src/intel/compiler/brw_fs_builder.h

   1 /* -*- c++ -*- */
   2 /*
   3  * Copyright © 2010-2015 Intel Corporation
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  */
  24
  25 #ifndef BRW_FS_BUILDER_H
  26 #define BRW_FS_BUILDER_H
  27
  28 #include "brw_ir_fs.h"
  29 #include "brw_shader.h"
  30
  31 namespace brw {
  32    /**
  33     * Toolbox to assemble an FS IR program out of individual instructions.
  34     *
  35     * This object is meant to have an interface consistent with
  36     * brw::vec4_builder.  They cannot be fully interchangeable because
  37     * brw::fs_builder generates scalar code while brw::vec4_builder generates
  38     * vector code.
  39     */
  40    class fs_builder {
  41    public:
  42       /** Type used in this IR to represent a source of an instruction. */
  43       typedef fs_reg src_reg;
  44
  45       /** Type used in this IR to represent the destination of an instruction. */
  46       typedef fs_reg dst_reg;
  47
  48       /** Type used in this IR to represent an instruction. */
  49       typedef fs_inst instruction;
  50
  51       /**
  52        * Construct an fs_builder that inserts instructions into \p shader.
  53        * \p dispatch_width gives the native execution width of the program.
  54        */
  55       fs_builder(backend_shader *shader,
  56                  unsigned dispatch_width) :
  57          shader(shader), block(NULL), cursor(NULL),
  58          _dispatch_width(dispatch_width),
  59          _group(0),
  60          force_writemask_all(false),
  61          annotation()
  62       {
  63       }
  64
  65       /**
  66        * Construct an fs_builder that inserts instructions into \p shader
  67        * before instruction \p inst in basic block \p block.  The default
  68        * execution controls and debug annotation are initialized from the
  69        * instruction passed as argument.
  70        */
  71       fs_builder(backend_shader *shader, bblock_t *block, fs_inst *inst) :
  72          shader(shader), block(block), cursor(inst),
  73          _dispatch_width(inst->exec_size),
  74          _group(inst->group),
  75          force_writemask_all(inst->force_writemask_all)
  76       {
  77          annotation.str = inst->annotation;
  78          annotation.ir = inst->ir;
  79       }
  80
  81       /**
  82        * Construct an fs_builder that inserts instructions before \p cursor in
  83        * basic block \p block, inheriting other code generation parameters
  84        * from this.
  85        */
  86       fs_builder
  87       at(bblock_t *block, exec_node *cursor) const
  88       {
  89          fs_builder bld = *this;
  90          bld.block = block;
  91          bld.cursor = cursor;
  92          return bld;
  93       }
  94
  95       /**
  96        * Construct an fs_builder appending instructions at the end of the
  97        * instruction list of the shader, inheriting other code generation
  98        * parameters from this.
  99        */
 100       fs_builder
 101       at_end() const
 102       {
 103          return at(NULL, (exec_node *)&shader->instructions.tail_sentinel);
 104       }
 105
 106       /**
 107        * Construct a builder specifying the default SIMD width and group of
 108        * channel enable signals, inheriting other code generation parameters
 109        * from this.
 110        *
 111        * \p n gives the default SIMD width, \p i gives the slot group used for
 112        * predication and control flow masking in multiples of \p n channels.
 113        */
 114       fs_builder
 115       group(unsigned n, unsigned i) const
 116       {
 117          fs_builder bld = *this;
 118
 119          if (n <= dispatch_width() && i < dispatch_width() / n) {
 120             bld._group += i * n;
 121          } else {
 122             /* The requested channel group isn't a subset of the channel group
 123              * of this builder, which means that the resulting instructions
 124              * would use (potentially undefined) channel enable signals not
 125              * specified by the parent builder.  That's only valid if the
 126              * instruction doesn't have per-channel semantics, in which case
 127              * we should clear off the default group index in order to prevent
 128              * emitting instructions with channel group not aligned to their
 129              * own execution size.
 130              */
 131             assert(force_writemask_all);
 132             bld._group = 0;
 133          }
 134
 135          bld._dispatch_width = n;
 136          return bld;
 137       }
 138
 139       /**
 140        * Alias for group() with width equal to eight.
