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 "util/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(const struct brw_device_info *devinfo,
  58                     struct brw_vue_map *vue_map,
  59                     GLbitfield64 slots_valid)
  60 {
  61    vue_map->slots_valid = slots_valid;
  62    int i;
  63
  64    /* gl_Layer and gl_ViewportIndex don't get their own varying slots -- they
  65     * are stored in the first VUE slot (VARYING_SLOT_PSIZ).
  66     */
  67    slots_valid &= ~(VARYING_BIT_LAYER | VARYING_BIT_VIEWPORT);
  68
  69    /* Make sure that the values we store in vue_map->varying_to_slot and
  70     * vue_map->slot_to_varying won't overflow the signed chars that are used
  71     * to store them.  Note that since vue_map->slot_to_varying sometimes holds
  72     * values equal to BRW_VARYING_SLOT_COUNT, we need to ensure that
  73     * BRW_VARYING_SLOT_COUNT is <= 127, not 128.
  74     */
  75    STATIC_ASSERT(BRW_VARYING_SLOT_COUNT <= 127);
  76
  77    vue_map->num_slots = 0;
  78    for (i = 0; i < BRW_VARYING_SLOT_COUNT; ++i) {
  79       vue_map->varying_to_slot[i] = -1;
  80       vue_map->slot_to_varying[i] = BRW_VARYING_SLOT_COUNT;
  81    }
  82
  83    /* VUE header: format depends on chip generation and whether clipping is
  84     * enabled.
  85     */
  86    if (devinfo->gen < 6) {
  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    } else {
  99       /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
 100        * dword 0-3 of the header is indices, point width, clip flags.
 101        * dword 4-7 is the 4D space position
 102        * dword 8-15 of the vertex header is the user clip distance if
 103        * enabled.
 104        * dword 8-11 or 16-19 is the first vertex element data we fill.
 105        */
 106       assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
 107       assign_vue_slot(vue_map, VARYING_SLOT_POS);
 108       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0))
 109          assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST0);
 110       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1))
 111          assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST1);
 112
 113       /* front and back colors need to be consecutive so that we can use
 114        * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
 115        * two-sided color.
 116        */
 117       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL0))
 118          assign_vue_slot(vue_map, VARYING_SLOT_COL0);
 119       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC0))
 120          assign_vue_slot(vue_map, VARYING_SLOT_BFC0);
 121       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL1))
 122          assign_vue_slot(vue_map, VARYING_SLOT_COL1);
 123       if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC1))
 124          assign_vue_slot(vue_map, VARYING_SLOT_BFC1);
 125    }
 126
 127    /* The hardware doesn't care about the rest of the vertex outputs, so just
 128     * assign them contiguously.  Don't reassign outputs that already have a
 129     * slot.
 130     *
 131     * We generally don't need to assign a slot for VARYING_SLOT_CLIP_VERTEX,
 132     * since it's encoded as the clip distances by emit_clip_distances().
 133     * However, it may be output by transform feedback, and we'd rather not
 134     * recompute state when TF changes, so we just always include it.
 135     */
 136    for (int i = 0; i < VARYING_SLOT_MAX; ++i) {
 137       if ((slots_valid & BITFIELD64_BIT(i)) &&
 138           vue_map->varying_to_slot[i] == -1) {
 139          assign_vue_slot(vue_map, i);
 140       }
 141    }
 142 }
 143
 144
 145 /**
 146  * Decide which set of clip planes should be used when clipping via
 147  * gl_Position or gl_ClipVertex.
 148  */
 149 gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx)
 150 {
 151    if (ctx->_Shader->CurrentProgram[MESA_SHADER_VERTEX]) {
 152       /* There is currently a GLSL vertex shader, so clip according to GLSL
 153        * rules, which means compare gl_ClipVertex (or gl_Position, if
 154        * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
 155        * that were stored in EyeUserPlane at the time the clip planes were
 156        * specified.
 157        */
 158       return ctx->Transform.EyeUserPlane;
 159    } else {
 160       /* Either we are using fixed function or an ARB vertex program.  In
 161        * either case the clip planes are going to be compared against
 162        * gl_Position (which is in clip coordinates) so we have to clip using
 163        * _ClipUserPlane, which was transformed into clip coordinates by Mesa
 164        * core.
