src/intel/compiler/brw_fs.h

   1 /*
   2  * Copyright © 2010 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eric Anholt <eric@anholt.net>
  25  *
  26  */
  27
  28 #ifndef BRW_FS_H
  29 #define BRW_FS_H
  30
  31 #include "brw_shader.h"
  32 #include "brw_ir_fs.h"
  33 #include "brw_fs_builder.h"
  34 #include "compiler/nir/nir.h"
  35
  36 struct bblock_t;
  37 namespace {
  38    struct acp_entry;
  39 }
  40
  41 namespace brw {
  42    class fs_live_variables;
  43 }
  44
  45 struct brw_gs_compile;
  46
  47 static inline fs_reg
  48 offset(const fs_reg &reg, const brw::fs_builder &bld, unsigned delta)
  49 {
  50    return offset(reg, bld.dispatch_width(), delta);
  51 }
  52
  53 #define UBO_START ((1 << 16) - 4)
  54
  55 struct shader_stats {
  56    const char *scheduler_mode;
  57    unsigned promoted_constants;
  58 };
  59
  60 /**
  61  * The fragment shader front-end.
  62  *
  63  * Translates either GLSL IR or Mesa IR (for ARB_fragment_program) into FS IR.
  64  */
  65 class fs_visitor : public backend_shader
  66 {
  67 public:
  68    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  69               void *mem_ctx,
  70               const brw_base_prog_key *key,
  71               struct brw_stage_prog_data *prog_data,
  72               const nir_shader *shader,
  73               unsigned dispatch_width,
  74               int shader_time_index,
  75               const struct brw_vue_map *input_vue_map = NULL);
  76    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  77               void *mem_ctx,
  78               struct brw_gs_compile *gs_compile,
  79               struct brw_gs_prog_data *prog_data,
  80               const nir_shader *shader,
  81               int shader_time_index);
  82    void init();
  83    ~fs_visitor();
  84
  85    fs_reg vgrf(const glsl_type *const type);
  86    void import_uniforms(fs_visitor *v);
  87
  88    void VARYING_PULL_CONSTANT_LOAD(const brw::fs_builder &bld,
  89                                    const fs_reg &dst,
  90                                    const fs_reg &surf_index,
  91                                    const fs_reg &varying_offset,
  92                                    uint32_t const_offset);
  93    void DEP_RESOLVE_MOV(const brw::fs_builder &bld, int grf);
  94
  95    bool run_fs(bool allow_spilling, bool do_rep_send);
  96    bool run_vs();
  97    bool run_tcs();
  98    bool run_tes();
  99    bool run_gs();
 100    bool run_cs(unsigned min_dispatch_width);
 101    void optimize();
 102    void allocate_registers(unsigned min_dispatch_width, bool allow_spilling);
 103    void setup_fs_payload_gen4();
 104    void setup_fs_payload_gen6();
 105    void setup_vs_payload();
 106    void setup_gs_payload();
 107    void setup_cs_payload();
 108    bool fixup_sends_duplicate_payload();
 109    void fixup_3src_null_dest();
 110    void assign_curb_setup();
 111    void assign_urb_setup();
 112    void convert_attr_sources_to_hw_regs(fs_inst *inst);
 113    void assign_vs_urb_setup();
 114    void assign_tcs_urb_setup();
 115    void assign_tes_urb_setup();
 116    void assign_gs_urb_setup();
 117    bool assign_regs(bool allow_spilling, bool spill_all);
 118    void assign_regs_trivial();
 119    void calculate_payload_ranges(int payload_node_count,
 120                                  int *payload_last_use_ip);
 121    void split_virtual_grfs();
 122    bool compact_virtual_grfs();
 123    void assign_constant_locations();
 124    bool get_pull_locs(const fs_reg &src, unsigned *out_surf_index,
 125                       unsigned *out_pull_index);
 126    void lower_constant_loads();
 127    void invalidate_live_intervals();
 128    void calculate_live_intervals();
 129    void calculate_register_pressure();
 130    void validate();
 131    bool opt_algebraic();
 132    bool opt_redundant_discard_jumps();
 133    bool opt_cse();
 134    bool opt_cse_local(bblock_t *block);
 135    bool opt_copy_propagation();
 136    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 137    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 138    bool opt_copy_propagation_local(void *mem_ctx, bblock_t *block,