 141        */
 142       fs_builder
 143       half(unsigned i) const
 144       {
 145          return group(8, i);
 146       }
 147
 148       /**
 149        * Construct a builder with per-channel control flow execution masking
 150        * disabled if \p b is true.  If control flow execution masking is
 151        * already disabled this has no effect.
 152        */
 153       fs_builder
 154       exec_all(bool b = true) const
 155       {
 156          fs_builder bld = *this;
 157          if (b)
 158             bld.force_writemask_all = true;
 159          return bld;
 160       }
 161
 162       /**
 163        * Construct a builder with the given debug annotation info.
 164        */
 165       fs_builder
 166       annotate(const char *str, const void *ir = NULL) const
 167       {
 168          fs_builder bld = *this;
 169          bld.annotation.str = str;
 170          bld.annotation.ir = ir;
 171          return bld;
 172       }
 173
 174       /**
 175        * Get the SIMD width in use.
 176        */
 177       unsigned
 178       dispatch_width() const
 179       {
 180          return _dispatch_width;
 181       }
 182
 183       /**
 184        * Get the channel group in use.
 185        */
 186       unsigned
 187       group() const
 188       {
 189          return _group;
 190       }
 191
 192       /**
 193        * Allocate a virtual register of natural vector size (one for this IR)
 194        * and SIMD width.  \p n gives the amount of space to allocate in
 195        * dispatch_width units (which is just enough space for one logical
 196        * component in this IR).
 197        */
 198       dst_reg
 199       vgrf(enum brw_reg_type type, unsigned n = 1) const
 200       {
 201          assert(dispatch_width() <= 32);
 202
 203          if (n > 0)
 204             return dst_reg(VGRF, shader->alloc.allocate(
 205                               DIV_ROUND_UP(n * type_sz(type) * dispatch_width(),
 206                                            REG_SIZE)),
 207                            type);
 208          else
 209             return retype(null_reg_ud(), type);
 210       }
 211
 212       /**
 213        * Create a null register of floating type.
 214        */
 215       dst_reg
 216       null_reg_f() const
 217       {
 218          return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_F));
 219       }
 220
 221       dst_reg
 222       null_reg_df() const
 223       {
 224          return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_DF));
 225       }
 226
 227       /**
 228        * Create a null register of signed integer type.
 229        */
 230       dst_reg
 231       null_reg_d() const
 232       {
 233          return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
 234       }
 235
 236       /**
 237        * Create a null register of unsigned integer type.
 238        */
 239       dst_reg
 240       null_reg_ud() const
 241       {
 242          return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD));
 243       }
 244
 245       /**
 246        * Get the mask of SIMD channels enabled by dispatch and not yet
 247        * disabled by discard.
 248        */
 249       src_reg
 250       sample_mask_reg() const
 251       {
 252          if (shader->stage != MESA_SHADER_FRAGMENT) {
 253             return brw_imm_d(0xffffffff);
 254          } else if (brw_wm_prog_data(shader->stage_prog_data)->uses_kill) {
 255             return brw_flag_reg(0, 1);
 256          } else {
 257             assert(shader->devinfo->gen >= 6 && dispatch_width() <= 16);
 258             return retype(brw_vec1_grf((_group >= 16 ? 2 : 1), 7),
 259                           BRW_REGISTER_TYPE_UD);
 260          }
 261       }
 262
 263       /**
 264        * Insert an instruction into the program.
 265        */
 266       instruction *
 267       emit(const instruction &inst) const
 268       {
 269          return emit(new(shader->mem_ctx) instruction(inst));
 270       }
 271
 272       /**
 273        * Create and insert a nullary control instruction into the program.
 274        */
 275       instruction *
 276       emit(enum opcode opcode) const
 277       {
 278          return emit(instruction(opcode, dispatch_width()));
 279       }
 280
 281       /**
 282        * Create and insert a nullary instruction into the program.
 283        */
 284       instruction *
 285       emit(enum opcode opcode, const dst_reg &dst) const
 286       {
 287          return emit(instruction(opcode, dispatch_width(), dst));
 288       }
 289
 290       /**
 291        * Create and insert a unary instruction into the program.
