src/gallium/drivers/r300/r300_state_derived.c

   1 /*
   2  * Copyright 2008 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 #include "draw/draw_context.h"
  24
  25 #include "util/u_math.h"
  26 #include "util/u_memory.h"
  27
  28 #include "r300_context.h"
  29 #include "r300_fs.h"
  30 #include "r300_screen.h"
  31 #include "r300_state_derived.h"
  32 #include "r300_state_inlines.h"
  33 #include "r300_vs.h"
  34
  35 /* r300_state_derived: Various bits of state which are dependent upon
  36  * currently bound CSO data. */
  37
  38 struct r300_shader_key {
  39     struct r300_vertex_shader* vs;
  40     struct r300_fragment_shader* fs;
  41 };
  42
  43 struct r300_shader_derived_value {
  44     struct r300_vertex_format* vformat;
  45     struct r300_rs_block* rs_block;
  46 };
  47
  48 unsigned r300_shader_key_hash(void* key) {
  49     struct r300_shader_key* shader_key = (struct r300_shader_key*)key;
  50     unsigned vs = (intptr_t)shader_key->vs;
  51     unsigned fs = (intptr_t)shader_key->fs;
  52
  53     return (vs << 16) | (fs & 0xffff);
  54 }
  55
  56 int r300_shader_key_compare(void* key1, void* key2) {
  57     struct r300_shader_key* shader_key1 = (struct r300_shader_key*)key1;
  58     struct r300_shader_key* shader_key2 = (struct r300_shader_key*)key2;
  59
  60     return (shader_key1->vs == shader_key2->vs) &&
  61         (shader_key1->fs == shader_key2->fs);
  62 }
  63
  64 /* Set up the vs_tab and routes. */
  65 static void r300_vs_tab_routes(struct r300_context* r300,
  66                                struct r300_vertex_info* vformat)
  67 {
  68     struct vertex_info* vinfo = &vformat->vinfo;
  69     int* tab = vformat->vs_tab;
  70     boolean pos = FALSE, psize = FALSE, fog = FALSE;
  71     int i, texs = 0, cols = 0;
  72     struct tgsi_shader_info* info = &r300->fs->info;
  73
  74     /* XXX One day we should figure out how to handle a different number of
  75      * VS outputs and FS inputs, as well as a different number of vertex streams
  76      * and VS inputs. It's definitely one of the sources of hardlocks. */
  77
  78     for (i = 0; i < info->num_inputs; i++) {
  79         switch (info->input_semantic_name[i]) {
  80             case TGSI_SEMANTIC_POSITION:
  81                 pos = TRUE;
  82                 tab[i] = 0;
  83                 break;
  84             case TGSI_SEMANTIC_COLOR:
  85                 tab[i] = 2 + cols;
  86                 cols++;
  87                 break;
  88             case TGSI_SEMANTIC_PSIZE:
  89                 assert(psize == FALSE);
  90                 psize = TRUE;
  91                 tab[i] = 15;
  92                 break;
  93             case TGSI_SEMANTIC_FOG:
  94                 assert(fog == FALSE);
  95                 fog = TRUE;
  96                 /* Fall through */
  97             case TGSI_SEMANTIC_GENERIC:
  98                 tab[i] = 6 + texs;
  99                 texs++;
 100                 break;
 101             default:
 102                 debug_printf("r300: Unknown vertex input %d\n",
 103                     info->input_semantic_name[i]);
 104                 break;
 105         }
 106     }
 107
 108     /* XXX magic */
 109     assert(texs <= 8);
 110
 111     /* Do the actual vertex_info setup.
 112      *
 113      * vertex_info has four uints of hardware-specific data in it.
