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 static void r500_emit_index_offset(struct r300_context *r300, int index_bias)
 118 {
 119     CS_LOCALS(r300);
 120
 121     if (r300->screen->caps.is_r500 &&
 122         r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0)) {
 123         BEGIN_CS(2);
 124         OUT_CS_REG(R500_VAP_INDEX_OFFSET,
 125                    (index_bias & 0xFFFFFF) | (index_bias < 0 ? 1<<24 : 0));
 126         END_CS;
 127     } else {
 128         if (index_bias) {
 129             fprintf(stderr, "r300: Non-zero index bias is unsupported "
 130                             "on this hardware.\n");
 131             assert(0);
 132         }
 133     }
 134 }
 135
 136 enum r300_prepare_flags {
 137     PREP_FIRST_DRAW     = (1 << 0),
 138     PREP_VALIDATE_VBOS  = (1 << 1),
 139     PREP_EMIT_AOS       = (1 << 2),
 140     PREP_INDEXED        = (1 << 3)
 141 };
 142
 143 /* Check if the requested number of dwords is available in the CS and
 144  * if not, flush. Then validate buffers and emit dirty state.
 145  * Return TRUE if flush occured. */
 146 static void r300_prepare_for_rendering(struct r300_context *r300,
 147                                        enum r300_prepare_flags flags,
 148                                        struct pipe_resource *index_buffer,
 149                                        unsigned cs_dwords,
 150                                        unsigned aos_offset,
 151                                        int index_bias)
 152 {
 153     boolean flushed = FALSE;
 154     boolean first_draw = flags & PREP_FIRST_DRAW;
 155     boolean emit_aos = flags & PREP_EMIT_AOS;
 156
 157     /* Stencil ref fallback. */
 158     if (r300->stencil_ref_bf_fallback) {
 159         cs_dwords = cs_dwords * 2 + 10;
 160     }
 161
 162     /* Add dirty state, index offset, and AOS. */
 163     if (first_draw) {
 164         cs_dwords += r300_get_num_dirty_dwords(r300);
 165
 166         if (r300->screen->caps.is_r500)
 167             cs_dwords += 2; /* emit_index_offset */
 168
 169         if (emit_aos)
 170             cs_dwords += 55; /* emit_aos */
 171     }
 172
 173     /* Emitted in flush. */
 174     cs_dwords += 26; /* emit_query_end */
 175
 176     /* Reserve requested CS space. */
 177     if (!r300->rws->check_cs(r300->rws, cs_dwords)) {
 178         r300->context.flush(&r300->context, 0, NULL);
 179         flushed = TRUE;
 180     }
 181
 182     /* Validate buffers and emit dirty state if needed. */
 183     if (first_draw || flushed) {
 184         r300_emit_buffer_validate(r300, flags & PREP_VALIDATE_VBOS, index_buffer);
 185         r300_emit_dirty_state(r300);
 186         r500_emit_index_offset(r300, index_bias);
 187         if (emit_aos)
 188             r300_emit_aos(r300, aos_offset, flags & PREP_INDEXED);
 189     }
 190 }
 191
 192 static boolean immd_is_good_idea(struct r300_context *r300,
 193                                  unsigned count)
 194 {
 195     struct pipe_vertex_element* velem;
 196     struct pipe_vertex_buffer* vbuf;
 197     boolean checked[PIPE_MAX_ATTRIBS] = {0};
 198     unsigned vertex_element_count = r300->velems->count;
 199     unsigned i, vbi;
 200
 201     if (DBG_ON(r300, DBG_NO_IMMD)) {
 202         return FALSE;
 203     }
 204
 205     if (r300->draw) {
 206         return FALSE;
 207     }
 208
 209     if (count > 10) {
 210         return FALSE;
 211     }
 212
 213     /* We shouldn't map buffers referenced by CS, busy buffers,
 214      * and ones placed in VRAM. */
 215     /* XXX Check for VRAM buffers. */
 216     for (i = 0; i < vertex_element_count; i++) {
 217         velem = &r300->velems->velem[i];
 218         vbi = velem->vertex_buffer_index;
 219
 220         if (!checked[vbi]) {
 221             vbuf = &r300->vertex_buffer[vbi];
 222
 223             if (r300_buffer_is_referenced(&r300->context,
 224                                           vbuf->buffer,
 225                                           R300_REF_CS | R300_REF_HW)) {
 226                 /* It's a very bad idea to map it... */
 227                 return FALSE;
 228             }
 229             checked[vbi] = TRUE;
 230         }
 231     }
 232     return TRUE;
 233 }
 234
 235 /*****************************************************************************
 236  * The emission of draw packets for r500. Older GPUs may use these functions *
 237  * after resolving fallback issues (e.g. stencil ref two-sided).             *
 238  ****************************************************************************/
 239
 240 void r500_emit_draw_arrays_immediate(struct r300_context *r300,
 241                                      unsigned mode,
 242                                      unsigned start,
 243                                      unsigned count)
 244 {
 245     struct pipe_vertex_element* velem;
 246     struct pipe_vertex_buffer* vbuf;
 247     unsigned vertex_element_count = r300->velems->count;
 248     unsigned i, v, vbi, dw, elem_offset, dwords;
