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 vert_result)
  45 {
  46    /* Make sure this vert_result hasn't been assigned a slot already */
  47    assert (vue_map->vert_result_to_slot[vert_result] == -1);
  48
  49    vue_map->vert_result_to_slot[vert_result] = vue_map->num_slots;
  50    vue_map->slot_to_vert_result[vue_map->num_slots++] = vert_result;
  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 static void
  61 brw_compute_vue_map(struct brw_vs_compile *c)
  62 {
  63    struct brw_context *brw = c->func.brw;
  64    const struct intel_context *intel = &brw->intel;
  65    struct brw_vue_map *vue_map = &c->prog_data.vue_map;
  66    GLbitfield64 outputs_written = c->prog_data.outputs_written;
  67    int i;
  68
  69    vue_map->num_slots = 0;
  70    for (i = 0; i < BRW_VERT_RESULT_MAX; ++i) {
  71       vue_map->vert_result_to_slot[i] = -1;
  72       vue_map->slot_to_vert_result[i] = BRW_VERT_RESULT_MAX;
  73    }
  74
  75    /* VUE header: format depends on chip generation and whether clipping is
  76     * enabled.
  77     */
  78    switch (intel->gen) {
  79    case 4:
  80       /* There are 8 dwords in VUE header pre-Ironlake:
  81        * dword 0-3 is indices, point width, clip flags.
  82        * dword 4-7 is ndc position
  83        * dword 8-11 is the first vertex data.
  84        */
  85       assign_vue_slot(vue_map, VERT_RESULT_PSIZ);
  86       assign_vue_slot(vue_map, BRW_VERT_RESULT_NDC);
  87       assign_vue_slot(vue_map, VERT_RESULT_HPOS);
  88       break;
  89    case 5:
  90       /* There are 20 DWs (D0-D19) in VUE header on Ironlake:
  91        * dword 0-3 of the header is indices, point width, clip flags.
  92        * dword 4-7 is the ndc position
  93        * dword 8-11 of the vertex header is the 4D space position
  94        * dword 12-19 of the vertex header is the user clip distance.
  95        * dword 20-23 is a pad so that the vertex element data is aligned
  96        * dword 24-27 is the first vertex data we fill.
  97        *
  98        * Note: future pipeline stages expect 4D space position to be
  99        * contiguous with the other vert_results, so we make dword 24-27 a
 100        * duplicate copy of the 4D space position.
 101        */
 102       assign_vue_slot(vue_map, VERT_RESULT_PSIZ);
 103       assign_vue_slot(vue_map, BRW_VERT_RESULT_NDC);
 104       assign_vue_slot(vue_map, BRW_VERT_RESULT_HPOS_DUPLICATE);
 105       assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST0);
 106       assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST1);
 107       assign_vue_slot(vue_map, BRW_VERT_RESULT_PAD);
 108       assign_vue_slot(vue_map, VERT_RESULT_HPOS);
 109       break;
 110    case 6:
 111    case 7:
 112       /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
 113        * dword 0-3 of the header is indices, point width, clip flags.
 114        * dword 4-7 is the 4D space position
 115        * dword 8-15 of the vertex header is the user clip distance if
 116        * enabled.
 117        * dword 8-11 or 16-19 is the first vertex element data we fill.
 118        */
 119       assign_vue_slot(vue_map, VERT_RESULT_PSIZ);
 120       assign_vue_slot(vue_map, VERT_RESULT_HPOS);
 121       if (c->key.userclip_active) {
 122          assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST0);
 123          assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST1);
 124       }
 125       /* front and back colors need to be consecutive so that we can use
 126        * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
 127        * two-sided color.
