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 extern "C" {
  29
  30 #include <sys/types.h>
  31
  32 #include "main/macros.h"
  33 #include "main/shaderobj.h"
  34 #include "main/uniforms.h"
  35 #include "program/prog_parameter.h"
  36 #include "program/prog_print.h"
  37 #include "program/prog_optimize.h"
  38 #include "program/register_allocate.h"
  39 #include "program/sampler.h"
  40 #include "program/hash_table.h"
  41 #include "brw_context.h"
  42 #include "brw_eu.h"
  43 #include "brw_wm.h"
  44 }
  45 #include "../glsl/glsl_types.h"
  46 #include "../glsl/ir.h"
  47
  48 enum register_file {
  49    ARF = BRW_ARCHITECTURE_REGISTER_FILE,
  50    GRF = BRW_GENERAL_REGISTER_FILE,
  51    MRF = BRW_MESSAGE_REGISTER_FILE,
  52    IMM = BRW_IMMEDIATE_VALUE,
  53    FIXED_HW_REG, /* a struct brw_reg */
  54    UNIFORM, /* prog_data->params[hw_reg] */
  55    BAD_FILE
  56 };
  57
  58 enum fs_opcodes {
  59    FS_OPCODE_FB_WRITE = 256,
  60    FS_OPCODE_RCP,
  61    FS_OPCODE_RSQ,
  62    FS_OPCODE_SQRT,
  63    FS_OPCODE_EXP2,
  64    FS_OPCODE_LOG2,
  65    FS_OPCODE_POW,
  66    FS_OPCODE_SIN,
  67    FS_OPCODE_COS,
  68    FS_OPCODE_DDX,
  69    FS_OPCODE_DDY,
  70    FS_OPCODE_PIXEL_X,
  71    FS_OPCODE_PIXEL_Y,
  72    FS_OPCODE_CINTERP,
  73    FS_OPCODE_LINTERP,
  74    FS_OPCODE_TEX,
  75    FS_OPCODE_TXB,
  76    FS_OPCODE_TXD,
  77    FS_OPCODE_TXL,
  78    FS_OPCODE_DISCARD,
  79    FS_OPCODE_SPILL,
  80    FS_OPCODE_UNSPILL,
  81    FS_OPCODE_PULL_CONSTANT_LOAD,
  82 };
  83
  84
  85 class fs_reg {
  86 public:
  87    /* Callers of this ralloc-based new need not call delete. It's
  88     * easier to just ralloc_free 'ctx' (or any of its ancestors). */
  89    static void* operator new(size_t size, void *ctx)
  90    {
  91       void *node;
  92
  93       node = ralloc_size(ctx, size);
  94       assert(node != NULL);
  95
  96       return node;
  97    }
  98
  99    void init()
 100    {
 101       memset(this, 0, sizeof(*this));
 102       this->hw_reg = -1;
 103       this->smear = -1;
 104    }
 105
 106    /** Generic unset register constructor. */
 107    fs_reg()
 108    {
 109       init();
 110       this->file = BAD_FILE;
 111    }
 112
 113    /** Immediate value constructor. */
 114    fs_reg(float f)
 115    {
 116       init();
 117       this->file = IMM;
 118       this->type = BRW_REGISTER_TYPE_F;
 119       this->imm.f = f;
 120    }
 121
 122    /** Immediate value constructor. */
 123    fs_reg(int32_t i)
 124    {
 125       init();
 126       this->file = IMM;
 127       this->type = BRW_REGISTER_TYPE_D;
 128       this->imm.i = i;
 129    }
 130
 131    /** Immediate value constructor. */
 132    fs_reg(uint32_t u)
 133    {
 134       init();
 135       this->file = IMM;
 136       this->type = BRW_REGISTER_TYPE_UD;
 137       this->imm.u = u;
 138    }
 139
 140    /** Fixed brw_reg Immediate value constructor. */
 141    fs_reg(struct brw_reg fixed_hw_reg)
 142    {
 143       init();
 144       this->file = FIXED_HW_REG;
 145       this->fixed_hw_reg = fixed_hw_reg;
 146       this->type = fixed_hw_reg.type;
 147    }
 148
 149    fs_reg(enum register_file file, int hw_reg);
 150    fs_reg(enum register_file file, int hw_reg, uint32_t type);
 151    fs_reg(class fs_visitor *v, const struct glsl_type *type);
 152
 153    bool equals(fs_reg *r)
 154    {
 155       return (file == r->file &&
 156               reg == r->reg &&
 157               reg_offset == r->reg_offset &&
 158               hw_reg == r->hw_reg &&
 159               type == r->type &&
 160               negate == r->negate &&
 161               abs == r->abs &&
 162               memcmp(&fixed_hw_reg, &r->fixed_hw_reg,
