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_state_derived.h"
  42
  43 static uint32_t r300_translate_primitive(unsigned prim)
  44 {
  45     switch (prim) {
  46         case PIPE_PRIM_POINTS:
  47             return R300_VAP_VF_CNTL__PRIM_POINTS;
  48         case PIPE_PRIM_LINES:
  49             return R300_VAP_VF_CNTL__PRIM_LINES;
  50         case PIPE_PRIM_LINE_LOOP:
  51             return R300_VAP_VF_CNTL__PRIM_LINE_LOOP;
  52         case PIPE_PRIM_LINE_STRIP:
  53             return R300_VAP_VF_CNTL__PRIM_LINE_STRIP;
  54         case PIPE_PRIM_TRIANGLES:
  55             return R300_VAP_VF_CNTL__PRIM_TRIANGLES;
  56         case PIPE_PRIM_TRIANGLE_STRIP:
  57             return R300_VAP_VF_CNTL__PRIM_TRIANGLE_STRIP;
  58         case PIPE_PRIM_TRIANGLE_FAN:
  59             return R300_VAP_VF_CNTL__PRIM_TRIANGLE_FAN;
  60         case PIPE_PRIM_QUADS:
  61             return R300_VAP_VF_CNTL__PRIM_QUADS;
  62         case PIPE_PRIM_QUAD_STRIP:
  63             return R300_VAP_VF_CNTL__PRIM_QUAD_STRIP;
  64         case PIPE_PRIM_POLYGON:
  65             return R300_VAP_VF_CNTL__PRIM_POLYGON;
  66         default:
  67             return 0;
  68     }
  69 }
  70
  71 static uint32_t r300_provoking_vertex_fixes(struct r300_context *r300,
  72                                             unsigned mode)
  73 {
  74     struct r300_rs_state* rs = (struct r300_rs_state*)r300->rs_state.state;
  75     uint32_t color_control = rs->color_control;
  76
  77     /* By default (see r300_state.c:r300_create_rs_state) color_control is
  78      * initialized to provoking the first vertex.
  79      *
  80      * Triangle fans must be reduced to the second vertex, not the first, in
  81      * Gallium flatshade-first mode, as per the GL spec.
  82      * (http://www.opengl.org/registry/specs/ARB/provoking_vertex.txt)
  83      *
  84      * Quads never provoke correctly in flatshade-first mode. The first
  85      * vertex is never considered as provoking, so only the second, third,
  86      * and fourth vertices can be selected, and both "third" and "last" modes
  87      * select the fourth vertex. This is probably due to D3D lacking quads.
  88      *
  89      * Similarly, polygons reduce to the first, not the last, vertex, when in
  90      * "last" mode, and all other modes start from the second vertex.
  91      *
  92      * ~ C.
  93      */
  94
  95     if (rs->rs.flatshade_first) {
  96         switch (mode) {
  97             case PIPE_PRIM_TRIANGLE_FAN:
  98                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_SECOND;
  99                 break;
 100             case PIPE_PRIM_QUADS:
 101             case PIPE_PRIM_QUAD_STRIP:
 102             case PIPE_PRIM_POLYGON:
 103                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_LAST;
 104                 break;
 105             default:
 106                 color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_FIRST;
 107                 break;
 108         }
 109     } else {
 110         color_control |= R300_GA_COLOR_CONTROL_PROVOKING_VERTEX_LAST;
 111     }
 112
 113     return color_control;
 114 }
 115
 116 static void r500_emit_index_offset(struct r300_context *r300, int index_bias)
 117 {
 118     CS_LOCALS(r300);
 119
 120     if (r300->screen->caps.is_r500 &&
 121         r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0)) {
 122         BEGIN_CS(2);
 123         OUT_CS_REG(R500_VAP_INDEX_OFFSET,
 124                    (index_bias & 0xFFFFFF) | (index_bias < 0 ? 1<<24 : 0));
 125         END_CS;
 126     } else {
 127         if (index_bias) {
 128             fprintf(stderr, "r300: Non-zero index bias is unsupported "
 129                             "on this hardware.\n");
 130             assert(0);
 131         }
 132     }
 133 }
 134
 135 enum r300_prepare_flags {
 136     PREP_FIRST_DRAW     = (1 << 0),
 137     PREP_VALIDATE_VBOS  = (1 << 1),
 138     PREP_EMIT_AOS       = (1 << 2),
 139     PREP_EMIT_AOS_SWTCL = (1 << 3),
 140     PREP_INDEXED        = (1 << 4)
 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     boolean emit_aos_swtcl = flags & PREP_EMIT_AOS_SWTCL;
 157
 158     /* Stencil ref fallback. */
 159     if (r300->stencil_ref_bf_fallback) {
 160         cs_dwords = cs_dwords * 2 + 10;
 161     }
 162
 163     /* Add dirty state, index offset, and AOS. */
 164     if (first_draw) {
 165         cs_dwords += r300_get_num_dirty_dwords(r300);
 166
 167         if (r300->screen->caps.is_r500)
 168             cs_dwords += 2; /* emit_index_offset */
 169
 170         if (emit_aos)
 171             cs_dwords += 55; /* emit_aos */
 172
 173         if (emit_aos_swtcl)
 174             cs_dwords += 7; /* emit_aos_swtcl */
 175     }
 176
