src/mesa/drivers/dri/i965/brw_wm.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 #include "brw_context.h"
  33 #include "brw_wm.h"
  34 #include "brw_state.h"
  35 #include "main/enums.h"
  36 #include "main/formats.h"
  37 #include "main/fbobject.h"
  38 #include "main/samplerobj.h"
  39 #include "program/prog_parameter.h"
  40
  41 #include "glsl/ralloc.h"
  42
  43 /**
  44  * Return a bitfield where bit n is set if barycentric interpolation mode n
  45  * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
  46  */
  47 static unsigned
  48 brw_compute_barycentric_interp_modes(struct brw_context *brw,
  49                                      bool shade_model_flat,
  50                                      const struct gl_fragment_program *fprog)
  51 {
  52    unsigned barycentric_interp_modes = 0;
  53    int attr;
  54
  55    /* Loop through all fragment shader inputs to figure out what interpolation
  56     * modes are in use, and set the appropriate bits in
  57     * barycentric_interp_modes.
  58     */
  59    for (attr = 0; attr < VARYING_SLOT_MAX; ++attr) {
  60       enum glsl_interp_qualifier interp_qualifier =
  61          fprog->InterpQualifier[attr];
  62       bool is_centroid = fprog->IsCentroid & BITFIELD64_BIT(attr);
  63       bool is_gl_Color = attr == VARYING_SLOT_COL0 || attr == VARYING_SLOT_COL1;
  64
  65       /* Ignore unused inputs. */
  66       if (!(fprog->Base.InputsRead & BITFIELD64_BIT(attr)))
  67          continue;
  68
  69       /* Ignore WPOS and FACE, because they don't require interpolation. */
  70       if (attr == VARYING_SLOT_POS || attr == VARYING_SLOT_FACE)
  71          continue;
  72
  73       /* Determine the set (or sets) of barycentric coordinates needed to
  74        * interpolate this variable.  Note that when
  75        * brw->needs_unlit_centroid_workaround is set, centroid interpolation
  76        * uses PIXEL interpolation for unlit pixels and CENTROID interpolation
  77        * for lit pixels, so we need both sets of barycentric coordinates.
  78        */
  79       if (interp_qualifier == INTERP_QUALIFIER_NOPERSPECTIVE) {
  80          if (is_centroid) {
  81             barycentric_interp_modes |=
  82                1 << BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC;
  83          }
  84          if (!is_centroid || brw->needs_unlit_centroid_workaround) {
  85             barycentric_interp_modes |=
  86                1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
  87          }
  88       } else if (interp_qualifier == INTERP_QUALIFIER_SMOOTH ||
  89                  (!(shade_model_flat && is_gl_Color) &&
  90                   interp_qualifier == INTERP_QUALIFIER_NONE)) {
  91          if (is_centroid) {
  92             barycentric_interp_modes |=
  93                1 << BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
  94          }
  95          if (!is_centroid || brw->needs_unlit_centroid_workaround) {
  96             barycentric_interp_modes |=
  97                1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
  98          }
  99       }
 100    }
 101
 102    return barycentric_interp_modes;
 103 }
 104
 105 bool
 106 brw_wm_prog_data_compare(const void *in_a, const void *in_b)
 107 {
 108    const struct brw_wm_prog_data *a = in_a;
 109    const struct brw_wm_prog_data *b = in_b;
 110
 111    /* Compare all the struct up to the pointers. */
 112    if (memcmp(a, b, offsetof(struct brw_wm_prog_data, param)))
 113       return false;
 114
 115    if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
 116       return false;
 117
 118    if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
 119       return false;
 120
 121    return true;
 122 }
 123
 124 void
 125 brw_wm_prog_data_free(const void *in_prog_data)
 126 {
 127    const struct brw_wm_prog_data *prog_data = in_prog_data;
 128
 129    ralloc_free((void *)prog_data->param);
 130    ralloc_free((void *)prog_data->pull_param);
 131 }
 132
 133 /**
 134  * All Mesa program -> GPU code generation goes through this function.
 135  * Depending on the instructions used (i.e. flow control instructions)
 136  * we'll use one of two code generators.