 139                                    exec_list *acp);
 140    bool opt_drop_redundant_mov_to_flags();
 141    bool opt_register_renaming();
 142    bool opt_bank_conflicts();
 143    unsigned bank_conflict_cycles(const fs_inst *inst) const;
 144    bool register_coalesce();
 145    bool compute_to_mrf();
 146    bool eliminate_find_live_channel();
 147    bool dead_code_eliminate();
 148    bool remove_duplicate_mrf_writes();
 149    bool remove_extra_rounding_modes();
 150
 151    bool opt_sampler_eot();
 152    bool virtual_grf_interferes(int a, int b);
 153    void schedule_instructions(instruction_scheduler_mode mode);
 154    void insert_gen4_send_dependency_workarounds();
 155    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 156                                                     fs_inst *inst);
 157    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 158                                                      fs_inst *inst);
 159    void vfail(const char *msg, va_list args);
 160    void fail(const char *msg, ...);
 161    void limit_dispatch_width(unsigned n, const char *msg);
 162    void lower_uniform_pull_constant_loads();
 163    bool lower_load_payload();
 164    bool lower_pack();
 165    bool lower_regioning();
 166    bool lower_logical_sends();
 167    bool lower_integer_multiplication();
 168    bool lower_minmax();
 169    bool lower_simd_width();
 170    bool opt_combine_constants();
 171
 172    void emit_dummy_fs();
 173    void emit_repclear_shader();
 174    void emit_fragcoord_interpolation(fs_reg wpos);
 175    fs_reg *emit_frontfacing_interpolation();
 176    fs_reg *emit_samplepos_setup();
 177    fs_reg *emit_sampleid_setup();
 178    fs_reg *emit_samplemaskin_setup();
 179    void emit_interpolation_setup_gen4();
 180    void emit_interpolation_setup_gen6();
 181    void compute_sample_position(fs_reg dst, fs_reg int_sample_pos);
 182    fs_reg emit_mcs_fetch(const fs_reg &coordinate, unsigned components,
 183                          const fs_reg &texture,
 184                          const fs_reg &texture_handle);
 185    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 186    fs_reg resolve_source_modifiers(const fs_reg &src);
 187    void emit_discard_jump();
 188    void emit_fsign(const class brw::fs_builder &, const nir_alu_instr *instr,
 189                    fs_reg result, fs_reg *op, unsigned fsign_src);
 190    bool opt_peephole_sel();
 191    bool opt_peephole_csel();
 192    bool opt_peephole_predicated_break();
 193    bool opt_saturate_propagation();
 194    bool opt_cmod_propagation();
 195    bool opt_zero_samples();
 196
 197    void set_tcs_invocation_id();
 198
 199    void emit_nir_code();
 200    void nir_setup_outputs();
 201    void nir_setup_uniforms();
 202    void nir_emit_system_values();
 203    void nir_emit_impl(nir_function_impl *impl);
 204    void nir_emit_cf_list(exec_list *list);
 205    void nir_emit_if(nir_if *if_stmt);
 206    void nir_emit_loop(nir_loop *loop);
 207    void nir_emit_block(nir_block *block);
 208    void nir_emit_instr(nir_instr *instr);
 209    void nir_emit_alu(const brw::fs_builder &bld, nir_alu_instr *instr,
 210                      bool need_dest);
 211    bool try_emit_b2fi_of_inot(const brw::fs_builder &bld, fs_reg result,
 212                               nir_alu_instr *instr);
 213    void nir_emit_load_const(const brw::fs_builder &bld,
 214                             nir_load_const_instr *instr);
 215    void nir_emit_vs_intrinsic(const brw::fs_builder &bld,
 216                               nir_intrinsic_instr *instr);
 217    void nir_emit_tcs_intrinsic(const brw::fs_builder &bld,
 218                                nir_intrinsic_instr *instr);
 219    void nir_emit_gs_intrinsic(const brw::fs_builder &bld,
 220                               nir_intrinsic_instr *instr);
 221    void nir_emit_fs_intrinsic(const brw::fs_builder &bld,
 222                               nir_intrinsic_instr *instr);
