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

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics (http://www.tungstengraphics.com) 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 <keith@tungstengraphics.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  * Note that consumers of this map using cache keys must include
  57  * prog_data->userclip and prog_data->outputs_written in their key
  58  * (generated by CACHE_NEW_VS_PROG).
  59  */
  60 void
  61 brw_compute_vue_map(struct brw_context *brw, struct brw_vue_map *vue_map,
  62                     GLbitfield64 slots_valid, bool userclip_active)
  63 {
  64    vue_map->slots_valid = slots_valid;
  65    int i;
  66
  67    /* Make sure that the values we store in vue_map->varying_to_slot and
  68     * vue_map->slot_to_varying won't overflow the signed chars that are used
  69     * to store them.  Note that since vue_map->slot_to_varying sometimes holds
  70     * values equal to BRW_VARYING_SLOT_COUNT, we need to ensure that
  71     * BRW_VARYING_SLOT_COUNT is <= 127, not 128.
  72     */
  73    STATIC_ASSERT(BRW_VARYING_SLOT_COUNT <= 127);
  74
  75    vue_map->num_slots = 0;
  76    for (i = 0; i < BRW_VARYING_SLOT_COUNT; ++i) {
  77       vue_map->varying_to_slot[i] = -1;
  78       vue_map->slot_to_varying[i] = BRW_VARYING_SLOT_COUNT;
  79    }
  80
  81    /* VUE header: format depends on chip generation and whether clipping is
  82     * enabled.
  83     */
  84    switch (brw->gen) {
  85    case 4:
  86    case 5:
  87       /* There are 8 dwords in VUE header pre-Ironlake:
  88        * dword 0-3 is indices, point width, clip flags.
  89        * dword 4-7 is ndc position
  90        * dword 8-11 is the first vertex data.
  91        *
  92        * On Ironlake the VUE header is nominally 20 dwords, but the hardware
  93        * will accept the same header layout as Gen4 [and should be a bit faster]
  94        */
  95       assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
  96       assign_vue_slot(vue_map, BRW_VARYING_SLOT_NDC);
  97       assign_vue_slot(vue_map, VARYING_SLOT_POS);
  98       break;
  99    case 6:
 100    case 7:
 101       /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
 102        * dword 0-3 of the header is indices, point width, clip flags.
 103        * dword 4-7 is the 4D space position
 104        * dword 8-15 of the vertex header is the user clip distance if
 105        * enabled.
 106        * dword 8-11 or 16-19 is the first vertex element data we fill.
 107        */
 108       assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
 109       assign_vue_slot(vue_map, VARYING_SLOT_POS);
 110       if (userclip_active) {
 111          assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST0);
 112          assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST1);
 113       }
 114       /* front and back colors need to be consecutive so that we can use
 115        * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
 116        * two-sided color.
 117        */
 118       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL0))
 119          assign_vue_slot(vue_map, VARYING_SLOT_COL0);
 120       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC0))
 121          assign_vue_slot(vue_map, VARYING_SLOT_BFC0);
 122       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL1))
 123          assign_vue_slot(vue_map, VARYING_SLOT_COL1);
 124       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC1))
 125          assign_vue_slot(vue_map, VARYING_SLOT_BFC1);
 126       break;
 127    default:
 128       assert (!"VUE map not known for this chip generation");
 129       break;
 130    }
 131
 132    /* The hardware doesn't care about the rest of the vertex outputs, so just
 133     * assign them contiguously.  Don't reassign outputs that already have a
 134     * slot.
 135     *
 136     * We generally don't need to assign a slot for VARYING_SLOT_CLIP_VERTEX,
 137     * since it's encoded as the clip distances by emit_clip_distances().
 138     * However, it may be output by transform feedback, and we'd rather not
 139     * recompute state when TF changes, so we just always include it.
 140     */
 141    for (int i = 0; i < VARYING_SLOT_MAX; ++i) {
 142       if ((slots_valid & BITFIELD64_BIT(i)) &&
 143           vue_map->varying_to_slot[i] == -1) {
 144          assign_vue_slot(vue_map, i);
 145       }
 146    }
 147 }
 148
 149
 150 /**
 151  * Decide which set of clip planes should be used when clipping via
 152  * gl_Position or gl_ClipVertex.
 153  */
 154 gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx)
 155 {
 156    if (ctx->Shader.CurrentVertexProgram) {
 157       /* There is currently a GLSL vertex shader, so clip according to GLSL
 158        * rules, which means compare gl_ClipVertex (or gl_Position, if
 159        * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
 160        * that were stored in EyeUserPlane at the time the clip planes were
 161        * specified.
