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/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 /**
  66  * The fragment shader front-end.
  67  *
  68  * Translates either GLSL IR or Mesa IR (for ARB_fragment_program) into FS IR.
  69  */
  70 class fs_visitor : public backend_shader
  71 {
  72 public:
  73    fs_visitor(const struct brw_compiler *compiler, void *log_data,
  74               void *mem_ctx,
  75               gl_shader_stage stage,
  76               const void *key,
  77               struct brw_stage_prog_data *prog_data,
  78               struct gl_shader_program *shader_prog,
  79               struct gl_program *prog,
  80               unsigned dispatch_width,
  81               int shader_time_index);
  82
  83    ~fs_visitor();
  84
  85    fs_reg vgrf(const glsl_type *const type);
  86    void import_uniforms(fs_visitor *v);
  87    void setup_uniform_clipplane_values(gl_clip_plane *clip_planes);
  88    void compute_clip_distance(gl_clip_plane *clip_planes);
  89
  90    uint32_t gather_channel(int orig_chan, uint32_t sampler);
  91    void swizzle_result(ir_texture_opcode op, int dest_components,
  92                        fs_reg orig_val, uint32_t sampler);
  93
  94    int type_size(const struct glsl_type *type);
  95    fs_inst *get_instruction_generating_reg(fs_inst *start,
  96                                            fs_inst *end,
  97                                            const fs_reg &reg);
  98
  99    void VARYING_PULL_CONSTANT_LOAD(const brw::fs_builder &bld,
 100                                    const fs_reg &dst,
 101                                    const fs_reg &surf_index,
 102                                    const fs_reg &varying_offset,
 103                                    uint32_t const_offset);
 104    void DEP_RESOLVE_MOV(const brw::fs_builder &bld, int grf);
 105
 106    bool run_fs(bool do_rep_send);
 107    bool run_vs(gl_clip_plane *clip_planes);
 108    bool run_cs();
 109    void optimize();
 110    void allocate_registers();
 111    void assign_binding_table_offsets();
 112    void setup_payload_gen4();
 113    void setup_payload_gen6();
 114    void setup_vs_payload();
 115    void setup_cs_payload();
 116    void fixup_3src_null_dest();
 117    void assign_curb_setup();
 118    void calculate_urb_setup();
 119    void assign_urb_setup();
 120    void assign_vs_urb_setup();
 121    bool assign_regs(bool allow_spilling);
 122    void assign_regs_trivial();
 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 move_uniform_array_access_to_pull_constants();
 130    void assign_constant_locations();
 131    void demote_pull_constants();
 132    void invalidate_live_intervals();
 133    void calculate_live_intervals();
 134    void calculate_register_pressure();
 135    bool opt_algebraic();
 136    bool opt_redundant_discard_jumps();
 137    bool opt_cse();
 138    bool opt_cse_local(bblock_t *block);
 139    bool opt_copy_propagate();
 140    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 141    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 142    bool opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
 143                                  exec_list *acp);
 144    bool opt_register_renaming();
 145    bool register_coalesce();
 146    bool compute_to_mrf();
 147    bool eliminate_find_live_channel();
 148    bool dead_code_eliminate();
 149    bool remove_duplicate_mrf_writes();
 150
 151    bool opt_sampler_eot();
 152    bool virtual_grf_interferes(int a, int b);
 153    void schedule_instructions(instruction_scheduler_mode mode);
 154    void insert_gen4_send_dependency_workarounds();
 155    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 156                                                     fs_inst *inst);
 157    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 158                                                      fs_inst *inst);
 159    void vfail(const char *msg, va_list args);
 160    void fail(const char *msg, ...);
 161    void no16(const char *msg);
 162    void lower_uniform_pull_constant_loads();
 163    bool lower_load_payload();
 164    bool lower_integer_multiplication();
 165    bool opt_combine_constants();
 166
 167    void emit_dummy_fs();
 168    void emit_repclear_shader();
 169    fs_reg *emit_fragcoord_interpolation(bool pixel_center_integer,
 170                                         bool origin_upper_left);
 171    fs_inst *emit_linterp(const fs_reg &attr, const fs_reg &interp,
 172                          glsl_interp_qualifier interpolation_mode,
 173                          bool is_centroid, bool is_sample);
 174    fs_reg *emit_frontfacing_interpolation();