 223    void nir_emit_cs_intrinsic(const brw::fs_builder &bld,
 224                               nir_intrinsic_instr *instr);
 225    fs_reg get_nir_image_intrinsic_image(const brw::fs_builder &bld,
 226                                         nir_intrinsic_instr *instr);
 227    fs_reg get_nir_ssbo_intrinsic_index(const brw::fs_builder &bld,
 228                                        nir_intrinsic_instr *instr);
 229    void nir_emit_intrinsic(const brw::fs_builder &bld,
 230                            nir_intrinsic_instr *instr);
 231    void nir_emit_tes_intrinsic(const brw::fs_builder &bld,
 232                                nir_intrinsic_instr *instr);
 233    void nir_emit_ssbo_atomic(const brw::fs_builder &bld,
 234                              int op, nir_intrinsic_instr *instr);
 235    void nir_emit_ssbo_atomic_float(const brw::fs_builder &bld,
 236                                    int op, nir_intrinsic_instr *instr);
 237    void nir_emit_shared_atomic(const brw::fs_builder &bld,
 238                                int op, nir_intrinsic_instr *instr);
 239    void nir_emit_shared_atomic_float(const brw::fs_builder &bld,
 240                                      int op, nir_intrinsic_instr *instr);
 241    void nir_emit_global_atomic(const brw::fs_builder &bld,
 242                                int op, nir_intrinsic_instr *instr);
 243    void nir_emit_global_atomic_float(const brw::fs_builder &bld,
 244                                      int op, nir_intrinsic_instr *instr);
 245    void nir_emit_texture(const brw::fs_builder &bld,
 246                          nir_tex_instr *instr);
 247    void nir_emit_jump(const brw::fs_builder &bld,
 248                       nir_jump_instr *instr);
 249    fs_reg get_nir_src(const nir_src &src);
 250    fs_reg get_nir_src_imm(const nir_src &src);
 251    fs_reg get_nir_dest(const nir_dest &dest);
 252    fs_reg get_indirect_offset(nir_intrinsic_instr *instr);
 253    fs_reg get_tcs_single_patch_icp_handle(const brw::fs_builder &bld,
 254                                           nir_intrinsic_instr *instr);
 255    fs_reg get_tcs_eight_patch_icp_handle(const brw::fs_builder &bld,
 256                                          nir_intrinsic_instr *instr);
 257    struct brw_reg get_tcs_output_urb_handle();
 258
 259    void emit_percomp(const brw::fs_builder &bld, const fs_inst &inst,
 260                      unsigned wr_mask);
 261
 262    bool optimize_extract_to_float(nir_alu_instr *instr,
 263                                   const fs_reg &result);
 264    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 265                                      const fs_reg &result);
 266
 267    void emit_alpha_test();
 268    fs_inst *emit_single_fb_write(const brw::fs_builder &bld,
 269                                  fs_reg color1, fs_reg color2,
 270                                  fs_reg src0_alpha, unsigned components);
 271    void emit_alpha_to_coverage_workaround(const fs_reg &src0_alpha);
 272    void emit_fb_writes();
 273    fs_inst *emit_non_coherent_fb_read(const brw::fs_builder &bld,
 274                                       const fs_reg &dst, unsigned target);
 275    void emit_urb_writes(const fs_reg &gs_vertex_count = fs_reg());
 276    void set_gs_stream_control_data_bits(const fs_reg &vertex_count,
 277                                         unsigned stream_id);
 278    void emit_gs_control_data_bits(const fs_reg &vertex_count);
 279    void emit_gs_end_primitive(const nir_src &vertex_count_nir_src);
 280    void emit_gs_vertex(const nir_src &vertex_count_nir_src,
 281                        unsigned stream_id);
 282    void emit_gs_thread_end();
 283    void emit_gs_input_load(const fs_reg &dst, const nir_src &vertex_src,
 284                            unsigned base_offset, const nir_src &offset_src,
 285                            unsigned num_components, unsigned first_component);
 286    void emit_cs_terminate();
 287    fs_reg *emit_cs_work_group_id_setup();
 288
 289    void emit_barrier();
 290
 291    void emit_shader_time_begin();
 292    void emit_shader_time_end();
