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    fs_inst *get_instruction_generating_reg(fs_inst *start,
  87                                            fs_inst *end,
  88                                            const fs_reg &reg);
  89
  90    void VARYING_PULL_CONSTANT_LOAD(const brw::fs_builder &bld,
  91                                    const fs_reg &dst,
  92                                    const fs_reg &surf_index,
  93                                    const fs_reg &varying_offset,
  94                                    uint32_t const_offset);
  95    void DEP_RESOLVE_MOV(const brw::fs_builder &bld, int grf);
  96
  97    bool run_fs(bool allow_spilling, bool do_rep_send);
  98    bool run_vs();
  99    bool run_tcs_single_patch();
 100    bool run_tes();
 101    bool run_gs();
 102    bool run_cs(unsigned min_dispatch_width);
 103    void optimize();
 104    void allocate_registers(unsigned min_dispatch_width, bool allow_spilling);
 105    void setup_fs_payload_gen4();
 106    void setup_fs_payload_gen6();
 107    void setup_vs_payload();
 108    void setup_gs_payload();
 109    void setup_cs_payload();
 110    void fixup_3src_null_dest();
 111    void assign_curb_setup();
 112    void calculate_urb_setup();
 113    void assign_urb_setup();
 114    void convert_attr_sources_to_hw_regs(fs_inst *inst);
 115    void assign_vs_urb_setup();
 116    void assign_tcs_single_patch_urb_setup();
 117    void assign_tes_urb_setup();
 118    void assign_gs_urb_setup();
 119    bool assign_regs(bool allow_spilling, bool spill_all);
 120    void assign_regs_trivial();
 121    void calculate_payload_ranges(int payload_node_count,
 122                                  int *payload_last_use_ip);
 123    void setup_payload_interference(struct ra_graph *g, int payload_reg_count,
 124                                    int first_payload_node);
 125    int choose_spill_reg(struct ra_graph *g);
 126    void spill_reg(int spill_reg);
 127    void split_virtual_grfs();
 128    bool compact_virtual_grfs();
 129    void assign_constant_locations();
 130    bool get_pull_locs(const fs_reg &src, unsigned *out_surf_index,
 131                       unsigned *out_pull_index);
 132    void lower_constant_loads();
 133    void invalidate_live_intervals();
 134    void calculate_live_intervals();
 135    void calculate_register_pressure();
 136    void validate();
 137    bool opt_algebraic();
 138    bool opt_redundant_discard_jumps();
 139    bool opt_cse();
 140    bool opt_cse_local(bblock_t *block);
 141    bool opt_copy_propagation();
 142    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 143    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 144    bool opt_copy_propagation_local(void *mem_ctx, bblock_t *block,
 145                                    exec_list *acp);
 146    bool opt_drop_redundant_mov_to_flags();
 147    bool opt_register_renaming();
 148    bool register_coalesce();
 149    bool compute_to_mrf();
 150    bool eliminate_find_live_channel();
 151    bool dead_code_eliminate();
 152    bool remove_duplicate_mrf_writes();
 153    bool remove_extra_rounding_modes();
 154
 155    bool opt_sampler_eot();
 156    bool virtual_grf_interferes(int a, int b);
 157    void schedule_instructions(instruction_scheduler_mode mode);
 158    void insert_gen4_send_dependency_workarounds();
 159    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 160                                                     fs_inst *inst);
 161    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 162                                                      fs_inst *inst);
 163    void vfail(const char *msg, va_list args);
 164    void fail(const char *msg, ...);
 165    void limit_dispatch_width(unsigned n, const char *msg);
 166    void lower_uniform_pull_constant_loads();
 167    bool lower_load_payload();
 168    bool lower_pack();
 169    bool lower_conversions();
 170    bool lower_logical_sends();
 171    bool lower_integer_multiplication();
 172    bool lower_minmax();
 173    bool lower_simd_width();
 174    bool opt_combine_constants();
 175
 176    void emit_dummy_fs();
 177    void emit_repclear_shader();
 178    void emit_fragcoord_interpolation(fs_reg wpos);
 179    fs_reg *emit_frontfacing_interpolation();
 180    fs_reg *emit_samplepos_setup();
 181    fs_reg *emit_sampleid_setup();
 182    fs_reg *emit_samplemaskin_setup();
 183    void emit_interpolation_setup_gen4();
 184    void emit_interpolation_setup_gen6();
 185    void compute_sample_position(fs_reg dst, fs_reg int_sample_pos);
