src/gallium/drivers/lima/ir/pp/ppir.h

   1 /*
   2  * Copyright (c) 2017 Lima Project
   3  * Copyright (c) 2013 Connor Abbott
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a copy
   6  * of this software and associated documentation files (the "Software"), to deal
   7  * in the Software without restriction, including without limitation the rights
   8  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
   9  * copies of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the 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 THE
  18  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  21  * THE SOFTWARE.
  22  *
  23  */
  24
  25 #ifndef LIMA_IR_PP_PPIR_H
  26 #define LIMA_IR_PP_PPIR_H
  27
  28 #include "util/u_math.h"
  29 #include "util/list.h"
  30 #include "util/set.h"
  31
  32 #include "ir/lima_ir.h"
  33
  34 typedef enum {
  35    ppir_op_mov,
  36    ppir_op_abs,
  37    ppir_op_neg,
  38    ppir_op_sat,
  39    ppir_op_add,
  40
  41    ppir_op_ddx,
  42    ppir_op_ddy,
  43
  44    ppir_op_mul,
  45    ppir_op_rcp,
  46
  47    ppir_op_sin_lut,
  48    ppir_op_cos_lut,
  49
  50    ppir_op_sum3,
  51    ppir_op_sum4,
  52
  53    ppir_op_normalize2,
  54    ppir_op_normalize3,
  55    ppir_op_normalize4,
  56
  57    ppir_op_select,
  58
  59    ppir_op_sin,
  60    ppir_op_cos,
  61    ppir_op_tan,
  62    ppir_op_asin,
  63    ppir_op_acos,
  64
  65    ppir_op_atan,
  66    ppir_op_atan2,
  67    ppir_op_atan_pt1,
  68    ppir_op_atan2_pt1,
  69    ppir_op_atan_pt2,
  70
  71    ppir_op_exp,
  72    ppir_op_log,
  73    ppir_op_exp2,
  74    ppir_op_log2,
  75    ppir_op_sqrt,
  76    ppir_op_rsqrt,
  77
  78    ppir_op_sign,
  79    ppir_op_floor,
  80    ppir_op_ceil,
  81    ppir_op_fract,
  82    ppir_op_mod,
  83    ppir_op_min,
  84    ppir_op_max,
  85    ppir_op_trunc,
  86
  87    ppir_op_and,
  88    ppir_op_or,
  89    ppir_op_xor,
  90
  91    ppir_op_lt,
  92    ppir_op_gt,
  93    ppir_op_le,
  94    ppir_op_ge,
  95    ppir_op_eq,
  96    ppir_op_ne,
  97    ppir_op_not,
  98
  99    ppir_op_load_uniform,
 100    ppir_op_load_varying,
 101    ppir_op_load_coords,
 102    ppir_op_load_coords_reg,
 103    ppir_op_load_fragcoord,
 104    ppir_op_load_pointcoord,
 105    ppir_op_load_frontface,
 106    ppir_op_load_texture,
 107    ppir_op_load_temp,
 108
 109    ppir_op_store_temp,
 110
 111    ppir_op_const,
 112
 113    ppir_op_discard,
 114    ppir_op_branch,
 115
 116    ppir_op_undef,
 117    ppir_op_dummy,
 118
 119    ppir_op_num,
 120 } ppir_op;
 121
 122 typedef enum {
 123    ppir_node_type_alu,
 124    ppir_node_type_const,
 125    ppir_node_type_load,
 126    ppir_node_type_store,
 127    ppir_node_type_load_texture,
 128    ppir_node_type_discard,
 129    ppir_node_type_branch,
 130 } ppir_node_type;
 131
 132 typedef struct {
 133    char *name;
 134    ppir_node_type type;
 135    int *slots;
 136 } ppir_op_info;
 137
 138 extern const ppir_op_info ppir_op_infos[];
 139
 140 typedef enum {
 141    ppir_dep_src,
 142    ppir_dep_write_after_read,
 143    ppir_dep_sequence,
 144 } ppir_dep_type;
 145
 146 typedef struct {
 147    void *pred, *succ;
 148    ppir_dep_type type;
 149    struct list_head pred_link;
 150    struct list_head succ_link;
 151 } ppir_dep;
 152
 153 typedef struct ppir_node {