 292        */
 293       instruction *
 294       emit(enum opcode opcode, const dst_reg &dst, const src_reg &src0) const
 295       {
 296          switch (opcode) {
 297          case SHADER_OPCODE_RCP:
 298          case SHADER_OPCODE_RSQ:
 299          case SHADER_OPCODE_SQRT:
 300          case SHADER_OPCODE_EXP2:
 301          case SHADER_OPCODE_LOG2:
 302          case SHADER_OPCODE_SIN:
 303          case SHADER_OPCODE_COS:
 304             return emit(instruction(opcode, dispatch_width(), dst,
 305                                     fix_math_operand(src0)));
 306
 307          default:
 308             return emit(instruction(opcode, dispatch_width(), dst, src0));
 309          }
 310       }
 311
 312       /**
 313        * Create and insert a binary instruction into the program.
 314        */
 315       instruction *
 316       emit(enum opcode opcode, const dst_reg &dst, const src_reg &src0,
 317            const src_reg &src1) const
 318       {
 319          switch (opcode) {
 320          case SHADER_OPCODE_POW:
 321          case SHADER_OPCODE_INT_QUOTIENT:
 322          case SHADER_OPCODE_INT_REMAINDER:
 323             return emit(instruction(opcode, dispatch_width(), dst,
 324                                     fix_math_operand(src0),
 325                                     fix_math_operand(fix_byte_src(src1))));
 326
 327          default:
 328             return emit(instruction(opcode, dispatch_width(), dst,
 329                                     src0, fix_byte_src(src1)));
 330
 331          }
 332       }
 333
 334       /**
 335        * Create and insert a ternary instruction into the program.
 336        */
 337       instruction *
 338       emit(enum opcode opcode, const dst_reg &dst, const src_reg &src0,
 339            const src_reg &src1, const src_reg &src2) const
 340       {
 341          switch (opcode) {
 342          case BRW_OPCODE_BFE:
 343          case BRW_OPCODE_BFI2:
 344          case BRW_OPCODE_MAD:
 345          case BRW_OPCODE_LRP:
 346             return emit(instruction(opcode, dispatch_width(), dst,
 347                                     fix_3src_operand(src0),
 348                                     fix_3src_operand(fix_byte_src(src1)),
 349                                     fix_3src_operand(fix_byte_src(src2))));
 350
 351          default:
 352             return emit(instruction(opcode, dispatch_width(), dst,
 353                                     src0, fix_byte_src(src1), fix_byte_src(src2)));
 354          }
 355       }
 356
 357       /**
 358        * Create and insert an instruction with a variable number of sources
 359        * into the program.
 360        */
 361       instruction *
 362       emit(enum opcode opcode, const dst_reg &dst, const src_reg srcs[],
 363            unsigned n) const
 364       {
 365          return emit(instruction(opcode, dispatch_width(), dst, srcs, n));
 366       }
 367
 368       /**
 369        * Insert a preallocated instruction into the program.
 370        */
 371       instruction *
 372       emit(instruction *inst) const
 373       {
 374          assert(inst->exec_size <= 32);
 375          assert(inst->exec_size == dispatch_width() ||
 376                 force_writemask_all);
 377
 378          inst->group = _group;
 379          inst->force_writemask_all = force_writemask_all;
 380          inst->annotation = annotation.str;
 381          inst->ir = annotation.ir;
 382
 383          if (block)
 384             static_cast<instruction *>(cursor)->insert_before(block, inst);
 385          else
 386             cursor->insert_before(inst);
 387
 388          return inst;
 389       }
 390
 391       /**
 392        * Select \p src0 if the comparison of both sources with the given
 393        * conditional mod evaluates to true, otherwise select \p src1.
 394        *
 395        * Generally useful to get the minimum or maximum of two values.
 396        */
 397       instruction *
 398       emit_minmax(const dst_reg &dst, const src_reg &src0,
 399                   const src_reg &src1, brw_conditional_mod mod) const
 400       {
 401          assert(mod == BRW_CONDITIONAL_GE || mod == BRW_CONDITIONAL_L);
 402
 403          /* In some cases we can't have bytes as operand for src1, so use the
 404           * same type for both operand.
