src/mesa/drivers/dri/i965/brw_wm.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 #include "brw_context.h"
  27 #include "brw_wm.h"
  28 #include "brw_state.h"
  29 #include "main/enums.h"
  30 #include "main/formats.h"
  31 #include "main/fbobject.h"
  32 #include "main/samplerobj.h"
  33 #include "main/framebuffer.h"
  34 #include "program/prog_parameter.h"
  35 #include "program/program.h"
  36 #include "intel_mipmap_tree.h"
  37 #include "intel_image.h"
  38 #include "compiler/brw_nir.h"
  39 #include "brw_program.h"
  40
  41 #include "util/ralloc.h"
  42
  43 static void
  44 assign_fs_binding_table_offsets(const struct gen_device_info *devinfo,
  45                                 const struct gl_program *prog,
  46                                 const struct brw_wm_prog_key *key,
  47                                 struct brw_wm_prog_data *prog_data)
  48 {
  49    uint32_t next_binding_table_offset = 0;
  50
  51    /* If there are no color regions, we still perform an FB write to a null
  52     * renderbuffer, which we place at surface index 0.
  53     */
  54    prog_data->binding_table.render_target_start = next_binding_table_offset;
  55    next_binding_table_offset += MAX2(key->nr_color_regions, 1);
  56
  57    next_binding_table_offset =
  58       brw_assign_common_binding_table_offsets(devinfo, prog, &prog_data->base,
  59                                               next_binding_table_offset);
  60
  61    if (prog->nir->info.outputs_read && !key->coherent_fb_fetch) {
  62       prog_data->binding_table.render_target_read_start =
  63          next_binding_table_offset;
  64       next_binding_table_offset += key->nr_color_regions;
  65    }
  66 }
  67
  68 static void
  69 brw_wm_debug_recompile(struct brw_context *brw, struct gl_program *prog,
  70                        const struct brw_wm_prog_key *key)
  71 {
  72    perf_debug("Recompiling fragment shader for program %d\n", prog->Id);
  73
  74    bool found = false;
  75    const struct brw_wm_prog_key *old_key =
  76       brw_find_previous_compile(&brw->cache, BRW_CACHE_FS_PROG,
  77                                 key->program_string_id);
  78
  79    if (!old_key) {
  80       perf_debug("  Didn't find previous compile in the shader cache for debug\n");
  81       return;
  82    }
  83
  84    found |= key_debug(brw, "alphatest, computed depth, depth test, or "
  85                       "depth write",
  86                       old_key->iz_lookup, key->iz_lookup);
  87    found |= key_debug(brw, "depth statistics",
  88                       old_key->stats_wm, key->stats_wm);
  89    found |= key_debug(brw, "flat shading",
  90                       old_key->flat_shade, key->flat_shade);
  91    found |= key_debug(brw, "per-sample interpolation",
  92                       old_key->persample_interp, key->persample_interp);
  93    found |= key_debug(brw, "number of color buffers",
  94                       old_key->nr_color_regions, key->nr_color_regions);
  95    found |= key_debug(brw, "MRT alpha test or alpha-to-coverage",
  96                       old_key->replicate_alpha, key->replicate_alpha);
  97    found |= key_debug(brw, "fragment color clamping",
  98                       old_key->clamp_fragment_color, key->clamp_fragment_color);
  99    found |= key_debug(brw, "multisampled FBO",
 100                       old_key->multisample_fbo, key->multisample_fbo);
 101    found |= key_debug(brw, "line smoothing",
 102                       old_key->line_aa, key->line_aa);
 103    found |= key_debug(brw, "input slots valid",
 104                       old_key->input_slots_valid, key->input_slots_valid);
 105    found |= key_debug(brw, "mrt alpha test function",
 106                       old_key->alpha_test_func, key->alpha_test_func);
 107    found |= key_debug(brw, "mrt alpha test reference value",
 108                       old_key->alpha_test_ref, key->alpha_test_ref);
 109    found |= key_debug(brw, "force dual color blending",
 110                       old_key->force_dual_color_blend,
 111                       key->force_dual_color_blend);
 112
 113    found |= brw_debug_recompile_sampler_key(brw, &old_key->tex, &key->tex);
 114
 115    if (!found) {
 116       perf_debug("  Something else\n");
 117    }
 118 }
 119
 120 /**
 121  * All Mesa program -> GPU code generation goes through this function.