 293    void SHADER_TIME_ADD(const brw::fs_builder &bld,
 294                         int shader_time_subindex,
 295                         fs_reg value);
 296
 297    fs_reg get_timestamp(const brw::fs_builder &bld);
 298
 299    fs_reg interp_reg(int location, int channel);
 300
 301    int implied_mrf_writes(fs_inst *inst) const;
 302
 303    virtual void dump_instructions();
 304    virtual void dump_instructions(const char *name);
 305    void dump_instruction(backend_instruction *inst);
 306    void dump_instruction(backend_instruction *inst, FILE *file);
 307
 308    const brw_base_prog_key *const key;
 309    const struct brw_sampler_prog_key_data *key_tex;
 310
 311    struct brw_gs_compile *gs_compile;
 312
 313    struct brw_stage_prog_data *prog_data;
 314
 315    const struct brw_vue_map *input_vue_map;
 316
 317    int *virtual_grf_start;
 318    int *virtual_grf_end;
 319    brw::fs_live_variables *live_intervals;
 320
 321    int *regs_live_at_ip;
 322
 323    /** Number of uniform variable components visited. */
 324    unsigned uniforms;
 325
 326    /** Byte-offset for the next available spot in the scratch space buffer. */
 327    unsigned last_scratch;
 328
 329    /**
 330     * Array mapping UNIFORM register numbers to the pull parameter index,
 331     * or -1 if this uniform register isn't being uploaded as a pull constant.
 332     */
 333    int *pull_constant_loc;
 334
 335    /**
 336     * Array mapping UNIFORM register numbers to the push parameter index,
 337     * or -1 if this uniform register isn't being uploaded as a push constant.
 338     */
 339    int *push_constant_loc;
 340
 341    fs_reg subgroup_id;
 342    fs_reg frag_depth;
 343    fs_reg frag_stencil;
 344    fs_reg sample_mask;
 345    fs_reg outputs[VARYING_SLOT_MAX];
 346    fs_reg dual_src_output;
 347    int first_non_payload_grf;
 348    /** Either BRW_MAX_GRF or GEN7_MRF_HACK_START */
 349    unsigned max_grf;
 350
 351    fs_reg *nir_locals;
 352    fs_reg *nir_ssa_values;
 353    fs_reg *nir_system_values;
 354
 355    bool failed;
 356    char *fail_msg;
 357
 358    /** Register numbers for thread payload fields. */
 359    struct thread_payload {
 360       uint8_t subspan_coord_reg[2];
 361       uint8_t source_depth_reg[2];
 362       uint8_t source_w_reg[2];
 363       uint8_t aa_dest_stencil_reg[2];
 364       uint8_t dest_depth_reg[2];
 365       uint8_t sample_pos_reg[2];
 366       uint8_t sample_mask_in_reg[2];
 367       uint8_t barycentric_coord_reg[BRW_BARYCENTRIC_MODE_COUNT][2];
 368       uint8_t local_invocation_id_reg[2];
 369
 370       /** The number of thread payload registers the hardware will supply. */
 371       uint8_t num_regs;
 372    } payload;
 373
 374    bool source_depth_to_render_target;
 375    bool runtime_check_aads_emit;
 376
 377    fs_reg pixel_x;
 378    fs_reg pixel_y;
 379    fs_reg wpos_w;
 380    fs_reg pixel_w;
 381    fs_reg delta_xy[BRW_BARYCENTRIC_MODE_COUNT];
 382    fs_reg shader_start_time;
 383    fs_reg final_gs_vertex_count;
 384    fs_reg control_data_bits;
 385    fs_reg invocation_id;
 386
 387    unsigned grf_used;
 388    bool spilled_any_registers;
 389
 390    const unsigned dispatch_width; /**< 8, 16 or 32 */
 391    unsigned max_dispatch_width;
 392
 393    int shader_time_index;
 394
 395    struct shader_stats shader_stats;
 396
 397    brw::fs_builder bld;
 398
 399 private:
 400    fs_reg prepare_alu_destination_and_sources(const brw::fs_builder &bld,
 401                                               nir_alu_instr *instr,
 402                                               fs_reg *op,
 403                                               bool need_dest);
 404
 405    void resolve_inot_sources(const brw::fs_builder &bld, nir_alu_instr *instr,
 406                              fs_reg *op);
 407    void lower_mul_dword_inst(fs_inst *inst, bblock_t *block);
 408    void lower_mul_qword_inst(fs_inst *inst, bblock_t *block);
 409    void lower_mulh_inst(fs_inst *inst, bblock_t *block);
 410 };
 411
 412 /**
 413  * The fragment shader code generator.
