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/formats.h"
  36 #include "main/fbobject.h"
  37 #include "main/samplerobj.h"
  38 #include "program/prog_parameter.h"
  39
  40 #include "glsl/ralloc.h"
  41
  42 /** Return number of src args for given instruction */
  43 GLuint brw_wm_nr_args( GLuint opcode )
  44 {
  45    switch (opcode) {
  46    case WM_FRONTFACING:
  47    case WM_PIXELXY:
  48       return 0;
  49    case WM_CINTERP:
  50    case WM_WPOSXY:
  51    case WM_DELTAXY:
  52       return 1;
  53    case WM_LINTERP:
  54    case WM_PIXELW:
  55       return 2;
  56    case WM_FB_WRITE:
  57    case WM_PINTERP:
  58       return 3;
  59    default:
  60       assert(opcode < MAX_OPCODE);
  61       return _mesa_num_inst_src_regs(opcode);
  62    }
  63 }
  64
  65
  66 GLuint brw_wm_is_scalar_result( GLuint opcode )
  67 {
  68    switch (opcode) {
  69    case OPCODE_COS:
  70    case OPCODE_EX2:
  71    case OPCODE_LG2:
  72    case OPCODE_POW:
  73    case OPCODE_RCP:
  74    case OPCODE_RSQ:
  75    case OPCODE_SIN:
  76    case OPCODE_DP2:
  77    case OPCODE_DP3:
  78    case OPCODE_DP4:
  79    case OPCODE_DPH:
  80    case OPCODE_DST:
  81       return 1;
  82
  83    default:
  84       return 0;
  85    }
  86 }
  87
  88
  89 /**
  90  * Do GPU code generation for non-GLSL shader.  non-GLSL shaders have
  91  * no flow control instructions so we can more readily do SSA-style
  92  * optimizations.
  93  */
  94 static void
  95 brw_wm_non_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c)
  96 {
  97    /* Augment fragment program.  Add instructions for pre- and
  98     * post-fragment-program tasks such as interpolation and fogging.
  99     */
 100    brw_wm_pass_fp(c);
 101
 102    /* Translate to intermediate representation.  Build register usage
 103     * chains.
 104     */
 105    brw_wm_pass0(c);
 106
 107    /* Dead code removal.
 108     */
 109    brw_wm_pass1(c);
 110
 111    /* Register allocation.
 112     * Divide by two because we operate on 16 pixels at a time and require
 113     * two GRF entries for each logical shader register.
 114     */
 115    c->grf_limit = BRW_WM_MAX_GRF / 2;
 116
 117    brw_wm_pass2(c);
 118
 119    /* how many general-purpose registers are used */
 120    c->prog_data.reg_blocks = brw_register_blocks(c->max_wm_grf);
 121
 122    /* Emit GEN4 code.
 123     */
 124    brw_wm_emit(c);
 125 }
 126
 127
 128 /**
 129  * Return a bitfield where bit n is set if barycentric interpolation mode n
 130  * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
 131  */
 132 static unsigned
 133 brw_compute_barycentric_interp_modes(bool shade_model_flat,
 134                                      const struct gl_fragment_program *fprog)
 135 {
 136    unsigned barycentric_interp_modes = 0;
 137    int attr;
 138
 139    /* Loop through all fragment shader inputs to figure out what interpolation
 140     * modes are in use, and set the appropriate bits in
 141     * barycentric_interp_modes.
 142     */
 143    for (attr = 0; attr < FRAG_ATTRIB_MAX; ++attr) {
 144       enum glsl_interp_qualifier interp_qualifier =
 145          fprog->InterpQualifier[attr];
 146       bool is_gl_Color = attr == FRAG_ATTRIB_COL0 || attr == FRAG_ATTRIB_COL1;
 147
 148       /* Ignore unused inputs. */
 149       if (!(fprog->Base.InputsRead & BITFIELD64_BIT(attr)))
 150          continue;
 151
 152       /* Ignore WPOS and FACE, because they don't require interpolation. */
 153       if (attr == FRAG_ATTRIB_WPOS || attr == FRAG_ATTRIB_FACE)
 154          continue;
 155
 156       if (interp_qualifier == INTERP_QUALIFIER_NOPERSPECTIVE) {
 157          barycentric_interp_modes |=
 158             1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
 159       } else if (interp_qualifier == INTERP_QUALIFIER_SMOOTH ||
 160                  (!(shade_model_flat && is_gl_Color) &&
 161                   interp_qualifier == INTERP_QUALIFIER_NONE)) {
 162          barycentric_interp_modes |=
 163             1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
 164       }
 165    }
 166
 167    return barycentric_interp_modes;
 168 }
 169
 170
 171 void
 172 brw_wm_payload_setup(struct brw_context *brw,
 173                      struct brw_wm_compile *c)
 174 {
 175    struct intel_context *intel = &brw->intel;
 176    bool uses_depth = (c->fp->program.Base.InputsRead &
 177                       (1 << FRAG_ATTRIB_WPOS)) != 0;
 178    unsigned barycentric_interp_modes = c->prog_data.barycentric_interp_modes;
 179    int i;
 180
 181    if (intel->gen >= 6) {
 182       /* R0-1: masks, pixel X/Y coordinates. */
 183       c->nr_payload_regs = 2;
 184       /* R2: only for 32-pixel dispatch.*/
 185
 186       /* R3-26: barycentric interpolation coordinates.  These appear in the
 187        * same order that they appear in the brw_wm_barycentric_interp_mode
 188        * enum.  Each set of coordinates occupies 2 registers if dispatch width
 189        * == 8 and 4 registers if dispatch width == 16.  Coordinates only
 190        * appear if they were enabled using the "Barycentric Interpolation
 191        * Mode" bits in WM_STATE.