 137  */
 138 bool do_wm_prog(struct brw_context *brw,
 139                 struct gl_shader_program *prog,
 140                 struct brw_fragment_program *fp,
 141                 struct brw_wm_prog_key *key)
 142 {
 143    struct brw_wm_compile *c;
 144    const GLuint *program;
 145    struct gl_shader *fs = NULL;
 146    GLuint program_size;
 147
 148    if (prog)
 149       fs = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
 150
 151    c = rzalloc(NULL, struct brw_wm_compile);
 152
 153    /* Allocate the references to the uniforms that will end up in the
 154     * prog_data associated with the compiled program, and which will be freed
 155     * by the state cache.
 156     */
 157    int param_count;
 158    if (fs) {
 159       param_count = fs->num_uniform_components;
 160    } else {
 161       param_count = fp->program.Base.Parameters->NumParameters * 4;
 162    }
 163    /* The backend also sometimes adds params for texture size. */
 164    param_count += 2 * BRW_MAX_TEX_UNIT;
 165    c->prog_data.param = rzalloc_array(NULL, const float *, param_count);
 166    c->prog_data.pull_param = rzalloc_array(NULL, const float *, param_count);
 167
 168    memcpy(&c->key, key, sizeof(*key));
 169
 170    c->prog_data.barycentric_interp_modes =
 171       brw_compute_barycentric_interp_modes(brw, c->key.flat_shade,
 172                                            &fp->program);
 173
 174    program = brw_wm_fs_emit(brw, c, &fp->program, prog, &program_size);
 175    if (program == NULL)
 176       return false;
 177
 178    /* Scratch space is used for register spilling */
 179    if (c->last_scratch) {
 180       perf_debug("Fragment shader triggered register spilling.  "
 181                  "Try reducing the number of live scalar values to "
 182                  "improve performance.\n");
 183
 184       c->prog_data.total_scratch = brw_get_scratch_size(c->last_scratch);
 185
 186       brw_get_scratch_bo(brw, &brw->wm.base.scratch_bo,
 187                          c->prog_data.total_scratch * brw->max_wm_threads);
 188    }
 189
 190    if (unlikely(INTEL_DEBUG & DEBUG_WM))
 191       fprintf(stderr, "\n");
 192
 193    brw_upload_cache(&brw->cache, BRW_WM_PROG,
 194                     &c->key, sizeof(c->key),
 195                     program, program_size,
 196                     &c->prog_data, sizeof(c->prog_data),
 197                     &brw->wm.base.prog_offset, &brw->wm.prog_data);
 198
 199    ralloc_free(c);
 200
 201    return true;
 202 }
 203
 204 static bool
 205 key_debug(struct brw_context *brw, const char *name, int a, int b)
 206 {
 207    if (a != b) {
 208       perf_debug("  %s %d->%d\n", name, a, b);
 209       return true;
 210    } else {
 211       return false;
 212    }
 213 }
 214
 215 bool
 216 brw_debug_recompile_sampler_key(struct brw_context *brw,
 217                                 const struct brw_sampler_prog_key_data *old_key,
 218                                 const struct brw_sampler_prog_key_data *key)
 219 {
 220    bool found = false;
 221
 222    for (unsigned int i = 0; i < MAX_SAMPLERS; i++) {
 223       found |= key_debug(brw, "EXT_texture_swizzle or DEPTH_TEXTURE_MODE",
 224                          old_key->swizzles[i], key->swizzles[i]);
 225    }
 226    found |= key_debug(brw, "GL_CLAMP enabled on any texture unit's 1st coordinate",
 227                       old_key->gl_clamp_mask[0], key->gl_clamp_mask[0]);
 228    found |= key_debug(brw, "GL_CLAMP enabled on any texture unit's 2nd coordinate",
 229                       old_key->gl_clamp_mask[1], key->gl_clamp_mask[1]);
 230    found |= key_debug(brw, "GL_CLAMP enabled on any texture unit's 3rd coordinate",
 231                       old_key->gl_clamp_mask[2], key->gl_clamp_mask[2]);
 232    found |= key_debug(brw, "GL_MESA_ycbcr texturing\n",
 233                       old_key->yuvtex_mask, key->yuvtex_mask);
 234    found |= key_debug(brw, "GL_MESA_ycbcr UV swapping\n",
 235                       old_key->yuvtex_swap_mask, key->yuvtex_swap_mask);