 128        */
 129       if (outputs_written & BITFIELD64_BIT(VERT_RESULT_COL0))
 130          assign_vue_slot(vue_map, VERT_RESULT_COL0);
 131       if (outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC0))
 132          assign_vue_slot(vue_map, VERT_RESULT_BFC0);
 133       if (outputs_written & BITFIELD64_BIT(VERT_RESULT_COL1))
 134          assign_vue_slot(vue_map, VERT_RESULT_COL1);
 135       if (outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC1))
 136          assign_vue_slot(vue_map, VERT_RESULT_BFC1);
 137       break;
 138    default:
 139       assert (!"VUE map not known for this chip generation");
 140       break;
 141    }
 142
 143    /* The hardware doesn't care about the rest of the vertex outputs, so just
 144     * assign them contiguously.  Don't reassign outputs that already have a
 145     * slot.
 146     *
 147     * Also, prior to Gen6, don't assign a slot for VERT_RESULT_CLIP_VERTEX,
 148     * since it is unsupported.  In Gen6 and above, VERT_RESULT_CLIP_VERTEX may
 149     * be needed for transform feedback; since we don't want to have to
 150     * recompute the VUE map (and everything that depends on it) when transform
 151     * feedback is enabled or disabled, just go ahead and assign a slot for it.
 152     */
 153    for (int i = 0; i < VERT_RESULT_MAX; ++i) {
 154       if (intel->gen < 6 && i == VERT_RESULT_CLIP_VERTEX)
 155          continue;
 156       if ((outputs_written & BITFIELD64_BIT(i)) &&
 157           vue_map->vert_result_to_slot[i] == -1) {
 158          assign_vue_slot(vue_map, i);
 159       }
 160    }
 161 }
 162
 163
 164 /**
 165  * Decide which set of clip planes should be used when clipping via
 166  * gl_Position or gl_ClipVertex.
 167  */
 168 gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx)
 169 {
 170    if (ctx->Shader.CurrentVertexProgram) {
 171       /* There is currently a GLSL vertex shader, so clip according to GLSL
 172        * rules, which means compare gl_ClipVertex (or gl_Position, if
 173        * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
 174        * that were stored in EyeUserPlane at the time the clip planes were
 175        * specified.
 176        */
 177       return ctx->Transform.EyeUserPlane;
 178    } else {
 179       /* Either we are using fixed function or an ARB vertex program.  In
 180        * either case the clip planes are going to be compared against
 181        * gl_Position (which is in clip coordinates) so we have to clip using
 182        * _ClipUserPlane, which was transformed into clip coordinates by Mesa
 183        * core.
 184        */
 185       return ctx->Transform._ClipUserPlane;
 186    }
 187 }
 188
 189 bool
 190 brw_vs_prog_data_compare(const void *in_a, const void *in_b,
 191                          int aux_size, const void *in_key)
 192 {
 193    const struct brw_vs_prog_data *a = in_a;
 194    const struct brw_vs_prog_data *b = in_b;
 195
 196    /* Compare all the struct up to the pointers. */
 197    if (memcmp(a, b, offsetof(struct brw_vs_prog_data, param)))
 198       return false;
 199
 200    if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
 201       return false;
 202
 203    if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
 204       return false;
 205
 206    return true;
 207 }
 208
 209 static bool
 210 do_vs_prog(struct brw_context *brw,
 211            struct gl_shader_program *prog,
 212            struct brw_vertex_program *vp,
 213            struct brw_vs_prog_key *key)
 214 {
 215    struct intel_context *intel = &brw->intel;
 216    GLuint program_size;
 217    const GLuint *program;
 218    struct brw_vs_compile c;
 219    void *mem_ctx;
 220    int i;
 221    struct gl_shader *vs = NULL;
 222
 223    if (prog)
 224       vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
 225
 226    memset(&c, 0, sizeof(c));
 227    memcpy(&c.key, key, sizeof(*key));
 228
 229    mem_ctx = ralloc_context(NULL);
 230
 231    brw_init_compile(brw, &c.func, mem_ctx);
 232    c.vp = vp;
 233
 234    /* Allocate the references to the uniforms that will end up in the
 235     * prog_data associated with the compiled program, and which will be freed
 236     * by the state cache.