 249
 250     /* Size of the vertex, in dwords. */
 251     unsigned vertex_size = 0;
 252
 253     /* Offsets of the attribute, in dwords, from the start of the vertex. */
 254     unsigned offset[PIPE_MAX_ATTRIBS];
 255
 256     /* Size of the vertex element, in dwords. */
 257     unsigned size[PIPE_MAX_ATTRIBS];
 258
 259     /* Stride to the same attrib in the next vertex in the vertex buffer,
 260      * in dwords. */
 261     unsigned stride[PIPE_MAX_ATTRIBS] = {0};
 262
 263     /* Mapped vertex buffers. */
 264     uint32_t* map[PIPE_MAX_ATTRIBS] = {0};
 265     struct pipe_transfer* transfer[PIPE_MAX_ATTRIBS] = {NULL};
 266
 267     CS_LOCALS(r300);
 268
 269     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
 270     for (i = 0; i < vertex_element_count; i++) {
 271         velem = &r300->velems->velem[i];
 272         offset[i] = velem->src_offset / 4;
 273         size[i] = util_format_get_blocksize(velem->src_format) / 4;
 274         vertex_size += size[i];
 275         vbi = velem->vertex_buffer_index;
 276
 277         /* Map the buffer. */
 278         if (!map[vbi]) {
 279             vbuf = &r300->vertex_buffer[vbi];
 280             map[vbi] = (uint32_t*)pipe_buffer_map(&r300->context,
 281                                                   vbuf->buffer,
 282                                                   PIPE_TRANSFER_READ,
 283                                                   &transfer[vbi]);
 284             map[vbi] += vbuf->buffer_offset / 4;
 285             stride[vbi] = vbuf->stride / 4;
 286         }
 287     }
 288
 289     dwords = 9 + count * vertex_size;
 290
 291     r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0);
 292
 293     BEGIN_CS(dwords);
 294     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 295             r300_provoking_vertex_fixes(r300, mode));
 296     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
 297     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 298     OUT_CS(count - 1);
 299     OUT_CS(0);
 300     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
 301     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
 302             r300_translate_primitive(mode));
 303
 304     /* Emit vertices. */
 305     for (v = 0; v < count; v++) {
 306         for (i = 0; i < vertex_element_count; i++) {
 307             velem = &r300->velems->velem[i];
 308             vbi = velem->vertex_buffer_index;
 309             elem_offset = offset[i] + stride[vbi] * (v + start);
 310
 311             for (dw = 0; dw < size[i]; dw++) {
 312                 OUT_CS(map[vbi][elem_offset + dw]);
 313             }
 314         }
 315     }
 316     END_CS;
 317
 318     /* Unmap buffers. */
 319     for (i = 0; i < vertex_element_count; i++) {
 320         vbi = r300->velems->velem[i].vertex_buffer_index;
 321
 322         if (map[vbi]) {
 323             vbuf = &r300->vertex_buffer[vbi];
 324             pipe_buffer_unmap(&r300->context, vbuf->buffer, transfer[vbi]);
 325             map[vbi] = NULL;
 326         }
 327     }
 328 }
 329
 330 void r500_emit_draw_arrays(struct r300_context *r300,
 331                            unsigned mode,
 332                            unsigned count)
 333 {
 334     boolean alt_num_verts = count > 65535;
 335     CS_LOCALS(r300);
 336
 337     if (count >= (1 << 24)) {
 338         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 339                 "refusing to render.\n", count);
 340         return;
 341     }
 342
 343     BEGIN_CS(7 + (alt_num_verts ? 2 : 0));
 344     if (alt_num_verts) {
 345         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 346     }
 347     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 348             r300_provoking_vertex_fixes(r300, mode));
 349     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 350     OUT_CS(count - 1);
 351     OUT_CS(0);
 352     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 353     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 354            r300_translate_primitive(mode) |
 355            (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 356     END_CS;
 357 }
 358
 359 void r500_emit_draw_elements(struct r300_context *r300,
 360                              struct pipe_resource* indexBuffer,
 361                              unsigned indexSize,
 362                              unsigned minIndex,
 363                              unsigned maxIndex,
 364                              unsigned mode,
 365                              unsigned start,
 366                              unsigned count)
 367 {
 368     uint32_t count_dwords;
 369     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 370     boolean alt_num_verts = count > 65535;
 371     CS_LOCALS(r300);
 372
 373     if (count >= (1 << 24)) {
 374         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 375                 "refusing to render.\n", count);
 376         return;
 377     }
 378
 379     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 380