 162        */
 163       return ctx->Transform.EyeUserPlane;
 164    } else {
 165       /* Either we are using fixed function or an ARB vertex program.  In
 166        * either case the clip planes are going to be compared against
 167        * gl_Position (which is in clip coordinates) so we have to clip using
 168        * _ClipUserPlane, which was transformed into clip coordinates by Mesa
 169        * core.
 170        */
 171       return ctx->Transform._ClipUserPlane;
 172    }
 173 }
 174
 175
 176 bool
 177 brw_vs_prog_data_compare(const void *in_a, const void *in_b,
 178                          int aux_size, const void *in_key)
 179 {
 180    const struct brw_vs_prog_data *a = in_a;
 181    const struct brw_vs_prog_data *b = in_b;
 182
 183    /* Compare the base vec4 structure. */
 184    if (!brw_vec4_prog_data_compare(&a->base, &b->base))
 185       return false;
 186
 187    /* Compare the rest of the struct. */
 188    const unsigned offset = sizeof(struct brw_vec4_prog_data);
 189    if (memcmp(((char *) &a) + offset, ((char *) &b) + offset,
 190               sizeof(struct brw_vs_prog_data) - offset)) {
 191       return false;
 192    }
 193
 194    return true;
 195 }
 196
 197 static bool
 198 do_vs_prog(struct brw_context *brw,
 199            struct gl_shader_program *prog,
 200            struct brw_vertex_program *vp,
 201            struct brw_vs_prog_key *key)
 202 {
 203    GLuint program_size;
 204    const GLuint *program;
 205    struct brw_vs_compile c;
 206    struct brw_vs_prog_data prog_data;
 207    void *mem_ctx;
 208    int i;
 209    struct gl_shader *vs = NULL;
 210
 211    if (prog)
 212       vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
 213
 214    memset(&c, 0, sizeof(c));
 215    memcpy(&c.key, key, sizeof(*key));
 216    memset(&prog_data, 0, sizeof(prog_data));
 217
 218    mem_ctx = ralloc_context(NULL);
 219
 220    c.vp = vp;
 221
 222    /* Allocate the references to the uniforms that will end up in the
 223     * prog_data associated with the compiled program, and which will be freed
 224     * by the state cache.
 225     */
 226    int param_count;
 227    if (vs) {
 228       /* We add padding around uniform values below vec4 size, with the worst
 229        * case being a float value that gets blown up to a vec4, so be
 230        * conservative here.
 231        */
 232       param_count = vs->num_uniform_components * 4;
 233
 234    } else {
 235       param_count = vp->program.Base.Parameters->NumParameters * 4;
 236    }
 237    /* We also upload clip plane data as uniforms */
 238    param_count += MAX_CLIP_PLANES * 4;
 239
 240    prog_data.base.param = rzalloc_array(NULL, const float *, param_count);
 241    prog_data.base.pull_param = rzalloc_array(NULL, const float *, param_count);
 242
 243    GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
 244    prog_data.inputs_read = vp->program.Base.InputsRead;
 245
 246    if (c.key.copy_edgeflag) {
 247       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_EDGE);
 248       prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
 249    }
 250
 251    if (brw->gen < 6) {
 252       /* Put dummy slots into the VUE for the SF to put the replaced
 253        * point sprite coords in.  We shouldn't need these dummy slots,
 254        * which take up precious URB space, but it would mean that the SF
 255        * doesn't get nice aligned pairs of input coords into output
 256        * coords, which would be a pain to handle.
 257        */
 258       for (i = 0; i < 8; i++) {
 259          if (c.key.point_coord_replace & (1 << i))
 260             outputs_written |= BITFIELD64_BIT(VARYING_SLOT_TEX0 + i);
 261       }
 262
 263       /* if back colors are written, allocate slots for front colors too */
 264       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC0))
 265          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL0);
 266       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC1))
 267          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL1);
 268
 269       if (c.key.base.userclip_active) {
 270          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
 271          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
 272       }
 273    }
 274
 275    brw_compute_vue_map(brw, &prog_data.base.vue_map, outputs_written,
 276                        c.key.base.userclip_active);
 277
 278    if (0) {
 279       _mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG,
 280                                true);
 281    }
 282
 283    /* Emit GEN4 code.
