940
 941 static struct vbuf_render* r300_render_create(struct r300_context* r300)
 942 {
 943     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
 944
 945     r300render->r300 = r300;
 946
 947     /* XXX find real numbers plz */
 948     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
 949     r300render->base.max_indices = 16 * 1024;
 950
 951     r300render->base.get_vertex_info = r300_render_get_vertex_info;
 952     r300render->base.allocate_vertices = r300_render_allocate_vertices;
 953     r300render->base.map_vertices = r300_render_map_vertices;
 954     r300render->base.unmap_vertices = r300_render_unmap_vertices;
 955     r300render->base.set_primitive = r300_render_set_primitive;
 956     if (r300->screen->caps.is_r500) {
 957         r300render->base.draw = r500_render_draw;
 958         r300render->base.draw_arrays = r500_render_draw_arrays;
 959     } else {
 960         r300render->base.draw = r300_render_draw;
 961         r300render->base.draw_arrays = r300_render_draw_arrays;
 962     }
 963     r300render->base.release_vertices = r300_render_release_vertices;
 964     r300render->base.destroy = r300_render_destroy;
 965
 966     r300render->vbo = NULL;
 967     r300render->vbo_size = 0;
 968     r300render->vbo_offset = 0;
 969
 970     return &r300render->base;
 971 }
 972
 973 struct draw_stage* r300_draw_stage(struct r300_context* r300)
 974 {
 975     struct vbuf_render* render;
 976     struct draw_stage* stage;
 977
 978     render = r300_render_create(r300);
 979
 980     if (!render) {
 981         return NULL;
 982     }
 983
 984     stage = draw_vbuf_stage(r300->draw, render);
 985
 986     if (!stage) {
 987         render->destroy(render);
 988         return NULL;
 989     }
 990
 991     draw_set_render(r300->draw, render);
 992
 993     return stage;
 994 }