 163                      sizeof(fixed_hw_reg)) == 0 &&
 164               smear == r->smear &&
 165               imm.u == r->imm.u);
 166    }
 167
 168    /** Register file: ARF, GRF, MRF, IMM. */
 169    enum register_file file;
 170    /** virtual register number.  0 = fixed hw reg */
 171    int reg;
 172    /** Offset within the virtual register. */
 173    int reg_offset;
 174    /** HW register number.  Generally unset until register allocation. */
 175    int hw_reg;
 176    /** Register type.  BRW_REGISTER_TYPE_* */
 177    int type;
 178    bool negate;
 179    bool abs;
 180    bool sechalf;
 181    struct brw_reg fixed_hw_reg;
 182    int smear; /* -1, or a channel of the reg to smear to all channels. */
 183
 184    /** Value for file == BRW_IMMMEDIATE_FILE */
 185    union {
 186       int32_t i;
 187       uint32_t u;
 188       float f;
 189    } imm;
 190 };
 191
 192 static const fs_reg reg_undef;
 193 static const fs_reg reg_null_f(ARF, BRW_ARF_NULL, BRW_REGISTER_TYPE_F);
 194 static const fs_reg reg_null_d(ARF, BRW_ARF_NULL, BRW_REGISTER_TYPE_D);
 195
 196 class fs_inst : public exec_node {
 197 public:
 198    /* Callers of this ralloc-based new need not call delete. It's
 199     * easier to just ralloc_free 'ctx' (or any of its ancestors). */
 200    static void* operator new(size_t size, void *ctx)
 201    {
 202       void *node;
 203
 204       node = rzalloc_size(ctx, size);
 205       assert(node != NULL);
 206
 207       return node;
 208    }
 209
 210    void init()
 211    {
 212       memset(this, 0, sizeof(*this));
 213       this->opcode = BRW_OPCODE_NOP;
 214       this->conditional_mod = BRW_CONDITIONAL_NONE;
 215
 216       this->dst = reg_undef;
 217       this->src[0] = reg_undef;
 218       this->src[1] = reg_undef;
 219       this->src[2] = reg_undef;
 220    }
 221
 222    fs_inst()
 223    {
 224       init();
 225    }
 226
 227    fs_inst(int opcode)
 228    {
 229       init();
 230       this->opcode = opcode;
 231    }
 232
 233    fs_inst(int opcode, fs_reg dst)
 234    {
 235       init();
 236       this->opcode = opcode;
 237       this->dst = dst;
 238
 239       if (dst.file == GRF)
 240          assert(dst.reg_offset >= 0);
 241    }
 242
 243    fs_inst(int opcode, fs_reg dst, fs_reg src0)
 244    {
 245       init();
 246       this->opcode = opcode;
 247       this->dst = dst;
 248       this->src[0] = src0;
 249
 250       if (dst.file == GRF)
 251          assert(dst.reg_offset >= 0);
 252       if (src[0].file == GRF)
 253          assert(src[0].reg_offset >= 0);
 254    }
 255
 256    fs_inst(int opcode, fs_reg dst, fs_reg src0, fs_reg src1)
 257    {
 258       init();
 259       this->opcode = opcode;
 260       this->dst = dst;
 261       this->src[0] = src0;
 262       this->src[1] = src1;
 263
 264       if (dst.file == GRF)
 265          assert(dst.reg_offset >= 0);
 266       if (src[0].file == GRF)
 267          assert(src[0].reg_offset >= 0);
 268       if (src[1].file == GRF)
 269          assert(src[1].reg_offset >= 0);
 270    }
 271
 272    fs_inst(int opcode, fs_reg dst, fs_reg src0, fs_reg src1, fs_reg src2)
 273    {
 274       init();
 275       this->opcode = opcode;
 276       this->dst = dst;
 277       this->src[0] = src0;
 278       this->src[1] = src1;
 279       this->src[2] = src2;
 280
 281       if (dst.file == GRF)
 282          assert(dst.reg_offset >= 0);
 283       if (src[0].file == GRF)
 284          assert(src[0].reg_offset >= 0);
 285       if (src[1].file == GRF)
 286          assert(src[1].reg_offset >= 0);
 287       if (src[2].file == GRF)
 288          assert(src[2].reg_offset >= 0);
 289    }
 290
