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 static uint32_t r300_translate_primitive(unsigned prim)
  45 {
  46     switch (prim) {
  47         case PIPE_PRIM_POINTS:
  48             return R300_VAP_VF_CNTL__PRIM_POINTS;
  49         case PIPE_PRIM_LINES:
  50             return R300_VAP_VF_CNTL__PRIM_LINES;
  51         case PIPE_PRIM_LINE_LOOP:
  52             return R300_VAP_VF_CNTL__PRIM_LINE_LOOP;
  53         case PIPE_PRIM_LINE_STRIP:
  54             return R300_VAP_VF_CNTL__PRIM_LINE_STRIP;
  55         case PIPE_PRIM_TRIANGLES:
  56             return R300_VAP_VF_CNTL__PRIM_TRIANGLES;
  57         case PIPE_PRIM_TRIANGLE_STRIP:
  58             return R300_VAP_VF_CNTL__PRIM_TRIANGLE_STRIP;
  59         case PIPE_PRIM_TRIANGLE_FAN:
  60             return R300_VAP_VF_CNTL__PRIM_TRIANGLE_FAN;
  61         case PIPE_PRIM_QUADS:
  62             return R300_VAP_VF_CNTL__PRIM_QUADS;
  63         case PIPE_PRIM_QUAD_STRIP:
  64             return R300_VAP_VF_CNTL__PRIM_QUAD_STRIP;
  65         case PIPE_PRIM_POLYGON:
  66             return R300_VAP_VF_CNTL__PRIM_POLYGON;
  67         default:
  68             return 0;
  69     }
  70 }
  71
  72 static uint32_t r300_provoking_vertex_fixes(struct r300_context *r300,
  73                                             unsigned mode)
  74 {
  75     struct r300_rs_state* rs = (struct r300_rs_state*)r300->rs_state.state;
  76     uint32_t color_control = rs->color_control;
  77
  78     /* By default (see r300_state.c:r300_create_rs_state) color_control is
  79      * initialized to provoking the first vertex.
  80      *
  81      * Triangle fans must be reduced to the second vertex, not the first, in
  82      * Gallium flatshade-first mode, as per the GL spec.
  83      * (http://www.opengl.org/registry/specs/ARB/provoking_vertex.txt)
  84      *
  85      * Quads never provoke correctly in flatshade-first mode. The first
  86      * vertex is never considered as provoking, so only the second, third,
  87      * and fourth vertices can be selected, and both "third" and "last" modes
  88      * select the fourth vertex. This is probably due to D3D lacking quads.
  89      *
  90      * Similarly, polygons reduce to the first, not the last, vertex, when in
  91      * "last" mode, and all other modes start from the second vertex.
  92      *
  93      * ~ C.
  94      */
  95
  96     if (rs->rs.flatshade_first) {
  97         switch (mode) {
  98             case PIPE_PRIM_TRIANGLE_FAN:
  99                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_SECOND;
 100                 break;
 101             case PIPE_PRIM_QUADS:
 102             case PIPE_PRIM_QUAD_STRIP:
 103             case PIPE_PRIM_POLYGON:
 104                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_LAST;
 105                 break;
 106             default:
 107                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_FIRST;
 108                 break;
 109         }
 110     } else {
 111         color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_LAST;
 112     }
 113
 114     return color_control;
 115 }
 116
 117 /* Check if the requested number of dwords is available in the CS and
 118  * if not, flush. Return TRUE if the flush occured. */
 119 static boolean r300_reserve_cs_space(struct r300_context *r300,
 120                                      unsigned dwords)
 121 {
 122     if (!r300->rws->check_cs(r300->rws, dwords)) {
 123         r300->context.flush(&r300->context, 0, NULL);
 124         return TRUE;
 125     }
 126     return FALSE;
 127 }
 128
 129 static boolean immd_is_good_idea(struct r300_context *r300,
 130                                  unsigned count)
 131 {
 132     struct pipe_vertex_element* velem;
 133     struct pipe_vertex_buffer* vbuf;
 134     boolean checked[PIPE_MAX_ATTRIBS] = {0};
 135     unsigned vertex_element_count = r300->velems->count;
 136     unsigned i, vbi;
 137
 138     if (count > 10 || DBG_ON(r300, DBG_NO_IMMD)) {
 139         return FALSE;
 140     }
 141
 142     /* We shouldn't map buffers referenced by CS, busy buffers,
 143      * and ones placed in VRAM. */
 144     /* XXX Check for VRAM buffers. */
 145     for (i = 0; i < vertex_element_count; i++) {
 146         velem = &r300->velems->velem[i];
 147         vbi = velem->vertex_buffer_index;
 148
 149         if (!checked[vbi]) {
 150             vbuf = &r300->vertex_buffer[vbi];
 151
 152             if (r300_buffer_is_referenced(&r300->context,
 153                                           vbuf->buffer,
 154                                           R300_REF_CS | R300_REF_HW)) {