 165        */
 166       return ctx->Transform._ClipUserPlane;
 167    }
 168 }
 169
 170
 171 bool
 172 brw_vs_prog_data_compare(const void *in_a, const void *in_b)
 173 {
 174    const struct brw_vs_prog_data *a = in_a;
 175    const struct brw_vs_prog_data *b = in_b;
 176
 177    /* Compare the base structure. */
 178    if (!brw_stage_prog_data_compare(&a->base.base, &b->base.base))
 179       return false;
 180
 181    /* Compare the rest of the struct. */
 182    const unsigned offset = sizeof(struct brw_stage_prog_data);
 183    if (memcmp(((char *) a) + offset, ((char *) b) + offset,
 184               sizeof(struct brw_vs_prog_data) - offset)) {
 185       return false;
 186    }
 187
 188    return true;
 189 }
 190
 191 bool
 192 brw_compile_vs_prog(struct brw_context *brw,
 193                     struct gl_shader_program *prog,
 194                     struct brw_vertex_program *vp,
 195                     struct brw_vs_prog_key *key)
 196 {
 197    GLuint program_size;
 198    const GLuint *program;
 199    struct brw_vs_compile c;
 200    struct brw_vs_prog_data prog_data;
 201    struct brw_stage_prog_data *stage_prog_data = &prog_data.base.base;
 202    void *mem_ctx;
 203    int i;
 204    struct gl_shader *vs = NULL;
 205
 206    if (prog)
 207       vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
 208
 209    memset(&c, 0, sizeof(c));
 210    memcpy(&c.key, key, sizeof(*key));
 211    memset(&prog_data, 0, sizeof(prog_data));
 212
 213    /* Use ALT floating point mode for ARB programs so that 0^0 == 1. */
 214    if (!prog)
 215       stage_prog_data->use_alt_mode = true;
 216
 217    mem_ctx = ralloc_context(NULL);
 218
 219    c.vp = vp;
 220
 221    /* Allocate the references to the uniforms that will end up in the
 222     * prog_data associated with the compiled program, and which will be freed
 223     * by the state cache.
 224     */
 225    int param_count;
 226    if (vs) {
 227       /* We add padding around uniform values below vec4 size, with the worst
 228        * case being a float value that gets blown up to a vec4, so be
 229        * conservative here.
 230        */
 231       param_count = vs->num_uniform_components * 4;
 232
 233    } else {
 234       param_count = vp->program.Base.Parameters->NumParameters * 4;
 235    }
 236    /* vec4_visitor::setup_uniform_clipplane_values() also uploads user clip
 237     * planes as uniforms.
 238     */
 239    param_count += c.key.base.nr_userclip_plane_consts * 4;
 240
 241    stage_prog_data->param =
 242       rzalloc_array(NULL, const gl_constant_value *, param_count);
 243    stage_prog_data->pull_param =
 244       rzalloc_array(NULL, const gl_constant_value *, param_count);
 245    stage_prog_data->nr_params = param_count;
 246
 247    GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
 248    prog_data.inputs_read = vp->program.Base.InputsRead;
 249
 250    if (c.key.copy_edgeflag) {
 251       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_EDGE);
 252       prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
 253    }
 254
 255    if (brw->gen < 6) {
 256       /* Put dummy slots into the VUE for the SF to put the replaced
 257        * point sprite coords in.  We shouldn't need these dummy slots,
 258        * which take up precious URB space, but it would mean that the SF
 259        * doesn't get nice aligned pairs of input coords into output
 260        * coords, which would be a pain to handle.
 261        */
 262       for (i = 0; i < 8; i++) {
 263          if (c.key.point_coord_replace & (1 << i))
 264             outputs_written |= BITFIELD64_BIT(VARYING_SLOT_TEX0 + i);
 265       }
 266
 267       /* if back colors are written, allocate slots for front colors too */
 268       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC0))
 269          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL0);
 270       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC1))
 271          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL1);
 272    }
 273
 274    /* In order for legacy clipping to work, we need to populate the clip
 275     * distance varying slots whenever clipping is enabled, even if the vertex
 276     * shader doesn't write to gl_ClipDistance.
 277     */
 278    if (c.key.base.userclip_active) {
 279       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
 280       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
 281    }
 282
 283    brw_compute_vue_map(brw->intelScreen->devinfo,
 284                        &prog_data.base.vue_map, outputs_written);
 285
 286    if (0) {
 287       _mesa_fprint_program_opt(stderr, &c.vp->program.Base, PROG_PRINT_DEBUG,
 288                                true);
 289    }
 290
 291    /* Emit GEN4 code.