 291    bool equals(fs_inst *inst)
 292    {
 293       return (opcode == inst->opcode &&
 294               dst.equals(&inst->dst) &&
 295               src[0].equals(&inst->src[0]) &&
 296               src[1].equals(&inst->src[1]) &&
 297               src[2].equals(&inst->src[2]) &&
 298               saturate == inst->saturate &&
 299               predicated == inst->predicated &&
 300               conditional_mod == inst->conditional_mod &&
 301               mlen == inst->mlen &&
 302               base_mrf == inst->base_mrf &&
 303               sampler == inst->sampler &&
 304               target == inst->target &&
 305               eot == inst->eot &&
 306               header_present == inst->header_present &&
 307               shadow_compare == inst->shadow_compare &&
 308               offset == inst->offset);
 309    }
 310
 311    bool is_tex()
 312    {
 313       return (opcode == FS_OPCODE_TEX ||
 314               opcode == FS_OPCODE_TXB ||
 315               opcode == FS_OPCODE_TXD ||
 316               opcode == FS_OPCODE_TXL);
 317    }
 318
 319    bool is_math()
 320    {
 321       return (opcode == FS_OPCODE_RCP ||
 322               opcode == FS_OPCODE_RSQ ||
 323               opcode == FS_OPCODE_SQRT ||
 324               opcode == FS_OPCODE_EXP2 ||
 325               opcode == FS_OPCODE_LOG2 ||
 326               opcode == FS_OPCODE_SIN ||
 327               opcode == FS_OPCODE_COS ||
 328               opcode == FS_OPCODE_POW);
 329    }
 330
 331    int opcode; /* BRW_OPCODE_* or FS_OPCODE_* */
 332    fs_reg dst;
 333    fs_reg src[3];
 334    bool saturate;
 335    bool predicated;
 336    bool predicate_inverse;
 337    int conditional_mod; /**< BRW_CONDITIONAL_* */
 338
 339    int mlen; /**< SEND message length */
 340    int base_mrf; /**< First MRF in the SEND message, if mlen is nonzero. */
 341    int sampler;
 342    int target; /**< MRT target. */
 343    bool eot;
 344    bool header_present;
 345    bool shadow_compare;
 346    bool force_uncompressed;
 347    bool force_sechalf;
 348    uint32_t offset; /* spill/unspill offset */
 349
 350    /** @{
 351     * Annotation for the generated IR.  One of the two can be set.
 352     */
 353    ir_instruction *ir;
 354    const char *annotation;
 355    /** @} */
 356 };
 357
 358 class fs_visitor : public ir_visitor
 359 {
 360 public:
 361
 362    fs_visitor(struct brw_wm_compile *c, struct gl_shader_program *prog,
 363               struct brw_shader *shader)
 364    {
 365       this->c = c;
 366       this->p = &c->func;
 367       this->brw = p->brw;
 368       this->fp = prog->FragmentProgram;
 369       this->prog = prog;
 370       this->intel = &brw->intel;
 371       this->ctx = &intel->ctx;
 372       this->mem_ctx = ralloc_context(NULL);
 373       this->shader = shader;
 374       this->failed = false;
 375       this->variable_ht = hash_table_ctor(0,
 376                                           hash_table_pointer_hash,
 377                                           hash_table_pointer_compare);
 378
 379       /* There's a question that appears to be left open in the spec:
 380        * How do implicit dst conversions interact with the CMP
 381        * instruction or conditional mods?  On gen6, the instruction:
 382        *
 383        * CMP null<d> src0<f> src1<f>
 384        *
 385        * will do src1 - src0 and compare that result as if it was an
 386        * integer.  On gen4, it will do src1 - src0 as float, convert
 387        * the result to int, and compare as int.  In between, it
 388        * appears that it does src1 - src0 and does the compare in the
 389        * execution type so dst type doesn't matter.