 192        */
 193       for (i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
 194          if (barycentric_interp_modes & (1 << i)) {
 195             c->barycentric_coord_reg[i] = c->nr_payload_regs;
 196             c->nr_payload_regs += 2;
 197             if (c->dispatch_width == 16) {
 198                c->nr_payload_regs += 2;
 199             }
 200          }
 201       }
 202
 203       /* R27: interpolated depth if uses source depth */
 204       if (uses_depth) {
 205          c->source_depth_reg = c->nr_payload_regs;
 206          c->nr_payload_regs++;
 207          if (c->dispatch_width == 16) {
 208             /* R28: interpolated depth if not 8-wide. */
 209             c->nr_payload_regs++;
 210          }
 211       }
 212       /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W.
 213        */
 214       if (uses_depth) {
 215          c->source_w_reg = c->nr_payload_regs;
 216          c->nr_payload_regs++;
 217          if (c->dispatch_width == 16) {
 218             /* R30: interpolated W if not 8-wide. */
 219             c->nr_payload_regs++;
 220          }
 221       }
 222       /* R31: MSAA position offsets. */
 223       /* R32-: bary for 32-pixel. */
 224       /* R58-59: interp W for 32-pixel. */
 225
 226       if (c->fp->program.Base.OutputsWritten &
 227           BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
 228          c->source_depth_to_render_target = true;
 229          c->computes_depth = true;
 230       }
 231    } else {
 232       brw_wm_lookup_iz(intel, c);
 233    }
 234 }
 235
 236 /**
 237  * All Mesa program -> GPU code generation goes through this function.
 238  * Depending on the instructions used (i.e. flow control instructions)
 239  * we'll use one of two code generators.
 240  */
 241 bool do_wm_prog(struct brw_context *brw,
 242                 struct gl_shader_program *prog,
 243                 struct brw_fragment_program *fp,
 244                 struct brw_wm_prog_key *key)
 245 {
 246    struct intel_context *intel = &brw->intel;
 247    struct brw_wm_compile *c;
 248    const GLuint *program;
 249    GLuint program_size;
 250
 251    c = brw->wm.compile_data;
 252    if (c == NULL) {
 253       brw->wm.compile_data = rzalloc(NULL, struct brw_wm_compile);
 254       c = brw->wm.compile_data;
 255       if (c == NULL) {
 256          /* Ouch - big out of memory problem.  Can't continue
 257           * without triggering a segfault, no way to signal,
 258           * so just return.