 122  * Depending on the instructions used (i.e. flow control instructions)
 123  * we'll use one of two code generators.
 124  */
 125 static bool
 126 brw_codegen_wm_prog(struct brw_context *brw,
 127                     struct brw_program *fp,
 128                     struct brw_wm_prog_key *key,
 129                     struct brw_vue_map *vue_map)
 130 {
 131    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 132    struct gl_context *ctx = &brw->ctx;
 133    void *mem_ctx = ralloc_context(NULL);
 134    struct brw_wm_prog_data prog_data;
 135    const GLuint *program;
 136    GLuint program_size;
 137    bool start_busy = false;
 138    double start_time = 0;
 139
 140    memset(&prog_data, 0, sizeof(prog_data));
 141
 142    /* Use ALT floating point mode for ARB programs so that 0^0 == 1. */
 143    if (fp->program.is_arb_asm)
 144       prog_data.base.use_alt_mode = true;
 145
 146    assign_fs_binding_table_offsets(devinfo, &fp->program, key, &prog_data);
 147
 148    /* Allocate the references to the uniforms that will end up in the
 149     * prog_data associated with the compiled program, and which will be freed
 150     * by the state cache.
 151     */
 152    int param_count = fp->program.nir->num_uniforms / 4;
 153    prog_data.base.nr_image_params = fp->program.info.num_images;
 154    /* The backend also sometimes adds params for texture size. */
 155    param_count += 2 * ctx->Const.Program[MESA_SHADER_FRAGMENT].MaxTextureImageUnits;
 156    prog_data.base.param =
 157       rzalloc_array(NULL, const gl_constant_value *, param_count);
 158    prog_data.base.pull_param =
 159       rzalloc_array(NULL, const gl_constant_value *, param_count);
 160    prog_data.base.image_param =
 161       rzalloc_array(NULL, struct brw_image_param,
 162                     prog_data.base.nr_image_params);
 163    prog_data.base.nr_params = param_count;
 164
 165    if (!fp->program.is_arb_asm) {
 166       brw_nir_setup_glsl_uniforms(fp->program.nir, &fp->program,
 167                                   &prog_data.base, true);
 168       brw_nir_analyze_ubo_ranges(brw->screen->compiler, fp->program.nir,
 169                                  prog_data.base.ubo_ranges);
 170    } else {
 171       brw_nir_setup_arb_uniforms(fp->program.nir, &fp->program,
 172                                  &prog_data.base);
 173
 174       if (unlikely(INTEL_DEBUG & DEBUG_WM))
 175          brw_dump_arb_asm("fragment", &fp->program);
 176    }
 177
 178    if (unlikely(brw->perf_debug)) {
 179       start_busy = (brw->batch.last_bo &&
 180                     brw_bo_busy(brw->batch.last_bo));
 181       start_time = get_time();
 182    }
 183
 184    int st_index8 = -1, st_index16 = -1;
 185    if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
 186       st_index8 = brw_get_shader_time_index(brw, &fp->program, ST_FS8,
 187                                             !fp->program.is_arb_asm);
 188       st_index16 = brw_get_shader_time_index(brw, &fp->program, ST_FS16,
 189                                              !fp->program.is_arb_asm);
 190    }
 191
 192    char *error_str = NULL;
 193    program = brw_compile_fs(brw->screen->compiler, brw, mem_ctx,
 194                             key, &prog_data, fp->program.nir,
 195                             &fp->program, st_index8, st_index16,
 196                             true, false, vue_map,
 197                             &program_size, &error_str);
 198
 199    if (program == NULL) {
 200       if (!fp->program.is_arb_asm) {
 201          fp->program.sh.data->LinkStatus = linking_failure;
 202          ralloc_strcat(&fp->program.sh.data->InfoLog, error_str);
 203       }
 204
 205       _mesa_problem(NULL, "Failed to compile fragment shader: %s\n", error_str);
 206
 207       ralloc_free(mem_ctx);
 208       return false;
 209    }
 210
 211    if (unlikely(brw->perf_debug)) {