 414  *
 415  * Translates FS IR to actual i965 assembly code.
 416  */
 417 class fs_generator
 418 {
 419 public:
 420    fs_generator(const struct brw_compiler *compiler, void *log_data,
 421                 void *mem_ctx,
 422                 struct brw_stage_prog_data *prog_data,
 423                 struct shader_stats shader_stats,
 424                 bool runtime_check_aads_emit,
 425                 gl_shader_stage stage);
 426    ~fs_generator();
 427
 428    void enable_debug(const char *shader_name);
 429    int generate_code(const cfg_t *cfg, int dispatch_width,
 430                      struct brw_compile_stats *stats);
 431    const unsigned *get_assembly();
 432
 433 private:
 434    void fire_fb_write(fs_inst *inst,
 435                       struct brw_reg payload,
 436                       struct brw_reg implied_header,
 437                       GLuint nr);
 438    void generate_send(fs_inst *inst,
 439                       struct brw_reg dst,
 440                       struct brw_reg desc,
 441                       struct brw_reg ex_desc,
 442                       struct brw_reg payload,
 443                       struct brw_reg payload2);
 444    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 445    void generate_fb_read(fs_inst *inst, struct brw_reg dst,
 446                          struct brw_reg payload);
 447    void generate_urb_read(fs_inst *inst, struct brw_reg dst, struct brw_reg payload);
 448    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 449    void generate_cs_terminate(fs_inst *inst, struct brw_reg payload);
 450    void generate_barrier(fs_inst *inst, struct brw_reg src);
 451    bool generate_linterp(fs_inst *inst, struct brw_reg dst,
 452                          struct brw_reg *src);
 453    void generate_tex(fs_inst *inst, struct brw_reg dst,
 454                      struct brw_reg surface_index,
 455                      struct brw_reg sampler_index);
 456    void generate_get_buffer_size(fs_inst *inst, struct brw_reg dst,
 457                                  struct brw_reg src,
 458                                  struct brw_reg surf_index);
 459    void generate_ddx(const fs_inst *inst,
 460                      struct brw_reg dst, struct brw_reg src);
 461    void generate_ddy(const fs_inst *inst,
 462                      struct brw_reg dst, struct brw_reg src);
 463    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 464    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 465    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 466    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 467                                             struct brw_reg index,
 468                                             struct brw_reg offset);
 469    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 470                                                  struct brw_reg dst,
 471                                                  struct brw_reg surf_index,
 472                                                  struct brw_reg payload);
 473    void generate_varying_pull_constant_load_gen4(fs_inst *inst,
 474                                                  struct brw_reg dst,
 475                                                  struct brw_reg index);
 476    void generate_mov_dispatch_to_flags(fs_inst *inst);
 477
 478    void generate_pixel_interpolator_query(fs_inst *inst,
 479                                           struct brw_reg dst,
 480                                           struct brw_reg src,
 481                                           struct brw_reg msg_data,
 482                                           unsigned msg_type);
 483
 484    void generate_set_sample_id(fs_inst *inst,
 485                                struct brw_reg dst,
 486                                struct brw_reg src0,