 292     */
 293    program = brw_vs_emit(brw, prog, &c, &prog_data, mem_ctx, &program_size);
 294    if (program == NULL) {
 295       ralloc_free(mem_ctx);
 296       return false;
 297    }
 298
 299    /* Scratch space is used for register spilling */
 300    if (c.base.last_scratch) {
 301       perf_debug("Vertex shader triggered register spilling.  "
 302                  "Try reducing the number of live vec4 values to "
 303                  "improve performance.\n");
 304
 305       prog_data.base.base.total_scratch
 306          = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
 307
 308       brw_get_scratch_bo(brw, &brw->vs.base.scratch_bo,
 309                          prog_data.base.base.total_scratch *
 310                          brw->max_vs_threads);
 311    }
 312
 313    brw_upload_cache(&brw->cache, BRW_CACHE_VS_PROG,
 314                     &c.key, sizeof(c.key),
 315                     program, program_size,
 316                     &prog_data, sizeof(prog_data),
 317                     &brw->vs.base.prog_offset, &brw->vs.prog_data);
 318    ralloc_free(mem_ctx);
 319
 320    return true;
 321 }
 322
 323 static bool
 324 key_debug(struct brw_context *brw, const char *name, int a, int b)
 325 {
 326    if (a != b) {
 327       perf_debug("  %s %d->%d\n", name, a, b);
 328       return true;
 329    }
 330    return false;
 331 }
 332
 333 void
 334 brw_vs_debug_recompile(struct brw_context *brw,
 335                        struct gl_shader_program *prog,
 336                        const struct brw_vs_prog_key *key)
 337 {
 338    struct brw_cache_item *c = NULL;
 339    const struct brw_vs_prog_key *old_key = NULL;
 340    bool found = false;
 341
 342    perf_debug("Recompiling vertex shader for program %d\n", prog->Name);
 343
 344    for (unsigned int i = 0; i < brw->cache.size; i++) {
 345       for (c = brw->cache.items[i]; c; c = c->next) {
 346          if (c->cache_id == BRW_CACHE_VS_PROG) {
 347             old_key = c->key;
 348
 349             if (old_key->base.program_string_id == key->base.program_string_id)
 350                break;
 351          }
 352       }
 353       if (c)
 354          break;
 355    }
 356
 357    if (!c) {
 358       perf_debug("  Didn't find previous compile in the shader cache for "
 359                  "debug\n");
 360       return;
 361    }
 362
 363    for (unsigned int i = 0; i < VERT_ATTRIB_MAX; i++) {
 364       found |= key_debug(brw, "Vertex attrib w/a flags",
 365                          old_key->gl_attrib_wa_flags[i],
 366                          key->gl_attrib_wa_flags[i]);
 367    }
 368
 369    found |= key_debug(brw, "user clip flags",
 370                       old_key->base.userclip_active, key->base.userclip_active);
 371
 372    found |= key_debug(brw, "user clipping planes as push constants",
 373                       old_key->base.nr_userclip_plane_consts,
 374                       key->base.nr_userclip_plane_consts);
 375
 376    found |= key_debug(brw, "copy edgeflag",
 377                       old_key->copy_edgeflag, key->copy_edgeflag);
 378    found |= key_debug(brw, "PointCoord replace",
 379                       old_key->point_coord_replace, key->point_coord_replace);
 380    found |= key_debug(brw, "vertex color clamping",
 381                       old_key->clamp_vertex_color, key->clamp_vertex_color);
 382
 383    found |= brw_debug_recompile_sampler_key(brw, &old_key->base.tex,
 384                                             &key->base.tex);
 385
 386    if (!found) {
 387       perf_debug("  Something else\n");
 388    }
 389 }
 390
 391
 392 void
 393 brw_setup_vue_key_clip_info(struct brw_context *brw,
 394                             struct brw_vue_prog_key *key,
 395                             bool program_uses_clip_distance)
 396 {
 397    struct gl_context *ctx = &brw->ctx;
 398
 399    key->userclip_active = (ctx->Transform.ClipPlanesEnabled != 0);
 400    if (key->userclip_active && !program_uses_clip_distance) {
 401       key->nr_userclip_plane_consts
 402          = _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1;
 403    }
 404 }
 405
 406 static bool
 407 brw_vs_state_dirty(struct brw_context *brw)
 408 {
 409    return brw_state_dirty(brw,
 410                           _NEW_BUFFERS |
 411                           _NEW_LIGHT |
 412                           _NEW_POINT |
 413                           _NEW_POLYGON |
 414                           _NEW_TEXTURE |
 415                           _NEW_TRANSFORM,
 416                           BRW_NEW_VERTEX_PROGRAM |
 417                           BRW_NEW_VS_ATTRIB_WORKAROUNDS);
 418 }
 419
 420 static void
 421 brw_vs_populate_key(struct brw_context *brw,
 422                     struct brw_vs_prog_key *key)
 423 {
 424    struct gl_context *ctx = &brw->ctx;
 425    /* BRW_NEW_VERTEX_PROGRAM */
 426    struct brw_vertex_program *vp =
 427       (struct brw_vertex_program *)brw->vertex_program;
 428    struct gl_program *prog = (struct gl_program *) brw->vertex_program;
 429    int i;
 430
 431    memset(key, 0, sizeof(*key));
 432
 433    /* Just upload the program verbatim for now.  Always send it all
 434     * the inputs it asks for, whether they are varying or not.
