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