 259           */
 260          return false;
 261       }
 262    } else {
 263       void *instruction = c->instruction;
 264       void *prog_instructions = c->prog_instructions;
 265       void *vreg = c->vreg;
 266       void *refs = c->refs;
 267       memset(c, 0, sizeof(*brw->wm.compile_data));
 268       c->instruction = instruction;
 269       c->prog_instructions = prog_instructions;
 270       c->vreg = vreg;
 271       c->refs = refs;
 272    }
 273    memcpy(&c->key, key, sizeof(*key));
 274
 275    c->fp = fp;
 276    c->env_param = brw->intel.ctx.FragmentProgram.Parameters;
 277
 278    brw_init_compile(brw, &c->func, c);
 279
 280    c->prog_data.barycentric_interp_modes =
 281       brw_compute_barycentric_interp_modes(c->key.flat_shade, &fp->program);
 282
 283    if (prog && prog->_LinkedShaders[MESA_SHADER_FRAGMENT]) {
 284       if (!brw_wm_fs_emit(brw, c, prog))
 285          return false;
 286    } else {
 287       if (!c->instruction) {
 288          c->instruction = rzalloc_array(c, struct brw_wm_instruction, BRW_WM_MAX_INSN);
 289          c->prog_instructions = rzalloc_array(c, struct prog_instruction, BRW_WM_MAX_INSN);
 290          c->vreg = rzalloc_array(c, struct brw_wm_value, BRW_WM_MAX_VREG);
 291          c->refs = rzalloc_array(c, struct brw_wm_ref, BRW_WM_MAX_REF);
 292       }
 293
 294       /* Fallback for fixed function and ARB_fp shaders. */
 295       c->dispatch_width = 16;
 296       brw_wm_payload_setup(brw, c);
 297       brw_wm_non_glsl_emit(brw, c);
 298       c->prog_data.dispatch_width = 16;
 299    }
 300
 301    /* Scratch space is used for register spilling */
 302    if (c->last_scratch) {
 303       c->prog_data.total_scratch = brw_get_scratch_size(c->last_scratch);
 304
 305       brw_get_scratch_bo(intel, &brw->wm.scratch_bo,
 306                          c->prog_data.total_scratch * brw->max_wm_threads);
 307    }
 308
 309    if (unlikely(INTEL_DEBUG & DEBUG_WM))
 310       fprintf(stderr, "\n");
 311
 312    /* get the program
 313     */
 314    program = brw_get_program(&c->func, &program_size);
 315
 316    brw_upload_cache(&brw->cache, BRW_WM_PROG,
 317                     &c->key, sizeof(c->key),
 318                     program, program_size,
 319                     &c->prog_data, sizeof(c->prog_data),
 320                     &brw->wm.prog_offset, &brw->wm.prog_data);
 321
 322    return true;
 323 }
 324
 325 void
 326 brw_populate_sampler_prog_key_data(struct gl_context *ctx,
 327                                    struct brw_sampler_prog_key_data *key,
 328                                    int i)
 329 {
 330    const struct gl_texture_unit *unit = &ctx->Texture.Unit[i];
 331
 332    if (unit->_ReallyEnabled && unit->_Current->Target != GL_TEXTURE_BUFFER) {
 333       const struct gl_texture_object *t = unit->_Current;
 334       const struct gl_texture_image *img = t->Image[0][t->BaseLevel];
 335       struct gl_sampler_object *sampler = _mesa_get_samplerobj(ctx, i);
 336       int swizzles[SWIZZLE_NIL + 1] = {
 337          SWIZZLE_X,
 338          SWIZZLE_Y,
 339          SWIZZLE_Z,
 340          SWIZZLE_W,
 341          SWIZZLE_ZERO,
 342          SWIZZLE_ONE,
 343          SWIZZLE_NIL
 344       };
 345
 346       if (img->_BaseFormat == GL_DEPTH_COMPONENT ||
 347           img->_BaseFormat == GL_DEPTH_STENCIL) {
 348          if (sampler->CompareMode == GL_COMPARE_R_TO_TEXTURE_ARB)
 349             key->compare_funcs[i] = sampler->CompareFunc;
 350
 351          /* We handle GL_DEPTH_TEXTURE_MODE here instead of as surface format
 352           * overrides because shadow comparison always returns the result of
 353           * the comparison in all channels anyway.