 154    struct list_head list;
 155    ppir_op op;
 156    ppir_node_type type;
 157    int index;
 158    char name[16];
 159    bool printed;
 160    struct ppir_instr *instr;
 161    int instr_pos;
 162    struct ppir_block *block;
 163    bool is_end;
 164
 165    /* for scheduler */
 166    struct list_head succ_list;
 167    struct list_head pred_list;
 168 } ppir_node;
 169
 170 typedef enum {
 171    ppir_pipeline_reg_const0,
 172    ppir_pipeline_reg_const1,
 173    ppir_pipeline_reg_sampler,
 174    ppir_pipeline_reg_uniform,
 175    ppir_pipeline_reg_vmul,
 176    ppir_pipeline_reg_fmul,
 177    ppir_pipeline_reg_discard, /* varying load */
 178 } ppir_pipeline;
 179
 180 typedef struct ppir_reg {
 181    struct list_head list;
 182    int index;
 183    int regalloc_index;
 184    int num_components;
 185
 186    /* whether this reg has to start from the x component
 187     * of a full physical reg, this is true for reg used
 188     * in load/store instr which has no swizzle field */
 189    bool is_head;
 190    bool spilled;
 191    bool undef;
 192 } ppir_reg;
 193
 194 typedef enum {
 195    ppir_target_ssa,
 196    ppir_target_pipeline,
 197    ppir_target_register,
 198 } ppir_target;
 199
 200 typedef struct ppir_src {
 201    ppir_target type;
 202    ppir_node *node;
 203
 204    union {
 205       ppir_reg *ssa;
 206       ppir_reg *reg;
 207       ppir_pipeline pipeline;
 208    };
 209
 210    uint8_t swizzle[4];
 211    bool absolute, negate;
 212 } ppir_src;
 213
 214 typedef enum {
 215    ppir_outmod_none,
 216    ppir_outmod_clamp_fraction,
 217    ppir_outmod_clamp_positive,
 218    ppir_outmod_round,
 219 } ppir_outmod;
 220
 221 typedef struct ppir_dest {
 222    ppir_target type;
 223
 224    union {
 225       ppir_reg ssa;
 226       ppir_reg *reg;
 227       ppir_pipeline pipeline;
 228    };
 229
 230    ppir_outmod modifier;
 231    unsigned write_mask : 4;
 232 } ppir_dest;
 233
 234 typedef struct {
 235    ppir_node node;
 236    ppir_dest dest;
 237    ppir_src src[3];
 238    int num_src;
 239    int shift : 3; /* Only used for ppir_op_mul */
 240 } ppir_alu_node;
 241
 242 typedef struct ppir_const {
 243    union fi value[4];
 244    int num;
 245 } ppir_const;
 246
 247 typedef struct {
 248    ppir_node node;
 249    ppir_const constant;
 250    ppir_dest dest;
 251 } ppir_const_node;
 252
 253 typedef struct {
 254    ppir_node node;
 255    int index;
 256    int num_components;
 257    ppir_dest dest;
 258    ppir_src src;
 259    int num_src;
 260 } ppir_load_node;
 261
 262 typedef struct {
 263    ppir_node node;
 264    int index;
 265    int num_components;
 266    ppir_src src;
 267 } ppir_store_node;
 268
 269 typedef struct {
 270    ppir_node node;
 271    ppir_dest dest;
 272    ppir_src src[2]; /* src[0] temporarily stores src_coords,
 273                        not to be used after lowering */
 274    int num_src;
 275    int sampler;
 276    int sampler_dim;
 277    bool lod_bias_en;
 278    bool explicit_lod;
 279 } ppir_load_texture_node;
 280
 281 typedef struct {
 282    ppir_node node;
 283 } ppir_discard_node;
 284
 285 enum ppir_instr_slot {
 286    PPIR_INSTR_SLOT_VARYING,
 287    PPIR_INSTR_SLOT_TEXLD,
 288    PPIR_INSTR_SLOT_UNIFORM,
 289    PPIR_INSTR_SLOT_ALU_VEC_MUL,
 290    PPIR_INSTR_SLOT_ALU_SCL_MUL,
