src/mesa/drivers/dri/i965/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 #pragma once
  29
  30 #include "brw_shader.h"
  31 #include "brw_ir_fs.h"
  32 #include "brw_fs_builder.h"
  33 #include "glsl/ir.h"
  34 #include "glsl/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(fs_reg reg, const brw::fs_builder& bld, unsigned delta)
  49 {
  50    switch (reg.file) {
  51    case BAD_FILE:
  52       break;
  53    case ARF:
  54    case FIXED_GRF:
  55    case MRF:
  56    case VGRF:
  57    case ATTR:
  58       return byte_offset(reg,
  59                          delta * reg.component_size(bld.dispatch_width()));
  60    case UNIFORM:
  61       reg.reg_offset += delta;
  62       break;
  63    case IMM:
  64       assert(delta == 0);
  65    }
  66    return reg;
  67 }
  68
  69 /**
  70  * The fragment shader front-end.
  71  *
  72  * Translates either GLSL IR or Mesa IR (for ARB_fragment_program) into FS IR.
  73  */
  74 class fs_visitor : public backend_shader
  75 {
  76 public:
  77    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  78               void *mem_ctx,
  79               const void *key,
  80               struct brw_stage_prog_data *prog_data,
  81               struct gl_program *prog,
  82               const nir_shader *shader,
  83               unsigned dispatch_width,
  84               int shader_time_index);
  85    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  86               void *mem_ctx,
  87               struct brw_gs_compile *gs_compile,
  88               struct brw_gs_prog_data *prog_data,
  89               const nir_shader *shader,
  90               int shader_time_index);
  91    void init();
  92    ~fs_visitor();
  93
  94    fs_reg vgrf(const glsl_type *const type);
  95    void import_uniforms(fs_visitor *v);
  96    void setup_uniform_clipplane_values(gl_clip_plane *clip_planes);
  97    void compute_clip_distance(gl_clip_plane *clip_planes);
  98
  99    fs_inst *get_instruction_generating_reg(fs_inst *start,
 100                                            fs_inst *end,
 101                                            const fs_reg &reg);
 102
 103    void VARYING_PULL_CONSTANT_LOAD(const brw::fs_builder &bld,
 104                                    const fs_reg &dst,
 105                                    const fs_reg &surf_index,
 106                                    const fs_reg &varying_offset,
 107                                    uint32_t const_offset);
 108    void DEP_RESOLVE_MOV(const brw::fs_builder &bld, int grf);
 109
 110    bool run_fs(bool do_rep_send);
 111    bool run_vs(gl_clip_plane *clip_planes);
 112    bool run_gs();
 113    bool run_cs();
 114    void optimize();
 115    void allocate_registers();
 116    void setup_payload_gen4();
 117    void setup_payload_gen6();
 118    void setup_vs_payload();
 119    void setup_gs_payload();
 120    void setup_cs_payload();
 121    void fixup_3src_null_dest();
 122    void assign_curb_setup();
 123    void calculate_urb_setup();
 124    void assign_urb_setup();
 125    void convert_attr_sources_to_hw_regs(fs_inst *inst);
 126    void assign_vs_urb_setup();
 127    void assign_gs_urb_setup();
 128    bool assign_regs(bool allow_spilling);
 129    void assign_regs_trivial();
 130    void calculate_payload_ranges(int payload_node_count,
 131                                  int *payload_last_use_ip);
 132    void setup_payload_interference(struct ra_graph *g, int payload_reg_count,
 133                                    int first_payload_node);
 134    int choose_spill_reg(struct ra_graph *g);
 135    void spill_reg(int spill_reg);
 136    void split_virtual_grfs();
 137    bool compact_virtual_grfs();
 138    void assign_constant_locations();
 139    void lower_constant_loads();
 140    void invalidate_live_intervals();
 141    void calculate_live_intervals();
 142    void calculate_register_pressure();
 143    void validate();
 144    bool opt_algebraic();
 145    bool opt_redundant_discard_jumps();
 146    bool opt_cse();
 147    bool opt_cse_local(bblock_t *block);