 354           */
 355          switch (sampler->DepthMode) {
 356          case GL_ALPHA:
 357             swizzles[0] = SWIZZLE_ZERO;
 358             swizzles[1] = SWIZZLE_ZERO;
 359             swizzles[2] = SWIZZLE_ZERO;
 360             swizzles[3] = SWIZZLE_X;
 361             break;
 362          case GL_LUMINANCE:
 363             swizzles[0] = SWIZZLE_X;
 364             swizzles[1] = SWIZZLE_X;
 365             swizzles[2] = SWIZZLE_X;
 366             swizzles[3] = SWIZZLE_ONE;
 367             break;
 368          case GL_INTENSITY:
 369             swizzles[0] = SWIZZLE_X;
 370             swizzles[1] = SWIZZLE_X;
 371             swizzles[2] = SWIZZLE_X;
 372             swizzles[3] = SWIZZLE_X;
 373             break;
 374          case GL_RED:
 375             swizzles[0] = SWIZZLE_X;
 376             swizzles[1] = SWIZZLE_ZERO;
 377             swizzles[2] = SWIZZLE_ZERO;
 378             swizzles[3] = SWIZZLE_ONE;
 379             break;
 380          }
 381       }
 382
 383       if (img->InternalFormat == GL_YCBCR_MESA) {
 384          key->yuvtex_mask |= 1 << i;
 385          if (img->TexFormat == MESA_FORMAT_YCBCR)
 386              key->yuvtex_swap_mask |= 1 << i;
 387       }
 388
 389       key->swizzles[i] =
 390          MAKE_SWIZZLE4(swizzles[GET_SWZ(t->_Swizzle, 0)],
 391                        swizzles[GET_SWZ(t->_Swizzle, 1)],
 392                        swizzles[GET_SWZ(t->_Swizzle, 2)],
 393                        swizzles[GET_SWZ(t->_Swizzle, 3)]);
 394
 395       if (sampler->MinFilter != GL_NEAREST &&
 396           sampler->MagFilter != GL_NEAREST) {
 397          if (sampler->WrapS == GL_CLAMP)
 398             key->gl_clamp_mask[0] |= 1 << i;
 399          if (sampler->WrapT == GL_CLAMP)
 400             key->gl_clamp_mask[1] |= 1 << i;
 401          if (sampler->WrapR == GL_CLAMP)
 402             key->gl_clamp_mask[2] |= 1 << i;
 403       }
 404    }
 405    else {
 406       key->swizzles[i] = SWIZZLE_NOOP;
 407    }
 408 }
 409
 410 static void brw_wm_populate_key( struct brw_context *brw,
 411                                  struct brw_wm_prog_key *key )
 412 {
 413    struct gl_context *ctx = &brw->intel.ctx;
 414    struct intel_context *intel = &brw->intel;
 415    /* BRW_NEW_FRAGMENT_PROGRAM */
 416    const struct brw_fragment_program *fp =
 417       (struct brw_fragment_program *)brw->fragment_program;
 418    const struct gl_program *prog = (struct gl_program *) brw->fragment_program;
 419    GLuint lookup = 0;
 420    GLuint line_aa;
 421    GLuint i;
 422
 423    /* As a temporary measure we assume that all programs use dFdy() (and hence
 424     * need to be compiled differently depending on whether we're rendering to
 425     * an FBO).  FIXME: set this bool correctly based on the contents of the
 426     * program.
 427     */
 428    bool program_uses_dfdy = true;
 429
 430    memset(key, 0, sizeof(*key));
 431
 432    /* Build the index for table lookup
 433     */
 434    /* _NEW_COLOR */
 435    key->alpha_test = ctx->Color.AlphaEnabled;
 436
 437    if (intel->gen < 6) {
 438       if (fp->program.UsesKill || ctx->Color.AlphaEnabled)
 439          lookup |= IZ_PS_KILL_ALPHATEST_BIT;
 440
 441       if (fp->program.Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
 442          lookup |= IZ_PS_COMPUTES_DEPTH_BIT;
 443
 444       /* _NEW_DEPTH */
 445       if (ctx->Depth.Test)
 446          lookup |= IZ_DEPTH_TEST_ENABLE_BIT;
 447
 448       if (ctx->Depth.Test && ctx->Depth.Mask) /* ?? */
 449          lookup |= IZ_DEPTH_WRITE_ENABLE_BIT;
 450
 451       /* _NEW_STENCIL */
 452       if (ctx->Stencil._Enabled) {
 453          lookup |= IZ_STENCIL_TEST_ENABLE_BIT;
 454
 455          if (ctx->Stencil.WriteMask[0] ||
 456              ctx->Stencil.WriteMask[ctx->Stencil._BackFace])
 457             lookup |= IZ_STENCIL_WRITE_ENABLE_BIT;
 458       }
 459       key->iz_lookup = lookup;
 460    }
 461
 462    line_aa = AA_NEVER;
 463
 464    /* _NEW_LINE, _NEW_POLYGON, BRW_NEW_REDUCED_PRIMITIVE */
 465    if (ctx->Line.SmoothFlag) {
 466       if (brw->intel.reduced_primitive == GL_LINES) {
 467          line_aa = AA_ALWAYS;
 468       }
 469       else if (brw->intel.reduced_primitive == GL_TRIANGLES) {
 470          if (ctx->Polygon.FrontMode == GL_LINE) {