 291    PPIR_INSTR_SLOT_ALU_VEC_ADD,
 292    PPIR_INSTR_SLOT_ALU_SCL_ADD,
 293    PPIR_INSTR_SLOT_ALU_COMBINE,
 294    PPIR_INSTR_SLOT_STORE_TEMP,
 295    PPIR_INSTR_SLOT_BRANCH,
 296    PPIR_INSTR_SLOT_NUM,
 297    PPIR_INSTR_SLOT_END,
 298    PPIR_INSTR_SLOT_ALU_START = PPIR_INSTR_SLOT_ALU_VEC_MUL,
 299    PPIR_INSTR_SLOT_ALU_END = PPIR_INSTR_SLOT_ALU_COMBINE,
 300 };
 301
 302 struct ppir_liveness {
 303    ppir_reg *reg;
 304    unsigned mask : 4;
 305 };
 306
 307 typedef struct ppir_instr {
 308    struct list_head list;
 309    int index;
 310    bool printed;
 311    int seq; /* command sequence after schedule */
 312
 313    ppir_node *slots[PPIR_INSTR_SLOT_NUM];
 314    ppir_const constant[2];
 315    bool is_end;
 316
 317    /* for scheduler */
 318    struct list_head succ_list;
 319    struct list_head pred_list;
 320    float reg_pressure;
 321    int est; /* earliest start time */
 322    int parent_index;
 323    bool scheduled;
 324    int offset;
 325    int encode_size;
 326
 327    /* for liveness analysis */
 328    struct ppir_liveness *live_in;
 329    struct ppir_liveness *live_out;
 330    /* live_internal is to mark registers only live within an
 331     * instruction, without propagation */
 332    struct ppir_liveness *live_internal;
 333    struct set *live_in_set;
 334    struct set *live_out_set;
 335    struct set *live_internal_set;
 336 } ppir_instr;
 337
 338 typedef struct ppir_block {
 339    struct list_head list;
 340    struct list_head node_list;
 341    struct list_head instr_list;
 342
 343    struct ppir_block *successors[2];
 344
 345    struct ppir_compiler *comp;
 346
 347    /* for scheduler */
 348    int sched_instr_index;
 349    int sched_instr_base;
 350    int index;
 351
 352    /* for liveness analysis */
 353    struct ppir_liveness *live_in;
 354    struct ppir_liveness *live_out;
 355    struct set *live_in_set;
 356    struct set *live_out_set;
 357 } ppir_block;
 358
 359 typedef struct {
 360    ppir_node node;
 361    ppir_src src[2];
 362    int num_src;
 363    bool cond_gt;
 364    bool cond_eq;
 365    bool cond_lt;
 366    bool negate;
 367    ppir_block *target;
 368 } ppir_branch_node;
 369
 370 struct ra_regs;
 371 struct lima_fs_shader_state;
 372
 373 typedef struct ppir_compiler {
 374    struct list_head block_list;
 375    struct hash_table_u64 *blocks;
 376    int cur_index;
 377    int cur_instr_index;
 378
 379    struct list_head reg_list;
 380
 381    /* array for searching ssa/reg node */
 382    ppir_node **var_nodes;
 383    unsigned reg_base;
 384
 385    struct ra_regs *ra;
 386    struct lima_fs_shader_state *prog;
 387    bool uses_discard;
 388
 389    /* for scheduler */
 390    int sched_instr_base;
 391
 392    /* for regalloc spilling debug */
 393    int force_spilling;
 394
 395    /* shaderdb */
 396    int num_loops;
 397    int num_spills;
 398    int num_fills;
 399
 400    ppir_block *discard_block;
 401    ppir_block *current_block;
 402    ppir_block *loop_break_block;
 403    ppir_block *loop_cont_block;
 404 } ppir_compiler;
 405
 406 void *ppir_node_create(ppir_block *block, ppir_op op, int index, unsigned mask);
 407 void ppir_node_add_dep(ppir_node *succ, ppir_node *pred, ppir_dep_type type);
 408 void ppir_node_remove_dep(ppir_dep *dep);
 409 void ppir_node_delete(ppir_node *node);
