src/amd/compiler/aco_insert_waitcnt.cpp

   1 /*
   2  * Copyright © 2018 Valve 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  */
  24
  25 #include <algorithm>
  26 #include <map>
  27 #include <stack>
  28 #include <math.h>
  29
  30 #include "aco_ir.h"
  31 #include "vulkan/radv_shader.h"
  32
  33 namespace aco {
  34
  35 namespace {
  36
  37 /**
  38  * The general idea of this pass is:
  39  * The CFG is traversed in reverse postorder (forward) and loops are processed
  40  * several times until no progress is made.
  41  * Per BB two wait_ctx is maintained: an in-context and out-context.
  42  * The in-context is the joined out-contexts of the predecessors.
  43  * The context contains a map: gpr -> wait_entry
  44  * consisting of the information about the cnt values to be waited for.
  45  * Note: After merge-nodes, it might occur that for the same register
  46  *       multiple cnt values are to be waited for.
  47  *
  48  * The values are updated according to the encountered instructions:
  49  * - additional events increment the counter of waits of the same type
  50  * - or erase gprs with counters higher than to be waited for.
  51  */
  52
  53 // TODO: do a more clever insertion of wait_cnt (lgkm_cnt) when there is a load followed by a use of a previous load
  54
  55 /* Instructions of the same event will finish in-order except for smem
  56  * and maybe flat. Instructions of different events may not finish in-order. */
  57 enum wait_event : uint16_t {
  58    event_smem = 1 << 0,
  59    event_lds = 1 << 1,
  60    event_gds = 1 << 2,
  61    event_vmem = 1 << 3,
  62    event_vmem_store = 1 << 4, /* GFX10+ */
  63    event_flat = 1 << 5,
  64    event_exp_pos = 1 << 6,
  65    event_exp_param = 1 << 7,
  66    event_exp_mrt_null = 1 << 8,
  67    event_gds_gpr_lock = 1 << 9,
  68    event_vmem_gpr_lock = 1 << 10,
  69    event_sendmsg = 1 << 11,
  70    num_events = 12,
  71 };
  72
  73 enum counter_type : uint8_t {
  74    counter_exp = 1 << 0,
  75    counter_lgkm = 1 << 1,
  76    counter_vm = 1 << 2,
  77    counter_vs = 1 << 3,
  78    num_counters = 4,
  79 };
  80
  81 static const uint16_t exp_events = event_exp_pos | event_exp_param | event_exp_mrt_null | event_gds_gpr_lock | event_vmem_gpr_lock;
  82 static const uint16_t lgkm_events = event_smem | event_lds | event_gds | event_flat | event_sendmsg;
  83 static const uint16_t vm_events = event_vmem | event_flat;
  84 static const uint16_t vs_events = event_vmem_store;
  85
  86 uint8_t get_counters_for_event(wait_event ev)
  87 {
  88    switch (ev) {
  89    case event_smem:
  90    case event_lds:
  91    case event_gds:
  92    case event_sendmsg:
  93       return counter_lgkm;
  94    case event_vmem:
  95       return counter_vm;
  96    case event_vmem_store:
  97       return counter_vs;
  98    case event_flat:
  99       return counter_vm | counter_lgkm;
 100    case event_exp_pos:
 101    case event_exp_param:
 102    case event_exp_mrt_null:
 103    case event_gds_gpr_lock:
 104    case event_vmem_gpr_lock:
 105       return counter_exp;
 106    default:
 107       return 0;
 108    }
 109 }
 110
 111 uint16_t get_events_for_counter(counter_type ctr)
 112 {
 113    switch (ctr) {
 114    case counter_exp:
 115       return exp_events;
 116    case counter_lgkm:
 117       return lgkm_events;
 118    case counter_vm:
 119       return vm_events;
 120    case counter_vs:
 121       return vs_events;
 122    }
 123    return 0;
 124 }
 125
 126 struct wait_imm {
 127    static const uint8_t unset_counter = 0xff;
 128
 129    uint8_t vm;
 130    uint8_t exp;
 131    uint8_t lgkm;
 132    uint8_t vs;
 133
 134    wait_imm() :
 135       vm(unset_counter), exp(unset_counter), lgkm(unset_counter), vs(unset_counter) {}