 237     */
 238    int param_count;
 239    if (vs) {
 240       /* We add padding around uniform values below vec4 size, with the worst
 241        * case being a float value that gets blown up to a vec4, so be
 242        * conservative here.
 243        */
 244       param_count = vs->num_uniform_components * 4;
 245
 246    } else {
 247       param_count = vp->program.Base.Parameters->NumParameters * 4;
 248    }
 249    /* We also upload clip plane data as uniforms */
 250    param_count += MAX_CLIP_PLANES * 4;
 251
 252    c.prog_data.param = rzalloc_array(NULL, const float *, param_count);
 253    c.prog_data.pull_param = rzalloc_array(NULL, const float *, param_count);
 254
 255    c.prog_data.outputs_written = vp->program.Base.OutputsWritten;
 256    c.prog_data.inputs_read = vp->program.Base.InputsRead;
 257
 258    if (c.key.copy_edgeflag) {
 259       c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_EDGE);
 260       c.prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
 261    }
 262
 263    /* Put dummy slots into the VUE for the SF to put the replaced
 264     * point sprite coords in.  We shouldn't need these dummy slots,
 265     * which take up precious URB space, but it would mean that the SF
 266     * doesn't get nice aligned pairs of input coords into output
 267     * coords, which would be a pain to handle.
 268     */
 269    for (i = 0; i < 8; i++) {
 270       if (c.key.point_coord_replace & (1 << i))
 271          c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_TEX0 + i);
 272    }
 273
 274    brw_compute_vue_map(&c);
 275
 276    if (0) {
 277       _mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG,
 278                                true);
 279    }
 280
 281    /* Emit GEN4 code.
 282     */
 283    if (!brw_vs_emit(prog, &c)) {
 284       ralloc_free(mem_ctx);
 285       return false;
 286    }
 287
 288    if (c.prog_data.nr_pull_params)
 289       c.prog_data.num_surfaces = 1;
 290    if (c.vp->program.Base.SamplersUsed)
 291       c.prog_data.num_surfaces = SURF_INDEX_VS_TEXTURE(BRW_MAX_TEX_UNIT);
 292    if (prog &&
 293        prog->_LinkedShaders[MESA_SHADER_VERTEX]->NumUniformBlocks) {
 294       c.prog_data.num_surfaces =
 295          SURF_INDEX_VS_UBO(prog->_LinkedShaders[MESA_SHADER_VERTEX]->NumUniformBlocks);
 296    }
 297
 298    /* Scratch space is used for register spilling */
 299    if (c.last_scratch) {
 300       perf_debug("Vertex shader triggered register spilling.  "
 301                  "Try reducing the number of live vec4 values to "
 302                  "improve performance.\n");
 303
 304       c.prog_data.total_scratch = brw_get_scratch_size(c.last_scratch*REG_SIZE);
 305
 306       brw_get_scratch_bo(intel, &brw->vs.scratch_bo,
 307                          c.prog_data.total_scratch * brw->max_vs_threads);
 308    }
 309
 310    /* get the program
 311     */
 312    program = brw_get_program(&c.func, &program_size);
 313
 314    brw_upload_cache(&brw->cache, BRW_VS_PROG,
 315                     &c.key, sizeof(c.key),
 316                     program, program_size,
 317                     &c.prog_data, sizeof(c.prog_data),
 318                     &brw->vs.prog_offset, &brw->vs.prog_data);
 319    ralloc_free(mem_ctx);
 320
