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
  34 extern "C" {
  35
  36 #include <sys/types.h>
  37
  38 #include "main/macros.h"
  39 #include "main/shaderobj.h"
  40 #include "main/uniforms.h"
  41 #include "program/prog_parameter.h"
  42 #include "program/prog_print.h"
  43 #include "program/prog_optimize.h"
  44 #include "util/register_allocate.h"
  45 #include "program/hash_table.h"
  46 #include "brw_context.h"
  47 #include "brw_eu.h"
  48 #include "brw_wm.h"
  49 #include "intel_asm_annotation.h"
  50 }
  51 #include "glsl/nir/glsl_types.h"
  52 #include "glsl/ir.h"
  53 #include "glsl/nir/nir.h"
  54 #include "program/sampler.h"
  55
  56 struct bblock_t;
  57 namespace {
  58    struct acp_entry;
  59 }
  60
  61 namespace brw {
  62    class fs_live_variables;
  63 }
  64
  65 struct brw_gs_compile;
  66
  67 static inline fs_reg
  68 offset(fs_reg reg, const brw::fs_builder& bld, unsigned delta)
  69 {
  70    switch (reg.file) {
  71    case BAD_FILE:
  72       break;
  73    case ARF:
  74    case FIXED_GRF:
  75    case MRF:
  76    case VGRF:
  77    case ATTR:
  78       return byte_offset(reg,
  79                          delta * reg.component_size(bld.dispatch_width()));
  80    case UNIFORM:
  81       reg.reg_offset += delta;
  82       break;
  83    case IMM:
  84       assert(delta == 0);
  85    }
  86    return reg;
  87 }
  88
  89 /**
  90  * The fragment shader front-end.
  91  *
  92  * Translates either GLSL IR or Mesa IR (for ARB_fragment_program) into FS IR.
  93  */
  94 class fs_visitor : public backend_shader
  95 {
  96 public:
  97    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  98               void *mem_ctx,
  99               const void *key,
 100               struct brw_stage_prog_data *prog_data,
 101               struct gl_program *prog,
 102               const nir_shader *shader,
 103               unsigned dispatch_width,
 104               int shader_time_index);
 105    fs_visitor(const struct brw_compiler *compiler, void *log_data,
 106               void *mem_ctx,
 107               struct brw_gs_compile *gs_compile,
 108               struct brw_gs_prog_data *prog_data,
 109               const nir_shader *shader,
 110               int shader_time_index);
 111    void init();
 112    ~fs_visitor();
 113
 114    fs_reg vgrf(const glsl_type *const type);
 115    void import_uniforms(fs_visitor *v);
 116    void setup_uniform_clipplane_values(gl_clip_plane *clip_planes);
 117    void compute_clip_distance(gl_clip_plane *clip_planes);
 118
 119    uint32_t gather_channel(int orig_chan, uint32_t sampler);
 120    void swizzle_result(ir_texture_opcode op, int dest_components,
 121                        fs_reg orig_val, uint32_t sampler);
 122
 123    fs_inst *get_instruction_generating_reg(fs_inst *start,
 124                                            fs_inst *end,
 125                                            const fs_reg &reg);
 126
 127    void VARYING_PULL_CONSTANT_LOAD(const brw::fs_builder &bld,
 128                                    const fs_reg &dst,
 129                                    const fs_reg &surf_index,
 130                                    const fs_reg &varying_offset,
 131                                    uint32_t const_offset);
 132    void DEP_RESOLVE_MOV(const brw::fs_builder &bld, int grf);
 133
 134    bool run_fs(bool do_rep_send);
 135    bool run_vs(gl_clip_plane *clip_planes);
 136    bool run_gs();
 137    bool run_cs();
 138    void optimize();
 139    void allocate_registers();
 140    void setup_payload_gen4();
 141    void setup_payload_gen6();
 142    void setup_vs_payload();
 143    void setup_gs_payload();
 144    void setup_cs_payload();
 145    void fixup_3src_null_dest();
 146    void assign_curb_setup();
 147    void calculate_urb_setup();
 148    void assign_urb_setup();
 149    void convert_attr_sources_to_hw_regs(fs_inst *inst);
 150    void assign_vs_urb_setup();
 151    void assign_gs_urb_setup();
 152    bool assign_regs(bool allow_spilling);
 153    void assign_regs_trivial();
 154    void calculate_payload_ranges(int payload_node_count,
 155                                  int *payload_last_use_ip);
 156    void setup_payload_interference(struct ra_graph *g, int payload_reg_count,
 157                                    int first_payload_node);
 158    int choose_spill_reg(struct ra_graph *g);
 159    void spill_reg(int spill_reg);
 160    void split_virtual_grfs();
 161    bool compact_virtual_grfs();
