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 /**
  56  * The fragment shader front-end.
  57  *
  58  * Translates either GLSL IR or Mesa IR (for ARB_fragment_program) into FS IR.
  59  */
  60 class fs_visitor : public backend_shader
  61 {
  62 public:
  63    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  64               void *mem_ctx,
  65               const void *key,
  66               struct brw_stage_prog_data *prog_data,
  67               struct gl_program *prog,
  68               const nir_shader *shader,
  69               unsigned dispatch_width,
  70               int shader_time_index,
  71               const struct brw_vue_map *input_vue_map = NULL);
  72    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  73               void *mem_ctx,
  74               struct brw_gs_compile *gs_compile,
  75               struct brw_gs_prog_data *prog_data,
  76               const nir_shader *shader,
  77               int shader_time_index);
  78    void init();
  79    ~fs_visitor();
  80
  81    fs_reg vgrf(const glsl_type *const type);
  82    void import_uniforms(fs_visitor *v);
  83    void setup_uniform_clipplane_values();
  84    void compute_clip_distance();
  85
  86    void VARYING_PULL_CONSTANT_LOAD(const brw::fs_builder &bld,
  87                                    const fs_reg &dst,
  88                                    const fs_reg &surf_index,
  89                                    const fs_reg &varying_offset,
  90                                    uint32_t const_offset);
  91    void DEP_RESOLVE_MOV(const brw::fs_builder &bld, int grf);
  92
  93    bool run_fs(bool allow_spilling, bool do_rep_send);
  94    bool run_vs();
  95    bool run_tcs_single_patch();
  96    bool run_tes();
  97    bool run_gs();
  98    bool run_cs(unsigned min_dispatch_width);
  99    void optimize();
 100    void allocate_registers(unsigned min_dispatch_width, bool allow_spilling);
 101    void setup_fs_payload_gen4();
 102    void setup_fs_payload_gen6();
 103    void setup_vs_payload();
 104    void setup_gs_payload();
 105    void setup_cs_payload();
 106    void fixup_3src_null_dest();
 107    void assign_curb_setup();
 108    void calculate_urb_setup();
 109    void assign_urb_setup();
 110    void convert_attr_sources_to_hw_regs(fs_inst *inst);
 111    void assign_vs_urb_setup();
 112    void assign_tcs_single_patch_urb_setup();
 113    void assign_tes_urb_setup();
 114    void assign_gs_urb_setup();
 115    bool assign_regs(bool allow_spilling, bool spill_all);
 116    void assign_regs_trivial();
 117    void calculate_payload_ranges(int payload_node_count,
 118                                  int *payload_last_use_ip);
 119    void setup_payload_interference(struct ra_graph *g, int payload_reg_count,
 120                                    int first_payload_node);
 121    int choose_spill_reg(struct ra_graph *g);
 122    void spill_reg(int spill_reg);
 123    void split_virtual_grfs();
 124    bool compact_virtual_grfs();
 125    void assign_constant_locations();
 126    bool get_pull_locs(const fs_reg &src, unsigned *out_surf_index,
 127                       unsigned *out_pull_index);
 128    void lower_constant_loads();
 129    void invalidate_live_intervals();
 130    void calculate_live_intervals();
 131    void calculate_register_pressure();
 132    void validate();
 133    bool opt_algebraic();
 134    bool opt_redundant_discard_jumps();
 135    bool opt_cse();
 136    bool opt_cse_local(bblock_t *block);
 137    bool opt_copy_propagation();
 138    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 139    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 140    bool opt_copy_propagation_local(void *mem_ctx, bblock_t *block,
 141                                    exec_list *acp);
 142    bool opt_drop_redundant_mov_to_flags();
 143    bool opt_register_renaming();
 144    bool opt_bank_conflicts();
 145    unsigned bank_conflict_cycles(const fs_inst *inst) const;
 146    bool register_coalesce();
 147    bool compute_to_mrf();