 321    return true;
 322 }
 323
 324 static bool
 325 key_debug(const char *name, int a, int b)
 326 {
 327    if (a != b) {
 328       perf_debug("  %s %d->%d\n", name, a, b);
 329       return true;
 330    }
 331    return false;
 332 }
 333
 334 void
 335 brw_vs_debug_recompile(struct brw_context *brw,
 336                        struct gl_shader_program *prog,
 337                        const struct brw_vs_prog_key *key)
 338 {
 339    struct brw_cache_item *c = NULL;
 340    const struct brw_vs_prog_key *old_key = NULL;
 341    bool found = false;
 342
 343    perf_debug("Recompiling vertex shader for program %d\n", prog->Name);
 344
 345    for (unsigned int i = 0; i < brw->cache.size; i++) {
 346       for (c = brw->cache.items[i]; c; c = c->next) {
 347          if (c->cache_id == BRW_VS_PROG) {
 348             old_key = c->key;
 349
 350             if (old_key->program_string_id == key->program_string_id)
 351                break;
 352          }
 353       }
 354       if (c)
 355          break;
 356    }
 357
 358    if (!c) {
 359       perf_debug("  Didn't find previous compile in the shader cache for "
 360                  "debug\n");
 361       return;
 362    }
 363
 364    for (unsigned int i = 0; i < VERT_ATTRIB_MAX; i++) {
 365       found |= key_debug("Vertex attrib w/a flags",
 366                          old_key->gl_attrib_wa_flags[i],
 367                          key->gl_attrib_wa_flags[i]);
 368    }
 369
 370    found |= key_debug("user clip flags",
 371                       old_key->userclip_active, key->userclip_active);
 372
 373    found |= key_debug("user clipping planes as push constants",
 374                       old_key->nr_userclip_plane_consts,
 375                       key->nr_userclip_plane_consts);
 376
 377    found |= key_debug("clip distance enable",
 378                       old_key->uses_clip_distance, key->uses_clip_distance);
 379    found |= key_debug("clip plane enable bitfield",
 380                       old_key->userclip_planes_enabled_gen_4_5,
 381                       key->userclip_planes_enabled_gen_4_5);
 382    found |= key_debug("copy edgeflag",
 383                       old_key->copy_edgeflag, key->copy_edgeflag);
 384    found |= key_debug("PointCoord replace",
 385                       old_key->point_coord_replace, key->point_coord_replace);
 386    found |= key_debug("vertex color clamping",
 387                       old_key->clamp_vertex_color, key->clamp_vertex_color);
 388
 389    found |= brw_debug_recompile_sampler_key(&old_key->tex, &key->tex);
 390
 391    if (!found) {
 392       perf_debug("  Something else\n");
 393    }
 394 }
 395
 396 static void brw_upload_vs_prog(struct brw_context *brw)
 397 {
 398    struct intel_context *intel = &brw->intel;
 399    struct gl_context *ctx = &intel->ctx;
 400    struct brw_vs_prog_key key;
 401    /* BRW_NEW_VERTEX_PROGRAM */
 402    struct brw_vertex_program *vp =
 403       (struct brw_vertex_program *)brw->vertex_program;
 404    struct gl_program *prog = (struct gl_program *) brw->vertex_program;
 405    int i;
 406
 407    memset(&key, 0, sizeof(key));
 408
 409    /* Just upload the program verbatim for now.  Always send it all
 410     * the inputs it asks for, whether they are varying or not.
