src/mesa/drivers/dri/i965/brw_vs.c

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a copy of this software and associated documentation files (the
   8  "Software"), to deal in the Software without restriction, including
   9  without limitation the rights to use, copy, modify, merge, publish,
  10  distribute, sublicense, and/or sell copies of the Software, and to
  11  permit persons to whom the Software is furnished to do so, subject to
  12  the following conditions:
  13
  14  The above copyright notice and this permission notice (including the
  15  next paragraph) shall be included in all copies or substantial
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keithw@vmware.com>
  30   */
  31
  32
  33 #include "main/compiler.h"
  34 #include "brw_context.h"
  35 #include "brw_vs.h"
  36 #include "brw_util.h"
  37 #include "brw_state.h"
  38 #include "program/prog_print.h"
  39 #include "program/prog_parameter.h"
  40
  41 #include "glsl/ralloc.h"
  42
  43 static inline void assign_vue_slot(struct brw_vue_map *vue_map,
  44                                    int varying)
  45 {
  46    /* Make sure this varying hasn't been assigned a slot already */
  47    assert (vue_map->varying_to_slot[varying] == -1);
  48
  49    vue_map->varying_to_slot[varying] = vue_map->num_slots;
  50    vue_map->slot_to_varying[vue_map->num_slots++] = varying;
  51 }
  52
  53 /**
  54  * Compute the VUE map for vertex shader program.
  55  */
  56 void
  57 brw_compute_vue_map(struct brw_context *brw, struct brw_vue_map *vue_map,
  58                     GLbitfield64 slots_valid)
  59 {
  60    vue_map->slots_valid = slots_valid;
  61    int i;
  62
  63    /* gl_Layer doesn't get its own varying slot--it's stored in the virst VUE
  64     * slot (VARYING_SLOT_PSIZ).
  65     */
  66    slots_valid &= ~VARYING_BIT_LAYER;
  67
  68    /* Make sure that the values we store in vue_map->varying_to_slot and
  69     * vue_map->slot_to_varying won't overflow the signed chars that are used
  70     * to store them.  Note that since vue_map->slot_to_varying sometimes holds
  71     * values equal to BRW_VARYING_SLOT_COUNT, we need to ensure that
  72     * BRW_VARYING_SLOT_COUNT is <= 127, not 128.
  73     */
  74    STATIC_ASSERT(BRW_VARYING_SLOT_COUNT <= 127);
  75
  76    vue_map->num_slots = 0;
  77    for (i = 0; i < BRW_VARYING_SLOT_COUNT; ++i) {
  78       vue_map->varying_to_slot[i] = -1;
  79       vue_map->slot_to_varying[i] = BRW_VARYING_SLOT_COUNT;
  80    }
  81
  82    /* VUE header: format depends on chip generation and whether clipping is
  83     * enabled.
  84     */
  85    switch (brw->gen) {
  86    case 4:
  87    case 5:
  88       /* There are 8 dwords in VUE header pre-Ironlake:
  89        * dword 0-3 is indices, point width, clip flags.
  90        * dword 4-7 is ndc position
  91        * dword 8-11 is the first vertex data.
  92        *
  93        * On Ironlake the VUE header is nominally 20 dwords, but the hardware
  94        * will accept the same header layout as Gen4 [and should be a bit faster]
  95        */
  96       assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
  97       assign_vue_slot(vue_map, BRW_VARYING_SLOT_NDC);
  98       assign_vue_slot(vue_map, VARYING_SLOT_POS);
  99       break;
 100    case 6:
 101    case 7:
 102       /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
 103        * dword 0-3 of the header is indices, point width, clip flags.
 104        * dword 4-7 is the 4D space position
 105        * dword 8-15 of the vertex header is the user clip distance if
 106        * enabled.
 107        * dword 8-11 or 16-19 is the first vertex element data we fill.
 108        */
 109       assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
 110       assign_vue_slot(vue_map, VARYING_SLOT_POS);
 111       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0))
 112          assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST0);
 113       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1))
 114          assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST1);
 115
 116       /* front and back colors need to be consecutive so that we can use
 117        * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
 118        * two-sided color.