 212       if (fp->compiled_once)
 213          brw_wm_debug_recompile(brw, &fp->program, key);
 214       fp->compiled_once = true;
 215
 216       if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
 217          perf_debug("FS compile took %.03f ms and stalled the GPU\n",
 218                     (get_time() - start_time) * 1000);
 219       }
 220    }
 221
 222    brw_alloc_stage_scratch(brw, &brw->wm.base,
 223                            prog_data.base.total_scratch,
 224                            devinfo->max_wm_threads);
 225
 226    if (unlikely((INTEL_DEBUG & DEBUG_WM) && fp->program.is_arb_asm))
 227       fprintf(stderr, "\n");
 228
 229    brw_upload_cache(&brw->cache, BRW_CACHE_FS_PROG,
 230                     key, sizeof(struct brw_wm_prog_key),
 231                     program, program_size,
 232                     &prog_data, sizeof(prog_data),
 233                     &brw->wm.base.prog_offset, &brw->wm.base.prog_data);
 234
 235    ralloc_free(mem_ctx);
 236
 237    return true;
 238 }
 239
 240 bool
 241 brw_debug_recompile_sampler_key(struct brw_context *brw,
 242                                 const struct brw_sampler_prog_key_data *old_key,
 243                                 const struct brw_sampler_prog_key_data *key)
 244 {
 245    bool found = false;
 246
 247    for (unsigned int i = 0; i < MAX_SAMPLERS; i++) {
 248       found |= key_debug(brw, "EXT_texture_swizzle or DEPTH_TEXTURE_MODE",
 249                          old_key->swizzles[i], key->swizzles[i]);
 250    }
 251    found |= key_debug(brw, "GL_CLAMP enabled on any texture unit's 1st coordinate",
 252                       old_key->gl_clamp_mask[0], key->gl_clamp_mask[0]);
 253    found |= key_debug(brw, "GL_CLAMP enabled on any texture unit's 2nd coordinate",
 254                       old_key->gl_clamp_mask[1], key->gl_clamp_mask[1]);
 255    found |= key_debug(brw, "GL_CLAMP enabled on any texture unit's 3rd coordinate",
 256                       old_key->gl_clamp_mask[2], key->gl_clamp_mask[2]);
 257    found |= key_debug(brw, "gather channel quirk on any texture unit",
 258                       old_key->gather_channel_quirk_mask, key->gather_channel_quirk_mask);
 259    found |= key_debug(brw, "compressed multisample layout",
 260                       old_key->compressed_multisample_layout_mask,
 261                       key->compressed_multisample_layout_mask);
 262    found |= key_debug(brw, "16x msaa",
 263                       old_key->msaa_16,
 264                       key->msaa_16);
 265
 266    found |= key_debug(brw, "y_uv image bound",
 267                       old_key->y_uv_image_mask,
 268                       key->y_uv_image_mask);
 269    found |= key_debug(brw, "y_u_v image bound",
 270                       old_key->y_u_v_image_mask,
 271                       key->y_u_v_image_mask);
 272    found |= key_debug(brw, "yx_xuxv image bound",
 273                       old_key->yx_xuxv_image_mask,
 274                       key->yx_xuxv_image_mask);
 275    found |= key_debug(brw, "xy_uxvx image bound",
 276                       old_key->xy_uxvx_image_mask,
 277                       key->xy_uxvx_image_mask);
 278
 279
 280    for (unsigned int i = 0; i < MAX_SAMPLERS; i++) {
 281       found |= key_debug(brw, "textureGather workarounds",
 282                          old_key->gen6_gather_wa[i], key->gen6_gather_wa[i]);
 283    }
 284
 285    return found;
 286 }
 287
 288 static uint8_t
 289 gen6_gather_workaround(GLenum internalformat)
 290 {
 291    switch (internalformat) {
 292    case GL_R8I: return WA_SIGN | WA_8BIT;
 293    case GL_R8UI: return WA_8BIT;
 294    case GL_R16I: return WA_SIGN | WA_16BIT;
 295    case GL_R16UI: return WA_16BIT;
 296    default:
 297       /* Note that even though GL_R32I and GL_R32UI have format overrides in
 298        * the surface state, there is no shader w/a required.