 162    void assign_constant_locations();
 163    void demote_pull_constants();
 164    void invalidate_live_intervals();
 165    void calculate_live_intervals();
 166    void calculate_register_pressure();
 167    void validate();
 168    bool opt_algebraic();
 169    bool opt_redundant_discard_jumps();
 170    bool opt_cse();
 171    bool opt_cse_local(bblock_t *block);
 172    bool opt_copy_propagate();
 173    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 174    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 175    bool opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
 176                                  exec_list *acp);
 177    bool opt_register_renaming();
 178    bool register_coalesce();
 179    bool compute_to_mrf();
 180    bool eliminate_find_live_channel();
 181    bool dead_code_eliminate();
 182    bool remove_duplicate_mrf_writes();
 183
 184    bool opt_sampler_eot();
 185    bool virtual_grf_interferes(int a, int b);
 186    void schedule_instructions(instruction_scheduler_mode mode);
 187    void insert_gen4_send_dependency_workarounds();
 188    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 189                                                     fs_inst *inst);
 190    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 191                                                      fs_inst *inst);
 192    void vfail(const char *msg, va_list args);
 193    void fail(const char *msg, ...);
 194    void no16(const char *msg);
 195    void lower_uniform_pull_constant_loads();
 196    bool lower_load_payload();
 197    bool lower_logical_sends();
 198    bool lower_integer_multiplication();
 199    bool lower_simd_width();
 200    bool opt_combine_constants();
 201
 202    void emit_dummy_fs();
 203    void emit_repclear_shader();
 204    fs_reg *emit_fragcoord_interpolation(bool pixel_center_integer,
 205                                         bool origin_upper_left);
 206    fs_inst *emit_linterp(const fs_reg &attr, const fs_reg &interp,
 207                          glsl_interp_qualifier interpolation_mode,
 208                          bool is_centroid, bool is_sample);
 209    fs_reg *emit_frontfacing_interpolation();
 210    fs_reg *emit_samplepos_setup();
 211    fs_reg *emit_sampleid_setup();
 212    void emit_general_interpolation(fs_reg attr, const char *name,
 213                                    const glsl_type *type,
 214                                    glsl_interp_qualifier interpolation_mode,
 215                                    int location, bool mod_centroid,
 216                                    bool mod_sample);
 217    fs_reg *emit_vs_system_value(int location);
 218    void emit_interpolation_setup_gen4();
 219    void emit_interpolation_setup_gen6();
 220    void compute_sample_position(fs_reg dst, fs_reg int_sample_pos);
 221    fs_reg rescale_texcoord(fs_reg coordinate, int coord_components,
 222                            bool is_rect, uint32_t sampler);
 223    void emit_texture(ir_texture_opcode op,
 224                      const glsl_type *dest_type,
 225                      fs_reg coordinate, int components,
 226                      fs_reg shadow_c,
 227                      fs_reg lod, fs_reg dpdy, int grad_components,
 228                      fs_reg sample_index,
 229                      fs_reg offset,
 230                      fs_reg mcs,
 231                      int gather_component,
 232                      bool is_cube_array,
 233                      bool is_rect,
 234                      uint32_t sampler,
 235                      fs_reg sampler_reg);
 236    fs_reg emit_mcs_fetch(const fs_reg &coordinate, unsigned components,
 237                          const fs_reg &sampler);
 238    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 239    fs_reg resolve_source_modifiers(const fs_reg &src);
 240    void emit_discard_jump();
 241    bool opt_peephole_sel();
 242    bool opt_peephole_predicated_break();
 243    bool opt_saturate_propagation();
 244    bool opt_cmod_propagation();
 245    bool opt_zero_samples();
 246    void emit_unspill(bblock_t *block, fs_inst *inst, fs_reg reg,
 247                      uint32_t spill_offset, int count);
 248    void emit_spill(bblock_t *block, fs_inst *inst, fs_reg reg,
 249                    uint32_t spill_offset, int count);
 250
 251    void emit_nir_code();
 252    void nir_setup_inputs();
 253    void nir_setup_outputs();
 254    void nir_setup_uniforms();
 255    void nir_emit_system_values();