 119        */
 120       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL0))
 121          assign_vue_slot(vue_map, VARYING_SLOT_COL0);
 122       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC0))
 123          assign_vue_slot(vue_map, VARYING_SLOT_BFC0);
 124       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL1))
 125          assign_vue_slot(vue_map, VARYING_SLOT_COL1);
 126       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC1))
 127          assign_vue_slot(vue_map, VARYING_SLOT_BFC1);
 128       break;
 129    default:
 130       assert (!"VUE map not known for this chip generation");
 131       break;
 132    }
 133
 134    /* The hardware doesn't care about the rest of the vertex outputs, so just
 135     * assign them contiguously.  Don't reassign outputs that already have a
 136     * slot.
 137     *
 138     * We generally don't need to assign a slot for VARYING_SLOT_CLIP_VERTEX,
 139     * since it's encoded as the clip distances by emit_clip_distances().
 140     * However, it may be output by transform feedback, and we'd rather not
 141     * recompute state when TF changes, so we just always include it.
 142     */
 143    for (int i = 0; i < VARYING_SLOT_MAX; ++i) {
 144       if ((slots_valid & BITFIELD64_BIT(i)) &&
 145           vue_map->varying_to_slot[i] == -1) {
 146          assign_vue_slot(vue_map, i);
 147       }
 148    }
 149 }
 150
 151
 152 /**
 153  * Decide which set of clip planes should be used when clipping via
 154  * gl_Position or gl_ClipVertex.
 155  */
 156 gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx)
 157 {
 158    if (ctx->Shader.CurrentVertexProgram) {
 159       /* There is currently a GLSL vertex shader, so clip according to GLSL
 160        * rules, which means compare gl_ClipVertex (or gl_Position, if
 161        * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
 162        * that were stored in EyeUserPlane at the time the clip planes were
 163        * specified.
 164        */
 165       return ctx->Transform.EyeUserPlane;
 166    } else {
 167       /* Either we are using fixed function or an ARB vertex program.  In
 168        * either case the clip planes are going to be compared against
 169        * gl_Position (which is in clip coordinates) so we have to clip using
 170        * _ClipUserPlane, which was transformed into clip coordinates by Mesa
 171        * core.
 172        */
 173       return ctx->Transform._ClipUserPlane;
 174    }
 175 }
 176
 177
 178 bool
 179 brw_vs_prog_data_compare(const void *in_a, const void *in_b)
 180 {
 181    const struct brw_vs_prog_data *a = in_a;
 182    const struct brw_vs_prog_data *b = in_b;
 183
 184    /* Compare the base vec4 structure. */
 185    if (!brw_vec4_prog_data_compare(&a->base, &b->base))
 186       return false;
 187
 188    /* Compare the rest of the struct. */
 189    const unsigned offset = sizeof(struct brw_vec4_prog_data);
 190    if (memcmp(((char *) a) + offset, ((char *) b) + offset,
 191               sizeof(struct brw_vs_prog_data) - offset)) {
 192       return false;
 193    }
 194
 195    return true;
 196 }
 197
 198 static bool
 199 do_vs_prog(struct brw_context *brw,
 200            struct gl_shader_program *prog,
 201            struct brw_vertex_program *vp,
 202            struct brw_vs_prog_key *key)
 203 {
 204    GLuint program_size;
 205    const GLuint *program;
 206    struct brw_vs_compile c;
 207    struct brw_vs_prog_data prog_data;
 208    void *mem_ctx;
 209    int i;
 210    struct gl_shader *vs = NULL;
 211
 212    if (prog)
 213       vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
 214
 215    memset(&c, 0, sizeof(c));
 216    memcpy(&c.key, key, sizeof(*key));
 217    memset(&prog_data, 0, sizeof(prog_data));
 218
 219    mem_ctx = ralloc_context(NULL);
 220
 221    c.vp = vp;
 222
 223    /* Allocate the references to the uniforms that will end up in the
 224     * prog_data associated with the compiled program, and which will be freed
 225     * by the state cache.
 226     */
 227    int param_count;
 228    if (vs) {
 229       /* We add padding around uniform values below vec4 size, with the worst
 230        * case being a float value that gets blown up to a vec4, so be
 231        * conservative here.