 148    bool opt_copy_propagate();
 149    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 150    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 151    bool opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
 152                                  exec_list *acp);
 153    bool opt_register_renaming();
 154    bool register_coalesce();
 155    bool compute_to_mrf();
 156    bool eliminate_find_live_channel();
 157    bool dead_code_eliminate();
 158    bool remove_duplicate_mrf_writes();
 159
 160    bool opt_sampler_eot();
 161    bool virtual_grf_interferes(int a, int b);
 162    void schedule_instructions(instruction_scheduler_mode mode);
 163    void insert_gen4_send_dependency_workarounds();
 164    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 165                                                     fs_inst *inst);
 166    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 167                                                      fs_inst *inst);
 168    void vfail(const char *msg, va_list args);
 169    void fail(const char *msg, ...);
 170    void no16(const char *msg);
 171    void lower_uniform_pull_constant_loads();
 172    bool lower_load_payload();
 173    bool lower_logical_sends();
 174    bool lower_integer_multiplication();
 175    bool lower_simd_width();
 176    bool opt_combine_constants();
 177
 178    void emit_dummy_fs();
 179    void emit_repclear_shader();
 180    fs_reg *emit_fragcoord_interpolation(bool pixel_center_integer,
 181                                         bool origin_upper_left);
 182    fs_inst *emit_linterp(const fs_reg &attr, const fs_reg &interp,
 183                          glsl_interp_qualifier interpolation_mode,
 184                          bool is_centroid, bool is_sample);
 185    fs_reg *emit_frontfacing_interpolation();
 186    fs_reg *emit_samplepos_setup();
 187    fs_reg *emit_sampleid_setup();
 188    void emit_general_interpolation(fs_reg *attr, const char *name,
 189                                    const glsl_type *type,
 190                                    glsl_interp_qualifier interpolation_mode,
 191                                    int *location, bool mod_centroid,
 192                                    bool mod_sample);
 193    fs_reg *emit_vs_system_value(int location);
 194    void emit_interpolation_setup_gen4();
 195    void emit_interpolation_setup_gen6();
 196    void compute_sample_position(fs_reg dst, fs_reg int_sample_pos);
 197    void emit_texture(ir_texture_opcode op,
 198                      const glsl_type *dest_type,
 199                      fs_reg coordinate, int components,
 200                      fs_reg shadow_c,
 201                      fs_reg lod, fs_reg dpdy, int grad_components,
 202                      fs_reg sample_index,
 203                      fs_reg offset,
 204                      fs_reg mcs,
 205                      int gather_component,
 206                      bool is_cube_array,
 207                      uint32_t sampler,
 208                      fs_reg sampler_reg);
 209    fs_reg emit_mcs_fetch(const fs_reg &coordinate, unsigned components,
 210                          const fs_reg &sampler);
 211    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 212    fs_reg resolve_source_modifiers(const fs_reg &src);
 213    void emit_discard_jump();
 214    bool opt_peephole_sel();
 215    bool opt_peephole_predicated_break();
 216    bool opt_saturate_propagation();
 217    bool opt_cmod_propagation();
 218    bool opt_zero_samples();
 219    void emit_unspill(bblock_t *block, fs_inst *inst, fs_reg reg,
 220                      uint32_t spill_offset, int count);
 221    void emit_spill(bblock_t *block, fs_inst *inst, fs_reg reg,
 222                    uint32_t spill_offset, int count);
 223
 224    void emit_nir_code();
 225    void nir_setup_inputs();
 226    void nir_setup_single_output_varying(fs_reg *reg, const glsl_type *type,
 227                                         unsigned *location);
 228    void nir_setup_outputs();
 229    void nir_setup_uniforms();