 405           */
 406          return set_condmod(mod, SEL(dst, fix_unsigned_negate(fix_byte_src(src0)),
 407                                      fix_unsigned_negate(fix_byte_src(src1))));
 408       }
 409
 410       /**
 411        * Copy any live channel from \p src to the first channel of the result.
 412        */
 413       src_reg
 414       emit_uniformize(const src_reg &src) const
 415       {
 416          /* FIXME: We use a vector chan_index and dst to allow constant and
 417           * copy propagration to move result all the way into the consuming
 418           * instruction (typically a surface index or sampler index for a
 419           * send). This uses 1 or 3 extra hw registers in 16 or 32 wide
 420           * dispatch. Once we teach const/copy propagation about scalars we
 421           * should go back to scalar destinations here.
 422           */
 423          const fs_builder ubld = exec_all();
 424          const dst_reg chan_index = vgrf(BRW_REGISTER_TYPE_UD);
 425          const dst_reg dst = vgrf(src.type);
 426
 427          ubld.emit(SHADER_OPCODE_FIND_LIVE_CHANNEL, chan_index)->flag_subreg = 2;
 428          ubld.emit(SHADER_OPCODE_BROADCAST, dst, src, component(chan_index, 0));
 429
 430          return src_reg(component(dst, 0));
 431       }
 432
 433       src_reg
 434       move_to_vgrf(const src_reg &src, unsigned num_components) const
 435       {
 436          src_reg *const src_comps = new src_reg[num_components];
 437          for (unsigned i = 0; i < num_components; i++)
 438             src_comps[i] = offset(src, dispatch_width(), i);
 439
 440          const dst_reg dst = vgrf(src.type, num_components);
 441          LOAD_PAYLOAD(dst, src_comps, num_components, 0);
 442
 443          delete[] src_comps;
 444
 445          return src_reg(dst);
 446       }
 447
 448       void
 449       emit_scan(enum opcode opcode, const dst_reg &tmp,
 450                 unsigned cluster_size, brw_conditional_mod mod) const
 451       {
 452          assert(dispatch_width() >= 8);
 453
 454          /* The instruction splitting code isn't advanced enough to split
 455           * these so we need to handle that ourselves.
 456           */
 457          if (dispatch_width() * type_sz(tmp.type) > 2 * REG_SIZE) {
 458             const unsigned half_width = dispatch_width() / 2;
 459             const fs_builder ubld = exec_all().group(half_width, 0);
 460             dst_reg left = tmp;
 461             dst_reg right = horiz_offset(tmp, half_width);
 462             ubld.emit_scan(opcode, left, cluster_size, mod);
 463             ubld.emit_scan(opcode, right, cluster_size, mod);
 464             if (cluster_size > half_width) {
 465                src_reg left_comp = component(left, half_width - 1);
 466                set_condmod(mod, ubld.emit(opcode, right, left_comp, right));
 467             }
 468             return;
 469          }
 470
 471          if (cluster_size > 1) {
 472             const fs_builder ubld = exec_all().group(dispatch_width() / 2, 0);
 473             const dst_reg left = horiz_stride(tmp, 2);
 474             const dst_reg right = horiz_stride(horiz_offset(tmp, 1), 2);
 475             set_condmod(mod, ubld.emit(opcode, right, left, right));
 476          }
 477
 478          if (cluster_size > 2) {
 479             if (type_sz(tmp.type) <= 4) {
 480                const fs_builder ubld =
 481                   exec_all().group(dispatch_width() / 4, 0);
 482                src_reg left = horiz_stride(horiz_offset(tmp, 1), 4);
 483
 484                dst_reg right = horiz_stride(horiz_offset(tmp, 2), 4);
 485                set_condmod(mod, ubld.emit(opcode, right, left, right));
 486
 487                right = horiz_stride(horiz_offset(tmp, 3), 4);
 488                set_condmod(mod, ubld.emit(opcode, right, left, right));
 489             } else {
 490                /* For 64-bit types, we have to do things differently because
 491                 * the code above would land us with destination strides that
 492                 * the hardware can't handle.  Fortunately, we'll only be
 493                 * 8-wide in that case and it's the same number of
 494                 * instructions.