 155                 /* It's a very bad idea to map it... */
 156                 return FALSE;
 157             }
 158             checked[vbi] = TRUE;
 159         }
 160     }
 161     return TRUE;
 162 }
 163
 164 /*****************************************************************************
 165  * The emission of draw packets for r500. Older GPUs may use these functions *
 166  * after resolving fallback issues (e.g. stencil ref two-sided).             *
 167  ****************************************************************************/
 168
 169 void r500_emit_draw_arrays_immediate(struct r300_context *r300,
 170                                      unsigned mode,
 171                                      unsigned start,
 172                                      unsigned count)
 173 {
 174     struct pipe_vertex_element* velem;
 175     struct pipe_vertex_buffer* vbuf;
 176     unsigned vertex_element_count = r300->velems->count;
 177     unsigned i, v, vbi, dw, elem_offset, dwords;
 178
 179     /* Size of the vertex, in dwords. */
 180     unsigned vertex_size = 0;
 181
 182     /* Offsets of the attribute, in dwords, from the start of the vertex. */
 183     unsigned offset[PIPE_MAX_ATTRIBS];
 184
 185     /* Size of the vertex element, in dwords. */
 186     unsigned size[PIPE_MAX_ATTRIBS];
 187
 188     /* Stride to the same attrib in the next vertex in the vertex buffer,
 189      * in dwords. */
 190     unsigned stride[PIPE_MAX_ATTRIBS] = {0};
 191
 192     /* Mapped vertex buffers. */
 193     uint32_t* map[PIPE_MAX_ATTRIBS] = {0};
 194     struct pipe_transfer* transfer[PIPE_MAX_ATTRIBS] = {NULL};
 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,
 210                                                   vbuf->buffer,
 211                                                   PIPE_TRANSFER_READ,
 212                                                   &transfer[vbi]);
 213             map[vbi] += vbuf->buffer_offset / 4;
 214             stride[vbi] = vbuf->stride / 4;
 215         }
 216     }
 217
 218     dwords = 9 + count * vertex_size;
 219
 220     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 221     r300_emit_buffer_validate(r300, FALSE, NULL);
 222     r300_emit_dirty_state(r300);
 223
 224     BEGIN_CS(dwords);
 225     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 226             r300_provoking_vertex_fixes(r300, mode));
 227     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
 228     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 229     OUT_CS(count - 1);
 230     OUT_CS(0);
 231     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
 232     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
 233             r300_translate_primitive(mode));
 234
 235     /* Emit vertices. */
 236     for (v = 0; v < count; v++) {
 237         for (i = 0; i < vertex_element_count; i++) {
 238             velem = &r300->velems->velem[i];
 239             vbi = velem->vertex_buffer_index;
 240             elem_offset = offset[i] + stride[vbi] * (v + start);
 241
 242             for (dw = 0; dw < size[i]; dw++) {
 243                 OUT_CS(map[vbi][elem_offset + dw]);
 244             }
 245         }
 246     }
 247     END_CS;
 248
 249     /* Unmap buffers. */
 250     for (i = 0; i < vertex_element_count; i++) {
 251         vbi = r300->velems->velem[i].vertex_buffer_index;
 252
 253         if (map[vbi]) {
 254             vbuf = &r300->vertex_buffer[vbi];
 255             pipe_buffer_unmap(&r300->context, vbuf->buffer, transfer[vbi]);
 256             map[vbi] = NULL;
 257         }
 258     }
 259 }
 260
 261 void r500_emit_draw_arrays(struct r300_context *r300,
 262                            unsigned mode,
 263                            unsigned count)
 264 {
 265     boolean alt_num_verts = count > 65535;
 266     CS_LOCALS(r300);
 267
 268     if (alt_num_verts) {
 269         if (count >= (1 << 24)) {
 270             fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 271                     "refusing to render.\n", count);
 272             return;
 273         }
 274         BEGIN_CS(9);
 275         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 276     } else {
 277         BEGIN_CS(7);
 278     }
 279     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 280             r300_provoking_vertex_fixes(r300, mode));
 281     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 282     OUT_CS(count - 1);
 283     OUT_CS(0);
 284     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 285     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 286            r300_translate_primitive(mode) |