 175    fs_reg *emit_samplepos_setup();
 176    fs_reg *emit_sampleid_setup();
 177    void emit_general_interpolation(fs_reg attr, const char *name,
 178                                    const glsl_type *type,
 179                                    glsl_interp_qualifier interpolation_mode,
 180                                    int location, bool mod_centroid,
 181                                    bool mod_sample);
 182    fs_reg *emit_vs_system_value(int location);
 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 rescale_texcoord(fs_reg coordinate, int coord_components,
 187                            bool is_rect, uint32_t sampler, int texunit);
 188    fs_inst *emit_texture_gen4(ir_texture_opcode op, fs_reg dst,
 189                               fs_reg coordinate, int coord_components,
 190                               fs_reg shadow_comp,
 191                               fs_reg lod, fs_reg lod2, int grad_components,
 192                               uint32_t sampler);
 193    fs_inst *emit_texture_gen4_simd16(ir_texture_opcode op, fs_reg dst,
 194                                      fs_reg coordinate, int vector_elements,
 195                                      fs_reg shadow_c, fs_reg lod,
 196                                      uint32_t sampler);
 197    fs_inst *emit_texture_gen5(ir_texture_opcode op, fs_reg dst,
 198                               fs_reg coordinate, int coord_components,
 199                               fs_reg shadow_comp,
 200                               fs_reg lod, fs_reg lod2, int grad_components,
 201                               fs_reg sample_index, uint32_t sampler,
 202                               bool has_offset);
 203    fs_inst *emit_texture_gen7(ir_texture_opcode op, fs_reg dst,
 204                               fs_reg coordinate, int coord_components,
 205                               fs_reg shadow_comp,
 206                               fs_reg lod, fs_reg lod2, int grad_components,
 207                               fs_reg sample_index, fs_reg mcs, fs_reg sampler,
 208                               fs_reg offset_value);
 209    void emit_texture(ir_texture_opcode op,
 210                      const glsl_type *dest_type,
 211                      fs_reg coordinate, int components,
 212                      fs_reg shadow_c,
 213                      fs_reg lod, fs_reg dpdy, int grad_components,
 214                      fs_reg sample_index,
 215                      fs_reg offset,
 216                      fs_reg mcs,
 217                      int gather_component,
 218                      bool is_cube_array,
 219                      bool is_rect,
 220                      uint32_t sampler,
 221                      fs_reg sampler_reg,
 222                      int texunit);
 223    fs_reg emit_mcs_fetch(fs_reg coordinate, int components, fs_reg sampler);
 224    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 225    void resolve_source_modifiers(fs_reg *src);
 226    void emit_discard_jump();
 227    bool try_replace_with_sel();
 228    bool opt_peephole_sel();
 229    bool opt_peephole_predicated_break();
 230    bool opt_saturate_propagation();
 231    bool opt_cmod_propagation();
 232    bool opt_zero_samples();
 233    void emit_unspill(bblock_t *block, fs_inst *inst, fs_reg reg,
 234                      uint32_t spill_offset, int count);
 235    void emit_spill(bblock_t *block, fs_inst *inst, fs_reg reg,
 236                    uint32_t spill_offset, int count);
 237
 238    void emit_nir_code();
 239    void nir_setup_inputs(nir_shader *shader);
 240    void nir_setup_outputs(nir_shader *shader);
 241    void nir_setup_uniforms(nir_shader *shader);
 242    void nir_setup_uniform(nir_variable *var);
 243    void nir_setup_builtin_uniform(nir_variable *var);
 244    void nir_emit_system_values(nir_shader *shader);
 245    void nir_emit_impl(nir_function_impl *impl);
 246    void nir_emit_cf_list(exec_list *list);
 247    void nir_emit_if(nir_if *if_stmt);
 248    void nir_emit_loop(nir_loop *loop);
 249    void nir_emit_block(nir_block *block);
 250    void nir_emit_instr(nir_instr *instr);
 251    void nir_emit_alu(const brw::fs_builder &bld, nir_alu_instr *instr);
 252    void nir_emit_undef(const brw::fs_builder &bld,
 253                        nir_ssa_undef_instr *instr);
 254    void nir_emit_intrinsic(const brw::fs_builder &bld,
 255                            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    void emit_percomp(const brw::fs_builder &bld, const fs_inst &inst,
 263                      unsigned wr_mask);
 264
 265    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 266                                      const fs_reg &result);