 299        */
 300       return 0;
 301    }
 302 }
 303
 304 void
 305 brw_populate_sampler_prog_key_data(struct gl_context *ctx,
 306                                    const struct gl_program *prog,
 307                                    struct brw_sampler_prog_key_data *key)
 308 {
 309    struct brw_context *brw = brw_context(ctx);
 310    GLbitfield mask = prog->SamplersUsed;
 311
 312    while (mask) {
 313       const int s = u_bit_scan(&mask);
 314
 315       key->swizzles[s] = SWIZZLE_NOOP;
 316
 317       int unit_id = prog->SamplerUnits[s];
 318       const struct gl_texture_unit *unit = &ctx->Texture.Unit[unit_id];
 319
 320       if (unit->_Current && unit->_Current->Target != GL_TEXTURE_BUFFER) {
 321          const struct gl_texture_object *t = unit->_Current;
 322          const struct gl_texture_image *img = t->Image[0][t->BaseLevel];
 323          struct gl_sampler_object *sampler = _mesa_get_samplerobj(ctx, unit_id);
 324
 325          const bool alpha_depth = t->DepthMode == GL_ALPHA &&
 326             (img->_BaseFormat == GL_DEPTH_COMPONENT ||
 327              img->_BaseFormat == GL_DEPTH_STENCIL);
 328
 329          /* Haswell handles texture swizzling as surface format overrides
 330           * (except for GL_ALPHA); all other platforms need MOVs in the shader.
 331           */
 332          if (alpha_depth || (brw->gen < 8 && !brw->is_haswell))
 333             key->swizzles[s] = brw_get_texture_swizzle(ctx, t);
 334
 335          if (brw->gen < 8 &&
 336              sampler->MinFilter != GL_NEAREST &&
 337              sampler->MagFilter != GL_NEAREST) {
 338             if (sampler->WrapS == GL_CLAMP)
 339                key->gl_clamp_mask[0] |= 1 << s;
 340             if (sampler->WrapT == GL_CLAMP)
 341                key->gl_clamp_mask[1] |= 1 << s;
 342             if (sampler->WrapR == GL_CLAMP)
 343                key->gl_clamp_mask[2] |= 1 << s;
 344          }
 345
 346          /* gather4 for RG32* is broken in multiple ways on Gen7. */
 347          if (brw->gen == 7 && prog->nir->info.uses_texture_gather) {
 348             switch (img->InternalFormat) {
 349             case GL_RG32I:
 350             case GL_RG32UI: {
 351                /* We have to override the format to R32G32_FLOAT_LD.
 352                 * This means that SCS_ALPHA and SCS_ONE will return 0x3f8
 353                 * (1.0) rather than integer 1.  This needs shader hacks.
 354                 *
 355                 * On Ivybridge, we whack W (alpha) to ONE in our key's
 356                 * swizzle.  On Haswell, we look at the original texture
 357                 * swizzle, and use XYZW with channels overridden to ONE,
 358                 * leaving normal texture swizzling to SCS.
 359                 */
 360                unsigned src_swizzle =
 361                   brw->is_haswell ? t->_Swizzle : key->swizzles[s];
 362                for (int i = 0; i < 4; i++) {
 363                   unsigned src_comp = GET_SWZ(src_swizzle, i);
 364                   if (src_comp == SWIZZLE_ONE || src_comp == SWIZZLE_W) {
 365                      key->swizzles[i] &= ~(0x7 << (3 * i));
 366                      key->swizzles[i] |= SWIZZLE_ONE << (3 * i);
 367                   }
 368                }
 369                /* fallthrough */
 370             }
 371             case GL_RG32F:
 372                /* The channel select for green doesn't work - we have to
 373                 * request blue.  Haswell can use SCS for this, but Ivybridge
 374                 * needs a shader workaround.
 375                 */
 376                if (!brw->is_haswell)
 377                   key->gather_channel_quirk_mask |= 1 << s;
 378                break;
 379             }
 380          }
 381
 382          /* Gen6's gather4 is broken for UINT/SINT; we treat them as
 383           * UNORM/FLOAT instead and fix it in the shader.
 384           */
 385          if (brw->gen == 6 && prog->nir->info.uses_texture_gather) {
 386             key->gen6_gather_wa[s] = gen6_gather_workaround(img->InternalFormat);
 387          }
 388
 389          /* If this is a multisample sampler, and uses the CMS MSAA layout,
 390           * then we need to emit slightly different code to first sample the
 391           * MCS surface.