 287            (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 288     END_CS;
 289 }
 290
 291 void r500_emit_draw_elements(struct r300_context *r300,
 292                              struct pipe_resource* indexBuffer,
 293                              unsigned indexSize,
 294                              int indexBias,
 295                              unsigned minIndex,
 296                              unsigned maxIndex,
 297                              unsigned mode,
 298                              unsigned start,
 299                              unsigned count)
 300 {
 301     uint32_t count_dwords;
 302     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 303     boolean alt_num_verts = count > 65535;
 304     CS_LOCALS(r300);
 305
 306     if (count >= (1 << 24)) {
 307         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 308                 "refusing to render.\n", count);
 309         return;
 310     }
 311
 312     assert(indexBias == 0);
 313
 314     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 315
 316     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 317         count, minIndex, maxIndex);
 318
 319     if (alt_num_verts) {
 320         BEGIN_CS(15);
 321         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 322     } else {
 323         BEGIN_CS(13);
 324     }
 325     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 326             r300_provoking_vertex_fixes(r300, mode));
 327     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 328     OUT_CS(maxIndex);
 329     OUT_CS(minIndex);
 330     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 331     if (indexSize == 4) {
 332         count_dwords = count;
 333         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 334                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 335                r300_translate_primitive(mode) |
 336                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 337     } else {
 338         count_dwords = (count + 1) / 2;
 339         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 340                r300_translate_primitive(mode) |
 341                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 342     }
 343
 344     /* INDX_BUFFER is a truly special packet3.
 345      * Unlike most other packet3, where the offset is after the count,
 346      * the order is reversed, so the relocation ends up carrying the
 347      * size of the indexbuf instead of the offset.
 348      */
 349     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 350     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 351            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 352     OUT_CS(offset_dwords << 2);
 353     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 354                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 355
 356     END_CS;
 357 }
 358
 359 /*****************************************************************************
 360  * The emission of draw packets for r300 which take care of the two-sided    *
 361  * stencil ref fallback and call r500's functions.                           *
 362  ****************************************************************************/
 363
 364 /* Set drawing for front faces. */
 365 static void r300_begin_stencil_ref_fallback(struct r300_context *r300)
 366 {
 367     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 368     CS_LOCALS(r300);
 369
 370     BEGIN_CS(2);
 371     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_BACK);
 372     END_CS;
 373 }
 374
 375 /* Set drawing for back faces. */
 376 static void r300_switch_stencil_ref_side(struct r300_context *r300)
 377 {
 378     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 379     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 380     CS_LOCALS(r300);
 381
 382     BEGIN_CS(4);
 383     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_FRONT);
 384     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 385                dsa->stencil_ref_bf | r300->stencil_ref.ref_value[1]);
 386     END_CS;
 387 }
 388
 389 /* Restore the original state. */
 390 static void r300_end_stencil_ref_fallback(struct r300_context *r300)
 391 {
 392     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 393     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 394     CS_LOCALS(r300);
 395
 396     BEGIN_CS(4);
 397     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode);
 398     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 399                dsa->stencil_ref_mask | r300->stencil_ref.ref_value[0]);