 232        */
 233       param_count = vs->num_uniform_components * 4;
 234
 235    } else {
 236       param_count = vp->program.Base.Parameters->NumParameters * 4;
 237    }
 238    /* vec4_visitor::setup_uniform_clipplane_values() also uploads user clip
 239     * planes as uniforms.
 240     */
 241    param_count += c.key.base.nr_userclip_plane_consts * 4;
 242
 243    prog_data.base.param = rzalloc_array(NULL, const float *, param_count);
 244    prog_data.base.pull_param = rzalloc_array(NULL, const float *, param_count);
 245
 246    GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
 247    prog_data.inputs_read = vp->program.Base.InputsRead;
 248
 249    if (c.key.copy_edgeflag) {
 250       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_EDGE);
 251       prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
 252    }
 253
 254    if (brw->gen < 6) {
 255       /* Put dummy slots into the VUE for the SF to put the replaced
 256        * point sprite coords in.  We shouldn't need these dummy slots,
 257        * which take up precious URB space, but it would mean that the SF
 258        * doesn't get nice aligned pairs of input coords into output
 259        * coords, which would be a pain to handle.
 260        */
 261       for (i = 0; i < 8; i++) {
 262          if (c.key.point_coord_replace & (1 << i))
 263             outputs_written |= BITFIELD64_BIT(VARYING_SLOT_TEX0 + i);
 264       }
 265
 266       /* if back colors are written, allocate slots for front colors too */
 267       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC0))
 268          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL0);
 269       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC1))
 270          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL1);
 271    }
 272
 273    /* In order for legacy clipping to work, we need to populate the clip
 274     * distance varying slots whenever clipping is enabled, even if the vertex
 275     * shader doesn't write to gl_ClipDistance.
 276     */
 277    if (c.key.base.userclip_active) {
 278       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
 279       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
 280    }
 281
 282    brw_compute_vue_map(brw, &prog_data.base.vue_map, outputs_written);
 283
 284    if (0) {
 285       _mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG,
 286                                true);
 287    }
 288
 289    /* Emit GEN4 code.
 290     */
 291    program = brw_vs_emit(brw, prog, &c, &prog_data, mem_ctx, &program_size);
 292    if (program == NULL) {
 293       ralloc_free(mem_ctx);
 294       return false;
 295    }
 296
 297    /* Scratch space is used for register spilling */
 298    if (c.base.last_scratch) {
 299       perf_debug("Vertex shader triggered register spilling.  "
 300                  "Try reducing the number of live vec4 values to "
 301                  "improve performance.\n");
 302
 303       prog_data.base.total_scratch
 304          = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
 305
 306       brw_get_scratch_bo(brw, &brw->vs.base.scratch_bo,
 307                          prog_data.base.total_scratch * brw->max_vs_threads);
 308    }
 309
 310    brw_upload_cache(&brw->cache, BRW_VS_PROG,
 311                     &c.key, sizeof(c.key),
 312                     program, program_size,
 313                     &prog_data, sizeof(prog_data),
 314                     &brw->vs.base.prog_offset, &brw->vs.prog_data);
 315    ralloc_free(mem_ctx);
 316
 317    return true;
 318 }
 319
 320 static bool
 321 key_debug(struct brw_context *brw, const char *name, int a, int b)
 322 {
 323    if (a != b) {
 324       perf_debug("  %s %d->%d\n", name, a, b);
 325       return true;
 326    }
 327    return false;
 328 }
 329
 330 void
 331 brw_vs_debug_recompile(struct brw_context *brw,
 332                        struct gl_shader_program *prog,
 333                        const struct brw_vs_prog_key *key)
 334 {
 335    struct brw_cache_item *c = NULL;
 336    const struct brw_vs_prog_key *old_key = NULL;
 337    bool found = false;
 338