 471             line_aa = AA_SOMETIMES;
 472
 473             if (ctx->Polygon.BackMode == GL_LINE ||
 474                 (ctx->Polygon.CullFlag &&
 475                  ctx->Polygon.CullFaceMode == GL_BACK))
 476                line_aa = AA_ALWAYS;
 477          }
 478          else if (ctx->Polygon.BackMode == GL_LINE) {
 479             line_aa = AA_SOMETIMES;
 480
 481             if ((ctx->Polygon.CullFlag &&
 482                  ctx->Polygon.CullFaceMode == GL_FRONT))
 483                line_aa = AA_ALWAYS;
 484          }
 485       }
 486    }
 487
 488    key->line_aa = line_aa;
 489    key->stats_wm = brw->intel.stats_wm;
 490
 491    /* BRW_NEW_WM_INPUT_DIMENSIONS */
 492    key->proj_attrib_mask = brw->wm.input_size_masks[4-1];
 493
 494    /* _NEW_LIGHT */
 495    key->flat_shade = (ctx->Light.ShadeModel == GL_FLAT);
 496
 497    /* _NEW_FRAG_CLAMP | _NEW_BUFFERS */
 498    key->clamp_fragment_color = ctx->Color._ClampFragmentColor;
 499
 500    /* _NEW_TEXTURE */
 501    for (i = 0; i < BRW_MAX_TEX_UNIT; i++) {
 502       if (prog->TexturesUsed[i])
 503          brw_populate_sampler_prog_key_data(ctx, &key->tex, i);
 504    }
 505
 506    /* _NEW_BUFFERS */
 507    /*
 508     * Include the draw buffer origin and height so that we can calculate
 509     * fragment position values relative to the bottom left of the drawable,
 510     * from the incoming screen origin relative position we get as part of our
 511     * payload.
 512     *
 513     * This is only needed for the WM_WPOSXY opcode when the fragment program
 514     * uses the gl_FragCoord input.
 515     *
 516     * We could avoid recompiling by including this as a constant referenced by
 517     * our program, but if we were to do that it would also be nice to handle
 518     * getting that constant updated at batchbuffer submit time (when we
 519     * hold the lock and know where the buffer really is) rather than at emit
 520     * time when we don't hold the lock and are just guessing.  We could also
 521     * just avoid using this as key data if the program doesn't use
 522     * fragment.position.
 523     *
 524     * For DRI2 the origin_x/y will always be (0,0) but we still need the
 525     * drawable height in order to invert the Y axis.
 526     */
 527    if (fp->program.Base.InputsRead & FRAG_BIT_WPOS) {
 528       key->drawable_height = ctx->DrawBuffer->Height;
 529    }
 530
 531    if ((fp->program.Base.InputsRead & FRAG_BIT_WPOS) || program_uses_dfdy) {
 532       key->render_to_fbo = _mesa_is_user_fbo(ctx->DrawBuffer);
 533    }
 534
 535    /* _NEW_BUFFERS */
 536    key->nr_color_regions = ctx->DrawBuffer->_NumColorDrawBuffers;
 537
 538    /* CACHE_NEW_VS_PROG */
 539    key->vp_outputs_written = brw->vs.prog_data->outputs_written;
 540
 541    /* The unique fragment program ID */
 542    key->program_string_id = fp->id;
 543 }
 544
 545
 546 static void
 547 brw_upload_wm_prog(struct brw_context *brw)
 548 {
 549    struct intel_context *intel = &brw->intel;
 550    struct gl_context *ctx = &intel->ctx;
 551    struct brw_wm_prog_key key;
 552    struct brw_fragment_program *fp = (struct brw_fragment_program *)
 553       brw->fragment_program;
 554
 555    brw_wm_populate_key(brw, &key);
 556
 557    if (!brw_search_cache(&brw->cache, BRW_WM_PROG,
 558                          &key, sizeof(key),
 559                          &brw->wm.prog_offset, &brw->wm.prog_data)) {
 560       bool success = do_wm_prog(brw, ctx->Shader._CurrentFragmentProgram, fp,
 561                                 &key);
 562       (void) success;
 563       assert(success);
 564    }
 565 }
 566
 567
 568 const struct brw_tracked_state brw_wm_prog = {
 569    .dirty = {
 570       .mesa  = (_NEW_COLOR |
 571                 _NEW_DEPTH |
 572                 _NEW_STENCIL |
 573                 _NEW_POLYGON |
 574                 _NEW_LINE |
 575                 _NEW_LIGHT |
 576                 _NEW_FRAG_CLAMP |
 577                 _NEW_BUFFERS |
 578                 _NEW_TEXTURE),
 579       .brw   = (BRW_NEW_FRAGMENT_PROGRAM |
 580                 BRW_NEW_WM_INPUT_DIMENSIONS |
 581                 BRW_NEW_REDUCED_PRIMITIVE),
 582       .cache = CACHE_NEW_VS_PROG,
 583    },
 584    .emit = brw_upload_wm_prog
 585 };
 586