 400     END_CS;
 401 }
 402
 403 void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 404                                      unsigned mode,
 405                                      unsigned start,
 406                                      unsigned count)
 407 {
 408     if (!r300->stencil_ref_bf_fallback) {
 409         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 410     } else {
 411         r300_begin_stencil_ref_fallback(r300);
 412         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 413         r300_switch_stencil_ref_side(r300);
 414         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 415         r300_end_stencil_ref_fallback(r300);
 416     }
 417 }
 418
 419 void r300_emit_draw_arrays(struct r300_context *r300,
 420                            unsigned mode,
 421                            unsigned count)
 422 {
 423     if (!r300->stencil_ref_bf_fallback) {
 424         r500_emit_draw_arrays(r300, mode, count);
 425     } else {
 426         r300_begin_stencil_ref_fallback(r300);
 427         r500_emit_draw_arrays(r300, mode, count);
 428         r300_switch_stencil_ref_side(r300);
 429         r500_emit_draw_arrays(r300, mode, count);
 430         r300_end_stencil_ref_fallback(r300);
 431     }
 432 }
 433
 434 void r300_emit_draw_elements(struct r300_context *r300,
 435                              struct pipe_resource* indexBuffer,
 436                              unsigned indexSize,
 437                              int indexBias,
 438                              unsigned minIndex,
 439                              unsigned maxIndex,
 440                              unsigned mode,
 441                              unsigned start,
 442                              unsigned count)
 443 {
 444     if (!r300->stencil_ref_bf_fallback) {
 445         r500_emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 446                                 minIndex, maxIndex, mode, start, count);
 447     } else {
 448         r300_begin_stencil_ref_fallback(r300);
 449         r500_emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 450                                 minIndex, maxIndex, mode, start, count);
 451         r300_switch_stencil_ref_side(r300);
 452         r500_emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 453                                 minIndex, maxIndex, mode, start, count);
 454         r300_end_stencil_ref_fallback(r300);
 455     }
 456 }
 457
 458 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 459                                     struct pipe_resource** elts,
 460                                     unsigned start,
 461                                     unsigned count)
 462 {
 463     struct pipe_context* context = &r300->context;
 464     struct pipe_screen* screen = r300->context.screen;
 465     struct pipe_resource* new_elts;
 466     unsigned char *in_map;
 467     unsigned short *out_map;
 468     struct pipe_transfer *src_transfer, *dst_transfer;
 469     unsigned i;
 470
 471     new_elts = pipe_buffer_create(screen,
 472                                   PIPE_BIND_INDEX_BUFFER,
 473                                   2 * count);
 474
 475     in_map = pipe_buffer_map(context, *elts, PIPE_TRANSFER_READ, &src_transfer);
 476     out_map = pipe_buffer_map(context, new_elts, PIPE_TRANSFER_WRITE, &dst_transfer);
 477
 478     in_map += start;
 479
 480     for (i = 0; i < count; i++) {
 481         *out_map = (unsigned short)*in_map;
 482         in_map++;
 483         out_map++;
 484     }
 485
 486     pipe_buffer_unmap(context, *elts, src_transfer);
 487     pipe_buffer_unmap(context, new_elts, dst_transfer);
 488
 489     *elts = new_elts;
 490 }
 491
 492 static void r300_align_ushort_elts(struct r300_context *r300,
 493                                    struct pipe_resource **elts,
 494                                    unsigned start, unsigned count)
 495 {
 496     struct pipe_context* context = &r300->context;
 497     struct pipe_transfer *in_transfer = NULL;
 498     struct pipe_transfer *out_transfer = NULL;
 499     struct pipe_resource* new_elts;
 500     unsigned short *in_map;
 501     unsigned short *out_map;
 502
 503     new_elts = pipe_buffer_create(context->screen,
 504                                   PIPE_BIND_INDEX_BUFFER,
 505                                   2 * count);
 506
 507     in_map = pipe_buffer_map(context, *elts,
 508                              PIPE_TRANSFER_READ, &in_transfer);
 509     out_map = pipe_buffer_map(context, new_elts,
 510                               PIPE_TRANSFER_WRITE, &out_transfer);
 511