 267
 268    void setup_color_payload(fs_reg *dst, fs_reg color, unsigned components,
 269                             unsigned exec_size, bool use_2nd_half);
 270    void emit_alpha_test();
 271    fs_inst *emit_single_fb_write(const brw::fs_builder &bld,
 272                                  fs_reg color1, fs_reg color2,
 273                                  fs_reg src0_alpha, unsigned components,
 274                                  unsigned exec_size, bool use_2nd_half = false);
 275    void emit_fb_writes();
 276    void emit_urb_writes();
 277    void emit_cs_terminate();
 278
 279    void emit_barrier();
 280
 281    void emit_shader_time_begin();
 282    void emit_shader_time_end();
 283    void SHADER_TIME_ADD(const brw::fs_builder &bld,
 284                         int shader_time_subindex,
 285                         fs_reg value);
 286
 287    void emit_untyped_atomic(unsigned atomic_op, unsigned surf_index,
 288                             fs_reg dst, fs_reg offset, fs_reg src0,
 289                             fs_reg src1);
 290
 291    void emit_untyped_surface_read(unsigned surf_index, fs_reg dst,
 292                                   fs_reg offset);
 293
 294    fs_reg get_timestamp(const brw::fs_builder &bld);
 295
 296    struct brw_reg interp_reg(int location, int channel);
 297    int implied_mrf_writes(fs_inst *inst);
 298
 299    virtual void dump_instructions();
 300    virtual void dump_instructions(const char *name);
 301    void dump_instruction(backend_instruction *inst);
 302    void dump_instruction(backend_instruction *inst, FILE *file);
 303
 304    const void *const key;
 305    const struct brw_sampler_prog_key_data *key_tex;
 306
 307    struct brw_stage_prog_data *prog_data;
 308    unsigned int sanity_param_count;
 309
 310    int *param_size;
 311
 312    int *virtual_grf_start;
 313    int *virtual_grf_end;
 314    brw::fs_live_variables *live_intervals;
 315
 316    int *regs_live_at_ip;
 317
 318    /** Number of uniform variable components visited. */
 319    unsigned uniforms;
 320
 321    /** Total number of direct uniforms we can get from NIR */
 322    unsigned num_direct_uniforms;
 323
 324    /** Byte-offset for the next available spot in the scratch space buffer. */
 325    unsigned last_scratch;
 326
 327    /**
 328     * Array mapping UNIFORM register numbers to the pull parameter index,
 329     * or -1 if this uniform register isn't being uploaded as a pull constant.
 330     */
 331    int *pull_constant_loc;
 332
 333    /**
 334     * Array mapping UNIFORM register numbers to the push parameter index,
 335     * or -1 if this uniform register isn't being uploaded as a push constant.
 336     */
 337    int *push_constant_loc;
 338
 339    fs_reg frag_depth;
 340    fs_reg sample_mask;
 341    fs_reg outputs[VARYING_SLOT_MAX];
 342    unsigned output_components[VARYING_SLOT_MAX];
 343    fs_reg dual_src_output;
 344    bool do_dual_src;
 345    int first_non_payload_grf;
 346    /** Either BRW_MAX_GRF or GEN7_MRF_HACK_START */
 347    unsigned max_grf;
 348
 349    fs_reg *nir_locals;
 350    fs_reg *nir_ssa_values;
 351    fs_reg *nir_globals;
 352    fs_reg nir_inputs;
 353    fs_reg nir_outputs;
 354    fs_reg *nir_system_values;
 355
 356    bool failed;
 357    char *fail_msg;
 358    bool simd16_unsupported;
 359    char *no16_msg;
 360
 361    /* Result of last visit() method. Still used by emit_texture() */
 362    fs_reg result;
 363
 364    /** Register numbers for thread payload fields. */
 365    struct {
 366       uint8_t source_depth_reg;
 367       uint8_t source_w_reg;
 368       uint8_t aa_dest_stencil_reg;
 369       uint8_t dest_depth_reg;
 370       uint8_t sample_pos_reg;
 371       uint8_t sample_mask_in_reg;
 372       uint8_t barycentric_coord_reg[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 373
 374       /** The number of thread payload registers the hardware will supply. */
 375       uint8_t num_regs;
 376    } payload;
 377
 378    bool source_depth_to_render_target;
 379    bool runtime_check_aads_emit;
 380
 381    fs_reg pixel_x;
 382    fs_reg pixel_y;
 383    fs_reg wpos_w;
 384    fs_reg pixel_w;
 385    fs_reg delta_xy[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 386    fs_reg shader_start_time;
 387    fs_reg userplane[MAX_CLIP_PLANES];
 388
 389    unsigned grf_used;
 390    bool spilled_any_registers;
 391
 392    const unsigned dispatch_width; /**< 8 or 16 */
 393
 394    int shader_time_index;
 395
 396    unsigned promoted_constants;
 397    brw::fs_builder bld;
 398 };
 399
 400 /**
 401  * The fragment shader code generator.
 402  *
 403  * Translates FS IR to actual i965 assembly code.