 114      * vinfo.hwfmt[0] is R300_VAP_VTX_STATE_CNTL
 115      * vinfo.hwfmt[1] is R300_VAP_VSM_VTX_ASSM
 116      * vinfo.hwfmt[2] is R300_VAP_OUTPUT_VTX_FMT_0
 117      * vinfo.hwfmt[3] is R300_VAP_OUTPUT_VTX_FMT_1 */
 118
 119     vinfo->hwfmt[0] = 0x5555; /* XXX this is classic Mesa bonghits */
 120
 121     /* We need to add vertex position attribute only for SW TCL case,
 122      * for HW TCL case it could be generated by vertex shader */
 123     if (!pos) {
 124         /* Make room for the position attribute
 125          * at the beginning of the tab. */
 126         for (i = 15; i > 0; i--) {
 127             tab[i] = tab[i-1];
 128         }
 129         tab[0] = 0;
 130     }
 131
 132     /* Position. */
 133     if (r300->draw) {
 134         draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE,
 135             draw_find_vs_output(r300->draw, TGSI_SEMANTIC_POSITION, 0));
 136     }
 137     vinfo->hwfmt[1] |= R300_INPUT_CNTL_POS;
 138     vinfo->hwfmt[2] |= R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT;
 139
 140     /* Point size. */
 141     if (psize) {
 142         if (r300->draw) {
 143             draw_emit_vertex_attr(vinfo, EMIT_1F_PSIZE, INTERP_POS,
 144                 draw_find_vs_output(r300->draw, TGSI_SEMANTIC_PSIZE, 0));
 145         }
 146         vinfo->hwfmt[2] |= R300_VAP_OUTPUT_VTX_FMT_0__PT_SIZE_PRESENT;
 147     }
 148
 149     /* Colors. */
 150     for (i = 0; i < cols; i++) {
 151         if (r300->draw) {
 152             draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_LINEAR,
 153                 draw_find_vs_output(r300->draw, TGSI_SEMANTIC_COLOR, i));
 154         }
 155         vinfo->hwfmt[1] |= R300_INPUT_CNTL_COLOR;
 156         vinfo->hwfmt[2] |= (R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << i);
 157     }
 158
 159     /* Init i right here, increment it if fog is enabled.
 160      * This gets around a double-increment problem. */
 161     i = 0;
 162
 163     /* Fog. This is a special-cased texcoord. */
 164     if (fog) {
 165         i++;
 166         if (r300->draw) {
 167             draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE,
 168                 draw_find_vs_output(r300->draw, TGSI_SEMANTIC_FOG, 0));
 169         }
 170         vinfo->hwfmt[1] |= (R300_INPUT_CNTL_TC0 << i);
 171         vinfo->hwfmt[3] |= (4 << (3 * i));
 172     }
 173
 174     /* Texcoords. */
 175     for (; i < texs; i++) {
 176         if (r300->draw) {
 177             draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE,
 178                 draw_find_vs_output(r300->draw, TGSI_SEMANTIC_GENERIC, i));
 179         }
 180         vinfo->hwfmt[1] |= (R300_INPUT_CNTL_TC0 << i);
 181         vinfo->hwfmt[3] |= (4 << (3 * i));
 182     }
 183
 184     draw_compute_vertex_size(vinfo);
 185 }
 186
 187 /* Update the PSC tables. */
 188 static void r300_vertex_psc(struct r300_context* r300,
 189                             struct r300_vertex_info* vformat)
 190 {
 191     uint16_t type, swizzle;
 192     enum pipe_format format;
 193     unsigned i;
 194
 195     /* Vertex shaders have no semantics on their inputs,
 196      * so PSC should just route stuff based on the vertex elements,
 197      * and not on attrib information. */
 198     DBG(r300, DBG_DRAW, "r300: vs expects %d attribs, routing %d elements"
 199             " in psc\n",
 200             r300->vs->info.num_inputs,
 201             r300->vertex_element_count);
 202
 203     for (i = 0; i < r300->vertex_element_count; i++) {
 204         format = r300->vertex_element[i].src_format;
 205
 206         type = r300_translate_vertex_data_type(format) |
 207             (i << R300_DST_VEC_LOC_SHIFT);
 208         swizzle = r300_translate_vertex_data_swizzle(format);
 209
 210         if (i % 2) {
 211             vformat->vap_prog_stream_cntl[i >> 1] |= type << 16;
 212             vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
 213         } else {
 214             vformat->vap_prog_stream_cntl[i >> 1] |= type;
 215             vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
 216         }
 217     }
 218
 219
 220     assert(i <= 15);
 221
 222     /* Set the last vector in the PSC. */
 223     if (i) {
 224         i -= 1;
 225     }
 226     vformat->vap_prog_stream_cntl[i >> 1] |=