 148    bool eliminate_find_live_channel();
 149    bool dead_code_eliminate();
 150    bool remove_duplicate_mrf_writes();
 151    bool remove_extra_rounding_modes();
 152
 153    bool opt_sampler_eot();
 154    bool virtual_grf_interferes(int a, int b);
 155    void schedule_instructions(instruction_scheduler_mode mode);
 156    void insert_gen4_send_dependency_workarounds();
 157    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 158                                                     fs_inst *inst);
 159    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 160                                                      fs_inst *inst);
 161    void vfail(const char *msg, va_list args);
 162    void fail(const char *msg, ...);
 163    void limit_dispatch_width(unsigned n, const char *msg);
 164    void lower_uniform_pull_constant_loads();
 165    bool lower_load_payload();
 166    bool lower_pack();
 167    bool lower_regioning();
 168    bool lower_conversions();
 169    bool lower_logical_sends();
 170    bool lower_integer_multiplication();
 171    bool lower_minmax();
 172    bool lower_simd_width();
 173    bool opt_combine_constants();
 174
 175    void emit_dummy_fs();
 176    void emit_repclear_shader();
 177    void emit_fragcoord_interpolation(fs_reg wpos);
 178    fs_reg *emit_frontfacing_interpolation();
 179    fs_reg *emit_samplepos_setup();
 180    fs_reg *emit_sampleid_setup();
 181    fs_reg *emit_samplemaskin_setup();
 182    void emit_interpolation_setup_gen4();
 183    void emit_interpolation_setup_gen6();
 184    void compute_sample_position(fs_reg dst, fs_reg int_sample_pos);
 185    fs_reg emit_mcs_fetch(const fs_reg &coordinate, unsigned components,
 186                          const fs_reg &sampler);
 187    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 188    fs_reg resolve_source_modifiers(const fs_reg &src);
 189    void emit_discard_jump();
 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 emit_nir_code();
 198    void nir_setup_outputs();
 199    void nir_setup_uniforms();
 200    void nir_emit_system_values();
 201    void nir_emit_impl(nir_function_impl *impl);
 202    void nir_emit_cf_list(exec_list *list);
 203    void nir_emit_if(nir_if *if_stmt);
 204    void nir_emit_loop(nir_loop *loop);
 205    void nir_emit_block(nir_block *block);
 206    void nir_emit_instr(nir_instr *instr);
 207    void nir_emit_alu(const brw::fs_builder &bld, nir_alu_instr *instr);
 208    void nir_emit_load_const(const brw::fs_builder &bld,
 209                             nir_load_const_instr *instr);
 210    void nir_emit_vs_intrinsic(const brw::fs_builder &bld,
 211                               nir_intrinsic_instr *instr);
 212    void nir_emit_tcs_intrinsic(const brw::fs_builder &bld,
 213                                nir_intrinsic_instr *instr);
 214    void nir_emit_gs_intrinsic(const brw::fs_builder &bld,
 215                               nir_intrinsic_instr *instr);
 216    void nir_emit_fs_intrinsic(const brw::fs_builder &bld,
 217                               nir_intrinsic_instr *instr);
 218    void nir_emit_cs_intrinsic(const brw::fs_builder &bld,
 219                               nir_intrinsic_instr *instr);
 220    fs_reg get_nir_image_intrinsic_image(const brw::fs_builder &bld,
 221                                         nir_intrinsic_instr *instr);
 222    fs_reg get_nir_ssbo_intrinsic_index(const brw::fs_builder &bld,
 223                                        nir_intrinsic_instr *instr);
 224    void nir_emit_intrinsic(const brw::fs_builder &bld,
 225                            nir_intrinsic_instr *instr);
 226    void nir_emit_tes_intrinsic(const brw::fs_builder &bld,
 227                                nir_intrinsic_instr *instr);
 228    void nir_emit_ssbo_atomic(const brw::fs_builder &bld,
 229                              int op, nir_intrinsic_instr *instr);
 230    void nir_emit_ssbo_atomic_float(const brw::fs_builder &bld,
 231                                    int op, nir_intrinsic_instr *instr);