 512     memcpy(out_map, in_map+start, 2 * count);
 513
 514     pipe_buffer_unmap(context, *elts, in_transfer);
 515     pipe_buffer_unmap(context, new_elts, out_transfer);
 516
 517     *elts = new_elts;
 518 }
 519
 520 /* This is the fast-path drawing & emission for HW TCL. */
 521 void r300_draw_range_elements(struct pipe_context* pipe,
 522                               struct pipe_resource* indexBuffer,
 523                               unsigned indexSize,
 524                               int indexBias,
 525                               unsigned minIndex,
 526                               unsigned maxIndex,
 527                               unsigned mode,
 528                               unsigned start,
 529                               unsigned count)
 530 {
 531     struct r300_context* r300 = r300_context(pipe);
 532     struct pipe_resource* orgIndexBuffer = indexBuffer;
 533     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 534                             count > 65536 &&
 535                             r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0);
 536     unsigned short_count;
 537
 538     if (r300->skip_rendering) {
 539         return;
 540     }
 541
 542     if (!u_trim_pipe_prim(mode, &count)) {
 543         return;
 544     }
 545
 546     if (indexSize == 1) {
 547         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 548         indexSize = 2;
 549         start = 0;
 550     } else if (indexSize == 2 && start % 2 != 0) {
 551         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 552         start = 0;
 553     }
 554
 555     r300_update_derived_state(r300);
 556
 557     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 558
 559     /* 128 dwords for emit_aos and emit_draw_elements */
 560     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 561     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 562     r300_emit_dirty_state(r300);
 563     r300_emit_aos(r300, 0, TRUE);
 564
 565     u_upload_flush(r300->upload_vb);
 566     u_upload_flush(r300->upload_ib);
 567     if (alt_num_verts || count <= 65535) {
 568         r300->emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 569                                  minIndex, maxIndex, mode, start, count);
 570     } else {
 571         do {
 572             short_count = MIN2(count, 65534);
 573             r300->emit_draw_elements(r300, indexBuffer, indexSize, indexBias,
 574                                      minIndex, maxIndex,
 575                                      mode, start, short_count);
 576
 577             start += short_count;
 578             count -= short_count;
 579
 580             /* 16 spare dwords are enough for emit_draw_elements. */
 581             if (count && r300_reserve_cs_space(r300, 16)) {
 582                 r300_emit_buffer_validate(r300, TRUE, indexBuffer);
 583                 r300_emit_dirty_state(r300);
 584                 r300_emit_aos(r300, 0, TRUE);
 585             }
 586         } while (count);
 587     }
 588
 589     if (indexBuffer != orgIndexBuffer) {
 590         pipe_resource_reference( &indexBuffer, NULL );
 591     }
 592 }
 593
 594 /* Simple helpers for context setup. Should probably be moved to util. */
 595 void r300_draw_elements(struct pipe_context* pipe,
 596                         struct pipe_resource* indexBuffer,
 597                         unsigned indexSize, int indexBias, unsigned mode,
 598                         unsigned start, unsigned count)
 599 {
 600     struct r300_context *r300 = r300_context(pipe);
 601
 602     pipe->draw_range_elements(pipe, indexBuffer, indexSize, indexBias,
 603                               0, r300->vertex_buffer_max_index,
 604                               mode, start, count);
 605 }
 606
 607 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 608                       unsigned start, unsigned count)
 609 {
 610     struct r300_context* r300 = r300_context(pipe);
 611     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 612                             count > 65536 &&
 613                             r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0);
 614     unsigned short_count;
 615
 616     if (r300->skip_rendering) {
 617         return;
 618     }
 619
 620     if (!u_trim_pipe_prim(mode, &count)) {
 621         return;
 622     }
 623
 624     r300_update_derived_state(r300);
 625
 626     if (immd_is_good_idea(r300, count)) {
 627         r300->emit_draw_arrays_immediate(r300, mode, start, count);
 628     } else {
 629         /* Make sure there are at least 128 spare dwords in the command buffer.