 410 void ppir_node_print_prog(ppir_compiler *comp);
 411 void ppir_node_replace_child(ppir_node *parent, ppir_node *old_child, ppir_node *new_child);
 412 void ppir_node_replace_all_succ(ppir_node *dst, ppir_node *src);
 413 void ppir_node_replace_pred(ppir_dep *dep, ppir_node *new_pred);
 414 ppir_dep *ppir_dep_for_pred(ppir_node *node, ppir_node *pred);
 415 /* Assumes that node successors are in the same block */
 416 ppir_node *ppir_node_insert_mov(ppir_node *node);
 417 ppir_node *ppir_node_insert_mov_all_blocks(ppir_node *node);
 418
 419 static inline bool ppir_node_is_root(ppir_node *node)
 420 {
 421    return list_is_empty(&node->succ_list);
 422 }
 423
 424 static inline bool ppir_node_is_leaf(ppir_node *node)
 425 {
 426    return list_is_empty(&node->pred_list);
 427 }
 428
 429 static inline bool ppir_node_has_single_succ(ppir_node *node)
 430 {
 431    return list_is_singular(&node->succ_list);
 432 }
 433
 434 bool ppir_node_has_single_src_succ(ppir_node *node);
 435
 436 static inline ppir_node *ppir_node_first_succ(ppir_node *node)
 437 {
 438    return list_first_entry(&node->succ_list, ppir_dep, succ_link)->succ;
 439 }
 440
 441 static inline bool ppir_node_has_single_pred(ppir_node *node)
 442 {
 443    return list_is_singular(&node->pred_list);
 444 }
 445
 446 static inline ppir_node *ppir_node_first_pred(ppir_node *node)
 447 {
 448    return list_first_entry(&node->pred_list, ppir_dep, pred_link)->pred;
 449 }
 450
 451 #define ppir_node_foreach_succ(node, dep) \
 452    list_for_each_entry(ppir_dep, dep, &node->succ_list, succ_link)
 453 #define ppir_node_foreach_succ_safe(node, dep) \
 454    list_for_each_entry_safe(ppir_dep, dep, &node->succ_list, succ_link)
 455 #define ppir_node_foreach_pred(node, dep) \
 456    list_for_each_entry(ppir_dep, dep, &node->pred_list, pred_link)
 457 #define ppir_node_foreach_pred_safe(node, dep) \
 458    list_for_each_entry_safe(ppir_dep, dep, &node->pred_list, pred_link)
 459
 460 #define ppir_node_to_alu(node) ((ppir_alu_node *)(node))
 461 #define ppir_node_to_const(node) ((ppir_const_node *)(node))
 462 #define ppir_node_to_load(node) ((ppir_load_node *)(node))
 463 #define ppir_node_to_store(node) ((ppir_store_node *)(node))
 464 #define ppir_node_to_load_texture(node) ((ppir_load_texture_node *)(node))
 465 #define ppir_node_to_discard(node) ((ppir_discard_node *)(node))
 466 #define ppir_node_to_branch(node) ((ppir_branch_node *)(node))
 467
 468 static inline ppir_dest *ppir_node_get_dest(ppir_node *node)
 469 {
 470    switch (node->type) {
 471    case ppir_node_type_alu:
 472       return &ppir_node_to_alu(node)->dest;
 473    case ppir_node_type_load:
 474       return &ppir_node_to_load(node)->dest;
 475    case ppir_node_type_const:
 476       return &ppir_node_to_const(node)->dest;
 477    case ppir_node_type_load_texture:
 478       return &ppir_node_to_load_texture(node)->dest;
 479    default:
 480       return NULL;
 481    }
 482 }
 483
 484 static inline int ppir_node_get_src_num(ppir_node *node)
 485 {
 486    switch (node->type) {
 487    case ppir_node_type_alu:
 488       return ppir_node_to_alu(node)->num_src;
 489    case ppir_node_type_branch:
 490       return ppir_node_to_branch(node)->num_src;
 491    case ppir_node_type_load:
 492       return ppir_node_to_load(node)->num_src;