 186    fs_reg emit_mcs_fetch(const fs_reg &coordinate, unsigned components,
 187                          const fs_reg &sampler);
 188    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 189    fs_reg resolve_source_modifiers(const fs_reg &src);
 190    void emit_discard_jump();
 191    bool opt_peephole_sel();
 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    void nir_emit_intrinsic(const brw::fs_builder &bld,
 221                            nir_intrinsic_instr *instr);
 222    void nir_emit_tes_intrinsic(const brw::fs_builder &bld,
 223                                nir_intrinsic_instr *instr);
 224    void nir_emit_ssbo_atomic(const brw::fs_builder &bld,
 225                              int op, nir_intrinsic_instr *instr);
 226    void nir_emit_shared_atomic(const brw::fs_builder &bld,
 227                                int op, nir_intrinsic_instr *instr);
 228    void nir_emit_texture(const brw::fs_builder &bld,
 229                          nir_tex_instr *instr);
 230    void nir_emit_jump(const brw::fs_builder &bld,
 231                       nir_jump_instr *instr);
 232    fs_reg get_nir_src(const nir_src &src);
 233    fs_reg get_nir_src_imm(const nir_src &src);
 234    fs_reg get_nir_dest(const nir_dest &dest);
 235    fs_reg get_nir_image_deref(const nir_deref_var *deref);
 236    fs_reg get_indirect_offset(nir_intrinsic_instr *instr);
 237    void emit_percomp(const brw::fs_builder &bld, const fs_inst &inst,
 238                      unsigned wr_mask);
 239
 240    bool optimize_extract_to_float(nir_alu_instr *instr,
 241                                   const fs_reg &result);
 242    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 243                                      const fs_reg &result);
 244
 245    void emit_alpha_test();
 246    fs_inst *emit_single_fb_write(const brw::fs_builder &bld,
 247                                  fs_reg color1, fs_reg color2,
 248                                  fs_reg src0_alpha, unsigned components);
 249    void emit_fb_writes();
 250    fs_inst *emit_non_coherent_fb_read(const brw::fs_builder &bld,
 251                                       const fs_reg &dst, unsigned target);
 252    void emit_urb_writes(const fs_reg &gs_vertex_count = fs_reg());
 253    void set_gs_stream_control_data_bits(const fs_reg &vertex_count,
 254                                         unsigned stream_id);
 255    void emit_gs_control_data_bits(const fs_reg &vertex_count);
 256    void emit_gs_end_primitive(const nir_src &vertex_count_nir_src);
 257    void emit_gs_vertex(const nir_src &vertex_count_nir_src,
 258                        unsigned stream_id);
 259    void emit_gs_thread_end();
 260    void emit_gs_input_load(const fs_reg &dst, const nir_src &vertex_src,
 261                            unsigned base_offset, const nir_src &offset_src,
 262                            unsigned num_components, unsigned first_component);
 263    void emit_cs_terminate();
 264    fs_reg *emit_cs_work_group_id_setup();
 265
 266    void emit_barrier();
 267
 268    void emit_shader_time_begin();
 269    void emit_shader_time_end();
 270    void SHADER_TIME_ADD(const brw::fs_builder &bld,
 271                         int shader_time_subindex,
 272                         fs_reg value);
 273
 274    fs_reg get_timestamp(const brw::fs_builder &bld);
 275
 276    struct brw_reg interp_reg(int location, int channel);
 277
 278    int implied_mrf_writes(fs_inst *inst);
 279
 280    virtual void dump_instructions();
 281    virtual void dump_instructions(const char *name);
 282    void dump_instruction(backend_instruction *inst);
 283    void dump_instruction(backend_instruction *inst, FILE *file);
 284
 285    const void *const key;
 286    const struct brw_sampler_prog_key_data *key_tex;
 287
 288    struct brw_gs_compile *gs_compile;
 289
 290    struct brw_stage_prog_data *prog_data;
 291    struct gl_program *prog;
 292
 293    const struct brw_vue_map *input_vue_map;
 294
 295    int *virtual_grf_start;
 296    int *virtual_grf_end;
 297    brw::fs_live_variables *live_intervals;
 298
 299    int *regs_live_at_ip;
 300
 301    /** Number of uniform variable components visited. */
 302    unsigned uniforms;
 303
 304    /** Byte-offset for the next available spot in the scratch space buffer. */
 305    unsigned last_scratch;
 306
 307    /**
 308     * Array mapping UNIFORM register numbers to the pull parameter index,
 309     * or -1 if this uniform register isn't being uploaded as a pull constant.