 232    void nir_emit_shared_atomic(const brw::fs_builder &bld,
 233                                int op, nir_intrinsic_instr *instr);
 234    void nir_emit_shared_atomic_float(const brw::fs_builder &bld,
 235                                      int op, nir_intrinsic_instr *instr);
 236    void nir_emit_texture(const brw::fs_builder &bld,
 237                          nir_tex_instr *instr);
 238    void nir_emit_jump(const brw::fs_builder &bld,
 239                       nir_jump_instr *instr);
 240    fs_reg get_nir_src(const nir_src &src);
 241    fs_reg get_nir_src_imm(const nir_src &src);
 242    fs_reg get_nir_dest(const nir_dest &dest);
 243    fs_reg get_indirect_offset(nir_intrinsic_instr *instr);
 244    void emit_percomp(const brw::fs_builder &bld, const fs_inst &inst,
 245                      unsigned wr_mask);
 246
 247    bool optimize_extract_to_float(nir_alu_instr *instr,
 248                                   const fs_reg &result);
 249    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 250                                      const fs_reg &result);
 251
 252    void emit_alpha_test();
 253    fs_inst *emit_single_fb_write(const brw::fs_builder &bld,
 254                                  fs_reg color1, fs_reg color2,
 255                                  fs_reg src0_alpha, unsigned components);
 256    void emit_fb_writes();
 257    fs_inst *emit_non_coherent_fb_read(const brw::fs_builder &bld,
 258                                       const fs_reg &dst, unsigned target);
 259    void emit_urb_writes(const fs_reg &gs_vertex_count = fs_reg());
 260    void set_gs_stream_control_data_bits(const fs_reg &vertex_count,
 261                                         unsigned stream_id);
 262    void emit_gs_control_data_bits(const fs_reg &vertex_count);
 263    void emit_gs_end_primitive(const nir_src &vertex_count_nir_src);
 264    void emit_gs_vertex(const nir_src &vertex_count_nir_src,
 265                        unsigned stream_id);
 266    void emit_gs_thread_end();
 267    void emit_gs_input_load(const fs_reg &dst, const nir_src &vertex_src,
 268                            unsigned base_offset, const nir_src &offset_src,
 269                            unsigned num_components, unsigned first_component);
 270    void emit_cs_terminate();
 271    fs_reg *emit_cs_work_group_id_setup();
 272
 273    void emit_barrier();
 274
 275    void emit_shader_time_begin();
 276    void emit_shader_time_end();
 277    void SHADER_TIME_ADD(const brw::fs_builder &bld,
 278                         int shader_time_subindex,
 279                         fs_reg value);
 280
 281    fs_reg get_timestamp(const brw::fs_builder &bld);
 282
 283    fs_reg interp_reg(int location, int channel);
 284
 285    int implied_mrf_writes(fs_inst *inst) const;
 286
 287    virtual void dump_instructions();
 288    virtual void dump_instructions(const char *name);
 289    void dump_instruction(backend_instruction *inst);
 290    void dump_instruction(backend_instruction *inst, FILE *file);
 291
 292    const void *const key;
 293    const struct brw_sampler_prog_key_data *key_tex;
 294
 295    struct brw_gs_compile *gs_compile;
 296
 297    struct brw_stage_prog_data *prog_data;
 298    struct gl_program *prog;
 299
 300    const struct brw_vue_map *input_vue_map;
 301
 302    int *virtual_grf_start;
 303    int *virtual_grf_end;
 304    brw::fs_live_variables *live_intervals;
 305
 306    int *regs_live_at_ip;
 307
 308    /** Number of uniform variable components visited. */
 309    unsigned uniforms;
 310
 311    /** Byte-offset for the next available spot in the scratch space buffer. */
 312    unsigned last_scratch;
 313
 314    /**
 315     * Array mapping UNIFORM register numbers to the pull parameter index,
 316     * or -1 if this uniform register isn't being uploaded as a pull constant.
 317     */
 318    int *pull_constant_loc;
 319
 320    /**
 321     * Array mapping UNIFORM register numbers to the push parameter index,
 322     * or -1 if this uniform register isn't being uploaded as a push constant.