 227         (R300_LAST_VEC << (i & 1 ? 16 : 0));
 228 }
 229
 230 /* Update the PSC tables for SW TCL, using Draw. */
 231 static void r300_swtcl_vertex_psc(struct r300_context* r300,
 232                                   struct r300_vertex_info* vformat)
 233 {
 234     struct vertex_info* vinfo = &vformat->vinfo;
 235     int* tab = vformat->vs_tab;
 236     uint16_t type, swizzle;
 237     enum pipe_format format;
 238     unsigned i, attrib_count;
 239
 240     /* For each Draw attribute, route it to the fragment shader according
 241      * to the tab. */
 242     attrib_count = vinfo->num_attribs;
 243     DBG(r300, DBG_DRAW, "r300: attrib count: %d\n", attrib_count);
 244     for (i = 0; i < attrib_count; i++) {
 245         DBG(r300, DBG_DRAW, "r300: attrib: offset %d, interp %d, size %d,"
 246                " tab %d\n", vinfo->attrib[i].src_index,
 247                vinfo->attrib[i].interp_mode, vinfo->attrib[i].emit,
 248                tab[i]);
 249     }
 250
 251     for (i = 0; i < attrib_count; i++) {
 252         /* Make sure we have a proper destination for our attribute. */
 253         assert(tab[i] != -1);
 254
 255         format = draw_translate_vinfo_format(vinfo->attrib[i].emit);
 256
 257         /* Obtain the type of data in this attribute. */
 258         type = r300_translate_vertex_data_type(format) |
 259             tab[i] << R300_DST_VEC_LOC_SHIFT;
 260
 261         /* Obtain the swizzle for this attribute. Note that the default
 262          * swizzle in the hardware is not XYZW! */
 263         swizzle = r300_translate_vertex_data_swizzle(format);
 264
 265         /* Add the attribute to the PSC table. */
 266         if (i & 1) {
 267             vformat->vap_prog_stream_cntl[i >> 1] |= type << 16;
 268             vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
 269         } else {
 270             vformat->vap_prog_stream_cntl[i >> 1] |= type;
 271             vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
 272         }
 273     }
 274
 275     /* Set the last vector in the PSC. */
 276     if (i) {
 277         i -= 1;
 278     }
 279     vformat->vap_prog_stream_cntl[i >> 1] |=
 280         (R300_LAST_VEC << (i & 1 ? 16 : 0));
 281 }
 282
 283 /* Set up the mappings from GB to US, for RS block. */
 284 static void r300_update_fs_tab(struct r300_context* r300,
 285                                struct r300_vertex_info* vformat)
 286 {
 287     struct tgsi_shader_info* info = &r300->fs->info;
 288     int i, cols = 0, texs = 0, cols_emitted = 0;
 289     int* tab = vformat->fs_tab;
 290
 291     for (i = 0; i < 16; i++) {
 292         tab[i] = -1;
 293     }
 294
 295     assert(info->num_inputs <= 16);
 296     for (i = 0; i < info->num_inputs; i++) {
 297         switch (info->input_semantic_name[i]) {
 298             case TGSI_SEMANTIC_COLOR:
 299                 tab[i] = INTERP_LINEAR;
 300                 cols++;
 301                 break;
 302             case TGSI_SEMANTIC_POSITION:
 303             case TGSI_SEMANTIC_PSIZE:
 304                 debug_printf("r300: Implementation error: Can't use "
 305                         "pos attribs in fragshader yet!\n");
 306                 /* Pass through for now */
 307             case TGSI_SEMANTIC_FOG:
 308             case TGSI_SEMANTIC_GENERIC:
 309                 tab[i] = INTERP_PERSPECTIVE;
 310                 break;
 311             default:
 312                 debug_printf("r300: Unknown vertex input %d\n",
 313                     info->input_semantic_name[i]);
 314                 break;
 315         }
 316     }
 317
 318     /* Now that we know where everything is... */
 319     DBG(r300, DBG_DRAW, "r300: fp input count: %d\n", info->num_inputs);
 320     for (i = 0; i < info->num_inputs; i++) {
 321         switch (tab[i]) {
 322             case INTERP_LINEAR:
 323                 DBG(r300, DBG_DRAW, "r300: attrib: "
 324                         "stack offset %d, color,    tab %d\n",
 325                         i, cols_emitted);
 326                 tab[i] = cols_emitted;
 327                 cols_emitted++;
 328                 break;
 329             case INTERP_PERSPECTIVE:
 330                 DBG(r300, DBG_DRAW, "r300: attrib: "
 331                         "stack offset %d, texcoord, tab %d\n",