 177     /* Emitted in flush. */
 178     cs_dwords += 26; /* emit_query_end */
 179
 180     /* Reserve requested CS space. */
 181     if (!r300_check_cs(r300, cs_dwords)) {
 182         r300->context.flush(&r300->context, 0, NULL);
 183         flushed = TRUE;
 184     }
 185
 186     /* Validate buffers and emit dirty state if needed. */
 187     if (first_draw || flushed) {
 188         r300_emit_buffer_validate(r300, flags & PREP_VALIDATE_VBOS, index_buffer);
 189         r300_emit_dirty_state(r300);
 190         r500_emit_index_offset(r300, index_bias);
 191         if (emit_aos)
 192             r300_emit_aos(r300, aos_offset, flags & PREP_INDEXED);
 193         if (emit_aos_swtcl)
 194             r300_emit_aos_swtcl(r300, flags & PREP_INDEXED);
 195     }
 196 }
 197
 198 static boolean immd_is_good_idea(struct r300_context *r300,
 199                                  unsigned count)
 200 {
 201     struct pipe_vertex_element* velem;
 202     struct pipe_vertex_buffer* vbuf;
 203     boolean checked[PIPE_MAX_ATTRIBS] = {0};
 204     unsigned vertex_element_count = r300->velems->count;
 205     unsigned i, vbi;
 206
 207     if (DBG_ON(r300, DBG_NO_IMMD)) {
 208         return FALSE;
 209     }
 210
 211     if (r300->draw) {
 212         return FALSE;
 213     }
 214
 215     if (count > 10) {
 216         return FALSE;
 217     }
 218
 219     /* We shouldn't map buffers referenced by CS, busy buffers,
 220      * and ones placed in VRAM. */
 221     /* XXX Check for VRAM buffers. */
 222     for (i = 0; i < vertex_element_count; i++) {
 223         velem = &r300->velems->velem[i];
 224         vbi = velem->vertex_buffer_index;
 225
 226         if (!checked[vbi]) {
 227             vbuf = &r300->vertex_buffer[vbi];
 228
 229             if (r300_buffer_is_referenced(&r300->context,
 230                                           vbuf->buffer,
 231                                           R300_REF_CS | R300_REF_HW)) {
 232                 /* It's a very bad idea to map it... */
 233                 return FALSE;
 234             }
 235             checked[vbi] = TRUE;
 236         }
 237     }
 238     return TRUE;
 239 }
 240
 241 /*****************************************************************************
 242  * The emission of draw packets for r500. Older GPUs may use these functions *
 243  * after resolving fallback issues (e.g. stencil ref two-sided).             *
 244  ****************************************************************************/
 245
 246 static void r500_emit_draw_arrays_immediate(struct r300_context *r300,
 247                                             unsigned mode,
 248                                             unsigned start,
 249                                             unsigned count)
 250 {
 251     struct pipe_vertex_element* velem;
 252     struct pipe_vertex_buffer* vbuf;
 253     unsigned vertex_element_count = r300->velems->count;
 254     unsigned i, v, vbi, dw, elem_offset, dwords;
 255
 256     /* Size of the vertex, in dwords. */
 257     unsigned vertex_size = 0;
 258
 259     /* Offsets of the attribute, in dwords, from the start of the vertex. */
 260     unsigned offset[PIPE_MAX_ATTRIBS];
 261
 262     /* Size of the vertex element, in dwords. */
 263     unsigned size[PIPE_MAX_ATTRIBS];
 264
 265     /* Stride to the same attrib in the next vertex in the vertex buffer,
 266      * in dwords. */
 267     unsigned stride[PIPE_MAX_ATTRIBS] = {0};
 268
 269     /* Mapped vertex buffers. */
 270     uint32_t* map[PIPE_MAX_ATTRIBS] = {0};
 271     struct pipe_transfer* transfer[PIPE_MAX_ATTRIBS] = {NULL};
 272
 273     CS_LOCALS(r300);
 274
 275     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
 276     for (i = 0; i < vertex_element_count; i++) {
 277         velem = &r300->velems->velem[i];
 278         offset[i] = velem->src_offset / 4;
 279         size[i] = util_format_get_blocksize(velem->src_format) / 4;
 280         vertex_size += size[i];
 281         vbi = velem->vertex_buffer_index;
 282
 283         /* Map the buffer. */
 284         if (!map[vbi]) {
 285             vbuf = &r300->vertex_buffer[vbi];
 286             map[vbi] = (uint32_t*)pipe_buffer_map(&r300->context,
 287                                                   vbuf->buffer,
 288                                                   PIPE_TRANSFER_READ,
 289                                                   &transfer[vbi]);
 290             map[vbi] += vbuf->buffer_offset / 4;
 291             stride[vbi] = vbuf->stride / 4;
 292         }
 293     }
 294
 295     dwords = 9 + count * vertex_size;
 296
 297     r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0);
 298
 299     BEGIN_CS(dwords);