 493    case ppir_node_type_load_texture:
 494       return ppir_node_to_load_texture(node)->num_src;
 495    case ppir_node_type_store:
 496       return 1;
 497    default:
 498       return 0;
 499    }
 500
 501    return 0;
 502 }
 503
 504 static inline ppir_src *ppir_node_get_src(ppir_node *node, int idx)
 505 {
 506    if (idx < 0 || idx >= ppir_node_get_src_num(node))
 507       return NULL;
 508
 509    switch (node->type) {
 510    case ppir_node_type_alu:
 511       return &ppir_node_to_alu(node)->src[idx];
 512    case ppir_node_type_branch:
 513       return &ppir_node_to_branch(node)->src[idx];
 514    case ppir_node_type_load_texture:
 515       return &ppir_node_to_load_texture(node)->src[idx];
 516    case ppir_node_type_load:
 517       return &ppir_node_to_load(node)->src;
 518    case ppir_node_type_store:
 519       return &ppir_node_to_store(node)->src;
 520    default:
 521       break;
 522    }
 523
 524    return NULL;
 525 }
 526
 527 static inline ppir_reg *ppir_src_get_reg(ppir_src *src)
 528 {
 529    switch (src->type) {
 530    case ppir_target_ssa:
 531       return src->ssa;
 532    case ppir_target_register:
 533       return src->reg;
 534    default:
 535       return NULL;
 536    }
 537 }
 538
 539 static inline ppir_reg *ppir_dest_get_reg(ppir_dest *dest)
 540 {
 541    switch (dest->type) {
 542    case ppir_target_ssa:
 543       return &dest->ssa;
 544    case ppir_target_register:
 545       return dest->reg;
 546    default:
 547       return NULL;
 548    }
 549 }
 550
 551 static inline void ppir_node_target_assign(ppir_src *src, ppir_node *node)
 552 {
 553    ppir_dest *dest = ppir_node_get_dest(node);
 554    src->type = dest->type;
 555    switch (src->type) {
 556    case ppir_target_ssa:
 557       src->ssa = &dest->ssa;
 558       src->node = node;
 559       break;
 560    case ppir_target_register:
 561       src->reg = dest->reg;
 562       /* Registers can be assigned from multiple nodes, so don't keep
 563        * pointer to the node here
 564        */
 565       src->node = NULL;
 566       break;
 567    case ppir_target_pipeline:
 568       src->pipeline = dest->pipeline;
 569       src->node = node;
 570       break;
 571    }
 572 }
 573
 574 static inline bool ppir_node_target_equal(ppir_src *src, ppir_dest *dest)
 575 {
 576    if (src->type != dest->type ||
 577        (src->type == ppir_target_ssa && src->ssa != &dest->ssa) ||
 578        (src->type == ppir_target_register && src->reg != dest->reg) ||
 579        (src->type == ppir_target_pipeline && src->pipeline != dest->pipeline))
 580       return false;
 581
 582    return true;
 583 }
 584
 585 static inline int ppir_target_get_src_reg_index(ppir_src *src)
 586 {
 587    switch (src->type) {
 588    case ppir_target_ssa:
 589       if (src->ssa)
 590          return src->ssa->index;
 591       break;
 592    case ppir_target_register:
 593       if (src->reg)
 594          return src->reg->index;
 595       break;
 596    case ppir_target_pipeline:
 597       if (src->pipeline == ppir_pipeline_reg_discard)
 598          return 15 * 4;
 599       return (src->pipeline + 12) * 4;
 600    }
 601
 602    return -1;
 603 }
 604
 605 static inline int ppir_target_get_dest_reg_index(ppir_dest *dest)
 606 {
 607    switch (dest->type) {