 332                         i, cols + texs);
 333                 tab[i] = cols + texs;
 334                 texs++;
 335                 break;
 336             case -1:
 337                 debug_printf("r300: Implementation error: Bad fp interp!\n");
 338             default:
 339                 break;
 340         }
 341     }
 342
 343 }
 344
 345 /* Set up the RS block. This is the part of the chipset that actually does
 346  * the rasterization of vertices into fragments. This is also the part of the
 347  * chipset that locks up if any part of it is even slightly wrong. */
 348 static void r300_update_rs_block(struct r300_context* r300,
 349                                  struct r300_rs_block* rs)
 350 {
 351     struct tgsi_shader_info* info = &r300->fs->info;
 352     int col_count = 0, fp_offset = 0, i, tex_count = 0;
 353     int rs_tex_comp = 0;
 354
 355     if (r300_screen(r300->context.screen)->caps->is_r500) {
 356         for (i = 0; i < info->num_inputs; i++) {
 357             switch (info->input_semantic_name[i]) {
 358                 case TGSI_SEMANTIC_COLOR:
 359                     rs->ip[col_count] |=
 360                         R500_RS_COL_PTR(col_count) |
 361                         R500_RS_COL_FMT(R300_RS_COL_FMT_RGBA);
 362                     col_count++;
 363                     break;
 364                 case TGSI_SEMANTIC_GENERIC:
 365                     rs->ip[tex_count] |=
 366                         R500_RS_SEL_S(rs_tex_comp) |
 367                         R500_RS_SEL_T(rs_tex_comp + 1) |
 368                         R500_RS_SEL_R(rs_tex_comp + 2) |
 369                         R500_RS_SEL_Q(rs_tex_comp + 3);
 370                     tex_count++;
 371                     rs_tex_comp += 4;
 372                     break;
 373                 default:
 374                     break;
 375             }
 376         }
 377
 378         /* Rasterize at least one color, or bad things happen. */
 379         if ((col_count == 0) && (tex_count == 0)) {
 380             rs->ip[0] |= R500_RS_COL_FMT(R300_RS_COL_FMT_0001);
 381             col_count++;
 382         }
 383
 384         for (i = 0; i < col_count; i++) {
 385             rs->inst[i] |= R500_RS_INST_COL_ID(i) |
 386                 R500_RS_INST_COL_CN_WRITE | R500_RS_INST_COL_ADDR(fp_offset);
 387             fp_offset++;
 388         }
 389
 390         for (i = 0; i < tex_count; i++) {
 391             rs->inst[i] |= R500_RS_INST_TEX_ID(i) |
 392                 R500_RS_INST_TEX_CN_WRITE | R500_RS_INST_TEX_ADDR(fp_offset);
 393             fp_offset++;
 394         }
 395
 396     } else {
 397         for (i = 0; i < info->num_inputs; i++) {
 398             switch (info->input_semantic_name[i]) {
 399                 case TGSI_SEMANTIC_COLOR:
 400                     rs->ip[col_count] |=
 401                         R300_RS_COL_PTR(col_count) |
 402                         R300_RS_COL_FMT(R300_RS_COL_FMT_RGBA);
 403                     col_count++;
 404                     break;
 405                 case TGSI_SEMANTIC_GENERIC:
 406                     rs->ip[tex_count] |=
 407                         R300_RS_TEX_PTR(rs_tex_comp) |
 408                         R300_RS_SEL_S(R300_RS_SEL_C0) |
 409                         R300_RS_SEL_T(R300_RS_SEL_C1) |
 410                         R300_RS_SEL_R(R300_RS_SEL_C2) |
 411                         R300_RS_SEL_Q(R300_RS_SEL_C3);
 412                     tex_count++;
 413                     rs_tex_comp+=4;
 414                     break;
 415                 default:
 416                     break;
 417             }
 418         }
 419
 420         /* Rasterize at least one color, or bad things happen. */
 421         if (col_count == 0) {
 422             rs->ip[0] |= R300_RS_COL_FMT(R300_RS_COL_FMT_0001);
 423             col_count++;
 424         }
 425
 426         if (tex_count == 0) {
 427             rs->ip[0] |=
 428                 R300_RS_SEL_S(R300_RS_SEL_K0) |
 429                 R300_RS_SEL_T(R300_RS_SEL_K0) |
 430                 R300_RS_SEL_R(R300_RS_SEL_K0) |
 431                 R300_RS_SEL_Q(R300_RS_SEL_K1);
 432         }
 433
 434         for (i = 0; i < col_count; i++) {
 435             rs->inst[i] |= R300_RS_INST_COL_ID(i) |