 236    found |= key_debug(brw, "gather channel quirk on any texture unit",
 237                       old_key->gather_channel_quirk_mask, key->gather_channel_quirk_mask);
 238
 239    return found;
 240 }
 241
 242 void
 243 brw_wm_debug_recompile(struct brw_context *brw,
 244                        struct gl_shader_program *prog,
 245                        const struct brw_wm_prog_key *key)
 246 {
 247    struct brw_cache_item *c = NULL;
 248    const struct brw_wm_prog_key *old_key = NULL;
 249    bool found = false;
 250
 251    perf_debug("Recompiling fragment shader for program %d\n", prog->Name);
 252
 253    for (unsigned int i = 0; i < brw->cache.size; i++) {
 254       for (c = brw->cache.items[i]; c; c = c->next) {
 255          if (c->cache_id == BRW_WM_PROG) {
 256             old_key = c->key;
 257
 258             if (old_key->program_string_id == key->program_string_id)
 259                break;
 260          }
 261       }
 262       if (c)
 263          break;
 264    }
 265
 266    if (!c) {
 267       perf_debug("  Didn't find previous compile in the shader cache for debug\n");
 268       return;
 269    }
 270
 271    found |= key_debug(brw, "alphatest, computed depth, depth test, or "
 272                       "depth write",
 273                       old_key->iz_lookup, key->iz_lookup);
 274    found |= key_debug(brw, "depth statistics",
 275                       old_key->stats_wm, key->stats_wm);
 276    found |= key_debug(brw, "flat shading",
 277                       old_key->flat_shade, key->flat_shade);
 278    found |= key_debug(brw, "number of color buffers",
 279                       old_key->nr_color_regions, key->nr_color_regions);
 280    found |= key_debug(brw, "MRT alpha test or alpha-to-coverage",
 281                       old_key->replicate_alpha, key->replicate_alpha);
 282    found |= key_debug(brw, "rendering to FBO",
 283                       old_key->render_to_fbo, key->render_to_fbo);
 284    found |= key_debug(brw, "fragment color clamping",
 285                       old_key->clamp_fragment_color, key->clamp_fragment_color);
 286    found |= key_debug(brw, "line smoothing",
 287                       old_key->line_aa, key->line_aa);
 288    found |= key_debug(brw, "renderbuffer height",
 289                       old_key->drawable_height, key->drawable_height);
 290    found |= key_debug(brw, "input slots valid",
 291                       old_key->input_slots_valid, key->input_slots_valid);
 292
 293    found |= brw_debug_recompile_sampler_key(brw, &old_key->tex, &key->tex);
 294
 295    if (!found) {
 296       perf_debug("  Something else\n");
 297    }
 298 }
 299
 300 void
 301 brw_populate_sampler_prog_key_data(struct gl_context *ctx,
 302                                    const struct gl_program *prog,
 303                                    unsigned sampler_count,
 304                                    struct brw_sampler_prog_key_data *key)
 305 {
 306    struct brw_context *brw = brw_context(ctx);
 307
 308    for (int s = 0; s < sampler_count; s++) {
 309       key->swizzles[s] = SWIZZLE_NOOP;
 310
 311       if (!(prog->SamplersUsed & (1 << s)))
 312          continue;
 313
 314       int unit_id = prog->SamplerUnits[s];
 315       const struct gl_texture_unit *unit = &ctx->Texture.Unit[unit_id];
 316
 317       if (unit->_ReallyEnabled && unit->_Current->Target != GL_TEXTURE_BUFFER) {
 318          const struct gl_texture_object *t = unit->_Current;
 319          const struct gl_texture_image *img = t->Image[0][t->BaseLevel];
 320          struct gl_sampler_object *sampler = _mesa_get_samplerobj(ctx, unit_id);
 321
 322          const bool alpha_depth = t->DepthMode == GL_ALPHA &&
 323             (img->_BaseFormat == GL_DEPTH_COMPONENT ||
 324              img->_BaseFormat == GL_DEPTH_STENCIL);
 325
 326          /* Haswell handles texture swizzling as surface format overrides
 327           * (except for GL_ALPHA); all other platforms need MOVs in the shader.