 300     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 301             r300_provoking_vertex_fixes(r300, mode));
 302     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
 303     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 304     OUT_CS(count - 1);
 305     OUT_CS(0);
 306     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
 307     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
 308             r300_translate_primitive(mode));
 309
 310     /* Emit vertices. */
 311     for (v = 0; v < count; v++) {
 312         for (i = 0; i < vertex_element_count; i++) {
 313             velem = &r300->velems->velem[i];
 314             vbi = velem->vertex_buffer_index;
 315             elem_offset = offset[i] + stride[vbi] * (v + start);
 316
 317             for (dw = 0; dw < size[i]; dw++) {
 318                 OUT_CS(map[vbi][elem_offset + dw]);
 319             }
 320         }
 321     }
 322     END_CS;
 323
 324     /* Unmap buffers. */
 325     for (i = 0; i < vertex_element_count; i++) {
 326         vbi = r300->velems->velem[i].vertex_buffer_index;
 327
 328         if (map[vbi]) {
 329             vbuf = &r300->vertex_buffer[vbi];
 330             pipe_buffer_unmap(&r300->context, vbuf->buffer, transfer[vbi]);
 331             map[vbi] = NULL;
 332         }
 333     }
 334 }
 335
 336 static void r500_emit_draw_arrays(struct r300_context *r300,
 337                                   unsigned mode,
 338                                   unsigned count)
 339 {
 340     boolean alt_num_verts = count > 65535;
 341     CS_LOCALS(r300);
 342
 343     if (count >= (1 << 24)) {
 344         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 345                 "refusing to render.\n", count);
 346         return;
 347     }
 348
 349     BEGIN_CS(7 + (alt_num_verts ? 2 : 0));
 350     if (alt_num_verts) {
 351         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 352     }
 353     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 354             r300_provoking_vertex_fixes(r300, mode));
 355     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 356     OUT_CS(count - 1);
 357     OUT_CS(0);
 358     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 359     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 360            r300_translate_primitive(mode) |
 361            (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 362     END_CS;
 363 }
 364
 365 static void r500_emit_draw_elements(struct r300_context *r300,
 366                                     struct pipe_resource* indexBuffer,
 367                                     unsigned indexSize,
 368                                     unsigned minIndex,
 369                                     unsigned maxIndex,
 370                                     unsigned mode,
 371                                     unsigned start,
 372                                     unsigned count)
 373 {
 374     uint32_t count_dwords;
 375     uint32_t offset_dwords = indexSize * start / sizeof(uint32_t);
 376     boolean alt_num_verts = count > 65535;
 377     CS_LOCALS(r300);
 378
 379     if (count >= (1 << 24)) {
 380         fprintf(stderr, "r300: Got a huge number of vertices: %i, "
 381                 "refusing to render.\n", count);
 382         return;
 383     }
 384
 385     maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
 386
 387     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
 388         count, minIndex, maxIndex);
 389
 390     BEGIN_CS(13 + (alt_num_verts ? 2 : 0));
 391     if (alt_num_verts) {
 392         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
 393     }
 394     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 395             r300_provoking_vertex_fixes(r300, mode));
 396     OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
 397     OUT_CS(maxIndex);
 398     OUT_CS(minIndex);
 399     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 400     if (indexSize == 4) {
 401         count_dwords = count;
 402         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 403                R300_VAP_VF_CNTL__INDEX_SIZE_32bit |
 404                r300_translate_primitive(mode) |
 405                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 406     } else {
 407         count_dwords = (count + 1) / 2;
 408         OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
 409                r300_translate_primitive(mode) |
 410                (alt_num_verts ? R500_VAP_VF_CNTL__USE_ALT_NUM_VERTS : 0));
 411     }
 412
 413     /* INDX_BUFFER is a truly special packet3.
 414      * Unlike most other packet3, where the offset is after the count,
 415      * the order is reversed, so the relocation ends up carrying the
 416      * size of the indexbuf instead of the offset.