 230    void nir_emit_system_values();
 231    void nir_emit_impl(nir_function_impl *impl);
 232    void nir_emit_cf_list(exec_list *list);
 233    void nir_emit_if(nir_if *if_stmt);
 234    void nir_emit_loop(nir_loop *loop);
 235    void nir_emit_block(nir_block *block);
 236    void nir_emit_instr(nir_instr *instr);
 237    void nir_emit_alu(const brw::fs_builder &bld, nir_alu_instr *instr);
 238    void nir_emit_load_const(const brw::fs_builder &bld,
 239                             nir_load_const_instr *instr);
 240    void nir_emit_undef(const brw::fs_builder &bld,
 241                        nir_ssa_undef_instr *instr);
 242    void nir_emit_vs_intrinsic(const brw::fs_builder &bld,
 243                               nir_intrinsic_instr *instr);
 244    void nir_emit_gs_intrinsic(const brw::fs_builder &bld,
 245                               nir_intrinsic_instr *instr);
 246    void nir_emit_fs_intrinsic(const brw::fs_builder &bld,
 247                               nir_intrinsic_instr *instr);
 248    void nir_emit_cs_intrinsic(const brw::fs_builder &bld,
 249                               nir_intrinsic_instr *instr);
 250    void nir_emit_intrinsic(const brw::fs_builder &bld,
 251                            nir_intrinsic_instr *instr);
 252    void nir_emit_ssbo_atomic(const brw::fs_builder &bld,
 253                              int op, nir_intrinsic_instr *instr);
 254    void nir_emit_shared_atomic(const brw::fs_builder &bld,
 255                                int op, nir_intrinsic_instr *instr);
 256    void nir_emit_texture(const brw::fs_builder &bld,
 257                          nir_tex_instr *instr);
 258    void nir_emit_jump(const brw::fs_builder &bld,
 259                       nir_jump_instr *instr);
 260    fs_reg get_nir_src(nir_src src);
 261    fs_reg get_nir_dest(nir_dest dest);
 262    fs_reg get_nir_image_deref(const nir_deref_var *deref);
 263    void emit_percomp(const brw::fs_builder &bld, const fs_inst &inst,
 264                      unsigned wr_mask);
 265
 266    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 267                                      const fs_reg &result);
 268
 269    void emit_alpha_test();
 270    fs_inst *emit_single_fb_write(const brw::fs_builder &bld,
 271                                  fs_reg color1, fs_reg color2,
 272                                  fs_reg src0_alpha, unsigned components);
 273    void emit_fb_writes();
 274    void emit_urb_writes(const fs_reg &gs_vertex_count = fs_reg());
 275    void set_gs_stream_control_data_bits(const fs_reg &vertex_count,
 276                                         unsigned stream_id);
 277    void emit_gs_control_data_bits(const fs_reg &vertex_count);
 278    void emit_gs_end_primitive(const nir_src &vertex_count_nir_src);
 279    void emit_gs_vertex(const nir_src &vertex_count_nir_src,
 280                        unsigned stream_id);
 281    void emit_gs_thread_end();
 282    void emit_gs_input_load(const fs_reg &dst, const nir_src &vertex_src,
 283                            unsigned base_offset, const nir_src &offset_src,
 284                            unsigned num_components);
 285    void emit_cs_terminate();
 286    fs_reg *emit_cs_local_invocation_id_setup();
 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    struct brw_reg interp_reg(int location, int channel);
 300
 301    int implied_mrf_writes(fs_inst *inst);
 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 void *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    struct gl_program *prog;
 315
 316    int *virtual_grf_start;
 317    int *virtual_grf_end;
 318    brw::fs_live_variables *live_intervals;
 319
 320    int *regs_live_at_ip;
 321
 322    /** Number of uniform variable components visited. */
 323    unsigned uniforms;
 324
 325    /** Byte-offset for the next available spot in the scratch space buffer. */
 326    unsigned last_scratch;
 327
 328    /**
 329     * Array mapping UNIFORM register numbers to the pull parameter index,
 330     * or -1 if this uniform register isn't being uploaded as a pull constant.