 392           */
 393          struct intel_texture_object *intel_tex =
 394             intel_texture_object((struct gl_texture_object *)t);
 395
 396          /* From gen9 onwards some single sampled buffers can also be
 397           * compressed. These don't need ld2dms sampling along with mcs fetch.
 398           */
 399          if (brw->gen >= 7 &&
 400              intel_tex->mt->msaa_layout == INTEL_MSAA_LAYOUT_CMS &&
 401              intel_tex->mt->num_samples > 1) {
 402             key->compressed_multisample_layout_mask |= 1 << s;
 403
 404             if (intel_tex->mt->num_samples >= 16) {
 405                assert(brw->gen >= 9);
 406                key->msaa_16 |= 1 << s;
 407             }
 408          }
 409
 410          if (t->Target == GL_TEXTURE_EXTERNAL_OES && intel_tex->planar_format) {
 411             switch (intel_tex->planar_format->components) {
 412             case __DRI_IMAGE_COMPONENTS_Y_UV:
 413                key->y_uv_image_mask |= 1 << s;
 414                break;
 415             case __DRI_IMAGE_COMPONENTS_Y_U_V:
 416                key->y_u_v_image_mask |= 1 << s;
 417                break;
 418             case __DRI_IMAGE_COMPONENTS_Y_XUXV:
 419                key->yx_xuxv_image_mask |= 1 << s;
 420                break;
 421             case __DRI_IMAGE_COMPONENTS_Y_UXVX:
 422                key->xy_uxvx_image_mask |= 1 << s;
 423                break;
 424             default:
 425                break;
 426             }
 427          }
 428
 429       }
 430    }
 431 }
 432
 433 static bool
 434 brw_wm_state_dirty(const struct brw_context *brw)
 435 {
 436    return brw_state_dirty(brw,
 437                           _NEW_BUFFERS |
 438                           _NEW_COLOR |
 439                           _NEW_DEPTH |
 440                           _NEW_FRAG_CLAMP |
 441                           _NEW_HINT |
 442                           _NEW_LIGHT |
 443                           _NEW_LINE |
 444                           _NEW_MULTISAMPLE |
 445                           _NEW_POLYGON |
 446                           _NEW_STENCIL |
 447                           _NEW_TEXTURE,
 448                           BRW_NEW_FRAGMENT_PROGRAM |
 449                           BRW_NEW_REDUCED_PRIMITIVE |
 450                           BRW_NEW_STATS_WM |
 451                           BRW_NEW_VUE_MAP_GEOM_OUT);
 452 }
 453
 454 void
 455 brw_wm_populate_key(struct brw_context *brw, struct brw_wm_prog_key *key)
 456 {
 457    struct gl_context *ctx = &brw->ctx;
 458    /* BRW_NEW_FRAGMENT_PROGRAM */
 459    const struct brw_program *fp = brw_program_const(brw->fragment_program);
 460    const struct gl_program *prog = (struct gl_program *) brw->fragment_program;
 461    GLuint lookup = 0;
 462    GLuint line_aa;
 463
 464    memset(key, 0, sizeof(*key));
 465
 466    /* Build the index for table lookup
 467     */
 468    if (brw->gen < 6) {
 469       /* _NEW_COLOR */
 470       if (prog->info.fs.uses_discard || ctx->Color.AlphaEnabled) {
 471          lookup |= BRW_WM_IZ_PS_KILL_ALPHATEST_BIT;
 472       }
 473
 474       if (prog->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
 475          lookup |= BRW_WM_IZ_PS_COMPUTES_DEPTH_BIT;
 476       }
 477
 478       /* _NEW_DEPTH */
 479       if (ctx->Depth.Test)
 480          lookup |= BRW_WM_IZ_DEPTH_TEST_ENABLE_BIT;
 481
 482       if (brw_depth_writes_enabled(brw))
 483          lookup |= BRW_WM_IZ_DEPTH_WRITE_ENABLE_BIT;
 484