 310     */
 311    int *pull_constant_loc;
 312
 313    /**
 314     * Array mapping UNIFORM register numbers to the push parameter index,
 315     * or -1 if this uniform register isn't being uploaded as a push constant.
 316     */
 317    int *push_constant_loc;
 318
 319    fs_reg subgroup_id;
 320    fs_reg frag_depth;
 321    fs_reg frag_stencil;
 322    fs_reg sample_mask;
 323    fs_reg outputs[VARYING_SLOT_MAX];
 324    fs_reg dual_src_output;
 325    int first_non_payload_grf;
 326    /** Either BRW_MAX_GRF or GEN7_MRF_HACK_START */
 327    unsigned max_grf;
 328
 329    fs_reg *nir_locals;
 330    fs_reg *nir_ssa_values;
 331    fs_reg *nir_system_values;
 332
 333    bool failed;
 334    char *fail_msg;
 335
 336    /** Register numbers for thread payload fields. */
 337    struct thread_payload {
 338       uint8_t source_depth_reg;
 339       uint8_t source_w_reg;
 340       uint8_t aa_dest_stencil_reg;
 341       uint8_t dest_depth_reg;
 342       uint8_t sample_pos_reg;
 343       uint8_t sample_mask_in_reg;
 344       uint8_t barycentric_coord_reg[BRW_BARYCENTRIC_MODE_COUNT];
 345       uint8_t local_invocation_id_reg;
 346
 347       /** The number of thread payload registers the hardware will supply. */
 348       uint8_t num_regs;
 349    } payload;
 350
 351    bool source_depth_to_render_target;
 352    bool runtime_check_aads_emit;
 353
 354    fs_reg pixel_x;
 355    fs_reg pixel_y;
 356    fs_reg wpos_w;
 357    fs_reg pixel_w;
 358    fs_reg delta_xy[BRW_BARYCENTRIC_MODE_COUNT];
 359    fs_reg shader_start_time;
 360    fs_reg userplane[MAX_CLIP_PLANES];
 361    fs_reg final_gs_vertex_count;
 362    fs_reg control_data_bits;
 363    fs_reg invocation_id;
 364
 365    unsigned grf_used;
 366    bool spilled_any_registers;
 367
 368    const unsigned dispatch_width; /**< 8, 16 or 32 */
 369    unsigned max_dispatch_width;
 370
 371    int shader_time_index;
 372
 373    unsigned promoted_constants;
 374    brw::fs_builder bld;
 375 };
 376
 377 /**
 378  * The fragment shader code generator.
 379  *
 380  * Translates FS IR to actual i965 assembly code.