 331     */
 332    int *pull_constant_loc;
 333
 334    /**
 335     * Array mapping UNIFORM register numbers to the push parameter index,
 336     * or -1 if this uniform register isn't being uploaded as a push constant.
 337     */
 338    int *push_constant_loc;
 339
 340    fs_reg frag_depth;
 341    fs_reg frag_stencil;
 342    fs_reg sample_mask;
 343    fs_reg outputs[VARYING_SLOT_MAX];
 344    unsigned output_components[VARYING_SLOT_MAX];
 345    fs_reg dual_src_output;
 346    bool do_dual_src;
 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_inputs;
 354    fs_reg nir_outputs;
 355    fs_reg *nir_system_values;
 356
 357    bool failed;
 358    char *fail_msg;
 359    bool simd16_unsupported;
 360    char *no16_msg;
 361
 362    /* Result of last visit() method. Still used by emit_texture() */
 363    fs_reg result;
 364
 365    /** Register numbers for thread payload fields. */
 366    struct thread_payload {
 367       uint8_t source_depth_reg;
 368       uint8_t source_w_reg;
 369       uint8_t aa_dest_stencil_reg;
 370       uint8_t dest_depth_reg;
 371       uint8_t sample_pos_reg;
 372       uint8_t sample_mask_in_reg;
 373       uint8_t barycentric_coord_reg[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 374       uint8_t local_invocation_id_reg;
 375
 376       /** The number of thread payload registers the hardware will supply. */
 377       uint8_t num_regs;
 378    } payload;
 379
 380    bool source_depth_to_render_target;
 381    bool runtime_check_aads_emit;
 382
 383    fs_reg pixel_x;
 384    fs_reg pixel_y;
 385    fs_reg wpos_w;
 386    fs_reg pixel_w;
 387    fs_reg delta_xy[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 388    fs_reg shader_start_time;
 389    fs_reg userplane[MAX_CLIP_PLANES];
 390    fs_reg final_gs_vertex_count;
 391    fs_reg control_data_bits;
 392
 393    unsigned grf_used;
 394    bool spilled_any_registers;
 395
 396    const unsigned dispatch_width; /**< 8 or 16 */
 397
 398    int shader_time_index;
 399
 400    unsigned promoted_constants;
 401    brw::fs_builder bld;
 402 };
 403
 404 /**
 405  * The fragment shader code generator.
 406  *
 407  * Translates FS IR to actual i965 assembly code.
 408  */
 409 class fs_generator
 410 {
 411 public:
 412    fs_generator(const struct brw_compiler *compiler, void *log_data,
 413                 void *mem_ctx,
 414                 const void *key,
 415                 struct brw_stage_prog_data *prog_data,
 416                 unsigned promoted_constants,
 417                 bool runtime_check_aads_emit,
 418                 const char *stage_abbrev);
 419    ~fs_generator();
 420
 421    void enable_debug(const char *shader_name);
 422    int generate_code(const cfg_t *cfg, int dispatch_width);
 423    const unsigned *get_assembly(unsigned int *assembly_size);
 424
 425 private:
 426    void fire_fb_write(fs_inst *inst,
 427                       struct brw_reg payload,
 428                       struct brw_reg implied_header,
 429                       GLuint nr);
 430    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 431    void generate_urb_read(fs_inst *inst, struct brw_reg dst, struct brw_reg payload);
 432    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 433    void generate_cs_terminate(fs_inst *inst, struct brw_reg payload);
 434    void generate_stencil_ref_packing(fs_inst *inst, struct brw_reg dst,
 435                                      struct brw_reg src);
 436    void generate_barrier(fs_inst *inst, struct brw_reg src);
 437    void generate_blorp_fb_write(fs_inst *inst);
 438    void generate_linterp(fs_inst *inst, struct brw_reg dst,
 439                          struct brw_reg *src);
 440    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
 441                      struct brw_reg sampler_index);
 442    void generate_get_buffer_size(fs_inst *inst, struct brw_reg dst,
 443                                  struct brw_reg src,
 444                                  struct brw_reg surf_index);
 445    void generate_math_gen6(fs_inst *inst,
 446                            struct brw_reg dst,
 447                            struct brw_reg src0,
 448                            struct brw_reg src1);
 449    void generate_math_gen4(fs_inst *inst,