 495                 */
 496                const fs_builder ubld = exec_all().group(2, 0);
 497
 498                for (unsigned i = 0; i < dispatch_width(); i += 4) {
 499                   src_reg left = component(tmp, i + 1);
 500                   dst_reg right = horiz_offset(tmp, i + 2);
 501                   set_condmod(mod, ubld.emit(opcode, right, left, right));
 502                }
 503             }
 504          }
 505
 506          if (cluster_size > 4) {
 507             const fs_builder ubld = exec_all().group(4, 0);
 508             src_reg left = component(tmp, 3);
 509             dst_reg right = horiz_offset(tmp, 4);
 510             set_condmod(mod, ubld.emit(opcode, right, left, right));
 511
 512             if (dispatch_width() > 8) {
 513                left = component(tmp, 8 + 3);
 514                right = horiz_offset(tmp, 8 + 4);
 515                set_condmod(mod, ubld.emit(opcode, right, left, right));
 516             }
 517          }
 518
 519          if (cluster_size > 8 && dispatch_width() > 8) {
 520             const fs_builder ubld = exec_all().group(8, 0);
 521             src_reg left = component(tmp, 7);
 522             dst_reg right = horiz_offset(tmp, 8);
 523             set_condmod(mod, ubld.emit(opcode, right, left, right));
 524          }
 525       }
 526
 527       /**
 528        * Assorted arithmetic ops.
 529        * @{
 530        */
 531 #define ALU1(op)                                        \
 532       instruction *                                     \
 533       op(const dst_reg &dst, const src_reg &src0) const \
 534       {                                                 \
 535          return emit(BRW_OPCODE_##op, dst, src0);       \
 536       }
 537
 538 #define ALU2(op)                                                        \
 539       instruction *                                                     \
 540       op(const dst_reg &dst, const src_reg &src0, const src_reg &src1) const \
 541       {                                                                 \
 542          return emit(BRW_OPCODE_##op, dst, src0, src1);                 \
 543       }
 544
 545 #define ALU2_ACC(op)                                                    \
 546       instruction *                                                     \
 547       op(const dst_reg &dst, const src_reg &src0, const src_reg &src1) const \
 548       {                                                                 \
 549          instruction *inst = emit(BRW_OPCODE_##op, dst, src0, src1);    \
 550          inst->writes_accumulator = true;                               \
 551          return inst;                                                   \
 552       }
 553
 554 #define ALU3(op)                                                        \
 555       instruction *                                                     \
 556       op(const dst_reg &dst, const src_reg &src0, const src_reg &src1,  \
 557          const src_reg &src2) const                                     \
 558       {                                                                 \
 559          return emit(BRW_OPCODE_##op, dst, src0, src1, src2);           \
 560       }
 561
 562       ALU2(ADD)
 563       ALU2_ACC(ADDC)
 564       ALU2(AND)
 565       ALU2(ASR)
 566       ALU2(AVG)
 567       ALU3(BFE)
 568       ALU2(BFI1)
 569       ALU3(BFI2)
 570       ALU1(BFREV)
 571       ALU1(CBIT)
 572       ALU2(CMPN)
 573       ALU1(DIM)
 574       ALU2(DP2)
 575       ALU2(DP3)
 576       ALU2(DP4)
 577       ALU2(DPH)
 578       ALU1(F16TO32)
 579       ALU1(F32TO16)
 580       ALU1(FBH)
 581       ALU1(FBL)
 582       ALU1(FRC)
 583       ALU2(LINE)
 584       ALU1(LZD)
 585       ALU2(MAC)
 586       ALU2_ACC(MACH)
 587       ALU3(MAD)
 588       ALU1(MOV)
 589       ALU2(MUL)
 590       ALU1(NOT)
 591       ALU2(OR)
 592       ALU2(PLN)
 593       ALU1(RNDD)
 594       ALU1(RNDE)
 595       ALU1(RNDU)
 596       ALU1(RNDZ)
 597       ALU2(ROL)
 598       ALU2(ROR)
 599       ALU2(SAD2)
 600       ALU2_ACC(SADA2)
 601       ALU2(SEL)
 602       ALU2(SHL)
 603       ALU2(SHR)
 604       ALU2_ACC(SUBB)
 605       ALU2(XOR)
 606
 607 #undef ALU3
 608 #undef ALU2_ACC
 609 #undef ALU2
 610 #undef ALU1
 611       /** @} */
 612
 613       /**
 614        * CMP: Sets the low bit of the destination channels with the result
 615        * of the comparison, while the upper bits are undefined, and updates
 616        * the flag register with the packed 16 bits of the result.