 381     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 382         count, minIndex, maxIndex);
 383
 384     BEGIN_CS(13 + (alt_num_verts ? 2 : 0));
 385     if (alt_num_verts) {
 386         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 387     }
 388     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 389             r300_provoking_vertex_fixes(r300, mode));
 390     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 391     OUT_CS(maxIndex);
 392     OUT_CS(minIndex);
 393     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 394     if (indexSize == 4) {
 395         count_dwords = count;
 396         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 397                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 398                r300_translate_primitive(mode) |
 399                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 400     } else {
 401         count_dwords = (count + 1) / 2;
 402         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 403                r300_translate_primitive(mode) |
 404                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 405     }
 406
 407     /* INDX_BUFFER is a truly special packet3.
 408      * Unlike most other packet3, where the offset is after the count,
 409      * the order is reversed, so the relocation ends up carrying the
 410      * size of the indexbuf instead of the offset.
 411      */
 412     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 413     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 414            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 415     OUT_CS(offset_dwords << 2);
 416     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 417                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 418
 419     END_CS;
 420 }
 421
 422 /*****************************************************************************
 423  * The emission of draw packets for r300 which take care of the two-sided    *
 424  * stencil ref fallback and call r500's functions.                           *
 425  ****************************************************************************/
 426
 427 /* Set drawing for front faces. */
 428 static void r300_begin_stencil_ref_fallback(struct r300_context *r300)
 429 {
 430     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 431     CS_LOCALS(r300);
 432
 433     BEGIN_CS(2);
 434     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_BACK);
 435     END_CS;
 436 }
 437
 438 /* Set drawing for back faces. */
 439 static void r300_switch_stencil_ref_side(struct r300_context *r300)
 440 {
 441     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 442     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 443     CS_LOCALS(r300);
 444
 445     BEGIN_CS(4);
 446     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_FRONT);
 447     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 448                dsa->stencil_ref_bf | r300->stencil_ref.ref_value[1]);
 449     END_CS;
 450 }
 451
 452 /* Restore the original state. */
 453 static void r300_end_stencil_ref_fallback(struct r300_context *r300)
 454 {
 455     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 456     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 457     CS_LOCALS(r300);
 458
 459     BEGIN_CS(4);
 460     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode);
 461     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 462                dsa->stencil_ref_mask | r300->stencil_ref.ref_value[0]);
 463     END_CS;
 464 }
 465
 466 void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 467                                      unsigned mode,
 468                                      unsigned start,
 469                                      unsigned count)
 470 {
 471     if (!r300->stencil_ref_bf_fallback) {
 472         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 473     } else {
 474         r300_begin_stencil_ref_fallback(r300);
 475         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 476         r300_switch_stencil_ref_side(r300);
 477         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 478         r300_end_stencil_ref_fallback(r300);
 479     }
 480 }
 481
 482 void r300_emit_draw_arrays(struct r300_context *r300,
 483                            unsigned mode,
 484                            unsigned count)
 485 {
 486     if (!r300->stencil_ref_bf_fallback) {
 487         r500_emit_draw_arrays(r300, mode, count);
 488     } else {
 489         r300_begin_stencil_ref_fallback(r300);
 490         r500_emit_draw_arrays(r300, mode, count);
 491         r300_switch_stencil_ref_side(r300);
 492         r500_emit_draw_arrays(r300, mode, count);
 493         r300_end_stencil_ref_fallback(r300);
 494     }
 495 }
 496
 497 void r300_emit_draw_elements(struct r300_context *r300,
 498                              struct pipe_resource* indexBuffer,
 499                              unsigned indexSize,