 417      */
 418     OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
 419     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
 420            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
 421     OUT_CS(offset_dwords << 2);
 422     OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
 423                      RADEON_GEM_DOMAIN_GTT, 0, 0);
 424
 425     END_CS;
 426 }
 427
 428 /*****************************************************************************
 429  * The emission of draw packets for r300 which take care of the two-sided    *
 430  * stencil ref fallback and call r500's functions.                           *
 431  ****************************************************************************/
 432
 433 /* Set drawing for front faces. */
 434 static void r300_begin_stencil_ref_fallback(struct r300_context *r300)
 435 {
 436     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 437     CS_LOCALS(r300);
 438
 439     BEGIN_CS(2);
 440     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode | R300_CULL_BACK);
 441     END_CS;
 442 }
 443
 444 /* Set drawing for back faces. */
 445 static void r300_switch_stencil_ref_side(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 | R300_CULL_FRONT);
 453     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 454                dsa->stencil_ref_bf | r300->stencil_ref.ref_value[1]);
 455     END_CS;
 456 }
 457
 458 /* Restore the original state. */
 459 static void r300_end_stencil_ref_fallback(struct r300_context *r300)
 460 {
 461     struct r300_rs_state *rs = (struct r300_rs_state*)r300->rs_state.state;
 462     struct r300_dsa_state *dsa = (struct r300_dsa_state*)r300->dsa_state.state;
 463     CS_LOCALS(r300);
 464
 465     BEGIN_CS(4);
 466     OUT_CS_REG(R300_SU_CULL_MODE, rs->cull_mode);
 467     OUT_CS_REG(R300_ZB_STENCILREFMASK,
 468                dsa->stencil_ref_mask | r300->stencil_ref.ref_value[0]);
 469     END_CS;
 470 }
 471
 472 static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 473                                             unsigned mode,
 474                                             unsigned start,
 475                                             unsigned count)
 476 {
 477     if (!r300->stencil_ref_bf_fallback) {
 478         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 479     } else {
 480         r300_begin_stencil_ref_fallback(r300);
 481         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 482         r300_switch_stencil_ref_side(r300);
 483         r500_emit_draw_arrays_immediate(r300, mode, start, count);
 484         r300_end_stencil_ref_fallback(r300);
 485     }
 486 }
 487
 488 static void r300_emit_draw_arrays(struct r300_context *r300,
 489                                   unsigned mode,
 490                                   unsigned count)
 491 {
 492     if (!r300->stencil_ref_bf_fallback) {
 493         r500_emit_draw_arrays(r300, mode, count);
 494     } else {
 495         r300_begin_stencil_ref_fallback(r300);
 496         r500_emit_draw_arrays(r300, mode, count);
 497         r300_switch_stencil_ref_side(r300);
 498         r500_emit_draw_arrays(r300, mode, count);
 499         r300_end_stencil_ref_fallback(r300);
 500     }
 501 }
 502
 503 static void r300_emit_draw_elements(struct r300_context *r300,
 504                                     struct pipe_resource* indexBuffer,
 505                                     unsigned indexSize,
 506                                     unsigned minIndex,
 507                                     unsigned maxIndex,
 508                                     unsigned mode,
 509                                     unsigned start,
 510                                     unsigned count)
 511 {
 512     if (!r300->stencil_ref_bf_fallback) {
 513         r500_emit_draw_elements(r300, indexBuffer, indexSize,
 514                                 minIndex, maxIndex, mode, start, count);
 515     } else {
 516         r300_begin_stencil_ref_fallback(r300);
 517         r500_emit_draw_elements(r300, indexBuffer, indexSize,
 518                                 minIndex, maxIndex, mode, start, count);
 519         r300_switch_stencil_ref_side(r300);
 520         r500_emit_draw_elements(r300, indexBuffer, indexSize,
 521                                 minIndex, maxIndex, mode, start, count);
 522         r300_end_stencil_ref_fallback(r300);
 523     }
 524 }
 525
 526 static void r300_shorten_ubyte_elts(struct r300_context* r300,
 527                                     struct pipe_resource** elts,
 528                                     unsigned start,
 529                                     unsigned count)
 530 {
 531     struct pipe_context* context = &r300->context;
 532     struct pipe_screen* screen = r300->context.screen;
 533     struct pipe_resource* new_elts;
 534     unsigned char *in_map;
 535     unsigned short *out_map;
 536     struct pipe_transfer *src_transfer, *dst_transfer;
 537     unsigned i;
 538
 539     new_elts = pipe_buffer_create(screen,
 540                                   PIPE_BIND_INDEX_BUFFER,
 541                                   2 * count);
 542
 543     in_map = pipe_buffer_map(context, *elts, PIPE_TRANSFER_READ, &src_transfer);
 544     out_map = pipe_buffer_map(context, new_elts, PIPE_TRANSFER_WRITE, &dst_transfer);
 545
 546     in_map += start;
 547
 548     for (i = 0; i < count; i++) {
 549         *out_map = (unsigned short)*in_map;
 550         in_map++;
 551         out_map++;
 552     }
 553
 554     pipe_buffer_unmap(context, *elts, src_transfer);
 555     pipe_buffer_unmap(context, new_elts, dst_transfer);
 556
 557     *elts = new_elts;
 558 }
 559
 560 static void r300_align_ushort_elts(struct r300_context *r300,
 561                                    struct pipe_resource **elts,
 562                                    unsigned start, unsigned count)
 563 {
 564     struct pipe_context* context = &r300->context;
 565     struct pipe_transfer *in_transfer = NULL;
 566     struct pipe_transfer *out_transfer = NULL;
 567     struct pipe_resource* new_elts;
 568     unsigned short *in_map;
 569     unsigned short *out_map;
 570
 571     new_elts = pipe_buffer_create(context->screen,
 572                                   PIPE_BIND_INDEX_BUFFER,
 573                                   2 * count);
 574
 575     in_map = pipe_buffer_map(context, *elts,
 576                              PIPE_TRANSFER_READ, &in_transfer);
 577     out_map = pipe_buffer_map(context, new_elts,
 578                               PIPE_TRANSFER_WRITE, &out_transfer);
 579
 580     memcpy(out_map, in_map+start, 2 * count);
 581
 582     pipe_buffer_unmap(context, *elts, in_transfer);
 583     pipe_buffer_unmap(context, new_elts, out_transfer);
 584
 585     *elts = new_elts;
 586 }
 587