 339    perf_debug("Recompiling vertex shader for program %d\n", prog->Name);
 340
 341    for (unsigned int i = 0; i < brw->cache.size; i++) {
 342       for (c = brw->cache.items[i]; c; c = c->next) {
 343          if (c->cache_id == BRW_VS_PROG) {
 344             old_key = c->key;
 345
 346             if (old_key->base.program_string_id == key->base.program_string_id)
 347                break;
 348          }
 349       }
 350       if (c)
 351          break;
 352    }
 353
 354    if (!c) {
 355       perf_debug("  Didn't find previous compile in the shader cache for "
 356                  "debug\n");
 357       return;
 358    }
 359
 360    for (unsigned int i = 0; i < VERT_ATTRIB_MAX; i++) {
 361       found |= key_debug(brw, "Vertex attrib w/a flags",
 362                          old_key->gl_attrib_wa_flags[i],
 363                          key->gl_attrib_wa_flags[i]);
 364    }
 365
 366    found |= key_debug(brw, "user clip flags",
 367                       old_key->base.userclip_active, key->base.userclip_active);
 368
 369    found |= key_debug(brw, "user clipping planes as push constants",
 370                       old_key->base.nr_userclip_plane_consts,
 371                       key->base.nr_userclip_plane_consts);
 372
 373    found |= key_debug(brw, "copy edgeflag",
 374                       old_key->copy_edgeflag, key->copy_edgeflag);
 375    found |= key_debug(brw, "PointCoord replace",
 376                       old_key->point_coord_replace, key->point_coord_replace);
 377    found |= key_debug(brw, "vertex color clamping",
 378                       old_key->base.clamp_vertex_color, key->base.clamp_vertex_color);
 379
 380    found |= brw_debug_recompile_sampler_key(brw, &old_key->base.tex,
 381                                             &key->base.tex);
 382
 383    if (!found) {
 384       perf_debug("  Something else\n");
 385    }
 386 }
 387
 388
 389 void
 390 brw_setup_vec4_key_clip_info(struct brw_context *brw,
 391                              struct brw_vec4_prog_key *key,
 392                              bool program_uses_clip_distance)
 393 {
 394    struct gl_context *ctx = &brw->ctx;
 395
 396    key->userclip_active = (ctx->Transform.ClipPlanesEnabled != 0);
 397    if (key->userclip_active && !program_uses_clip_distance) {
 398       key->nr_userclip_plane_consts
 399          = _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1;
 400    }
 401 }
 402
 403
 404 static void brw_upload_vs_prog(struct brw_context *brw)
 405 {
 406    struct gl_context *ctx = &brw->ctx;
 407    struct brw_vs_prog_key key;
 408    /* BRW_NEW_VERTEX_PROGRAM */
 409    struct brw_vertex_program *vp =
 410       (struct brw_vertex_program *)brw->vertex_program;
 411    struct gl_program *prog = (struct gl_program *) brw->vertex_program;
 412    int i;
 413
 414    memset(&key, 0, sizeof(key));
 415
 416    /* Just upload the program verbatim for now.  Always send it all
 417     * the inputs it asks for, whether they are varying or not.
 418     */
 419    key.base.program_string_id = vp->id;
 420    brw_setup_vec4_key_clip_info(brw, &key.base,
 421                                 vp->program.Base.UsesClipDistanceOut);
 422
 423    /* _NEW_POLYGON */
 424    if (brw->gen < 6) {
 425       key.copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL ||
 426                            ctx->Polygon.BackMode != GL_FILL);
 427    }
 428
 429    /* _NEW_LIGHT | _NEW_BUFFERS */
 430    key.base.clamp_vertex_color = ctx->Light._ClampVertexColor;
 431
 432    /* _NEW_POINT */
 433    if (brw->gen < 6 && ctx->Point.PointSprite) {
 434       for (i = 0; i < 8; i++) {
 435          if (ctx->Point.CoordReplace[i])
 436             key.point_coord_replace |= (1 << i);
 437       }
 438    }
 439
 440    /* _NEW_TEXTURE */
 441    brw_populate_sampler_prog_key_data(ctx, prog, brw->vs.base.sampler_count,
 442                                       &key.base.tex);
 443
 444    /* BRW_NEW_VERTICES */
 445    if (brw->gen < 8 && !brw->is_haswell) {
 446       /* Prior to Haswell, the hardware can't natively support GL_FIXED or
 447        * 2_10_10_10_REV vertex formats.  Set appropriate workaround flags.