 617        */
 618       instruction *
 619       CMP(const dst_reg &dst, const src_reg &src0, const src_reg &src1,
 620           brw_conditional_mod condition) const
 621       {
 622          /* Take the instruction:
 623           *
 624           * CMP null<d> src0<f> src1<f>
 625           *
 626           * Original gen4 does type conversion to the destination type
 627           * before comparison, producing garbage results for floating
 628           * point comparisons.
 629           *
 630           * The destination type doesn't matter on newer generations,
 631           * so we set the type to match src0 so we can compact the
 632           * instruction.
 633           */
 634          return set_condmod(condition,
 635                             emit(BRW_OPCODE_CMP, retype(dst, src0.type),
 636                                  fix_unsigned_negate(src0),
 637                                  fix_unsigned_negate(src1)));
 638       }
 639
 640       /**
 641        * Gen4 predicated IF.
 642        */
 643       instruction *
 644       IF(brw_predicate predicate) const
 645       {
 646          return set_predicate(predicate, emit(BRW_OPCODE_IF));
 647       }
 648
 649       /**
 650        * CSEL: dst = src2 <op> 0.0f ? src0 : src1
 651        */
 652       instruction *
 653       CSEL(const dst_reg &dst, const src_reg &src0, const src_reg &src1,
 654            const src_reg &src2, brw_conditional_mod condition) const
 655       {
 656          /* CSEL only operates on floats, so we can't do integer </<=/>=/>
 657           * comparisons.  Zero/non-zero (== and !=) comparisons almost work.
 658           * 0x80000000 fails because it is -0.0, and -0.0 == 0.0.
 659           */
 660          assert(src2.type == BRW_REGISTER_TYPE_F);
 661
 662          return set_condmod(condition,
 663                             emit(BRW_OPCODE_CSEL,
 664                                  retype(dst, BRW_REGISTER_TYPE_F),
 665                                  retype(src0, BRW_REGISTER_TYPE_F),
 666                                  retype(fix_byte_src(src1), BRW_REGISTER_TYPE_F),
 667                                  fix_byte_src(src2)));
 668       }
 669
 670       /**
 671        * Emit a linear interpolation instruction.
 672        */
 673       instruction *
 674       LRP(const dst_reg &dst, const src_reg &x, const src_reg &y,
 675           const src_reg &a) const
 676       {
 677          if (shader->devinfo->gen >= 6 && shader->devinfo->gen <= 10) {
 678             /* The LRP instruction actually does op1 * op0 + op2 * (1 - op0), so
 679              * we need to reorder the operands.
 680              */
 681             return emit(BRW_OPCODE_LRP, dst, a, y, x);
 682
 683          } else {
 684             /* We can't use the LRP instruction.  Emit x*(1-a) + y*a. */
 685             const dst_reg y_times_a = vgrf(dst.type);
 686             const dst_reg one_minus_a = vgrf(dst.type);
 687             const dst_reg x_times_one_minus_a = vgrf(dst.type);
 688
 689             MUL(y_times_a, y, a);
 690             ADD(one_minus_a, negate(a), brw_imm_f(1.0f));
 691             MUL(x_times_one_minus_a, x, src_reg(one_minus_a));
 692             return ADD(dst, src_reg(x_times_one_minus_a), src_reg(y_times_a));
 693          }
 694       }
 695
 696       /**
 697        * Collect a number of registers in a contiguous range of registers.