 381  */
 382 class fs_generator
 383 {
 384 public:
 385    fs_generator(const struct brw_compiler *compiler, void *log_data,
 386                 void *mem_ctx,
 387                 const void *key,
 388                 struct brw_stage_prog_data *prog_data,
 389                 unsigned promoted_constants,
 390                 bool runtime_check_aads_emit,
 391                 gl_shader_stage stage);
 392    ~fs_generator();
 393
 394    void enable_debug(const char *shader_name);
 395    int generate_code(const cfg_t *cfg, int dispatch_width);
 396    const unsigned *get_assembly(unsigned int *assembly_size);
 397
 398 private:
 399    void fire_fb_write(fs_inst *inst,
 400                       struct brw_reg payload,
 401                       struct brw_reg implied_header,
 402                       GLuint nr);
 403    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 404    void generate_fb_read(fs_inst *inst, struct brw_reg dst,
 405                          struct brw_reg payload);
 406    void generate_urb_read(fs_inst *inst, struct brw_reg dst, struct brw_reg payload);
 407    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 408    void generate_cs_terminate(fs_inst *inst, struct brw_reg payload);
 409    void generate_barrier(fs_inst *inst, struct brw_reg src);
 410    void generate_linterp(fs_inst *inst, struct brw_reg dst,
 411                          struct brw_reg *src);
 412    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
 413                      struct brw_reg surface_index,
 414                      struct brw_reg sampler_index);
 415    void generate_get_buffer_size(fs_inst *inst, struct brw_reg dst,
 416                                  struct brw_reg src,
 417                                  struct brw_reg surf_index);
 418    void generate_ddx(enum opcode op, struct brw_reg dst, struct brw_reg src);
 419    void generate_ddy(enum opcode op, struct brw_reg dst, struct brw_reg src);
 420    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 421    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 422    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 423    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 424                                             struct brw_reg index,
 425                                             struct brw_reg offset);
 426    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 427                                                  struct brw_reg dst,
 428                                                  struct brw_reg surf_index,
 429                                                  struct brw_reg payload);
 430    void generate_varying_pull_constant_load_gen4(fs_inst *inst,
 431                                                  struct brw_reg dst,
 432                                                  struct brw_reg index);
 433    void generate_varying_pull_constant_load_gen7(fs_inst *inst,
 434                                                  struct brw_reg dst,
 435                                                  struct brw_reg index,
 436                                                  struct brw_reg offset);
 437    void generate_mov_dispatch_to_flags(fs_inst *inst);
 438
 439    void generate_pixel_interpolator_query(fs_inst *inst,
 440                                           struct brw_reg dst,
 441                                           struct brw_reg src,
 442                                           struct brw_reg msg_data,
 443                                           unsigned msg_type);
 444
 445    void generate_set_sample_id(fs_inst *inst,
 446                                struct brw_reg dst,
 447                                struct brw_reg src0,
 448                                struct brw_reg src1);
 449
 450    void generate_discard_jump(fs_inst *inst);
 451
 452    void generate_pack_half_2x16_split(fs_inst *inst,
 453                                       struct brw_reg dst,
 454                                       struct brw_reg x,
 455                                       struct brw_reg y);
 456    void generate_unpack_half_2x16_split(fs_inst *inst,
 457                                         struct brw_reg dst,
 458                                         struct brw_reg src);
 459
 460    void generate_shader_time_add(fs_inst *inst,
 461                                  struct brw_reg payload,
 462                                  struct brw_reg offset,
 463                                  struct brw_reg value);
 464
 465    void generate_mov_indirect(fs_inst *inst,
 466                               struct brw_reg dst,
 467                               struct brw_reg reg,
 468                               struct brw_reg indirect_byte_offset);
 469
 470    bool patch_discard_jumps_to_fb_writes();
 471
 472    const struct brw_compiler *compiler;
 473    void *log_data; /* Passed to compiler->*_log functions */
 474
 475    const struct gen_device_info *devinfo;
 476
 477    struct brw_codegen *p;
 478    const void * const key;
 479    struct brw_stage_prog_data * const prog_data;
 480
 481    unsigned dispatch_width; /**< 8, 16 or 32 */
 482
 483    exec_list discard_halt_patches;
 484    unsigned promoted_constants;
 485    bool runtime_check_aads_emit;
 486    bool debug_flag;
 487    const char *shader_name;
 488    gl_shader_stage stage;
 489    void *mem_ctx;
 490 };
 491
 492 void shuffle_32bit_load_result_to_64bit_data(const brw::fs_builder &bld,
 493                                              const fs_reg &dst,
 494                                              const fs_reg &src,
 495                                              uint32_t components);
 496
 497 fs_reg shuffle_64bit_data_for_32bit_write(const brw::fs_builder &bld,
 498                                           const fs_reg &src,
 499                                           uint32_t components);
 500 fs_reg setup_imm_df(const brw::fs_builder &bld,
 501                     double v);
 502
 503 enum brw_barycentric_mode brw_barycentric_mode(enum glsl_interp_mode mode,
 504                                                nir_intrinsic_op op);
 505
 506 #endif /* BRW_FS_H */