 448        */
 449       for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 450          if (!(vp->program.Base.InputsRead & BITFIELD64_BIT(i)))
 451             continue;
 452
 453          uint8_t wa_flags = 0;
 454
 455          switch (brw->vb.inputs[i].glarray->Type) {
 456
 457          case GL_FIXED:
 458             wa_flags = brw->vb.inputs[i].glarray->Size;
 459             break;
 460
 461          case GL_INT_2_10_10_10_REV:
 462             wa_flags |= BRW_ATTRIB_WA_SIGN;
 463             /* fallthough */
 464
 465          case GL_UNSIGNED_INT_2_10_10_10_REV:
 466             if (brw->vb.inputs[i].glarray->Format == GL_BGRA)
 467                wa_flags |= BRW_ATTRIB_WA_BGRA;
 468
 469             if (brw->vb.inputs[i].glarray->Normalized)
 470                wa_flags |= BRW_ATTRIB_WA_NORMALIZE;
 471             else if (!brw->vb.inputs[i].glarray->Integer)
 472                wa_flags |= BRW_ATTRIB_WA_SCALE;
 473
 474             break;
 475          }
 476
 477          key.gl_attrib_wa_flags[i] = wa_flags;
 478       }
 479    }
 480
 481    if (!brw_search_cache(&brw->cache, BRW_VS_PROG,
 482                          &key, sizeof(key),
 483                          &brw->vs.base.prog_offset, &brw->vs.prog_data)) {
 484       bool success = do_vs_prog(brw, ctx->Shader.CurrentVertexProgram,
 485                                 vp, &key);
 486       (void) success;
 487       assert(success);
 488    }
 489    brw->vs.base.prog_data = &brw->vs.prog_data->base.base;
 490
 491    if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
 492               sizeof(brw->vue_map_geom_out)) != 0) {
 493       brw->vue_map_vs = brw->vs.prog_data->base.vue_map;
 494       brw->state.dirty.brw |= BRW_NEW_VUE_MAP_VS;
 495       if (brw->gen < 7) {
 496          /* No geometry shader support, so the VS VUE map is the VUE map for
 497           * the output of the "geometry" portion of the pipeline.
 498           */
 499          brw->vue_map_geom_out = brw->vue_map_vs;
 500          brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 501       }
 502    }
 503 }
 504
 505 /* See brw_vs.c:
 506  */
 507 const struct brw_tracked_state brw_vs_prog = {
 508    .dirty = {
 509       .mesa  = (_NEW_TRANSFORM | _NEW_POLYGON | _NEW_POINT | _NEW_LIGHT |
 510                 _NEW_TEXTURE |
 511                 _NEW_BUFFERS),
 512       .brw   = (BRW_NEW_VERTEX_PROGRAM |
 513                 BRW_NEW_VERTICES),
 514       .cache = 0
 515    },
 516    .emit = brw_upload_vs_prog
 517 };
 518
 519 bool
 520 brw_vs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
 521 {
 522    struct brw_context *brw = brw_context(ctx);
 523    struct brw_vs_prog_key key;
 524    uint32_t old_prog_offset = brw->vs.base.prog_offset;
 525    struct brw_vs_prog_data *old_prog_data = brw->vs.prog_data;
 526    bool success;
 527
 528    if (!prog->_LinkedShaders[MESA_SHADER_VERTEX])
 529       return true;
 530
 531    struct gl_vertex_program *vp = (struct gl_vertex_program *)
 532       prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program;
 533    struct brw_vertex_program *bvp = brw_vertex_program(vp);
 534
 535    memset(&key, 0, sizeof(key));
 536
 537    brw_vec4_setup_prog_key_for_precompile(ctx, &key.base, bvp->id, &vp->Base);
 538
 539    success = do_vs_prog(brw, prog, bvp, &key);
 540
 541    brw->vs.base.prog_offset = old_prog_offset;
 542    brw->vs.prog_data = old_prog_data;
 543
 544    return success;
 545 }
 546
 547
 548 void
 549 brw_vs_prog_data_free(const void *in_prog_data)
 550 {
 551    const struct brw_vs_prog_data *prog_data = in_prog_data;
 552
 553    brw_vec4_prog_data_free(&prog_data->base);
 554 }