 323     */
 324    int *push_constant_loc;
 325
 326    fs_reg subgroup_id;
 327    fs_reg frag_depth;
 328    fs_reg frag_stencil;
 329    fs_reg sample_mask;
 330    fs_reg outputs[VARYING_SLOT_MAX];
 331    fs_reg dual_src_output;
 332    int first_non_payload_grf;
 333    /** Either BRW_MAX_GRF or GEN7_MRF_HACK_START */
 334    unsigned max_grf;
 335
 336    fs_reg *nir_locals;
 337    fs_reg *nir_ssa_values;
 338    fs_reg *nir_system_values;
 339
 340    bool failed;
 341    char *fail_msg;
 342
 343    /** Register numbers for thread payload fields. */
 344    struct thread_payload {
 345       uint8_t subspan_coord_reg[2];
 346       uint8_t source_depth_reg[2];
 347       uint8_t source_w_reg[2];
 348       uint8_t aa_dest_stencil_reg[2];
 349       uint8_t dest_depth_reg[2];
 350       uint8_t sample_pos_reg[2];
 351       uint8_t sample_mask_in_reg[2];
 352       uint8_t barycentric_coord_reg[BRW_BARYCENTRIC_MODE_COUNT][2];
 353       uint8_t local_invocation_id_reg[2];
 354
 355       /** The number of thread payload registers the hardware will supply. */
 356       uint8_t num_regs;
 357    } payload;
 358
 359    bool source_depth_to_render_target;
 360    bool runtime_check_aads_emit;
 361
 362    fs_reg pixel_x;
 363    fs_reg pixel_y;
 364    fs_reg wpos_w;
 365    fs_reg pixel_w;
 366    fs_reg delta_xy[BRW_BARYCENTRIC_MODE_COUNT];
 367    fs_reg shader_start_time;
 368    fs_reg userplane[MAX_CLIP_PLANES];
 369    fs_reg final_gs_vertex_count;
 370    fs_reg control_data_bits;
 371    fs_reg invocation_id;
 372
 373    unsigned grf_used;
 374    bool spilled_any_registers;
 375
 376    const unsigned dispatch_width; /**< 8, 16 or 32 */
 377    unsigned max_dispatch_width;
 378
 379    int shader_time_index;
 380
 381    unsigned promoted_constants;
 382    brw::fs_builder bld;
 383 };
 384
 385 /**
 386  * The fragment shader code generator.
 387  *
 388  * Translates FS IR to actual i965 assembly code.
 389  */
 390 class fs_generator
 391 {
 392 public:
 393    fs_generator(const struct brw_compiler *compiler, void *log_data,
 394                 void *mem_ctx,
 395                 struct brw_stage_prog_data *prog_data,
 396                 unsigned promoted_constants,
 397                 bool runtime_check_aads_emit,
 398                 gl_shader_stage stage);
 399    ~fs_generator();
 400
 401    void enable_debug(const char *shader_name);
 402    int generate_code(const cfg_t *cfg, int dispatch_width);
 403    const unsigned *get_assembly();
 404
 405 private:
 406    void fire_fb_write(fs_inst *inst,
 407                       struct brw_reg payload,
 408                       struct brw_reg implied_header,
 409                       GLuint nr);
 410    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 411    void generate_fb_read(fs_inst *inst, struct brw_reg dst,
 412                          struct brw_reg payload);
 413    void generate_urb_read(fs_inst *inst, struct brw_reg dst, struct brw_reg payload);
 414    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 415    void generate_cs_terminate(fs_inst *inst, struct brw_reg payload);
 416    void generate_barrier(fs_inst *inst, struct brw_reg src);
 417    bool generate_linterp(fs_inst *inst, struct brw_reg dst,
 418                          struct brw_reg *src);
 419    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
 420                      struct brw_reg surface_index,
 421                      struct brw_reg sampler_index);
 422    void generate_get_buffer_size(fs_inst *inst, struct brw_reg dst,
 423                                  struct brw_reg src,
 424                                  struct brw_reg surf_index);
 425    void generate_ddx(const fs_inst *inst,
 426                      struct brw_reg dst, struct brw_reg src);
 427    void generate_ddy(const fs_inst *inst,
 428                      struct brw_reg dst, struct brw_reg src);
 429    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 430    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 431    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 432    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 433                                             struct brw_reg index,
 434                                             struct brw_reg offset);
 435    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 436                                                  struct brw_reg dst,
 437                                                  struct brw_reg surf_index,
 438                                                  struct brw_reg payload);
 439    void generate_varying_pull_constant_load_gen4(fs_inst *inst,
 440                                                  struct brw_reg dst,
 441                                                  struct brw_reg index);
 442    void generate_varying_pull_constant_load_gen7(fs_inst *inst,
 443                                                  struct brw_reg dst,