 404  */
 405 class fs_generator
 406 {
 407 public:
 408    fs_generator(const struct brw_compiler *compiler, void *log_data,
 409                 void *mem_ctx,
 410                 const void *key,
 411                 struct brw_stage_prog_data *prog_data,
 412                 struct gl_program *fp,
 413                 unsigned promoted_constants,
 414                 bool runtime_check_aads_emit,
 415                 const char *stage_abbrev);
 416    ~fs_generator();
 417
 418    void enable_debug(const char *shader_name);
 419    int generate_code(const cfg_t *cfg, int dispatch_width);
 420    const unsigned *get_assembly(unsigned int *assembly_size);
 421
 422 private:
 423    void fire_fb_write(fs_inst *inst,
 424                       struct brw_reg payload,
 425                       struct brw_reg implied_header,
 426                       GLuint nr);
 427    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 428    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 429    void generate_cs_terminate(fs_inst *inst, struct brw_reg payload);
 430    void generate_barrier(fs_inst *inst, struct brw_reg src);
 431    void generate_blorp_fb_write(fs_inst *inst);
 432    void generate_linterp(fs_inst *inst, struct brw_reg dst,
 433                          struct brw_reg *src);
 434    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
 435                      struct brw_reg sampler_index);
 436    void generate_math_gen6(fs_inst *inst,
 437                            struct brw_reg dst,
 438                            struct brw_reg src0,
 439                            struct brw_reg src1);
 440    void generate_math_gen4(fs_inst *inst,
 441                            struct brw_reg dst,
 442                            struct brw_reg src);
 443    void generate_math_g45(fs_inst *inst,
 444                           struct brw_reg dst,
 445                           struct brw_reg src);
 446    void generate_ddx(enum opcode op, struct brw_reg dst, struct brw_reg src);
 447    void generate_ddy(enum opcode op, struct brw_reg dst, struct brw_reg src,
 448                      bool negate_value);
 449    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 450    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 451    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 452    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 453                                             struct brw_reg index,
 454                                             struct brw_reg offset);
 455    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 456                                                  struct brw_reg dst,
 457                                                  struct brw_reg surf_index,
 458                                                  struct brw_reg offset);
 459    void generate_varying_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 460                                             struct brw_reg index,
 461                                             struct brw_reg offset);
 462    void generate_varying_pull_constant_load_gen7(fs_inst *inst,
 463                                                  struct brw_reg dst,
 464                                                  struct brw_reg index,
 465                                                  struct brw_reg offset);
 466    void generate_mov_dispatch_to_flags(fs_inst *inst);
 467
 468    void generate_pixel_interpolator_query(fs_inst *inst,
 469                                           struct brw_reg dst,
 470                                           struct brw_reg src,
 471                                           struct brw_reg msg_data,
 472                                           unsigned msg_type);
 473
 474    void generate_set_omask(fs_inst *inst,
 475                            struct brw_reg dst,
 476                            struct brw_reg sample_mask);
 477
 478    void generate_set_sample_id(fs_inst *inst,
 479                                struct brw_reg dst,
 480                                struct brw_reg src0,
 481                                struct brw_reg src1);
 482
 483    void generate_set_simd4x2_offset(fs_inst *inst,
 484                                     struct brw_reg dst,
 485                                     struct brw_reg offset);
 486    void generate_discard_jump(fs_inst *inst);
 487
 488    void generate_pack_half_2x16_split(fs_inst *inst,
 489                                       struct brw_reg dst,
 490                                       struct brw_reg x,
 491                                       struct brw_reg y);
 492    void generate_unpack_half_2x16_split(fs_inst *inst,
 493                                         struct brw_reg dst,
 494                                         struct brw_reg src);
 495
 496    void generate_shader_time_add(fs_inst *inst,
 497                                  struct brw_reg payload,
 498                                  struct brw_reg offset,
 499                                  struct brw_reg value);
 500
 501    bool patch_discard_jumps_to_fb_writes();
 502
 503    const struct brw_compiler *compiler;
 504    void *log_data; /* Passed to compiler->*_log functions */
 505
 506    const struct brw_device_info *devinfo;
 507
 508    struct brw_codegen *p;
 509    const void * const key;
 510    struct brw_stage_prog_data * const prog_data;
 511
 512    const struct gl_program *prog;
 513
 514    unsigned dispatch_width; /**< 8 or 16 */
 515
 516    exec_list discard_halt_patches;
 517    unsigned promoted_constants;
 518    bool runtime_check_aads_emit;
 519    bool debug_flag;
 520    const char *shader_name;
 521    const char *stage_abbrev;
 522    void *mem_ctx;
 523 };
 524
 525 bool brw_do_channel_expressions(struct exec_list *instructions);
 526 bool brw_do_vector_splitting(struct exec_list *instructions);
 527 void brw_setup_tex_for_precompile(struct brw_context *brw,
 528                                   struct brw_sampler_prog_key_data *tex,
 529                                   struct gl_program *prog);