 390        */
 391       if (this->intel->gen > 4)
 392          this->reg_null_cmp = reg_null_d;
 393       else
 394          this->reg_null_cmp = reg_null_f;
 395
 396       this->frag_color = NULL;
 397       this->frag_data = NULL;
 398       this->frag_depth = NULL;
 399       this->first_non_payload_grf = 0;
 400
 401       this->current_annotation = NULL;
 402       this->base_ir = NULL;
 403
 404       this->virtual_grf_sizes = NULL;
 405       this->virtual_grf_next = 1;
 406       this->virtual_grf_array_size = 0;
 407       this->virtual_grf_def = NULL;
 408       this->virtual_grf_use = NULL;
 409       this->live_intervals_valid = false;
 410
 411       this->kill_emitted = false;
 412       this->force_uncompressed_stack = 0;
 413       this->force_sechalf_stack = 0;
 414    }
 415
 416    ~fs_visitor()
 417    {
 418       ralloc_free(this->mem_ctx);
 419       hash_table_dtor(this->variable_ht);
 420    }
 421
 422    fs_reg *variable_storage(ir_variable *var);
 423    int virtual_grf_alloc(int size);
 424    void import_uniforms(struct hash_table *src_variable_ht);
 425
 426    void visit(ir_variable *ir);
 427    void visit(ir_assignment *ir);
 428    void visit(ir_dereference_variable *ir);
 429    void visit(ir_dereference_record *ir);
 430    void visit(ir_dereference_array *ir);
 431    void visit(ir_expression *ir);
 432    void visit(ir_texture *ir);
 433    void visit(ir_if *ir);
 434    void visit(ir_constant *ir);
 435    void visit(ir_swizzle *ir);
 436    void visit(ir_return *ir);
 437    void visit(ir_loop *ir);
 438    void visit(ir_loop_jump *ir);
 439    void visit(ir_discard *ir);
 440    void visit(ir_call *ir);
 441    void visit(ir_function *ir);
 442    void visit(ir_function_signature *ir);
 443
 444    void swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler);
 445
 446    fs_inst *emit(fs_inst inst);
 447
 448    fs_inst *emit(int opcode)
 449    {
 450       return emit(fs_inst(opcode));
 451    }
 452
 453    fs_inst *emit(int opcode, fs_reg dst)
 454    {
 455       return emit(fs_inst(opcode, dst));
 456    }
 457
 458    fs_inst *emit(int opcode, fs_reg dst, fs_reg src0)
 459    {
 460       return emit(fs_inst(opcode, dst, src0));
 461    }
 462
 463    fs_inst *emit(int opcode, fs_reg dst, fs_reg src0, fs_reg src1)
 464    {
 465       return emit(fs_inst(opcode, dst, src0, src1));
 466    }
 467
 468    fs_inst *emit(int opcode, fs_reg dst, fs_reg src0, fs_reg src1, fs_reg src2)
 469    {
 470       return emit(fs_inst(opcode, dst, src0, src1, src2));
 471    }
 472
 473    int type_size(const struct glsl_type *type);
 474
 475    bool run();
 476    void setup_paramvalues_refs();
 477    void assign_curb_setup();
 478    void calculate_urb_setup();
 479    void assign_urb_setup();
 480    bool assign_regs();
 481    void assign_regs_trivial();
 482    int choose_spill_reg(struct ra_graph *g);
 483    void spill_reg(int spill_reg);
 484    void split_virtual_grfs();
 485    void setup_pull_constants();
 486    void calculate_live_intervals();
 487    bool propagate_constants();
 488    bool register_coalesce();
 489    bool compute_to_mrf();
 490    bool dead_code_eliminate();
 491    bool remove_duplicate_mrf_writes();
 492    bool virtual_grf_interferes(int a, int b);
 493    void schedule_instructions();
 494    void fail(const char *msg, ...);
 495
 496    void push_force_uncompressed();
 497    void pop_force_uncompressed();
 498    void push_force_sechalf();
 499    void pop_force_sechalf();
 500
 501    void generate_code();
 502    void generate_fb_write(fs_inst *inst);
 503    void generate_pixel_xy(struct brw_reg dst, bool is_x);
 504    void generate_linterp(fs_inst *inst, struct brw_reg dst,
 505                          struct brw_reg *src);
 506    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src);
 507    void generate_math(fs_inst *inst, struct brw_reg dst, struct brw_reg *src);