 436                 R300_RS_INST_COL_CN_WRITE | R300_RS_INST_COL_ADDR(fp_offset);
 437             fp_offset++;
 438         }
 439
 440         for (i = 0; i < tex_count; i++) {
 441             rs->inst[i] |= R300_RS_INST_TEX_ID(i) |
 442                 R300_RS_INST_TEX_CN_WRITE | R300_RS_INST_TEX_ADDR(fp_offset);
 443             fp_offset++;
 444         }
 445     }
 446
 447     rs->count = (rs_tex_comp) | (col_count << R300_IC_COUNT_SHIFT) |
 448         R300_HIRES_EN;
 449
 450     rs->inst_count = MAX3(col_count - 1, tex_count - 1, 0);
 451 }
 452
 453 /* Update the vertex format. */
 454 static void r300_update_derived_shader_state(struct r300_context* r300)
 455 {
 456     struct r300_screen* r300screen = r300_screen(r300->context.screen);
 457     struct r300_vertex_info* vformat;
 458     struct r300_rs_block* rs_block;
 459     int i;
 460
 461     /*
 462     struct r300_shader_key* key;
 463     struct r300_shader_derived_value* value;
 464     key = CALLOC_STRUCT(r300_shader_key);
 465     key->vs = r300->vs;
 466     key->fs = r300->fs;
 467
 468     value = (struct r300_shader_derived_value*)
 469         util_hash_table_get(r300->shader_hash_table, (void*)key);
 470     if (value) {
 471         //vformat = value->vformat;
 472         rs_block = value->rs_block;
 473
 474         FREE(key);
 475     } else {
 476         rs_block = CALLOC_STRUCT(r300_rs_block);
 477         value = CALLOC_STRUCT(r300_shader_derived_value);
 478
 479         r300_update_rs_block(r300, rs_block);
 480
 481         //value->vformat = vformat;
 482         value->rs_block = rs_block;
 483         util_hash_table_set(r300->shader_hash_table,
 484             (void*)key, (void*)value);
 485     } */
 486
 487     /* XXX This will be refactored ASAP. */
 488     vformat = CALLOC_STRUCT(r300_vertex_info);
 489     rs_block = CALLOC_STRUCT(r300_rs_block);
 490
 491     for (i = 0; i < 16; i++) {
 492         vformat->vs_tab[i] = -1;
 493         vformat->fs_tab[i] = -1;
 494     }
 495
 496     r300_vs_tab_routes(r300, vformat);
 497
 498     if (r300screen->caps->has_tcl) {
 499         r300_vertex_psc(r300, vformat);
 500     } else {
 501         r300_swtcl_vertex_psc(r300, vformat);
 502     }
 503
 504     r300_update_fs_tab(r300, vformat);
 505
 506     r300_update_rs_block(r300, rs_block);
 507
 508     FREE(r300->vertex_info);
 509     FREE(r300->rs_block);
 510
 511     r300->vertex_info = vformat;
 512     r300->rs_block = rs_block;
 513     r300->dirty_state |= R300_NEW_RS_BLOCK;
 514 }
 515
 516 static void r300_update_ztop(struct r300_context* r300)
 517 {
 518     r300->ztop_state.z_buffer_top = R300_ZTOP_ENABLE;
 519
 520     /* This is important enough that I felt it warranted a comment.
 521      *
 522      * According to the docs, these are the conditions where ZTOP must be
 523      * disabled:
 524      * 1) Alpha testing enabled
 525      * 2) Texture kill instructions in fragment shader
 526      * 3) Chroma key culling enabled
 527      * 4) W-buffering enabled
 528      *
 529      * The docs claim that for the first three cases, if no ZS writes happen,
 530      * then ZTOP can be used.
 531      *
 532      * Additionally, the following conditions require disabled ZTOP:
 533      * ~) Depth writes in fragment shader
 534      * ~) Outstanding occlusion queries
 535      *
 536      * ~C.
 537      */
 538     if (r300->dsa_state->alpha_function) {
 539         r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
 540     } else if (r300->fs->info.uses_kill) {
 541         r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
 542     } else if (r300_fragment_shader_writes_depth(r300->fs)) {
 543         r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
 544     } else if (r300->query_current) {
 545         r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
 546     }
 547 }
 548
 549 void r300_update_derived_state(struct r300_context* r300)
 550 {
 551     if (r300->dirty_state &
 552         (R300_NEW_FRAGMENT_SHADER | R300_NEW_VERTEX_SHADER |
 553          R300_NEW_VERTEX_FORMAT)) {
 554         r300_update_derived_shader_state(r300);
 555     }
 556
 557     if (r300->dirty_state &
 558             (R300_NEW_DSA | R300_NEW_FRAGMENT_SHADER | R300_NEW_QUERY)) {
 559         r300_update_ztop(r300);
 560     }
 561 }