 500                              unsigned minIndex,
 501                              unsigned maxIndex,
 502                              unsigned mode,
 503                              unsigned start,
 504                              unsigned count)
 505 {
 506     if (!r300->stencil_ref_bf_fallback) {
 507         r500_emit_draw_elements(r300, indexBuffer, indexSize,
 508                                 minIndex, maxIndex, mode, start, count);
 509     } else {
 510         r300_begin_stencil_ref_fallback(r300);
 511         r500_emit_draw_elements(r300, indexBuffer, indexSize,
 512                                 minIndex, maxIndex, mode, start, count);
 513         r300_switch_stencil_ref_side(r300);
 514         r500_emit_draw_elements(r300, indexBuffer, indexSize,
 515                                 minIndex, maxIndex, mode, start, count);
 516         r300_end_stencil_ref_fallback(r300);
 517     }
 518 }
 519
 520 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 521                                     struct pipe_resource** elts,
 522                                     unsigned start,
 523                                     unsigned count)
 524 {
 525     struct pipe_context* context = &r300->context;
 526     struct pipe_screen* screen = r300->context.screen;
 527     struct pipe_resource* new_elts;
 528     unsigned char *in_map;
 529     unsigned short *out_map;
 530     struct pipe_transfer *src_transfer, *dst_transfer;
 531     unsigned i;
 532
 533     new_elts = pipe_buffer_create(screen,
 534                                   PIPE_BIND_INDEX_BUFFER,
 535                                   2 * count);
 536
 537     in_map = pipe_buffer_map(context, *elts, PIPE_TRANSFER_READ, &src_transfer);
 538     out_map = pipe_buffer_map(context, new_elts, PIPE_TRANSFER_WRITE, &dst_transfer);
 539
 540     in_map += start;
 541
 542     for (i = 0; i < count; i++) {
 543         *out_map = (unsigned short)*in_map;
 544         in_map++;
 545         out_map++;
 546     }
 547
 548     pipe_buffer_unmap(context, *elts, src_transfer);
 549     pipe_buffer_unmap(context, new_elts, dst_transfer);
 550
 551     *elts = new_elts;
 552 }
 553
 554 static void r300_align_ushort_elts(struct r300_context *r300,
 555                                    struct pipe_resource **elts,
 556                                    unsigned start, unsigned count)
 557 {
 558     struct pipe_context* context = &r300->context;
 559     struct pipe_transfer *in_transfer = NULL;
 560     struct pipe_transfer *out_transfer = NULL;
 561     struct pipe_resource* new_elts;
 562     unsigned short *in_map;
 563     unsigned short *out_map;
 564
 565     new_elts = pipe_buffer_create(context->screen,
 566                                   PIPE_BIND_INDEX_BUFFER,
 567                                   2 * count);
 568
 569     in_map = pipe_buffer_map(context, *elts,
 570                              PIPE_TRANSFER_READ, &in_transfer);
 571     out_map = pipe_buffer_map(context, new_elts,
 572                               PIPE_TRANSFER_WRITE, &out_transfer);
 573
 574     memcpy(out_map, in_map+start, 2 * count);
 575
 576     pipe_buffer_unmap(context, *elts, in_transfer);
 577     pipe_buffer_unmap(context, new_elts, out_transfer);
 578
 579     *elts = new_elts;
 580 }
 581
 582 /* This is the fast-path drawing & emission for HW TCL. */
 583 void r300_draw_range_elements(struct pipe_context* pipe,
 584                               struct pipe_resource* indexBuffer,
 585                               unsigned indexSize,
 586                               int indexBias,
 587                               unsigned minIndex,
 588                               unsigned maxIndex,
 589                               unsigned mode,
 590                               unsigned start,
 591                               unsigned count)
 592 {
 593     struct r300_context* r300 = r300_context(pipe);
 594     struct pipe_resource* orgIndexBuffer = indexBuffer;
 595     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 596                             count > 65536 &&
 597                             r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0);
 598     unsigned short_count;
 599
 600     if (r300->skip_rendering) {
 601         return;
 602     }
 603
 604     if (!u_trim_pipe_prim(mode, &count)) {
 605         return;
 606     }
 607
 608     if (indexSize == 1) {
 609         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 610         indexSize = 2;
 611         start = 0;
 612     } else if (indexSize == 2 && start % 2 != 0) {
 613         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 614         start = 0;
 615     }
 616
 617     r300_update_derived_state(r300);
 618     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 619
 620     /* 15 dwords for emit_draw_elements */
 621     r300_prepare_for_rendering(r300,
 622         PREP_FIRST_DRAW | PREP_VALIDATE_VBOS | PREP_EMIT_AOS | PREP_INDEXED,
 623         indexBuffer, 15, 0, indexBias);