 508    void generate_discard(fs_inst *inst);
 509    void generate_ddx(fs_inst *inst, struct brw_reg dst, struct brw_reg src);
 510    void generate_ddy(fs_inst *inst, struct brw_reg dst, struct brw_reg src);
 511    void generate_spill(fs_inst *inst, struct brw_reg src);
 512    void generate_unspill(fs_inst *inst, struct brw_reg dst);
 513    void generate_pull_constant_load(fs_inst *inst, struct brw_reg dst);
 514
 515    void emit_dummy_fs();
 516    fs_reg *emit_fragcoord_interpolation(ir_variable *ir);
 517    fs_reg *emit_frontfacing_interpolation(ir_variable *ir);
 518    fs_reg *emit_general_interpolation(ir_variable *ir);
 519    void emit_interpolation_setup_gen4();
 520    void emit_interpolation_setup_gen6();
 521    fs_inst *emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
 522                               int sampler);
 523    fs_inst *emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
 524                               int sampler);
 525    fs_inst *emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
 526                               int sampler);
 527    fs_inst *emit_math(fs_opcodes op, fs_reg dst, fs_reg src0);
 528    fs_inst *emit_math(fs_opcodes op, fs_reg dst, fs_reg src0, fs_reg src1);
 529    bool try_emit_saturate(ir_expression *ir);
 530    void emit_bool_to_cond_code(ir_rvalue *condition);
 531    void emit_if_gen6(ir_if *ir);
 532    void emit_unspill(fs_inst *inst, fs_reg reg, uint32_t spill_offset);
 533
 534    void emit_color_write(int index, int first_color_mrf, fs_reg color);
 535    void emit_fb_writes();
 536    void emit_assignment_writes(fs_reg &l, fs_reg &r,
 537                                const glsl_type *type, bool predicated);
 538
 539    struct brw_reg interp_reg(int location, int channel);
 540    int setup_uniform_values(int loc, const glsl_type *type);
 541    void setup_builtin_uniform_values(ir_variable *ir);
 542    int implied_mrf_writes(fs_inst *inst);
 543
 544    struct brw_context *brw;
 545    const struct gl_fragment_program *fp;
 546    struct intel_context *intel;
 547    struct gl_context *ctx;
 548    struct brw_wm_compile *c;
 549    struct brw_compile *p;
 550    struct brw_shader *shader;
 551    struct gl_shader_program *prog;
 552    void *mem_ctx;
 553    exec_list instructions;
 554
 555    /* Delayed setup of c->prog_data.params[] due to realloc of
 556     * ParamValues[] during compile.
 557     */
 558    int param_index[MAX_UNIFORMS * 4];
 559    int param_offset[MAX_UNIFORMS * 4];
 560
 561    int *virtual_grf_sizes;
 562    int virtual_grf_next;
 563    int virtual_grf_array_size;
 564    int *virtual_grf_def;
 565    int *virtual_grf_use;
 566    bool live_intervals_valid;
 567
 568    struct hash_table *variable_ht;
 569    ir_variable *frag_color, *frag_data, *frag_depth;
 570    int first_non_payload_grf;
 571    int urb_setup[FRAG_ATTRIB_MAX];
 572    bool kill_emitted;
 573
 574    /** @{ debug annotation info */
 575    const char *current_annotation;
 576    ir_instruction *base_ir;
 577    /** @} */
 578
 579    bool failed;
 580    char *fail_msg;
 581
 582    /* On entry to a visit() method, this is the storage for the
 583     * result.  On exit, the visit() called may have changed it, in
 584     * which case the parent must use the new storage instead.
 585     */
 586    fs_reg result;
 587
 588    fs_reg pixel_x;
 589    fs_reg pixel_y;
 590    fs_reg wpos_w;
 591    fs_reg pixel_w;
 592    fs_reg delta_x;
 593    fs_reg delta_y;
 594    fs_reg reg_null_cmp;
 595
 596    int grf_used;
 597
 598    int force_uncompressed_stack;
 599    int force_sechalf_stack;
 600 };
 601
 602 GLboolean brw_do_channel_expressions(struct exec_list *instructions);
 603 GLboolean brw_do_vector_splitting(struct exec_list *instructions);
 604 bool brw_fs_precompile(struct gl_context *ctx, struct gl_shader_program *prog);