 630          * (most of it being consumed by emit_aos) */
 631         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
 632         r300_emit_buffer_validate(r300, TRUE, NULL);
 633         r300_emit_dirty_state(r300);
 634
 635         if (alt_num_verts || count <= 65535) {
 636             r300_emit_aos(r300, start, FALSE);
 637             r300->emit_draw_arrays(r300, mode, count);
 638         } else {
 639             do {
 640                 short_count = MIN2(count, 65535);
 641                 r300_emit_aos(r300, start, FALSE);
 642                 r300->emit_draw_arrays(r300, mode, short_count);
 643
 644                 start += short_count;
 645                 count -= short_count;
 646
 647                 /* Again, we emit both AOS and draw_arrays so there should be
 648                  * at least 128 spare dwords. */
 649                 if (count && r300_reserve_cs_space(r300, 128)) {
 650                     r300_emit_buffer_validate(r300, TRUE, NULL);
 651                     r300_emit_dirty_state(r300);
 652                 }
 653             } while (count);
 654         }
 655         u_upload_flush(r300->upload_vb);
 656     }
 657 }
 658
 659 /****************************************************************************
 660  * The rest of this file is for SW TCL rendering only. Please be polite and *
 661  * keep these functions separated so that they are easier to locate. ~C.    *
 662  ***************************************************************************/
 663
 664 /* SW TCL arrays, using Draw. */
 665 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 666                                unsigned mode,
 667                                unsigned start,
 668                                unsigned count)
 669 {
 670     struct r300_context* r300 = r300_context(pipe);
 671     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 672     int i;
 673
 674     if (r300->skip_rendering) {
 675         return;
 676     }
 677
 678     if (!u_trim_pipe_prim(mode, &count)) {
 679         return;
 680     }
 681
 682     for (i = 0; i < r300->vertex_buffer_count; i++) {
 683         void* buf = pipe_buffer_map(pipe,
 684                                     r300->vertex_buffer[i].buffer,
 685                                     PIPE_TRANSFER_READ,
 686                                     &vb_transfer[i]);
 687         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 688     }
 689
 690     draw_set_mapped_element_buffer(r300->draw, 0, 0, NULL);
 691
 692     draw_arrays(r300->draw, mode, start, count);
 693
 694     for (i = 0; i < r300->vertex_buffer_count; i++) {
 695         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 696                           vb_transfer[i]);
 697         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 698     }
 699 }
 700
 701 /* SW TCL elements, using Draw. */
 702 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 703                                        struct pipe_resource* indexBuffer,
 704                                        unsigned indexSize,
 705                                        int indexBias,
 706                                        unsigned minIndex,
 707                                        unsigned maxIndex,
 708                                        unsigned mode,
 709                                        unsigned start,
 710                                        unsigned count)
 711 {
 712     struct r300_context* r300 = r300_context(pipe);
 713     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 714     struct pipe_transfer *ib_transfer;
 715     int i;
 716     void* indices;
 717
 718     if (r300->skip_rendering) {
 719         return;
 720     }
 721
 722     if (!u_trim_pipe_prim(mode, &count)) {
 723         return;
 724     }
 725
 726     for (i = 0; i < r300->vertex_buffer_count; i++) {
 727         void* buf = pipe_buffer_map(pipe,
 728                                     r300->vertex_buffer[i].buffer,
 729                                     PIPE_TRANSFER_READ,
 730                                     &vb_transfer[i]);
 731         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 732     }
 733
 734     indices = pipe_buffer_map(pipe, indexBuffer,
 735                               PIPE_TRANSFER_READ, &ib_transfer);
 736     draw_set_mapped_element_buffer_range(r300->draw, indexSize, indexBias,
 737                                          minIndex, maxIndex, indices);
 738
 739     draw_arrays(r300->draw, mode, start, count);
 740
 741     for (i = 0; i < r300->vertex_buffer_count; i++) {
 742         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 743                           vb_transfer[i]);