 485       /* _NEW_STENCIL | _NEW_BUFFERS */
 486       if (brw->stencil_enabled) {
 487          lookup |= BRW_WM_IZ_STENCIL_TEST_ENABLE_BIT;
 488
 489          if (ctx->Stencil.WriteMask[0] ||
 490              ctx->Stencil.WriteMask[ctx->Stencil._BackFace])
 491             lookup |= BRW_WM_IZ_STENCIL_WRITE_ENABLE_BIT;
 492       }
 493       key->iz_lookup = lookup;
 494    }
 495
 496    line_aa = BRW_WM_AA_NEVER;
 497
 498    /* _NEW_LINE, _NEW_POLYGON, BRW_NEW_REDUCED_PRIMITIVE */
 499    if (ctx->Line.SmoothFlag) {
 500       if (brw->reduced_primitive == GL_LINES) {
 501          line_aa = BRW_WM_AA_ALWAYS;
 502       }
 503       else if (brw->reduced_primitive == GL_TRIANGLES) {
 504          if (ctx->Polygon.FrontMode == GL_LINE) {
 505             line_aa = BRW_WM_AA_SOMETIMES;
 506
 507             if (ctx->Polygon.BackMode == GL_LINE ||
 508                 (ctx->Polygon.CullFlag &&
 509                  ctx->Polygon.CullFaceMode == GL_BACK))
 510                line_aa = BRW_WM_AA_ALWAYS;
 511          }
 512          else if (ctx->Polygon.BackMode == GL_LINE) {
 513             line_aa = BRW_WM_AA_SOMETIMES;
 514
 515             if ((ctx->Polygon.CullFlag &&
 516                  ctx->Polygon.CullFaceMode == GL_FRONT))
 517                line_aa = BRW_WM_AA_ALWAYS;
 518          }
 519       }
 520    }
 521
 522    key->line_aa = line_aa;
 523
 524    /* _NEW_HINT */
 525    key->high_quality_derivatives =
 526       ctx->Hint.FragmentShaderDerivative == GL_NICEST;
 527
 528    if (brw->gen < 6)
 529       key->stats_wm = brw->stats_wm;
 530
 531    /* _NEW_LIGHT */
 532    key->flat_shade = (ctx->Light.ShadeModel == GL_FLAT);
 533
 534    /* _NEW_FRAG_CLAMP | _NEW_BUFFERS */
 535    key->clamp_fragment_color = ctx->Color._ClampFragmentColor;
 536
 537    /* _NEW_TEXTURE */
 538    brw_populate_sampler_prog_key_data(ctx, prog, &key->tex);
 539
 540    /* _NEW_BUFFERS */
 541    key->nr_color_regions = ctx->DrawBuffer->_NumColorDrawBuffers;
 542
 543    /* _NEW_COLOR */
 544    key->force_dual_color_blend = brw->dual_color_blend_by_location &&
 545       (ctx->Color.BlendEnabled & 1) && ctx->Color.Blend[0]._UsesDualSrc;
 546
 547    /* _NEW_MULTISAMPLE, _NEW_COLOR, _NEW_BUFFERS */
 548    key->replicate_alpha = ctx->DrawBuffer->_NumColorDrawBuffers > 1 &&
 549       (_mesa_is_alpha_test_enabled(ctx) ||
 550        _mesa_is_alpha_to_coverage_enabled(ctx));
 551
 552    /* _NEW_BUFFERS _NEW_MULTISAMPLE */
 553    /* Ignore sample qualifier while computing this flag. */
 554    if (ctx->Multisample.Enabled) {
 555       key->persample_interp =
 556          ctx->Multisample.SampleShading &&
 557          (ctx->Multisample.MinSampleShadingValue *
 558           _mesa_geometric_samples(ctx->DrawBuffer) > 1);
 559
 560       key->multisample_fbo = _mesa_geometric_samples(ctx->DrawBuffer) > 1;
 561    }
 562
 563    /* BRW_NEW_VUE_MAP_GEOM_OUT */
 564    if (brw->gen < 6 || _mesa_bitcount_64(prog->info.inputs_read &
 565                                          BRW_FS_VARYING_INPUT_MASK) > 16) {
 566       key->input_slots_valid = brw->vue_map_geom_out.slots_valid;
 567    }
 568
 569    /* _NEW_COLOR | _NEW_BUFFERS */
 570    /* Pre-gen6, the hardware alpha test always used each render
 571     * target's alpha to do alpha test, as opposed to render target 0's alpha
 572     * like GL requires.  Fix that by building the alpha test into the
 573     * shader, and we'll skip enabling the fixed function alpha test.