 588 /* This is the fast-path drawing & emission for HW TCL. */
 589 static void r300_draw_range_elements(struct pipe_context* pipe,
 590                                      struct pipe_resource* indexBuffer,
 591                                      unsigned indexSize,
 592                                      int indexBias,
 593                                      unsigned minIndex,
 594                                      unsigned maxIndex,
 595                                      unsigned mode,
 596                                      unsigned start,
 597                                      unsigned count)
 598 {
 599     struct r300_context* r300 = r300_context(pipe);
 600     struct pipe_resource* orgIndexBuffer = indexBuffer;
 601     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 602                             count > 65536 &&
 603                             r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0);
 604     unsigned short_count;
 605
 606     if (r300->skip_rendering) {
 607         return;
 608     }
 609
 610     if (!u_trim_pipe_prim(mode, &count)) {
 611         return;
 612     }
 613
 614     if (indexSize == 1) {
 615         r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
 616         indexSize = 2;
 617         start = 0;
 618     } else if (indexSize == 2 && start % 2 != 0) {
 619         r300_align_ushort_elts(r300, &indexBuffer, start, count);
 620         start = 0;
 621     }
 622
 623     r300_update_derived_state(r300);
 624     r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
 625
 626     /* 15 dwords for emit_draw_elements */
 627     r300_prepare_for_rendering(r300,
 628         PREP_FIRST_DRAW | PREP_VALIDATE_VBOS | PREP_EMIT_AOS | PREP_INDEXED,
 629         indexBuffer, 15, 0, indexBias);
 630
 631     u_upload_flush(r300->upload_vb);
 632     u_upload_flush(r300->upload_ib);
 633     if (alt_num_verts || count <= 65535) {
 634         r300->emit_draw_elements(r300, indexBuffer, indexSize,
 635                                  minIndex, maxIndex, mode, start, count);
 636     } else {
 637         do {
 638             short_count = MIN2(count, 65534);
 639             r300->emit_draw_elements(r300, indexBuffer, indexSize,
 640                                      minIndex, maxIndex,
 641                                      mode, start, short_count);
 642
 643             start += short_count;
 644             count -= short_count;
 645
 646             /* 15 dwords for emit_draw_elements */
 647             if (count) {
 648                 r300_prepare_for_rendering(r300,
 649                     PREP_VALIDATE_VBOS | PREP_EMIT_AOS | PREP_INDEXED,
 650                     indexBuffer, 15, 0, indexBias);
 651             }
 652         } while (count);
 653     }
 654
 655     if (indexBuffer != orgIndexBuffer) {
 656         pipe_resource_reference( &indexBuffer, NULL );
 657     }
 658 }
 659
 660 /* Simple helpers for context setup. Should probably be moved to util. */
 661 static void r300_draw_elements(struct pipe_context* pipe,
 662                                struct pipe_resource* indexBuffer,
 663                                unsigned indexSize, int indexBias, unsigned mode,
 664                                unsigned start, unsigned count)
 665 {
 666     struct r300_context *r300 = r300_context(pipe);
 667
 668     pipe->draw_range_elements(pipe, indexBuffer, indexSize, indexBias,
 669                               0, r300->vertex_buffer_max_index,
 670                               mode, start, count);
 671 }
 672
 673 static void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
 674                              unsigned start, unsigned count)
 675 {
 676     struct r300_context* r300 = r300_context(pipe);
 677     boolean alt_num_verts = r300->screen->caps.is_r500 &&
 678                             count > 65536 &&
 679                             r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0);
 680     unsigned short_count;
 681
 682     if (r300->skip_rendering) {
 683         return;
 684     }
 685
 686     if (!u_trim_pipe_prim(mode, &count)) {
 687         return;
 688     }
 689
 690     r300_update_derived_state(r300);
 691
 692     if (immd_is_good_idea(r300, count)) {
 693         r300->emit_draw_arrays_immediate(r300, mode, start, count);
 694     } else {
 695         /* 9 spare dwords for emit_draw_arrays. */
 696         r300_prepare_for_rendering(r300, PREP_FIRST_DRAW | PREP_VALIDATE_VBOS | PREP_EMIT_AOS,
 697                                NULL, 9, start, 0);
 698
 699         if (alt_num_verts || count <= 65535) {
 700             r300->emit_draw_arrays(r300, mode, count);
 701         } else {
 702             do {
 703                 short_count = MIN2(count, 65535);
 704                 r300->emit_draw_arrays(r300, mode, short_count);
 705
 706                 start += short_count;
 707                 count -= short_count;
 708
 709                 /* 9 spare dwords for emit_draw_arrays. */
 710                 if (count) {
 711                     r300_prepare_for_rendering(r300,
 712                         PREP_VALIDATE_VBOS | PREP_EMIT_AOS, NULL, 9,
 713                         start, 0);
 714                 }
 715             } while (count);
 716         }
 717         u_upload_flush(r300->upload_vb);
 718     }
 719 }
 720
 721 /****************************************************************************
 722  * The rest of this file is for SW TCL rendering only. Please be polite and *
 723  * keep these functions separated so that they are easier to locate. ~C.    *
 724  ***************************************************************************/
 725
 726 /* SW TCL arrays, using Draw. */
 727 static void r300_swtcl_draw_arrays(struct pipe_context* pipe,
 728                                    unsigned mode,
 729                                    unsigned start,
 730                                    unsigned count)
 731 {
 732     struct r300_context* r300 = r300_context(pipe);
 733     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 734     int i;
 735
 736     if (r300->skip_rendering) {
 737         return;
 738     }
 739
 740     if (!u_trim_pipe_prim(mode, &count)) {
 741         return;
 742     }
 743
 744     r300_update_derived_state(r300);
 745
 746     for (i = 0; i < r300->vertex_buffer_count; i++) {
 747         void* buf = pipe_buffer_map(pipe,
 748                                     r300->vertex_buffer[i].buffer,
 749                                     PIPE_TRANSFER_READ,
 750                                     &vb_transfer[i]);
 751         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 752     }
 753
 754     draw_set_mapped_element_buffer(r300->draw, 0, 0, NULL);
 755
 756     draw_arrays(r300->draw, mode, start, count);
 757
 758     /* XXX Not sure whether this is the best fix.