 328           */
 329          if (!brw->is_haswell || alpha_depth)
 330             key->swizzles[s] = brw_get_texture_swizzle(ctx, t);
 331
 332          if (img->InternalFormat == GL_YCBCR_MESA) {
 333             key->yuvtex_mask |= 1 << s;
 334             if (img->TexFormat == MESA_FORMAT_YCBCR)
 335                 key->yuvtex_swap_mask |= 1 << s;
 336          }
 337
 338          if (sampler->MinFilter != GL_NEAREST &&
 339              sampler->MagFilter != GL_NEAREST) {
 340             if (sampler->WrapS == GL_CLAMP)
 341                key->gl_clamp_mask[0] |= 1 << s;
 342             if (sampler->WrapT == GL_CLAMP)
 343                key->gl_clamp_mask[1] |= 1 << s;
 344             if (sampler->WrapR == GL_CLAMP)
 345                key->gl_clamp_mask[2] |= 1 << s;
 346          }
 347
 348          /* gather4's channel select for green from RG32F is broken;
 349           * requires a shader w/a on IVB; fixable with just SCS on HSW. */
 350          if (brw->gen >= 7 && !brw->is_haswell && prog->UsesGather) {
 351             if (img->InternalFormat == GL_RG32F)
 352                key->gather_channel_quirk_mask |= 1 << s;
 353          }
 354       }
 355    }
 356 }
 357
 358 static void brw_wm_populate_key( struct brw_context *brw,
 359                                  struct brw_wm_prog_key *key )
 360 {
 361    struct gl_context *ctx = &brw->ctx;
 362    /* BRW_NEW_FRAGMENT_PROGRAM */
 363    const struct brw_fragment_program *fp =
 364       (struct brw_fragment_program *)brw->fragment_program;
 365    const struct gl_program *prog = (struct gl_program *) brw->fragment_program;
 366    GLuint lookup = 0;
 367    GLuint line_aa;
 368    bool program_uses_dfdy = fp->program.UsesDFdy;
 369
 370    memset(key, 0, sizeof(*key));
 371
 372    /* Build the index for table lookup
 373     */
 374    if (brw->gen < 6) {
 375       /* _NEW_COLOR */
 376       if (fp->program.UsesKill || ctx->Color.AlphaEnabled)
 377          lookup |= IZ_PS_KILL_ALPHATEST_BIT;
 378
 379       if (fp->program.Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
 380          lookup |= IZ_PS_COMPUTES_DEPTH_BIT;
 381
 382       /* _NEW_DEPTH */
 383       if (ctx->Depth.Test)
 384          lookup |= IZ_DEPTH_TEST_ENABLE_BIT;
 385
 386       if (ctx->Depth.Test && ctx->Depth.Mask) /* ?? */
 387          lookup |= IZ_DEPTH_WRITE_ENABLE_BIT;
 388
 389       /* _NEW_STENCIL | _NEW_BUFFERS */
 390       if (ctx->Stencil._Enabled) {
 391          lookup |= IZ_STENCIL_TEST_ENABLE_BIT;
 392
 393          if (ctx->Stencil.WriteMask[0] ||
 394              ctx->Stencil.WriteMask[ctx->Stencil._BackFace])
 395             lookup |= IZ_STENCIL_WRITE_ENABLE_BIT;
 396       }
 397       key->iz_lookup = lookup;
 398    }
 399
 400    line_aa = AA_NEVER;
 401
 402    /* _NEW_LINE, _NEW_POLYGON, BRW_NEW_REDUCED_PRIMITIVE */
 403    if (ctx->Line.SmoothFlag) {
 404       if (brw->reduced_primitive == GL_LINES) {
 405          line_aa = AA_ALWAYS;
 406       }
 407       else if (brw->reduced_primitive == GL_TRIANGLES) {
 408          if (ctx->Polygon.FrontMode == GL_LINE) {
 409             line_aa = AA_SOMETIMES;
 410
 411             if (ctx->Polygon.BackMode == GL_LINE ||
 412                 (ctx->Polygon.CullFlag &&
 413                  ctx->Polygon.CullFaceMode == GL_BACK))
 414                line_aa = AA_ALWAYS;
 415          }
 416          else if (ctx->Polygon.BackMode == GL_LINE) {
 417             line_aa = AA_SOMETIMES;
 418
 419             if ((ctx->Polygon.CullFlag &&
 420                  ctx->Polygon.CullFaceMode == GL_FRONT))
 421                line_aa = AA_ALWAYS;
 422          }
 423       }
 424    }
 425
 426    key->line_aa = line_aa;
 427
 428    /* _NEW_HINT */
 429    if (brw->disable_derivative_optimization) {
 430       key->high_quality_derivatives =
 431          ctx->Hint.FragmentShaderDerivative != GL_FASTEST;
 432    } else {
 433       key->high_quality_derivatives =
 434          ctx->Hint.FragmentShaderDerivative == GL_NICEST;
 435    }
 436
 437    if (brw->gen < 6)
 438       key->stats_wm = brw->stats_wm;
 439
 440    /* _NEW_LIGHT */
 441    key->flat_shade = (ctx->Light.ShadeModel == GL_FLAT);
 442
 443    /* _NEW_FRAG_CLAMP | _NEW_BUFFERS */
 444    key->clamp_fragment_color = ctx->Color._ClampFragmentColor;
 445
 446    /* _NEW_TEXTURE */
 447    brw_populate_sampler_prog_key_data(ctx, prog, brw->wm.base.sampler_count,
 448                                       &key->tex);
 449
 450    /* _NEW_BUFFERS */
 451    /*
 452     * Include the draw buffer origin and height so that we can calculate
 453     * fragment position values relative to the bottom left of the drawable,
 454     * from the incoming screen origin relative position we get as part of our
 455     * payload.