 256    void nir_emit_impl(nir_function_impl *impl);
 257    void nir_emit_cf_list(exec_list *list);
 258    void nir_emit_if(nir_if *if_stmt);
 259    void nir_emit_loop(nir_loop *loop);
 260    void nir_emit_block(nir_block *block);
 261    void nir_emit_instr(nir_instr *instr);
 262    void nir_emit_alu(const brw::fs_builder &bld, nir_alu_instr *instr);
 263    void nir_emit_load_const(const brw::fs_builder &bld,
 264                             nir_load_const_instr *instr);
 265    void nir_emit_undef(const brw::fs_builder &bld,
 266                        nir_ssa_undef_instr *instr);
 267    void nir_emit_vs_intrinsic(const brw::fs_builder &bld,
 268                               nir_intrinsic_instr *instr);
 269    void nir_emit_gs_intrinsic(const brw::fs_builder &bld,
 270                               nir_intrinsic_instr *instr);
 271    void nir_emit_fs_intrinsic(const brw::fs_builder &bld,
 272                               nir_intrinsic_instr *instr);
 273    void nir_emit_cs_intrinsic(const brw::fs_builder &bld,
 274                               nir_intrinsic_instr *instr);
 275    void nir_emit_intrinsic(const brw::fs_builder &bld,
 276                            nir_intrinsic_instr *instr);
 277    void nir_emit_ssbo_atomic(const brw::fs_builder &bld,
 278                              int op, nir_intrinsic_instr *instr);
 279    void nir_emit_texture(const brw::fs_builder &bld,
 280                          nir_tex_instr *instr);
 281    void nir_emit_jump(const brw::fs_builder &bld,
 282                       nir_jump_instr *instr);
 283    fs_reg get_nir_src(nir_src src);
 284    fs_reg get_nir_dest(nir_dest dest);
 285    fs_reg get_nir_image_deref(const nir_deref_var *deref);
 286    void emit_percomp(const brw::fs_builder &bld, const fs_inst &inst,
 287                      unsigned wr_mask);
 288
 289    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 290                                      const fs_reg &result);
 291
 292    void emit_alpha_test();
 293    fs_inst *emit_single_fb_write(const brw::fs_builder &bld,
 294                                  fs_reg color1, fs_reg color2,
 295                                  fs_reg src0_alpha, unsigned components);
 296    void emit_fb_writes();
 297    void emit_urb_writes(const fs_reg &gs_vertex_count = fs_reg());
 298    void set_gs_stream_control_data_bits(const fs_reg &vertex_count,
 299                                         unsigned stream_id);
 300    void emit_gs_control_data_bits(const fs_reg &vertex_count);
 301    void emit_gs_end_primitive(const nir_src &vertex_count_nir_src);
 302    void emit_gs_vertex(const nir_src &vertex_count_nir_src,
 303                        unsigned stream_id);
 304    void emit_gs_thread_end();
 305    void emit_gs_input_load(const fs_reg &dst, const nir_src &vertex_src,
 306                            unsigned offset, unsigned num_components);
 307    void emit_cs_terminate();
 308    fs_reg *emit_cs_local_invocation_id_setup();
 309    fs_reg *emit_cs_work_group_id_setup();
 310
 311    void emit_barrier();
 312
 313    void emit_shader_time_begin();
 314    void emit_shader_time_end();
 315    void SHADER_TIME_ADD(const brw::fs_builder &bld,
 316                         int shader_time_subindex,
 317                         fs_reg value);
 318
 319    fs_reg get_timestamp(const brw::fs_builder &bld);
 320
 321    struct brw_reg interp_reg(int location, int channel);
 322
 323    int implied_mrf_writes(fs_inst *inst);
 324
 325    virtual void dump_instructions();
 326    virtual void dump_instructions(const char *name);
 327    void dump_instruction(backend_instruction *inst);
 328    void dump_instruction(backend_instruction *inst, FILE *file);
 329
 330    const void *const key;
 331    const struct brw_sampler_prog_key_data *key_tex;
 332
 333    struct brw_gs_compile *gs_compile;
 334
 335    struct brw_stage_prog_data *prog_data;
 336    struct gl_program *prog;
 337
 338    int *param_size;
 339
 340    int *virtual_grf_start;
 341    int *virtual_grf_end;
 342    brw::fs_live_variables *live_intervals;
 343
 344    int *regs_live_at_ip;
 345
 346    /** Number of uniform variable components visited. */
 347    unsigned uniforms;
 348
 349    /** Byte-offset for the next available spot in the scratch space buffer. */
 350    unsigned last_scratch;
 351
 352    /**
 353     * Array mapping UNIFORM register numbers to the pull parameter index,
 354     * or -1 if this uniform register isn't being uploaded as a pull constant.