 698        */
 699       instruction *
 700       LOAD_PAYLOAD(const dst_reg &dst, const src_reg *src,
 701                    unsigned sources, unsigned header_size) const
 702       {
 703          instruction *inst = emit(SHADER_OPCODE_LOAD_PAYLOAD, dst, src, sources);
 704          inst->header_size = header_size;
 705          inst->size_written = header_size * REG_SIZE;
 706          for (unsigned i = header_size; i < sources; i++) {
 707             inst->size_written +=
 708                ALIGN(dispatch_width() * type_sz(src[i].type) * dst.stride,
 709                      REG_SIZE);
 710          }
 711
 712          return inst;
 713       }
 714
 715       instruction *
 716       UNDEF(const dst_reg &dst) const
 717       {
 718          assert(dst.file == VGRF);
 719          instruction *inst = emit(SHADER_OPCODE_UNDEF,
 720                                   retype(dst, BRW_REGISTER_TYPE_UD));
 721          inst->size_written = shader->alloc.sizes[dst.nr] * REG_SIZE;
 722
 723          return inst;
 724       }
 725
 726       backend_shader *shader;
 727
 728       /**
 729        * Byte sized operands are not supported for src1 on Gen11+.
 730        */
 731       src_reg
 732       fix_byte_src(const src_reg &src) const
 733       {
 734          if ((shader->devinfo->gen < 11 && !shader->devinfo->is_geminilake) ||
 735              type_sz(src.type) != 1)
 736             return src;
 737
 738          dst_reg temp = vgrf(src.type == BRW_REGISTER_TYPE_UB ?
 739                              BRW_REGISTER_TYPE_UD : BRW_REGISTER_TYPE_D);
 740          MOV(temp, src);
 741          return src_reg(temp);
 742       }
 743
 744    private:
 745       /**
 746        * Workaround for negation of UD registers.  See comment in
 747        * fs_generator::generate_code() for more details.
 748        */
 749       src_reg
 750       fix_unsigned_negate(const src_reg &src) const
 751       {
 752          if (src.type == BRW_REGISTER_TYPE_UD &&
 753              src.negate) {
 754             dst_reg temp = vgrf(BRW_REGISTER_TYPE_UD);
 755             MOV(temp, src);
 756             return src_reg(temp);
 757          } else {
 758             return src;
 759          }
 760       }
 761
 762       /**
 763        * Workaround for source register modes not supported by the ternary
 764        * instruction encoding.
 765        */
 766       src_reg
 767       fix_3src_operand(const src_reg &src) const
 768       {
 769          switch (src.file) {
 770          case FIXED_GRF:
 771             /* FINISHME: Could handle scalar region, other stride=1 regions */
 772             if (src.vstride != BRW_VERTICAL_STRIDE_8 ||
 773                 src.width != BRW_WIDTH_8 ||
 774                 src.hstride != BRW_HORIZONTAL_STRIDE_1)
 775                break;
 776             /* fallthrough */
 777          case ATTR:
 778          case VGRF:
 779          case UNIFORM:
 780          case IMM:
 781             return src;
 782          default:
 783             break;
 784          }
 785
 786          dst_reg expanded = vgrf(src.type);
 787          MOV(expanded, src);
 788          return expanded;
 789       }
 790
 791       /**
 792        * Workaround for source register modes not supported by the math
 793        * instruction.
 794        */
 795       src_reg
 796       fix_math_operand(const src_reg &src) const
 797       {
 798          /* Can't do hstride == 0 args on gen6 math, so expand it out. We
 799           * might be able to do better by doing execsize = 1 math and then
 800           * expanding that result out, but we would need to be careful with
 801           * masking.
 802           *
 803           * Gen6 hardware ignores source modifiers (negate and abs) on math
 804           * instructions, so we also move to a temp to set those up.
 805           *
 806           * Gen7 relaxes most of the above restrictions, but still can't use IMM
 807           * operands to math
 808           */
 809          if ((shader->devinfo->gen == 6 &&
 810               (src.file == IMM || src.file == UNIFORM ||
 811                src.abs || src.negate)) ||
 812              (shader->devinfo->gen == 7 && src.file == IMM)) {
 813             const dst_reg tmp = vgrf(src.type);
 814             MOV(tmp, src);
 815             return tmp;
 816          } else {
 817             return src;
 818          }
 819       }
 820
 821       bblock_t *block;
 822       exec_node *cursor;
 823
 824       unsigned _dispatch_width;
 825       unsigned _group;
 826       bool force_writemask_all;
 827
 828       /** Debug annotation info. */
 829       struct {
 830          const char *str;
 831          const void *ir;
 832       } annotation;
 833    };
 834 }
 835
 836 #endif