 744         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 745     }
 746
 747     pipe_buffer_unmap(pipe, indexBuffer,
 748                       ib_transfer);
 749     draw_set_mapped_element_buffer_range(r300->draw, 0, 0,
 750                                          start, start + count - 1,
 751                                          NULL);
 752 }
 753
 754 /* Object for rendering using Draw. */
 755 struct r300_render {
 756     /* Parent class */
 757     struct vbuf_render base;
 758
 759     /* Pipe context */
 760     struct r300_context* r300;
 761
 762     /* Vertex information */
 763     size_t vertex_size;
 764     unsigned prim;
 765     unsigned hwprim;
 766
 767     /* VBO */
 768     struct pipe_resource* vbo;
 769     size_t vbo_size;
 770     size_t vbo_offset;
 771     size_t vbo_max_used;
 772     void * vbo_ptr;
 773
 774    struct pipe_transfer *vbo_transfer;
 775 };
 776
 777 static INLINE struct r300_render*
 778 r300_render(struct vbuf_render* render)
 779 {
 780     return (struct r300_render*)render;
 781 }
 782
 783 static const struct vertex_info*
 784 r300_render_get_vertex_info(struct vbuf_render* render)
 785 {
 786     struct r300_render* r300render = r300_render(render);
 787     struct r300_context* r300 = r300render->r300;
 788
 789     r300_update_derived_state(r300);
 790
 791     return &r300->vertex_info;
 792 }
 793
 794 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 795                                                    ushort vertex_size,
 796                                                    ushort count)
 797 {
 798     struct r300_render* r300render = r300_render(render);
 799     struct r300_context* r300 = r300render->r300;
 800     struct pipe_screen* screen = r300->context.screen;
 801     size_t size = (size_t)vertex_size * (size_t)count;
 802
 803     if (size + r300render->vbo_offset > r300render->vbo_size)
 804     {
 805         pipe_resource_reference(&r300->vbo, NULL);
 806         r300render->vbo = pipe_buffer_create(screen,
 807                                              PIPE_BIND_VERTEX_BUFFER,
 808                                              R300_MAX_DRAW_VBO_SIZE);
 809         r300render->vbo_offset = 0;
 810         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 811     }
 812
 813     r300render->vertex_size = vertex_size;
 814     r300->vbo = r300render->vbo;
 815     r300->vbo_offset = r300render->vbo_offset;
 816
 817     return (r300render->vbo) ? TRUE : FALSE;
 818 }
 819
 820 static void* r300_render_map_vertices(struct vbuf_render* render)
 821 {
 822     struct r300_render* r300render = r300_render(render);
 823
 824     r300render->vbo_ptr = pipe_buffer_map(&r300render->r300->context,
 825                                           r300render->vbo,
 826                                           PIPE_TRANSFER_WRITE,
 827                                           &r300render->vbo_transfer);
 828
 829     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 830 }
 831
 832 static void r300_render_unmap_vertices(struct vbuf_render* render,
 833                                              ushort min,
 834                                              ushort max)
 835 {
 836     struct r300_render* r300render = r300_render(render);
 837     struct pipe_context* context = &r300render->r300->context;
 838     CS_LOCALS(r300render->r300);
 839     BEGIN_CS(2);
 840     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 841     END_CS;
 842
 843     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 844                                     r300render->vertex_size * (max + 1));
 845     pipe_buffer_unmap(context, r300render->vbo, r300render->vbo_transfer);
 846 }
 847
 848 static void r300_render_release_vertices(struct vbuf_render* render)
 849 {
 850     struct r300_render* r300render = r300_render(render);
 851
 852     r300render->vbo_offset += r300render->vbo_max_used;
 853     r300render->vbo_max_used = 0;
 854 }
 855
 856 static boolean r300_render_set_primitive(struct vbuf_render* render,
 857                                                unsigned prim)
 858 {
 859     struct r300_render* r300render = r300_render(render);
 860
 861     r300render->prim = prim;
 862     r300render->hwprim = r300_translate_primitive(prim);
 863
 864     return TRUE;
 865 }
 866
 867 static void r500_render_draw_arrays(struct vbuf_render* render,
 868                                     unsigned start,
 869                                     unsigned count)
 870 {
 871     struct r300_render* r300render = r300_render(render);
 872     struct r300_context* r300 = r300render->r300;