 759      * It prevents CS from being rejected and weird assertion failures. */
 760     draw_flush(r300->draw);
 761
 762     for (i = 0; i < r300->vertex_buffer_count; i++) {
 763         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 764                           vb_transfer[i]);
 765         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 766     }
 767 }
 768
 769 /* SW TCL elements, using Draw. */
 770 static void r300_swtcl_draw_range_elements(struct pipe_context* pipe,
 771                                            struct pipe_resource* indexBuffer,
 772                                            unsigned indexSize,
 773                                            int indexBias,
 774                                            unsigned minIndex,
 775                                            unsigned maxIndex,
 776                                            unsigned mode,
 777                                            unsigned start,
 778                                            unsigned count)
 779 {
 780     struct r300_context* r300 = r300_context(pipe);
 781     struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 782     struct pipe_transfer *ib_transfer;
 783     int i;
 784     void* indices;
 785
 786     if (r300->skip_rendering) {
 787         return;
 788     }
 789
 790     if (!u_trim_pipe_prim(mode, &count)) {
 791         return;
 792     }
 793
 794     r300_update_derived_state(r300);
 795
 796     for (i = 0; i < r300->vertex_buffer_count; i++) {
 797         void* buf = pipe_buffer_map(pipe,
 798                                     r300->vertex_buffer[i].buffer,
 799                                     PIPE_TRANSFER_READ,
 800                                     &vb_transfer[i]);
 801         draw_set_mapped_vertex_buffer(r300->draw, i, buf);
 802     }
 803
 804     indices = pipe_buffer_map(pipe, indexBuffer,
 805                               PIPE_TRANSFER_READ, &ib_transfer);
 806     draw_set_mapped_element_buffer_range(r300->draw, indexSize, indexBias,
 807                                          minIndex, maxIndex, indices);
 808
 809     draw_arrays(r300->draw, mode, start, count);
 810
 811     /* XXX Not sure whether this is the best fix.
 812      * It prevents CS from being rejected and weird assertion failures. */
 813     draw_flush(r300->draw);
 814
 815     for (i = 0; i < r300->vertex_buffer_count; i++) {
 816         pipe_buffer_unmap(pipe, r300->vertex_buffer[i].buffer,
 817                           vb_transfer[i]);
 818         draw_set_mapped_vertex_buffer(r300->draw, i, NULL);
 819     }
 820
 821     pipe_buffer_unmap(pipe, indexBuffer,
 822                       ib_transfer);
 823     draw_set_mapped_element_buffer_range(r300->draw, 0, 0,
 824                                          start, start + count - 1,
 825                                          NULL);
 826 }
 827
 828 /* Object for rendering using Draw. */
 829 struct r300_render {
 830     /* Parent class */
 831     struct vbuf_render base;
 832
 833     /* Pipe context */
 834     struct r300_context* r300;
 835
 836     /* Vertex information */
 837     size_t vertex_size;
 838     unsigned prim;
 839     unsigned hwprim;
 840
 841     /* VBO */
 842     struct pipe_resource* vbo;
 843     size_t vbo_size;
 844     size_t vbo_offset;
 845     size_t vbo_max_used;
 846     void * vbo_ptr;
 847
 848     struct pipe_transfer *vbo_transfer;
 849 };
 850
 851 static INLINE struct r300_render*
 852 r300_render(struct vbuf_render* render)
 853 {
 854     return (struct r300_render*)render;
 855 }
 856
 857 static const struct vertex_info*
 858 r300_render_get_vertex_info(struct vbuf_render* render)
 859 {
 860     struct r300_render* r300render = r300_render(render);
 861     struct r300_context* r300 = r300render->r300;
 862
 863     return &r300->vertex_info;
 864 }
 865
 866 static boolean r300_render_allocate_vertices(struct vbuf_render* render,
 867                                                    ushort vertex_size,
 868                                                    ushort count)
 869 {
 870     struct r300_render* r300render = r300_render(render);
 871     struct r300_context* r300 = r300render->r300;
 872     struct pipe_screen* screen = r300->context.screen;
 873     size_t size = (size_t)vertex_size * (size_t)count;
 874
 875     if (size + r300render->vbo_offset > r300render->vbo_size)
 876     {
 877         pipe_resource_reference(&r300->vbo, NULL);
 878         r300render->vbo = pipe_buffer_create(screen,
 879                                              PIPE_BIND_VERTEX_BUFFER,
 880                                              R300_MAX_DRAW_VBO_SIZE);
 881         r300render->vbo_offset = 0;
 882         r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