 411     */
 412    key.program_string_id = vp->id;
 413    key.userclip_active = (ctx->Transform.ClipPlanesEnabled != 0);
 414    key.uses_clip_distance = vp->program.UsesClipDistance;
 415    if (key.userclip_active && !key.uses_clip_distance) {
 416       if (intel->gen < 6) {
 417          key.nr_userclip_plane_consts
 418             = _mesa_bitcount_64(ctx->Transform.ClipPlanesEnabled);
 419          key.userclip_planes_enabled_gen_4_5
 420             = ctx->Transform.ClipPlanesEnabled;
 421       } else {
 422          key.nr_userclip_plane_consts
 423             = _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1;
 424       }
 425    }
 426
 427    /* _NEW_POLYGON */
 428    if (intel->gen < 6) {
 429       key.copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL ||
 430                            ctx->Polygon.BackMode != GL_FILL);
 431    }
 432
 433    /* _NEW_LIGHT | _NEW_BUFFERS */
 434    key.clamp_vertex_color = ctx->Light._ClampVertexColor;
 435
 436    /* _NEW_POINT */
 437    if (ctx->Point.PointSprite) {
 438       for (i = 0; i < 8; i++) {
 439          if (ctx->Point.CoordReplace[i])
 440             key.point_coord_replace |= (1 << i);
 441       }
 442    }
 443
 444    /* _NEW_TEXTURE */
 445    brw_populate_sampler_prog_key_data(ctx, prog, &key.tex);
 446
 447    /* BRW_NEW_VERTICES */
 448    for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 449       /* TODO: flag w/a for packed vertex formats here too */
 450       if (vp->program.Base.InputsRead & BITFIELD64_BIT(i) &&
 451           brw->vb.inputs[i].glarray->Type == GL_FIXED) {
 452          key.gl_attrib_wa_flags[i] = brw->vb.inputs[i].glarray->Size;
 453       }
 454    }
 455
 456    if (!brw_search_cache(&brw->cache, BRW_VS_PROG,
 457                          &key, sizeof(key),
 458                          &brw->vs.prog_offset, &brw->vs.prog_data)) {
 459       bool success = do_vs_prog(brw, ctx->Shader.CurrentVertexProgram,
 460                                 vp, &key);
 461
 462       assert(success);
 463    }
 464 }
 465
 466 /* See brw_vs.c:
 467  */
 468 const struct brw_tracked_state brw_vs_prog = {
 469    .dirty = {
 470       .mesa  = (_NEW_TRANSFORM | _NEW_POLYGON | _NEW_POINT | _NEW_LIGHT |
 471                 _NEW_TEXTURE |
 472                 _NEW_BUFFERS),
 473       .brw   = (BRW_NEW_VERTEX_PROGRAM |
 474                 BRW_NEW_VERTICES),
 475       .cache = 0
 476    },
 477    .emit = brw_upload_vs_prog
 478 };
 479
 480 bool
 481 brw_vs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
 482 {
 483    struct brw_context *brw = brw_context(ctx);
 484    struct brw_vs_prog_key key;
 485    uint32_t old_prog_offset = brw->vs.prog_offset;
 486    struct brw_vs_prog_data *old_prog_data = brw->vs.prog_data;
 487    bool success;
 488
 489    if (!prog->_LinkedShaders[MESA_SHADER_VERTEX])
 490       return true;
 491
 492    struct gl_vertex_program *vp = (struct gl_vertex_program *)
 493       prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program;
 494    struct brw_vertex_program *bvp = brw_vertex_program(vp);
 495
 496    memset(&key, 0, sizeof(key));
 497
 498    key.program_string_id = bvp->id;
 499    key.clamp_vertex_color = true;
 500
 501    for (int i = 0; i < MAX_SAMPLERS; i++) {
 502       if (vp->Base.ShadowSamplers & (1 << i)) {
 503          /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
 504          key.tex.swizzles[i] =
 505             MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_ONE);
 506       } else {
 507          /* Color sampler: assume no swizzling. */
 508          key.tex.swizzles[i] = SWIZZLE_XYZW;
 509       }
 510    }
 511
 512    success = do_vs_prog(brw, prog, bvp, &key);
 513
 514    brw->vs.prog_offset = old_prog_offset;
 515    brw->vs.prog_data = old_prog_data;
 516
 517    return success;
 518 }
 519
 520 void
 521 brw_vs_prog_data_free(const void *in_prog_data)
 522 {
 523    const struct brw_vs_prog_data *prog_data = in_prog_data;
 524
 525    ralloc_free((void *)prog_data->param);
 526    ralloc_free((void *)prog_data->pull_param);
 527 }