 444                                                  struct brw_reg index,
 445                                                  struct brw_reg offset);
 446    void generate_mov_dispatch_to_flags(fs_inst *inst);
 447
 448    void generate_pixel_interpolator_query(fs_inst *inst,
 449                                           struct brw_reg dst,
 450                                           struct brw_reg src,
 451                                           struct brw_reg msg_data,
 452                                           unsigned msg_type);
 453
 454    void generate_set_sample_id(fs_inst *inst,
 455                                struct brw_reg dst,
 456                                struct brw_reg src0,
 457                                struct brw_reg src1);
 458
 459    void generate_discard_jump(fs_inst *inst);
 460
 461    void generate_pack_half_2x16_split(fs_inst *inst,
 462                                       struct brw_reg dst,
 463                                       struct brw_reg x,
 464                                       struct brw_reg y);
 465    void generate_unpack_half_2x16_split(fs_inst *inst,
 466                                         struct brw_reg dst,
 467                                         struct brw_reg src);
 468
 469    void generate_shader_time_add(fs_inst *inst,
 470                                  struct brw_reg payload,
 471                                  struct brw_reg offset,
 472                                  struct brw_reg value);
 473
 474    void generate_mov_indirect(fs_inst *inst,
 475                               struct brw_reg dst,
 476                               struct brw_reg reg,
 477                               struct brw_reg indirect_byte_offset);
 478
 479    void generate_shuffle(fs_inst *inst,
 480                          struct brw_reg dst,
 481                          struct brw_reg src,
 482                          struct brw_reg idx);
 483
 484    void generate_quad_swizzle(const fs_inst *inst,
 485                               struct brw_reg dst, struct brw_reg src,
 486                               unsigned swiz);
 487
 488    bool patch_discard_jumps_to_fb_writes();
 489
 490    const struct brw_compiler *compiler;
 491    void *log_data; /* Passed to compiler->*_log functions */
 492
 493    const struct gen_device_info *devinfo;
 494
 495    struct brw_codegen *p;
 496    struct brw_stage_prog_data * const prog_data;
 497
 498    unsigned dispatch_width; /**< 8, 16 or 32 */
 499
 500    exec_list discard_halt_patches;
 501    unsigned promoted_constants;
 502    bool runtime_check_aads_emit;
 503    bool debug_flag;
 504    const char *shader_name;
 505    gl_shader_stage stage;
 506    void *mem_ctx;
 507 };
 508
 509 namespace brw {
 510    inline fs_reg
 511    fetch_payload_reg(const brw::fs_builder &bld, uint8_t regs[2],
 512                      brw_reg_type type = BRW_REGISTER_TYPE_F, unsigned n = 1)
 513    {
 514       if (!regs[0])
 515          return fs_reg();
 516
 517       if (bld.dispatch_width() > 16) {
 518          const fs_reg tmp = bld.vgrf(type, n);
 519          const brw::fs_builder hbld = bld.exec_all().group(16, 0);
 520          const unsigned m = bld.dispatch_width() / hbld.dispatch_width();
 521          fs_reg *const components = new fs_reg[n * m];
 522
 523          for (unsigned c = 0; c < n; c++) {
 524             for (unsigned g = 0; g < m; g++) {
 525                components[c * m + g] =
 526                   offset(retype(brw_vec8_grf(regs[g], 0), type), hbld, c);
 527             }
 528          }
 529
 530          hbld.LOAD_PAYLOAD(tmp, components, n * m, 0);
 531
 532          delete[] components;
 533          return tmp;
 534
 535       } else {
 536          return fs_reg(retype(brw_vec8_grf(regs[0], 0), type));
 537       }
 538    }
 539
 540    bool
 541    lower_src_modifiers(fs_visitor *v, bblock_t *block, fs_inst *inst, unsigned i);
 542 }
 543
 544 void shuffle_from_32bit_read(const brw::fs_builder &bld,
 545                              const fs_reg &dst,
 546                              const fs_reg &src,
 547                              uint32_t first_component,
 548                              uint32_t components);
 549
 550 fs_reg shuffle_for_32bit_write(const brw::fs_builder &bld,
 551                                const fs_reg &src,
 552                                uint32_t first_component,
 553                                uint32_t components);
 554
 555 fs_reg setup_imm_df(const brw::fs_builder &bld,
 556                     double v);
 557
 558 fs_reg setup_imm_b(const brw::fs_builder &bld,
 559                    int8_t v);
 560
 561 fs_reg setup_imm_ub(const brw::fs_builder &bld,
 562                    uint8_t v);
 563
 564 enum brw_barycentric_mode brw_barycentric_mode(enum glsl_interp_mode mode,
 565                                                nir_intrinsic_op op);
 566
 567 #endif /* BRW_FS_H */