 873
 874     CS_LOCALS(r300);
 875
 876     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
 877     r300_emit_buffer_validate(r300, FALSE, NULL);
 878     r300_emit_dirty_state(r300);
 879
 880     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 881
 882     BEGIN_CS(2);
 883     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 884     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 885            r300render->hwprim);
 886     END_CS;
 887 }
 888
 889 static void r500_render_draw(struct vbuf_render* render,
 890                              const ushort* indices,
 891                              uint count)
 892 {
 893     struct r300_render* r300render = r300_render(render);
 894     struct r300_context* r300 = r300render->r300;
 895     int i;
 896     unsigned dwords = 2 + (count+1)/2;
 897
 898     CS_LOCALS(r300);
 899
 900     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
 901     r300_emit_buffer_validate(r300, FALSE, NULL);
 902     r300_emit_dirty_state(r300);
 903
 904     BEGIN_CS(dwords);
 905     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 906     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 907            r300render->hwprim);
 908     for (i = 0; i < count-1; i += 2) {
 909         OUT_CS(indices[i+1] << 16 | indices[i]);
 910     }
 911     if (count % 2) {
 912         OUT_CS(indices[count-1]);
 913     }
 914     END_CS;
 915 }
 916
 917 static void r300_render_draw_arrays(struct vbuf_render* render,
 918                                     unsigned start,
 919                                     unsigned count)
 920 {
 921     struct r300_context* r300 = r300_render(render)->r300;
 922
 923     if (!r300->stencil_ref_bf_fallback) {
 924         r500_render_draw_arrays(render, start, count);
 925     } else {
 926         r300_begin_stencil_ref_fallback(r300);
 927         r500_render_draw_arrays(render, start, count);
 928         r300_switch_stencil_ref_side(r300);
 929         r500_render_draw_arrays(render, start, count);
 930         r300_end_stencil_ref_fallback(r300);
 931     }
 932 }
 933
 934 static void r300_render_draw(struct vbuf_render* render,
 935                              const ushort* indices,
 936                              uint count)
 937 {
 938     struct r300_context* r300 = r300_render(render)->r300;
 939
 940     if (!r300->stencil_ref_bf_fallback) {
 941         r500_render_draw(render, indices, count);
 942     } else {
 943         r300_begin_stencil_ref_fallback(r300);
 944         r500_render_draw(render, indices, count);
 945         r300_switch_stencil_ref_side(r300);
 946         r500_render_draw(render, indices, count);
 947         r300_end_stencil_ref_fallback(r300);
 948     }
 949 }
 950
 951 static void r300_render_destroy(struct vbuf_render* render)
 952 {
 953     FREE(render);
 954 }
 955
 956 static struct vbuf_render* r300_render_create(struct r300_context* r300)
 957 {
 958     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
 959
 960     r300render->r300 = r300;
 961
 962     /* XXX find real numbers plz */
 963     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
 964     r300render->base.max_indices = 16 * 1024;
 965
 966     r300render->base.get_vertex_info = r300_render_get_vertex_info;
 967     r300render->base.allocate_vertices = r300_render_allocate_vertices;
 968     r300render->base.map_vertices = r300_render_map_vertices;
 969     r300render->base.unmap_vertices = r300_render_unmap_vertices;
 970     r300render->base.set_primitive = r300_render_set_primitive;
 971     if (r300->screen->caps.is_r500) {
 972         r300render->base.draw = r500_render_draw;
 973         r300render->base.draw_arrays = r500_render_draw_arrays;
 974     } else {
 975         r300render->base.draw = r300_render_draw;
 976         r300render->base.draw_arrays = r300_render_draw_arrays;
 977     }
 978     r300render->base.release_vertices = r300_render_release_vertices;
 979     r300render->base.destroy = r300_render_destroy;
 980
 981     r300render->vbo = NULL;
 982     r300render->vbo_size = 0;
 983     r300render->vbo_offset = 0;
 984
 985     return &r300render->base;
 986 }
 987
 988 struct draw_stage* r300_draw_stage(struct r300_context* r300)
 989 {
 990     struct vbuf_render* render;
 991     struct draw_stage* stage;
 992
 993     render = r300_render_create(r300);
 994
 995     if (!render) {
 996         return NULL;
 997     }
 998
 999     stage = draw_vbuf_stage(r300->draw, render);
1000
1001     if (!stage) {
1002         render->destroy(render);
1003         return NULL;
1004     }
1005
1006     draw_set_render(r300->draw, render);
1007
1008     return stage;
1009 }