 624
 625     u_upload_flush(r300->upload_vb);
 626     u_upload_flush(r300->upload_ib);
 627     if (alt_num_verts || count <= 65535) {
 628         r300->emit_draw_elements(r300, indexBuffer, indexSize,
 629                                  minIndex, maxIndex, mode, start, count);
 630     } else {
 631         do {
 632             short_count = MIN2(count, 65534);
 633             r300->emit_draw_elements(r300, indexBuffer, indexSize,
 634                                      minIndex, maxIndex,
 635                                      mode, start, short_count);
 636
 637             start += short_count;
 638             count -= short_count;
 639
 640             /* 15 dwords for emit_draw_elements */
 641             if (count) {
 642                 r300_prepare_for_rendering(r300,
 643                     PREP_VALIDATE_VBOS | PREP_EMIT_AOS | PREP_INDEXED,
 644                     indexBuffer, 15, 0, indexBias);
 645             }
 646         } while (count);
 647     }
 648
 649     if (indexBuffer != orgIndexBuffer) {
 650         pipe_resource_reference( &indexBuffer, NULL );
 651     }
 652 }
 653
 654 /* Simple helpers for context setup. Should probably be moved to util. */
 655 void r300_draw_elements(struct pipe_context* pipe,
 656                         struct pipe_resource* indexBuffer,
 657                         unsigned indexSize, int indexBias, unsigned mode,
 658                         unsigned start, unsigned count)
 659 {
 660     struct r300_context *r300 = r300_context(pipe);
 661
 662     pipe->draw_range_elements(pipe, indexBuffer, indexSize, indexBias,
 663                               0, r300->vertex_buffer_max_index,
 664                               mode, start, count);
 665 }
 666
 667 void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 668                       unsigned start, unsigned count)
 669 {
 670     struct r300_context* r300 = r300_context(pipe);
 671     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 672                             count > 65536 &&
 673                             r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0);
 674     unsigned short_count;
 675
 676     if (r300->skip_rendering) {
 677         return;
 678     }
 679
 680     if (!u_trim_pipe_prim(mode, &count)) {
 681         return;
 682     }
 683
 684     r300_update_derived_state(r300);
 685
 686     if (immd_is_good_idea(r300, count)) {
 687         r300->emit_draw_arrays_immediate(r300, mode, start, count);
 688     } else {
 689         /* 9 spare dwords for emit_draw_arrays. */
 690         r300_prepare_for_rendering(r300, PREP_FIRST_DRAW | PREP_VALIDATE_VBOS | PREP_EMIT_AOS,
 691                                NULL, 9, start, 0);
 692
 693         if (alt_num_verts || count <= 65535) {
 694             r300->emit_draw_arrays(r300, mode, count);
 695         } else {
 696             do {
 697                 short_count = MIN2(count, 65535);
 698                 r300->emit_draw_arrays(r300, mode, short_count);
 699
 700                 start += short_count;
 701                 count -= short_count;
 702
 703                 /* 9 spare dwords for emit_draw_arrays. */
 704                 if (count) {
 705                     r300_prepare_for_rendering(r300,
 706                         PREP_VALIDATE_VBOS | PREP_EMIT_AOS, NULL, 9,
 707                         start, 0);
 708                 }
 709             } while (count);
 710         }
 711         u_upload_flush(r300->upload_vb);
 712     }
 713 }
 714
 715 /****************************************************************************
 716  * The rest of this file is for SW TCL rendering only. Please be polite and *
 717  * keep these functions separated so that they are easier to locate. ~C.    *
 718  ***************************************************************************/
 719
 720 /* SW TCL arrays, using Draw. */
 721 void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 722                                unsigned mode,
 723                                unsigned start,
 724                                unsigned count)
 725 {
 726     struct r300_context* r300 = r300_context(pipe);
 727     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 728     int i;
 729
 730     if (r300->skip_rendering) {
 731         return;
 732     }
 733
 734     if (!u_trim_pipe_prim(mode, &count)) {
 735         return;
 736     }
 737
 738     r300_update_derived_state(r300);
 739
 740     for (i = 0; i < r300->vertex_buffer_count; i++) {
 741         void* buf = pipe_buffer_map(pipe,
 742                                     r300->vertex_buffer[i].buffer,
 743                                     PIPE_TRANSFER_READ,
 744                                     &vb_transfer[i]);
 745         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 746     }
 747
 748     draw_set_mapped_element_buffer(r300->draw, 0, 0, NULL);
 749
 750     draw_arrays(r300->draw, mode, start, count);
 751
 752     /* XXX Not sure whether this is the best fix.