 883     }
 884
 885     r300render->vertex_size = vertex_size;
 886     r300->vbo = r300render->vbo;
 887     r300->vbo_offset = r300render->vbo_offset;
 888
 889     return (r300render->vbo) ? TRUE : FALSE;
 890 }
 891
 892 static void* r300_render_map_vertices(struct vbuf_render* render)
 893 {
 894     struct r300_render* r300render = r300_render(render);
 895
 896     assert(!r300render->vbo_transfer);
 897
 898     r300render->vbo_ptr = pipe_buffer_map(&r300render->r300->context,
 899                                           r300render->vbo,
 900                                           PIPE_TRANSFER_WRITE,
 901                                           &r300render->vbo_transfer);
 902
 903     return ((uint8_t*)r300render->vbo_ptr + r300render->vbo_offset);
 904 }
 905
 906 static void r300_render_unmap_vertices(struct vbuf_render* render,
 907                                              ushort min,
 908                                              ushort max)
 909 {
 910     struct r300_render* r300render = r300_render(render);
 911     struct pipe_context* context = &r300render->r300->context;
 912
 913     assert(r300render->vbo_transfer);
 914
 915     r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
 916                                     r300render->vertex_size * (max + 1));
 917     pipe_buffer_unmap(context, r300render->vbo, r300render->vbo_transfer);
 918
 919     r300render->vbo_transfer = NULL;
 920 }
 921
 922 static void r300_render_release_vertices(struct vbuf_render* render)
 923 {
 924     struct r300_render* r300render = r300_render(render);
 925
 926     r300render->vbo_offset += r300render->vbo_max_used;
 927     r300render->vbo_max_used = 0;
 928 }
 929
 930 static boolean r300_render_set_primitive(struct vbuf_render* render,
 931                                                unsigned prim)
 932 {
 933     struct r300_render* r300render = r300_render(render);
 934
 935     r300render->prim = prim;
 936     r300render->hwprim = r300_translate_primitive(prim);
 937
 938     return TRUE;
 939 }
 940
 941 static void r500_render_draw_arrays(struct vbuf_render* render,
 942                                     unsigned start,
 943                                     unsigned count)
 944 {
 945     struct r300_render* r300render = r300_render(render);
 946     struct r300_context* r300 = r300render->r300;
 947     uint8_t* ptr;
 948     unsigned i;
 949     unsigned dwords = 6;
 950
 951     CS_LOCALS(r300);
 952
 953     (void) i; (void) ptr;
 954
 955     r300_prepare_for_rendering(r300, PREP_FIRST_DRAW | PREP_EMIT_AOS_SWTCL,
 956                                NULL, dwords, 0, 0);
 957
 958     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 959
 960     /* Uncomment to dump all VBOs rendered through this interface.
 961      * Slow and noisy!
 962     ptr = pipe_buffer_map(&r300render->r300->context,
 963                           r300render->vbo, PIPE_TRANSFER_READ,
 964                           &r300render->vbo_transfer);
 965
 966     for (i = 0; i < count; i++) {
 967         printf("r300: Vertex %d\n", i);
 968         draw_dump_emitted_vertex(&r300->vertex_info, ptr);
 969         ptr += r300->vertex_info.size * 4;
 970         printf("\n");
 971     }
 972
 973     pipe_buffer_unmap(&r300render->r300->context, r300render->vbo,
 974         r300render->vbo_transfer);
 975     */
 976
 977     BEGIN_CS(dwords);
 978     OUT_CS_REG(R300_GA_COLOR_CONTROL,
 979             r300_provoking_vertex_fixes(r300, r300render->prim));
 980     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, count - 1);
 981     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 982     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
 983            r300render->hwprim);
 984     END_CS;
 985 }
 986
 987 static void r500_render_draw_elements(struct vbuf_render* render,
 988                                       const ushort* indices,
 989                                       uint count)
 990 {
 991     struct r300_render* r300render = r300_render(render);
 992     struct r300_context* r300 = r300render->r300;
 993     int i;
 994     unsigned dwords = 6 + (count+1)/2;
 995     unsigned max_index = (r300render->vbo_size - r300render->vbo_offset) /
 996                          (r300render->r300->vertex_info.size * 4) - 1;
 997
 998     CS_LOCALS(r300);
 999
1000     r300_prepare_for_rendering(r300,
1001         PREP_FIRST_DRAW | PREP_EMIT_AOS_SWTCL | PREP_INDEXED,
1002         NULL, dwords, 0, 0);
1003
1004     BEGIN_CS(dwords);
1005     OUT_CS_REG(R300_GA_COLOR_CONTROL,
1006             r300_provoking_vertex_fixes(r300, r300render->prim));
1007     OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max_index);
1008     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