 355     */
 356    int *pull_constant_loc;
 357
 358    /**
 359     * Array mapping UNIFORM register numbers to the push parameter index,
 360     * or -1 if this uniform register isn't being uploaded as a push constant.
 361     */
 362    int *push_constant_loc;
 363
 364    fs_reg frag_depth;
 365    fs_reg frag_stencil;
 366    fs_reg sample_mask;
 367    fs_reg outputs[VARYING_SLOT_MAX];
 368    unsigned output_components[VARYING_SLOT_MAX];
 369    fs_reg dual_src_output;
 370    bool do_dual_src;
 371    int first_non_payload_grf;
 372    /** Either BRW_MAX_GRF or GEN7_MRF_HACK_START */
 373    unsigned max_grf;
 374
 375    fs_reg *nir_locals;
 376    fs_reg *nir_ssa_values;
 377    fs_reg nir_inputs;
 378    fs_reg nir_outputs;
 379    fs_reg *nir_system_values;
 380
 381    bool failed;
 382    char *fail_msg;
 383    bool simd16_unsupported;
 384    char *no16_msg;
 385
 386    /* Result of last visit() method. Still used by emit_texture() */
 387    fs_reg result;
 388
 389    /** Register numbers for thread payload fields. */
 390    struct thread_payload {
 391       uint8_t source_depth_reg;
 392       uint8_t source_w_reg;
 393       uint8_t aa_dest_stencil_reg;
 394       uint8_t dest_depth_reg;
 395       uint8_t sample_pos_reg;
 396       uint8_t sample_mask_in_reg;
 397       uint8_t barycentric_coord_reg[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 398       uint8_t local_invocation_id_reg;
 399
 400       /** The number of thread payload registers the hardware will supply. */
 401       uint8_t num_regs;
 402    } payload;
 403
 404    bool source_depth_to_render_target;
 405    bool runtime_check_aads_emit;
 406
 407    fs_reg pixel_x;
 408    fs_reg pixel_y;
 409    fs_reg wpos_w;
 410    fs_reg pixel_w;
 411    fs_reg delta_xy[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 412    fs_reg shader_start_time;
 413    fs_reg userplane[MAX_CLIP_PLANES];
 414    fs_reg final_gs_vertex_count;
 415    fs_reg control_data_bits;
 416
 417    unsigned grf_used;
 418    bool spilled_any_registers;
 419
 420    const unsigned dispatch_width; /**< 8 or 16 */
 421
 422    int shader_time_index;
 423
 424    unsigned promoted_constants;
 425    brw::fs_builder bld;
 426 };
 427
 428 /**
 429  * The fragment shader code generator.
 430  *
 431  * Translates FS IR to actual i965 assembly code.