 753      * It prevents CS from being rejected and weird assertion failures. */
 754     draw_flush(r300->draw);
 755
 756     for (i = 0; i < r300->vertex_buffer_count; i++) {
 757         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 758                           vb_transfer[i]);
 759         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 760     }
 761 }
 762
 763 /* SW TCL elements, using Draw. */
 764 void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 765                                        struct pipe_resource* indexBuffer,
 766                                        unsigned indexSize,
 767                                        int indexBias,
 768                                        unsigned minIndex,
 769                                        unsigned maxIndex,
 770                                        unsigned mode,
 771                                        unsigned start,
 772                                        unsigned count)
 773 {
 774     struct r300_context* r300 = r300_context(pipe);
 775     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 776     struct pipe_transfer *ib_transfer;
 777     int i;
 778     void* indices;
 779
 780     if (r300->skip_rendering) {
 781         return;
 782     }
 783
 784     if (!u_trim_pipe_prim(mode, &count)) {
 785         return;
 786     }
 787
 788     r300_update_derived_state(r300);
 789
 790     for (i = 0; i < r300->vertex_buffer_count; i++) {
 791         void* buf = pipe_buffer_map(pipe,
 792                                     r300->vertex_buffer[i].buffer,
 793                                     PIPE_TRANSFER_READ,
 794                                     &vb_transfer[i]);
 795         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 796     }
 797
 798     indices = pipe_buffer_map(pipe, indexBuffer,
 799                               PIPE_TRANSFER_READ, &ib_transfer);
 800     draw_set_mapped_element_buffer_range(r300->draw, indexSize, indexBias,
 801                                          minIndex, maxIndex, indices);
 802
 803     draw_arrays(r300->draw, mode, start, count);
 804
 805     /* XXX Not sure whether this is the best fix.
 806      * It prevents CS from being rejected and weird assertion failures. */
 807     draw_flush(r300->draw);
 808
 809     for (i = 0; i < r300->vertex_buffer_count; i++) {
 810         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 811                           vb_transfer[i]);
 812         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 813     }
 814
 815     pipe_buffer_unmap(pipe, indexBuffer,
 816                       ib_transfer);
 817     draw_set_mapped_element_buffer_range(r300->draw, 0, 0,
 818                                          start, start + count - 1,
 819                                          NULL);
 820 }
 821
 822 /* Object for rendering using Draw. */
 823 struct r300_render {
 824     /* Parent class */
 825     struct vbuf_render base;
 826
 827     /* Pipe context */
 828     struct r300_context* r300;
 829
 830     /* Vertex information */
 831     size_t vertex_size;
 832     unsigned prim;
 833     unsigned hwprim;
 834
 835     /* VBO */
 836     struct pipe_resource* vbo;
 837     size_t vbo_size;
 838     size_t vbo_offset;
 839     size_t vbo_max_used;
 840     void * vbo_ptr;
 841
 842     struct pipe_transfer *vbo_transfer;
 843 };
 844
 845 static INLINE struct r300_render*
 846 r300_render(struct vbuf_render* render)
 847 {
 848     return (struct r300_render*)render;
 849 }
 850
 851 static const struct vertex_info*
 852 r300_render_get_vertex_info(struct vbuf_render* render)
 853 {
 854     struct r300_render* r300render = r300_render(render);