 284     */
 285    program = brw_vs_emit(brw, prog, &c, &prog_data, mem_ctx, &program_size);
 286    if (program == NULL) {
 287       ralloc_free(mem_ctx);
 288       return false;
 289    }
 290
 291    /* Scratch space is used for register spilling */
 292    if (c.base.last_scratch) {
 293       perf_debug("Vertex shader triggered register spilling.  "
 294                  "Try reducing the number of live vec4 values to "
 295                  "improve performance.\n");
 296
 297       prog_data.base.total_scratch
 298          = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
 299
 300       brw_get_scratch_bo(brw, &brw->vs.scratch_bo,
 301                          prog_data.base.total_scratch * brw->max_vs_threads);
 302    }
 303
 304    brw_upload_cache(&brw->cache, BRW_VS_PROG,
 305                     &c.key, sizeof(c.key),
 306                     program, program_size,
 307                     &prog_data, sizeof(prog_data),
 308                     &brw->vs.prog_offset, &brw->vs.prog_data);
 309    ralloc_free(mem_ctx);
 310
 311    return true;
 312 }
 313
 314 static bool
 315 key_debug(struct brw_context *brw, const char *name, int a, int b)
 316 {
 317    if (a != b) {
 318       perf_debug("  %s %d->%d\n", name, a, b);
 319       return true;
 320    }
 321    return false;
 322 }
 323
 324 void
 325 brw_vs_debug_recompile(struct brw_context *brw,
 326                        struct gl_shader_program *prog,
 327                        const struct brw_vs_prog_key *key)
 328 {
 329    struct brw_cache_item *c = NULL;
 330    const struct brw_vs_prog_key *old_key = NULL;
 331    bool found = false;
 332
 333    perf_debug("Recompiling vertex shader for program %d\n", prog->Name);
 334
 335    for (unsigned int i = 0; i < brw->cache.size; i++) {
 336       for (c = brw->cache.items[i]; c; c = c->next) {
 337          if (c->cache_id == BRW_VS_PROG) {
 338             old_key = c->key;
 339
 340             if (old_key->base.program_string_id == key->base.program_string_id)
 341                break;
 342          }
 343       }
 344       if (c)
 345          break;
 346    }
 347
 348    if (!c) {
 349       perf_debug("  Didn't find previous compile in the shader cache for "
 350                  "debug\n");
 351       return;
 352    }
 353
 354    for (unsigned int i = 0; i < VERT_ATTRIB_MAX; i++) {
 355       found |= key_debug(brw, "Vertex attrib w/a flags",
 356                          old_key->gl_attrib_wa_flags[i],
 357                          key->gl_attrib_wa_flags[i]);
 358    }
 359
 360    found |= key_debug(brw, "user clip flags",
 361                       old_key->base.userclip_active, key->base.userclip_active);
 362
 363    found |= key_debug(brw, "user clipping planes as push constants",
 364                       old_key->base.nr_userclip_plane_consts,
 365                       key->base.nr_userclip_plane_consts);
 366
 367    found |= key_debug(brw, "clip distance enable",
 368                       old_key->base.uses_clip_distance, key->base.uses_clip_distance);
 369    found |= key_debug(brw, "copy edgeflag",
 370                       old_key->copy_edgeflag, key->copy_edgeflag);
 371    found |= key_debug(brw, "PointCoord replace",
 372                       old_key->point_coord_replace, key->point_coord_replace);
 373    found |= key_debug(brw, "vertex color clamping",
 374                       old_key->base.clamp_vertex_color, key->base.clamp_vertex_color);
 375
 376    found |= brw_debug_recompile_sampler_key(brw, &old_key->base.tex,
 377                                             &key->base.tex);
 378
 379    if (!found) {
 380       perf_debug("  Something else\n");
 381    }
 382 }
 383
 384
 385 void
 386 brw_setup_vec4_key_clip_info(struct brw_context *brw,
 387                              struct brw_vec4_prog_key *key,
 388                              bool program_uses_clip_distance)
 389 {
 390    struct gl_context *ctx = &brw->ctx;
 391
 392    key->userclip_active = (ctx->Transform.ClipPlanesEnabled != 0);
 393    key->uses_clip_distance = program_uses_clip_distance;
 394    if (key->userclip_active && !key->uses_clip_distance) {
 395       key->nr_userclip_plane_consts
 396          = _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1;
 397    }
 398 }
 399
 400
 401 static void brw_upload_vs_prog(struct brw_context *brw)
 402 {
 403    struct gl_context *ctx = &brw->ctx;
 404    struct brw_vs_prog_key key;
 405    /* BRW_NEW_VERTEX_PROGRAM */
 406    struct brw_vertex_program *vp =
 407       (struct brw_vertex_program *)brw->vertex_program;
 408    struct gl_program *prog = (struct gl_program *) brw->vertex_program;
 409    int i;
 410
 411    memset(&key, 0, sizeof(key));
 412
 413    /* Just upload the program verbatim for now.  Always send it all
 414     * the inputs it asks for, whether they are varying or not.