 487                                struct brw_reg src1);
 488
 489    void generate_discard_jump(fs_inst *inst);
 490
 491    void generate_pack_half_2x16_split(fs_inst *inst,
 492                                       struct brw_reg dst,
 493                                       struct brw_reg x,
 494                                       struct brw_reg y);
 495
 496    void generate_shader_time_add(fs_inst *inst,
 497                                  struct brw_reg payload,
 498                                  struct brw_reg offset,
 499                                  struct brw_reg value);
 500
 501    void generate_mov_indirect(fs_inst *inst,
 502                               struct brw_reg dst,
 503                               struct brw_reg reg,
 504                               struct brw_reg indirect_byte_offset);
 505
 506    void generate_shuffle(fs_inst *inst,
 507                          struct brw_reg dst,
 508                          struct brw_reg src,
 509                          struct brw_reg idx);
 510
 511    void generate_quad_swizzle(const fs_inst *inst,
 512                               struct brw_reg dst, struct brw_reg src,
 513                               unsigned swiz);
 514
 515    bool patch_discard_jumps_to_fb_writes();
 516
 517    const struct brw_compiler *compiler;
 518    void *log_data; /* Passed to compiler->*_log functions */
 519
 520    const struct gen_device_info *devinfo;
 521
 522    struct brw_codegen *p;
 523    struct brw_stage_prog_data * const prog_data;
 524
 525    unsigned dispatch_width; /**< 8, 16 or 32 */
 526
 527    exec_list discard_halt_patches;
 528    struct shader_stats shader_stats;
 529    bool runtime_check_aads_emit;
 530    bool debug_flag;
 531    const char *shader_name;
 532    gl_shader_stage stage;
 533    void *mem_ctx;
 534 };
 535
 536 namespace brw {
 537    inline fs_reg
 538    fetch_payload_reg(const brw::fs_builder &bld, uint8_t regs[2],
 539                      brw_reg_type type = BRW_REGISTER_TYPE_F, unsigned n = 1)
 540    {
 541       if (!regs[0])
 542          return fs_reg();
 543
 544       if (bld.dispatch_width() > 16) {
 545          const fs_reg tmp = bld.vgrf(type, n);
 546          const brw::fs_builder hbld = bld.exec_all().group(16, 0);
 547          const unsigned m = bld.dispatch_width() / hbld.dispatch_width();
 548          fs_reg *const components = new fs_reg[n * m];
 549
 550          for (unsigned c = 0; c < n; c++) {
 551             for (unsigned g = 0; g < m; g++) {
 552                components[c * m + g] =
 553                   offset(retype(brw_vec8_grf(regs[g], 0), type), hbld, c);
 554             }
 555          }
 556
 557          hbld.LOAD_PAYLOAD(tmp, components, n * m, 0);
 558
 559          delete[] components;
 560          return tmp;
 561
 562       } else {
 563          return fs_reg(retype(brw_vec8_grf(regs[0], 0), type));
 564       }
 565    }
 566
 567    bool
 568    lower_src_modifiers(fs_visitor *v, bblock_t *block, fs_inst *inst, unsigned i);
 569 }
 570
 571 void shuffle_from_32bit_read(const brw::fs_builder &bld,
 572                              const fs_reg &dst,
 573                              const fs_reg &src,
 574                              uint32_t first_component,
 575                              uint32_t components);
 576
 577 fs_reg setup_imm_df(const brw::fs_builder &bld,
 578                     double v);
 579
 580 fs_reg setup_imm_b(const brw::fs_builder &bld,
 581                    int8_t v);
 582
 583 fs_reg setup_imm_ub(const brw::fs_builder &bld,
 584                    uint8_t v);
 585
 586 enum brw_barycentric_mode brw_barycentric_mode(enum glsl_interp_mode mode,
 587                                                nir_intrinsic_op op);
 588
 589 #endif /* BRW_FS_H */