 855     struct r300_context* r300 = r300render->r300;
 856
 857     return &r300->vertex_info;
 858 }
 859
 860 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 861                                                    ushort vertex_size,
 862                                                    ushort count)
 863 {
 864     struct r300_render* r300render = r300_render(render);
 865     struct r300_context* r300 = r300render->r300;
 866     struct pipe_screen* screen = r300->context.screen;
 867     size_t size = (size_t)vertex_size * (size_t)count;
 868
 869     if (size + r300render->vbo_offset > r300render->vbo_size)
 870     {
 871         pipe_resource_reference(&r300->vbo, NULL);
 872         r300render->vbo = pipe_buffer_create(screen,
 873                                              PIPE_BIND_VERTEX_BUFFER,
 874                                              R300_MAX_DRAW_VBO_SIZE);
 875         r300render->vbo_offset = 0;
 876         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 877     }
 878
 879     r300render->vertex_size = vertex_size;
 880     r300->vbo = r300render->vbo;
 881     r300->vbo_offset = r300render->vbo_offset;
 882
 883     return (r300render->vbo) ? TRUE : FALSE;
 884 }
 885
 886 static void* r300_render_map_vertices(struct vbuf_render* render)
 887 {
 888     struct r300_render* r300render = r300_render(render);
 889
 890     r300render->vbo_ptr = pipe_buffer_map(&r300render->r300->context,
 891                                           r300render->vbo,
 892                                           PIPE_TRANSFER_WRITE,
 893                                           &r300render->vbo_transfer);
 894
 895     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 896 }
 897
 898 static void r300_render_unmap_vertices(struct vbuf_render* render,
 899                                              ushort min,
 900                                              ushort max)
 901 {
 902     struct r300_render* r300render = r300_render(render);
 903     struct pipe_context* context = &r300render->r300->context;
 904     CS_LOCALS(r300render->r300);
 905     BEGIN_CS(2);
 906     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max);
 907     END_CS;
 908
 909     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 910                                     r300render->vertex_size * (max + 1));
 911     pipe_buffer_unmap(context, r300render->vbo, r300render->vbo_transfer);
 912 }
 913
 914 static void r300_render_release_vertices(struct vbuf_render* render)
 915 {
 916     struct r300_render* r300render = r300_render(render);
 917
 918     r300render->vbo_offset += r300render->vbo_max_used;
 919     r300render->vbo_max_used = 0;
 920 }
 921
 922 static boolean r300_render_set_primitive(struct vbuf_render* render,
 923                                                unsigned prim)
 924 {
 925     struct r300_render* r300render = r300_render(render);
 926
 927     r300render->prim = prim;
 928     r300render->hwprim = r300_translate_primitive(prim);
 929
 930     return TRUE;
 931 }
 932
 933 static void r500_render_draw_arrays(struct vbuf_render* render,
 934                                     unsigned start,
 935                                     unsigned count)
 936 {
 937     struct r300_render* r300render = r300_render(render);
 938     struct r300_context* r300 = r300render->r300;
 939     uint8_t* ptr;
 940     unsigned i;
 941
 942     CS_LOCALS(r300);
 943
 944     (void) i; (void) ptr;
 945
 946     r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, 4, 0, 0);
 947
 948     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 949
 950     /* Uncomment to dump all VBOs rendered through this interface.
 951      * Slow and noisy!