1009     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
1010            r300render->hwprim);
1011     for (i = 0; i < count-1; i += 2) {
1012         OUT_CS(indices[i+1] << 16 | indices[i]);
1013     }
1014     if (count % 2) {
1015         OUT_CS(indices[count-1]);
1016     }
1017     END_CS;
1018 }
1019
1020 static void r300_render_draw_arrays(struct vbuf_render* render,
1021                                     unsigned start,
1022                                     unsigned count)
1023 {
1024     struct r300_context* r300 = r300_render(render)->r300;
1025
1026     if (!r300->stencil_ref_bf_fallback) {
1027         r500_render_draw_arrays(render, start, count);
1028     } else {
1029         r300_begin_stencil_ref_fallback(r300);
1030         r500_render_draw_arrays(render, start, count);
1031         r300_switch_stencil_ref_side(r300);
1032         r500_render_draw_arrays(render, start, count);
1033         r300_end_stencil_ref_fallback(r300);
1034     }
1035 }
1036
1037 static void r300_render_draw_elements(struct vbuf_render* render,
1038                                       const ushort* indices,
1039                                       uint count)
1040 {
1041     struct r300_context* r300 = r300_render(render)->r300;
1042
1043     if (!r300->stencil_ref_bf_fallback) {
1044         r500_render_draw_elements(render, indices, count);
1045     } else {
1046         r300_begin_stencil_ref_fallback(r300);
1047         r500_render_draw_elements(render, indices, count);
1048         r300_switch_stencil_ref_side(r300);
1049         r500_render_draw_elements(render, indices, count);
1050         r300_end_stencil_ref_fallback(r300);
1051     }
1052 }
1053
1054 static void r300_render_destroy(struct vbuf_render* render)
1055 {
1056     FREE(render);
1057 }
1058
1059 static struct vbuf_render* r300_render_create(struct r300_context* r300)
1060 {
1061     struct r300_render* r300render = CALLOC_STRUCT(r300_render);
1062
1063     r300render->r300 = r300;
1064
1065     /* XXX find real numbers plz */
1066     r300render->base.max_vertex_buffer_bytes = 128 * 1024;
1067     r300render->base.max_indices = 16 * 1024;
1068
1069     r300render->base.get_vertex_info = r300_render_get_vertex_info;
1070     r300render->base.allocate_vertices = r300_render_allocate_vertices;
1071     r300render->base.map_vertices = r300_render_map_vertices;
1072     r300render->base.unmap_vertices = r300_render_unmap_vertices;
1073     r300render->base.set_primitive = r300_render_set_primitive;
1074     if (r300->screen->caps.is_r500) {
1075         r300render->base.draw_elements = r500_render_draw_elements;
1076         r300render->base.draw_arrays = r500_render_draw_arrays;
1077     } else {
1078         r300render->base.draw_elements = r300_render_draw_elements;
1079         r300render->base.draw_arrays = r300_render_draw_arrays;
1080     }
1081     r300render->base.release_vertices = r300_render_release_vertices;
1082     r300render->base.destroy = r300_render_destroy;
1083
1084     r300render->vbo = NULL;
1085     r300render->vbo_size = 0;
1086     r300render->vbo_offset = 0;
1087
1088     return &r300render->base;
1089 }
1090
1091 struct draw_stage* r300_draw_stage(struct r300_context* r300)
1092 {
1093     struct vbuf_render* render;
1094     struct draw_stage* stage;
1095
1096     render = r300_render_create(r300);
1097
1098     if (!render) {
1099         return NULL;
1100     }
1101
1102     stage = draw_vbuf_stage(r300->draw, render);
1103
1104     if (!stage) {
1105         render->destroy(render);
1106         return NULL;
1107     }
1108
1109     draw_set_render(r300->draw, render);
1110
1111     return stage;
1112 }
1113
1114 void r300_init_render_functions(struct r300_context *r300)
1115 {
1116     if (r300->screen->caps.has_tcl) {
1117         r300->context.draw_arrays = r300_draw_arrays;
1118         r300->context.draw_elements = r300_draw_elements;
1119         r300->context.draw_range_elements = r300_draw_range_elements;
1120
1121         if (r300->screen->caps.is_r500) {
1122             r300->emit_draw_arrays_immediate = r500_emit_draw_arrays_immediate;
1123             r300->emit_draw_arrays = r500_emit_draw_arrays;
1124             r300->emit_draw_elements = r500_emit_draw_elements;
1125         } else {
1126             r300->emit_draw_arrays_immediate = r300_emit_draw_arrays_immediate;
1127             r300->emit_draw_arrays = r300_emit_draw_arrays;
1128             r300->emit_draw_elements = r300_emit_draw_elements;
1129         }
1130     } else {
1131         r300->context.draw_arrays = r300_swtcl_draw_arrays;
1132         r300->context.draw_elements = r300_draw_elements;
1133         r300->context.draw_range_elements = r300_swtcl_draw_range_elements;
1134     }
1135 }