 608    case ppir_target_ssa:
 609       return dest->ssa.index;
 610    case ppir_target_register:
 611       return dest->reg->index;
 612    case ppir_target_pipeline:
 613       if (dest->pipeline == ppir_pipeline_reg_discard)
 614          return 15 * 4;
 615       return (dest->pipeline + 12) * 4;
 616    }
 617
 618    return -1;
 619 }
 620
 621 static inline int ppir_src_get_mask(ppir_src *src)
 622 {
 623    ppir_reg *reg = ppir_src_get_reg(src);
 624    int mask = 0;
 625
 626    for (int i = 0; i < reg->num_components; i++)
 627       mask |= (1 << src->swizzle[i]);
 628
 629    return mask;
 630 }
 631
 632 static inline bool ppir_target_is_scalar(ppir_dest *dest)
 633 {
 634    switch (dest->type) {
 635    case ppir_target_ssa:
 636       return dest->ssa.num_components == 1;
 637    case ppir_target_register:
 638       /* only one bit in mask is set */
 639       if ((dest->write_mask & 0x3) == 0x3 ||
 640           (dest->write_mask & 0x5) == 0x5 ||
 641           (dest->write_mask & 0x9) == 0x9 ||
 642           (dest->write_mask & 0x6) == 0x6 ||
 643           (dest->write_mask & 0xa) == 0xa ||
 644           (dest->write_mask & 0xc) == 0xc)
 645          return false;
 646       else
 647          return true;
 648    case ppir_target_pipeline:
 649       if (dest->pipeline == ppir_pipeline_reg_fmul)
 650          return true;
 651       else
 652          return false;
 653    default:
 654       return false;
 655    }
 656 }
 657
 658 static inline bool ppir_node_schedulable_slot(ppir_node *node,
 659                                               enum ppir_instr_slot slot)
 660 {
 661    int *slots = ppir_op_infos[node->op].slots;
 662    for (int i = 0; slots[i] != PPIR_INSTR_SLOT_END; i++)
 663       if (slots[i] == slot)
 664          return true;
 665
 666    return false;
 667 }
 668
 669 ppir_instr *ppir_instr_create(ppir_block *block);
 670 bool ppir_instr_insert_node(ppir_instr *instr, ppir_node *node);
 671 void ppir_instr_add_dep(ppir_instr *succ, ppir_instr *pred);
 672 void ppir_instr_print_list(ppir_compiler *comp);
 673 void ppir_instr_print_dep(ppir_compiler *comp);
 674 void ppir_instr_insert_mul_node(ppir_node *add, ppir_node *mul);
 675
 676 #define ppir_instr_foreach_succ(instr, dep) \
 677    list_for_each_entry(ppir_dep, dep, &instr->succ_list, succ_link)
 678 #define ppir_instr_foreach_succ_safe(instr, dep) \
 679    list_for_each_entry_safe(ppir_dep, dep, &instr->succ_list, succ_link)
 680 #define ppir_instr_foreach_pred(instr, dep) \
 681    list_for_each_entry(ppir_dep, dep, &instr->pred_list, pred_link)
 682 #define ppir_instr_foreach_pred_safe(instr, dep) \
 683    list_for_each_entry_safe(ppir_dep, dep, &instr->pred_list, pred_link)
 684
 685 static inline bool ppir_instr_is_root(ppir_instr *instr)
 686 {
 687    return list_is_empty(&instr->succ_list);
 688 }
 689
 690 static inline bool ppir_instr_is_leaf(ppir_instr *instr)
 691 {
 692    return list_is_empty(&instr->pred_list);
 693 }
 694
 695 bool ppir_lower_prog(ppir_compiler *comp);
 696 bool ppir_node_to_instr(ppir_compiler *comp);
 697 bool ppir_schedule_prog(ppir_compiler *comp);
 698 bool ppir_regalloc_prog(ppir_compiler *comp);
 699 bool ppir_codegen_prog(ppir_compiler *comp);
 700 void ppir_liveness_analysis(ppir_compiler *comp);
 701
 702 #endif