 952     ptr = pipe_buffer_map(&r300render->r300->context,
 953                           r300render->vbo, PIPE_TRANSFER_READ,
 954                           &r300render->vbo_transfer);
 955
 956     for (i = 0; i < count; i++) {
 957         printf("r300: Vertex %d\n", i);
 958         draw_dump_emitted_vertex(&r300->vertex_info, ptr);
 959         ptr += r300->vertex_info.size * 4;
 960         printf("\n");
 961     }
 962
 963     pipe_buffer_unmap(&r300render->r300->context, r300render->vbo,
 964         r300render->vbo_transfer);
 965     */
 966
 967     BEGIN_CS(4);
 968     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 969             r300_provoking_vertex_fixes(r300, r300render->prim));
 970     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 971     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 972            r300render->hwprim);
 973     END_CS;
 974 }
 975
 976 static void r500_render_draw_elements(struct vbuf_render* render,
 977                                       const ushort* indices,
 978                                       uint count)
 979 {
 980     struct r300_render* r300render = r300_render(render);
 981     struct r300_context* r300 = r300render->r300;
 982     int i;
 983     unsigned dwords = 4 + (count+1)/2;
 984
 985     CS_LOCALS(r300);
 986
 987     r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0);
 988
 989     BEGIN_CS(dwords);
 990     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 991             r300_provoking_vertex_fixes(r300, r300render->prim));
 992     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
 993     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 994            r300render->hwprim);
 995     for (i = 0; i < count-1; i += 2) {
 996         OUT_CS(indices[i+1] << 16 | indices[i]);
 997     }
 998     if (count % 2) {
 999         OUT_CS(indices[count-1]);
1000     }
1001     END_CS;
1002 }
1003
1004 static void r300_render_draw_arrays(struct vbuf_render* render,
1005                                     unsigned start,
1006                                     unsigned count)
1007 {
1008     struct r300_context* r300 = r300_render(render)->r300;
1009
1010     if (!r300->stencil_ref_bf_fallback) {
1011         r500_render_draw_arrays(render, start, count);
1012     } else {
1013         r300_begin_stencil_ref_fallback(r300);
1014         r500_render_draw_arrays(render, start, count);
1015         r300_switch_stencil_ref_side(r300);
1016         r500_render_draw_arrays(render, start, count);
1017         r300_end_stencil_ref_fallback(r300);
1018     }
1019 }
1020
1021 static void r300_render_draw_elements(struct vbuf_render* render,
1022                                       const ushort* indices,
1023                                       uint count)
1024 {
1025     struct r300_context* r300 = r300_render(render)->r300;
1026
1027     if (!r300->stencil_ref_bf_fallback) {
1028         r500_render_draw_elements(render, indices, count);
1029     } else {
1030         r300_begin_stencil_ref_fallback(r300);
1031         r500_render_draw_elements(render, indices, count);
1032         r300_switch_stencil_ref_side(r300);
1033         r500_render_draw_elements(render, indices, count);
1034         r300_end_stencil_ref_fallback(r300);
1035     }
1036 }
1037
1038 static void r300_render_destroy(struct vbuf_render* render)
1039 {
1040     FREE(render);
1041 }
1042
1043 static struct vbuf_render* r300_render_create(struct r300_context* r300)
1044 {
1045     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
1046
1047     r300render->r300 = r300;
1048
1049     /* XXX find real numbers plz */
1050     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
1051     r300render->base.max_indices = 16 * 1024;
1052
1053     r300render->base.get_vertex_info = r300_render_get_vertex_info;
1054     r300render->base.allocate_vertices = r300_render_allocate_vertices;
1055     r300render->base.map_vertices = r300_render_map_vertices;
1056     r300render->base.unmap_vertices = r300_render_unmap_vertices;
1057     r300render->base.set_primitive = r300_render_set_primitive;
1058     if (r300->screen->caps.is_r500) {
1059         r300render->base.draw_elements = r500_render_draw_elements;
1060         r300render->base.draw_arrays = r500_render_draw_arrays;
1061     } else {
1062         r300render->base.draw_elements = r300_render_draw_elements;
1063         r300render->base.draw_arrays = r300_render_draw_arrays;
1064     }
1065     r300render->base.release_vertices = r300_render_release_vertices;
1066     r300render->base.destroy = r300_render_destroy;
1067
1068     r300render->vbo = NULL;
1069     r300render->vbo_size = 0;
1070     r300render->vbo_offset = 0;
1071
1072     return &r300render->base;
1073 }
1074
1075 struct draw_stage* r300_draw_stage(struct r300_context* r300)
1076 {
1077     struct vbuf_render* render;
1078     struct draw_stage* stage;
1079
1080     render = r300_render_create(r300);
1081
1082     if (!render) {
1083         return NULL;
1084     }
1085
1086     stage = draw_vbuf_stage(r300->draw, render);
1087
1088     if (!stage) {
1089         render->destroy(render);
1090         return NULL;
1091     }
1092
1093     draw_set_render(r300->draw, render);
1094
1095     return stage;
1096 }