 450                            struct brw_reg dst,
 451                            struct brw_reg src);
 452    void generate_math_g45(fs_inst *inst,
 453                           struct brw_reg dst,
 454                           struct brw_reg src);
 455    void generate_ddx(enum opcode op, struct brw_reg dst, struct brw_reg src);
 456    void generate_ddy(enum opcode op, struct brw_reg dst, struct brw_reg src,
 457                      bool negate_value);
 458    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 459    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 460    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 461    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 462                                             struct brw_reg index,
 463                                             struct brw_reg offset);
 464    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 465                                                  struct brw_reg dst,
 466                                                  struct brw_reg surf_index,
 467                                                  struct brw_reg offset);
 468    void generate_varying_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 469                                             struct brw_reg index,
 470                                             struct brw_reg offset);
 471    void generate_varying_pull_constant_load_gen7(fs_inst *inst,
 472                                                  struct brw_reg dst,
 473                                                  struct brw_reg index,
 474                                                  struct brw_reg offset);
 475    void generate_mov_dispatch_to_flags(fs_inst *inst);
 476
 477    void generate_pixel_interpolator_query(fs_inst *inst,
 478                                           struct brw_reg dst,
 479                                           struct brw_reg src,
 480                                           struct brw_reg msg_data,
 481                                           unsigned msg_type);
 482
 483    void generate_set_sample_id(fs_inst *inst,
 484                                struct brw_reg dst,
 485                                struct brw_reg src0,
 486                                struct brw_reg src1);
 487
 488    void generate_set_simd4x2_offset(fs_inst *inst,
 489                                     struct brw_reg dst,
 490                                     struct brw_reg offset);
 491    void generate_discard_jump(fs_inst *inst);
 492
 493    void generate_pack_half_2x16_split(fs_inst *inst,
 494                                       struct brw_reg dst,
 495                                       struct brw_reg x,
 496                                       struct brw_reg y);
 497    void generate_unpack_half_2x16_split(fs_inst *inst,
 498                                         struct brw_reg dst,
 499                                         struct brw_reg src);
 500
 501    void generate_shader_time_add(fs_inst *inst,
 502                                  struct brw_reg payload,
 503                                  struct brw_reg offset,
 504                                  struct brw_reg value);
 505
 506    void generate_mov_indirect(fs_inst *inst,
 507                               struct brw_reg dst,
 508                               struct brw_reg reg,
 509                               struct brw_reg indirect_byte_offset);
 510
 511    bool patch_discard_jumps_to_fb_writes();
 512
 513    const struct brw_compiler *compiler;
 514    void *log_data; /* Passed to compiler->*_log functions */
 515
 516    const struct brw_device_info *devinfo;
 517
 518    struct brw_codegen *p;
 519    const void * const key;
 520    struct brw_stage_prog_data * const prog_data;
 521
 522    unsigned dispatch_width; /**< 8 or 16 */
 523
 524    exec_list discard_halt_patches;
 525    unsigned promoted_constants;
 526    bool runtime_check_aads_emit;
 527    bool debug_flag;
 528    const char *shader_name;
 529    const char *stage_abbrev;
 530    void *mem_ctx;
 531 };
 532
 533 bool brw_do_channel_expressions(struct exec_list *instructions);
 534 bool brw_do_vector_splitting(struct exec_list *instructions);