 435     */
 436    key->base.program_string_id = vp->id;
 437    brw_setup_vue_key_clip_info(brw, &key->base,
 438                                vp->program.Base.UsesClipDistanceOut);
 439
 440    /* _NEW_POLYGON */
 441    if (brw->gen < 6) {
 442       key->copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL ||
 443                             ctx->Polygon.BackMode != GL_FILL);
 444    }
 445
 446    if (prog->OutputsWritten & (VARYING_BIT_COL0 | VARYING_BIT_COL1 |
 447                                VARYING_BIT_BFC0 | VARYING_BIT_BFC1)) {
 448       /* _NEW_LIGHT | _NEW_BUFFERS */
 449       key->clamp_vertex_color = ctx->Light._ClampVertexColor;
 450    }
 451
 452    /* _NEW_POINT */
 453    if (brw->gen < 6 && ctx->Point.PointSprite) {
 454       for (i = 0; i < 8; i++) {
 455          if (ctx->Point.CoordReplace[i])
 456             key->point_coord_replace |= (1 << i);
 457       }
 458    }
 459
 460    /* _NEW_TEXTURE */
 461    brw_populate_sampler_prog_key_data(ctx, prog, brw->vs.base.sampler_count,
 462                                       &key->base.tex);
 463
 464    /* BRW_NEW_VS_ATTRIB_WORKAROUNDS */
 465    memcpy(key->gl_attrib_wa_flags, brw->vb.attrib_wa_flags,
 466           sizeof(brw->vb.attrib_wa_flags));
 467 }
 468
 469 void
 470 brw_upload_vs_prog(struct brw_context *brw)
 471 {
 472    struct gl_context *ctx = &brw->ctx;
 473    struct gl_shader_program **current = ctx->_Shader->CurrentProgram;
 474    struct brw_vs_prog_key key;
 475    /* BRW_NEW_VERTEX_PROGRAM */
 476    struct brw_vertex_program *vp =
 477       (struct brw_vertex_program *)brw->vertex_program;
 478
 479    if (!brw_vs_state_dirty(brw))
 480       return;
 481
 482    brw_vs_populate_key(brw, &key);
 483
 484    if (!brw_search_cache(&brw->cache, BRW_CACHE_VS_PROG,
 485                          &key, sizeof(key),
 486                          &brw->vs.base.prog_offset, &brw->vs.prog_data)) {
 487       bool success = brw_compile_vs_prog(brw, current[MESA_SHADER_VERTEX],
 488                                          vp, &key);
 489       (void) success;
 490       assert(success);
 491    }
 492    brw->vs.base.prog_data = &brw->vs.prog_data->base.base;
 493
 494    if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
 495               sizeof(brw->vue_map_geom_out)) != 0) {
 496       brw->vue_map_vs = brw->vs.prog_data->base.vue_map;
 497       brw->ctx.NewDriverState |= BRW_NEW_VUE_MAP_VS;
 498       if (brw->gen < 6) {
 499          /* No geometry shader support, so the VS VUE map is the VUE map for
 500           * the output of the "geometry" portion of the pipeline.
 501           */
 502          brw->vue_map_geom_out = brw->vue_map_vs;
 503          brw->ctx.NewDriverState |= BRW_NEW_VUE_MAP_GEOM_OUT;
 504       }
 505    }
 506 }
 507
 508 bool
 509 brw_vs_precompile(struct gl_context *ctx,
 510                   struct gl_shader_program *shader_prog,
 511                   struct gl_program *prog)
 512 {
 513    struct brw_context *brw = brw_context(ctx);
 514    struct brw_vs_prog_key key;
 515    uint32_t old_prog_offset = brw->vs.base.prog_offset;
 516    struct brw_vs_prog_data *old_prog_data = brw->vs.prog_data;
 517    bool success;
 518
 519    struct gl_vertex_program *vp = (struct gl_vertex_program *) prog;
 520    struct brw_vertex_program *bvp = brw_vertex_program(vp);
 521
 522    memset(&key, 0, sizeof(key));
 523
 524    brw_vue_setup_prog_key_for_precompile(ctx, &key.base, bvp->id, &vp->Base);
 525    key.clamp_vertex_color =
 526       (prog->OutputsWritten & (VARYING_BIT_COL0 | VARYING_BIT_COL1 |
 527                                VARYING_BIT_BFC0 | VARYING_BIT_BFC1));
 528
 529    success = brw_compile_vs_prog(brw, shader_prog, bvp, &key);
 530
 531    brw->vs.base.prog_offset = old_prog_offset;
 532    brw->vs.prog_data = old_prog_data;
 533
 534    return success;
 535 }