 432  */
 433 class fs_generator
 434 {
 435 public:
 436    fs_generator(const struct brw_compiler *compiler, void *log_data,
 437                 void *mem_ctx,
 438                 const void *key,
 439                 struct brw_stage_prog_data *prog_data,
 440                 unsigned promoted_constants,
 441                 bool runtime_check_aads_emit,
 442                 const char *stage_abbrev);
 443    ~fs_generator();
 444
 445    void enable_debug(const char *shader_name);
 446    int generate_code(const cfg_t *cfg, int dispatch_width);
 447    const unsigned *get_assembly(unsigned int *assembly_size);
 448
 449 private:
 450    void fire_fb_write(fs_inst *inst,
 451                       struct brw_reg payload,
 452                       struct brw_reg implied_header,
 453                       GLuint nr);
 454    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 455    void generate_urb_read(fs_inst *inst, struct brw_reg dst, struct brw_reg payload);
 456    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 457    void generate_cs_terminate(fs_inst *inst, struct brw_reg payload);
 458    void generate_stencil_ref_packing(fs_inst *inst, struct brw_reg dst,
 459                                      struct brw_reg src);
 460    void generate_barrier(fs_inst *inst, struct brw_reg src);
 461    void generate_blorp_fb_write(fs_inst *inst);
 462    void generate_linterp(fs_inst *inst, struct brw_reg dst,
 463                          struct brw_reg *src);
 464    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
 465                      struct brw_reg sampler_index);
 466    void generate_get_buffer_size(fs_inst *inst, struct brw_reg dst,
 467                                  struct brw_reg src,
 468                                  struct brw_reg surf_index);
 469    void generate_math_gen6(fs_inst *inst,
 470                            struct brw_reg dst,
 471                            struct brw_reg src0,
 472                            struct brw_reg src1);
 473    void generate_math_gen4(fs_inst *inst,
 474                            struct brw_reg dst,
 475                            struct brw_reg src);
 476    void generate_math_g45(fs_inst *inst,
 477                           struct brw_reg dst,
 478                           struct brw_reg src);
 479    void generate_ddx(enum opcode op, struct brw_reg dst, struct brw_reg src);
 480    void generate_ddy(enum opcode op, struct brw_reg dst, struct brw_reg src,
 481                      bool negate_value);
 482    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 483    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 484    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 485    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 486                                             struct brw_reg index,
 487                                             struct brw_reg offset);
 488    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 489                                                  struct brw_reg dst,
 490                                                  struct brw_reg surf_index,
 491                                                  struct brw_reg offset);
 492    void generate_varying_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 493                                             struct brw_reg index,
 494                                             struct brw_reg offset);
 495    void generate_varying_pull_constant_load_gen7(fs_inst *inst,
 496                                                  struct brw_reg dst,
 497                                                  struct brw_reg index,
 498                                                  struct brw_reg offset);
 499    void generate_mov_dispatch_to_flags(fs_inst *inst);
 500
 501    void generate_pixel_interpolator_query(fs_inst *inst,
 502                                           struct brw_reg dst,
 503                                           struct brw_reg src,
 504                                           struct brw_reg msg_data,
 505                                           unsigned msg_type);
 506
 507    void generate_set_sample_id(fs_inst *inst,
 508                                struct brw_reg dst,
 509                                struct brw_reg src0,
 510                                struct brw_reg src1);
 511
 512    void generate_set_simd4x2_offset(fs_inst *inst,
 513                                     struct brw_reg dst,
 514                                     struct brw_reg offset);
 515    void generate_discard_jump(fs_inst *inst);
 516
 517    void generate_pack_half_2x16_split(fs_inst *inst,
 518                                       struct brw_reg dst,
 519                                       struct brw_reg x,
 520                                       struct brw_reg y);
 521    void generate_unpack_half_2x16_split(fs_inst *inst,
 522                                         struct brw_reg dst,
 523                                         struct brw_reg src);
 524
 525    void generate_shader_time_add(fs_inst *inst,
 526                                  struct brw_reg payload,
 527                                  struct brw_reg offset,
 528                                  struct brw_reg value);
 529
 530    bool patch_discard_jumps_to_fb_writes();
 531
 532    const struct brw_compiler *compiler;
 533    void *log_data; /* Passed to compiler->*_log functions */
 534
 535    const struct brw_device_info *devinfo;
 536
 537    struct brw_codegen *p;
 538    const void * const key;
 539    struct brw_stage_prog_data * const prog_data;
 540
 541    unsigned dispatch_width; /**< 8 or 16 */
 542
 543    exec_list discard_halt_patches;
 544    unsigned promoted_constants;
 545    bool runtime_check_aads_emit;
 546    bool debug_flag;
 547    const char *shader_name;
 548    const char *stage_abbrev;
 549    void *mem_ctx;
 550 };
 551
 552 bool brw_do_channel_expressions(struct exec_list *instructions);
 553 bool brw_do_vector_splitting(struct exec_list *instructions);