 456     *
 457     * This is only needed for the WM_WPOSXY opcode when the fragment program
 458     * uses the gl_FragCoord input.
 459     *
 460     * We could avoid recompiling by including this as a constant referenced by
 461     * our program, but if we were to do that it would also be nice to handle
 462     * getting that constant updated at batchbuffer submit time (when we
 463     * hold the lock and know where the buffer really is) rather than at emit
 464     * time when we don't hold the lock and are just guessing.  We could also
 465     * just avoid using this as key data if the program doesn't use
 466     * fragment.position.
 467     *
 468     * For DRI2 the origin_x/y will always be (0,0) but we still need the
 469     * drawable height in order to invert the Y axis.
 470     */
 471    if (fp->program.Base.InputsRead & VARYING_BIT_POS) {
 472       key->drawable_height = ctx->DrawBuffer->Height;
 473    }
 474
 475    if ((fp->program.Base.InputsRead & VARYING_BIT_POS) || program_uses_dfdy) {
 476       key->render_to_fbo = _mesa_is_user_fbo(ctx->DrawBuffer);
 477    }
 478
 479    /* _NEW_BUFFERS */
 480    key->nr_color_regions = ctx->DrawBuffer->_NumColorDrawBuffers;
 481
 482    /* _NEW_MULTISAMPLE, _NEW_COLOR, _NEW_BUFFERS */
 483    key->replicate_alpha = ctx->DrawBuffer->_NumColorDrawBuffers > 1 &&
 484       (ctx->Multisample.SampleAlphaToCoverage || ctx->Color.AlphaEnabled);
 485
 486    /* BRW_NEW_VUE_MAP_GEOM_OUT */
 487    if (brw->gen < 6 || _mesa_bitcount_64(fp->program.Base.InputsRead &
 488                                          BRW_FS_VARYING_INPUT_MASK) > 16)
 489       key->input_slots_valid = brw->vue_map_geom_out.slots_valid;
 490
 491    /* The unique fragment program ID */
 492    key->program_string_id = fp->id;
 493 }
 494
 495
 496 static void
 497 brw_upload_wm_prog(struct brw_context *brw)
 498 {
 499    struct gl_context *ctx = &brw->ctx;
 500    struct brw_wm_prog_key key;
 501    struct brw_fragment_program *fp = (struct brw_fragment_program *)
 502       brw->fragment_program;
 503
 504    brw_wm_populate_key(brw, &key);
 505
 506    if (!brw_search_cache(&brw->cache, BRW_WM_PROG,
 507                          &key, sizeof(key),
 508                          &brw->wm.base.prog_offset, &brw->wm.prog_data)) {
 509       bool success = do_wm_prog(brw, ctx->Shader._CurrentFragmentProgram, fp,
 510                                 &key);
 511       (void) success;
 512       assert(success);
 513    }
 514 }
 515
 516
 517 const struct brw_tracked_state brw_wm_prog = {
 518    .dirty = {
 519       .mesa  = (_NEW_COLOR |
 520                 _NEW_DEPTH |
 521                 _NEW_STENCIL |
 522                 _NEW_POLYGON |
 523                 _NEW_LINE |
 524                 _NEW_HINT |
 525                 _NEW_LIGHT |
 526                 _NEW_FRAG_CLAMP |
 527                 _NEW_BUFFERS |
 528                 _NEW_TEXTURE |
 529                 _NEW_MULTISAMPLE),
 530       .brw   = (BRW_NEW_FRAGMENT_PROGRAM |
 531                 BRW_NEW_REDUCED_PRIMITIVE |
 532                 BRW_NEW_VUE_MAP_GEOM_OUT |
 533                 BRW_NEW_STATS_WM)
 534    },
 535    .emit = brw_upload_wm_prog
 536 };
 537