 415     */
 416    key.base.program_string_id = vp->id;
 417    brw_setup_vec4_key_clip_info(brw, &key.base, vp->program.UsesClipDistance);
 418
 419    /* _NEW_POLYGON */
 420    if (brw->gen < 6) {
 421       key.copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL ||
 422                            ctx->Polygon.BackMode != GL_FILL);
 423    }
 424
 425    /* _NEW_LIGHT | _NEW_BUFFERS */
 426    key.base.clamp_vertex_color = ctx->Light._ClampVertexColor;
 427
 428    /* _NEW_POINT */
 429    if (brw->gen < 6 && ctx->Point.PointSprite) {
 430       for (i = 0; i < 8; i++) {
 431          if (ctx->Point.CoordReplace[i])
 432             key.point_coord_replace |= (1 << i);
 433       }
 434    }
 435
 436    /* _NEW_TEXTURE */
 437    brw_populate_sampler_prog_key_data(ctx, prog, brw->vs.sampler_count,
 438                                       &key.base.tex);
 439
 440    /* BRW_NEW_VERTICES */
 441    if (brw->gen < 8 && !brw->is_haswell) {
 442       /* Prior to Haswell, the hardware can't natively support GL_FIXED or
 443        * 2_10_10_10_REV vertex formats.  Set appropriate workaround flags.
 444        */
 445       for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 446          if (!(vp->program.Base.InputsRead & BITFIELD64_BIT(i)))
 447             continue;
 448
 449          uint8_t wa_flags = 0;
 450
 451          switch (brw->vb.inputs[i].glarray->Type) {
 452
 453          case GL_FIXED:
 454             wa_flags = brw->vb.inputs[i].glarray->Size;
 455             break;
 456
 457          case GL_INT_2_10_10_10_REV:
 458             wa_flags |= BRW_ATTRIB_WA_SIGN;
 459             /* fallthough */
 460
 461          case GL_UNSIGNED_INT_2_10_10_10_REV:
 462             if (brw->vb.inputs[i].glarray->Format == GL_BGRA)
 463                wa_flags |= BRW_ATTRIB_WA_BGRA;
 464
 465             if (brw->vb.inputs[i].glarray->Normalized)
 466                wa_flags |= BRW_ATTRIB_WA_NORMALIZE;
 467             else if (!brw->vb.inputs[i].glarray->Integer)
 468                wa_flags |= BRW_ATTRIB_WA_SCALE;
 469
 470             break;
 471          }
 472
 473          key.gl_attrib_wa_flags[i] = wa_flags;
 474       }
 475    }
 476
 477    if (!brw_search_cache(&brw->cache, BRW_VS_PROG,
 478                          &key, sizeof(key),
 479                          &brw->vs.prog_offset, &brw->vs.prog_data)) {
 480       bool success = do_vs_prog(brw, ctx->Shader.CurrentVertexProgram,
 481                                 vp, &key);
 482
 483       assert(success);
 484    }
 485    if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
 486               sizeof(brw->vue_map_geom_out)) != 0) {
 487       brw->vue_map_geom_out = brw->vs.prog_data->base.vue_map;
 488       brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 489    }
 490 }
 491
 492 /* See brw_vs.c:
 493  */
 494 const struct brw_tracked_state brw_vs_prog = {
 495    .dirty = {
 496       .mesa  = (_NEW_TRANSFORM | _NEW_POLYGON | _NEW_POINT | _NEW_LIGHT |
 497                 _NEW_TEXTURE |
 498                 _NEW_BUFFERS),
 499       .brw   = (BRW_NEW_VERTEX_PROGRAM |
 500                 BRW_NEW_VERTICES),
 501       .cache = 0
 502    },
 503    .emit = brw_upload_vs_prog
 504 };
 505
 506 bool
 507 brw_vs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
 508 {
 509    struct brw_context *brw = brw_context(ctx);
 510    struct brw_vs_prog_key key;
 511    uint32_t old_prog_offset = brw->vs.prog_offset;
 512    struct brw_vs_prog_data *old_prog_data = brw->vs.prog_data;
 513    bool success;
 514
 515    if (!prog->_LinkedShaders[MESA_SHADER_VERTEX])
 516       return true;
 517
 518    struct gl_vertex_program *vp = (struct gl_vertex_program *)
 519       prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program;
 520    struct brw_vertex_program *bvp = brw_vertex_program(vp);
 521
 522    memset(&key, 0, sizeof(key));
 523
 524    key.base.program_string_id = bvp->id;
 525    key.base.clamp_vertex_color = ctx->API == API_OPENGL_COMPAT;
 526
 527    unsigned sampler_count = _mesa_fls(vp->Base.SamplersUsed);
 528    for (unsigned i = 0; i < sampler_count; i++) {
 529       if (vp->Base.ShadowSamplers & (1 << i)) {
 530          /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
 531          key.base.tex.swizzles[i] =
 532             MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_ONE);
 533       } else {
 534          /* Color sampler: assume no swizzling. */
 535          key.base.tex.swizzles[i] = SWIZZLE_XYZW;
 536       }
 537    }
 538
 539    success = do_vs_prog(brw, prog, bvp, &key);
 540
 541    brw->vs.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 }