 574     */
 575    if (brw->gen < 6 && ctx->DrawBuffer->_NumColorDrawBuffers > 1 &&
 576        ctx->Color.AlphaEnabled) {
 577       key->alpha_test_func = ctx->Color.AlphaFunc;
 578       key->alpha_test_ref = ctx->Color.AlphaRef;
 579    }
 580
 581    /* The unique fragment program ID */
 582    key->program_string_id = fp->id;
 583
 584    /* Whether reads from the framebuffer should behave coherently. */
 585    key->coherent_fb_fetch = ctx->Extensions.MESA_shader_framebuffer_fetch;
 586 }
 587
 588 void
 589 brw_upload_wm_prog(struct brw_context *brw)
 590 {
 591    struct brw_wm_prog_key key;
 592    struct brw_program *fp = (struct brw_program *) brw->fragment_program;
 593
 594    if (!brw_wm_state_dirty(brw))
 595       return;
 596
 597    brw_wm_populate_key(brw, &key);
 598
 599    if (!brw_search_cache(&brw->cache, BRW_CACHE_FS_PROG,
 600                          &key, sizeof(key),
 601                          &brw->wm.base.prog_offset,
 602                          &brw->wm.base.prog_data)) {
 603       bool success = brw_codegen_wm_prog(brw, fp, &key,
 604                                          &brw->vue_map_geom_out);
 605       (void) success;
 606       assert(success);
 607    }
 608 }
 609
 610 bool
 611 brw_fs_precompile(struct gl_context *ctx, struct gl_program *prog)
 612 {
 613    struct brw_context *brw = brw_context(ctx);
 614    struct brw_wm_prog_key key;
 615
 616    struct brw_program *bfp = brw_program(prog);
 617
 618    memset(&key, 0, sizeof(key));
 619
 620    uint64_t outputs_written = prog->info.outputs_written;
 621
 622    if (brw->gen < 6) {
 623       if (prog->info.fs.uses_discard)
 624          key.iz_lookup |= BRW_WM_IZ_PS_KILL_ALPHATEST_BIT;
 625
 626       if (outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
 627          key.iz_lookup |= BRW_WM_IZ_PS_COMPUTES_DEPTH_BIT;
 628
 629       /* Just assume depth testing. */
 630       key.iz_lookup |= BRW_WM_IZ_DEPTH_TEST_ENABLE_BIT;
 631       key.iz_lookup |= BRW_WM_IZ_DEPTH_WRITE_ENABLE_BIT;
 632    }
 633
 634    if (brw->gen < 6 || _mesa_bitcount_64(prog->info.inputs_read &
 635                                          BRW_FS_VARYING_INPUT_MASK) > 16) {
 636       key.input_slots_valid = prog->info.inputs_read | VARYING_BIT_POS;
 637    }
 638
 639    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 640
 641    key.nr_color_regions = _mesa_bitcount_64(outputs_written &
 642          ~(BITFIELD64_BIT(FRAG_RESULT_DEPTH) |
 643            BITFIELD64_BIT(FRAG_RESULT_STENCIL) |
 644            BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK)));
 645
 646    key.program_string_id = bfp->id;
 647
 648    /* Whether reads from the framebuffer should behave coherently. */
 649    key.coherent_fb_fetch = ctx->Extensions.MESA_shader_framebuffer_fetch;
 650
 651    uint32_t old_prog_offset = brw->wm.base.prog_offset;
 652    struct brw_stage_prog_data *old_prog_data = brw->wm.base.prog_data;
 653
 654    struct brw_vue_map vue_map;
 655    if (brw->gen < 6) {
 656       brw_compute_vue_map(&brw->screen->devinfo, &vue_map,
 657                           prog->info.inputs_read | VARYING_BIT_POS,
 658                           false);
 659    }
 660
 661    bool success = brw_codegen_wm_prog(brw, bfp, &key, &vue_map);
 662
 663    brw->wm.base.prog_offset = old_prog_offset;
 664    brw->wm.base.prog_data = old_prog_data;
 665
 666    return success;
 667 }