 136    wait_imm(uint16_t vm_, uint16_t exp_, uint16_t lgkm_, uint16_t vs_) :
 137       vm(vm_), exp(exp_), lgkm(lgkm_), vs(vs_) {}
 138
 139    wait_imm(enum chip_class chip, uint16_t packed) : vs(unset_counter)
 140    {
 141       vm = packed & 0xf;
 142       if (chip >= GFX9)
 143          vm |= (packed >> 10) & 0x30;
 144
 145       exp = (packed >> 4) & 0x7;
 146
 147       lgkm = (packed >> 8) & 0xf;
 148       if (chip >= GFX10)
 149          lgkm |= (packed >> 8) & 0x30;
 150    }
 151
 152    uint16_t pack(enum chip_class chip) const
 153    {
 154       uint16_t imm = 0;
 155       assert(exp == unset_counter || exp <= 0x7);
 156       switch (chip) {
 157       case GFX10:
 158          assert(lgkm == unset_counter || lgkm <= 0x3f);
 159          assert(vm == unset_counter || vm <= 0x3f);
 160          imm = ((vm & 0x30) << 10) | ((lgkm & 0x3f) << 8) | ((exp & 0x7) << 4) | (vm & 0xf);
 161          break;
 162       case GFX9:
 163          assert(lgkm == unset_counter || lgkm <= 0xf);
 164          assert(vm == unset_counter || vm <= 0x3f);
 165          imm = ((vm & 0x30) << 10) | ((lgkm & 0xf) << 8) | ((exp & 0x7) << 4) | (vm & 0xf);
 166          break;
 167       default:
 168          assert(lgkm == unset_counter || lgkm <= 0xf);
 169          assert(vm == unset_counter || vm <= 0xf);
 170          imm = ((lgkm & 0xf) << 8) | ((exp & 0x7) << 4) | (vm & 0xf);
 171          break;
 172       }
 173       if (chip < GFX9 && vm == wait_imm::unset_counter)
 174          imm |= 0xc000; /* should have no effect on pre-GFX9 and now we won't have to worry about the architecture when interpreting the immediate */
 175       if (chip < GFX10 && lgkm == wait_imm::unset_counter)
 176          imm |= 0x3000; /* should have no effect on pre-GFX10 and now we won't have to worry about the architecture when interpreting the immediate */
 177       return imm;
 178    }
 179
 180    bool combine(const wait_imm& other)
 181    {
 182       bool changed = other.vm < vm || other.exp < exp || other.lgkm < lgkm || other.vs < vs;
 183       vm = std::min(vm, other.vm);
 184       exp = std::min(exp, other.exp);
 185       lgkm = std::min(lgkm, other.lgkm);
 186       vs = std::min(vs, other.vs);
 187       return changed;
 188    }
 189
 190    bool empty() const
 191    {
 192       return vm == unset_counter && exp == unset_counter &&
 193              lgkm == unset_counter && vs == unset_counter;
 194    }
 195 };
 196
 197 struct wait_entry {
 198    wait_imm imm;
 199    uint16_t events; /* use wait_event notion */
 200    uint8_t counters; /* use counter_type notion */
 201    bool wait_on_read:1;
 202    bool logical:1;
 203
 204    wait_entry(wait_event event, wait_imm imm, bool logical, bool wait_on_read)
 205            : imm(imm), events(event), counters(get_counters_for_event(event)),
 206              wait_on_read(wait_on_read), logical(logical) {}
 207
 208    bool join(const wait_entry& other)
 209    {
 210       bool changed = (other.events & ~events) ||
 211                      (other.counters & ~counters) ||
 212                      (other.wait_on_read && !wait_on_read);
 213       events |= other.events;
 214       counters |= other.counters;
 215       changed |= imm.combine(other.imm);
 216       wait_on_read = wait_on_read || other.wait_on_read;
 217       assert(logical == other.logical);
 218       return changed;
 219    }
 220
 221    void remove_counter(counter_type counter)
 222    {
 223       counters &= ~counter;
 224
 225       if (counter == counter_lgkm) {
 226          imm.lgkm = wait_imm::unset_counter;
 227          events &= ~(event_smem | event_lds | event_gds | event_sendmsg);
 228       }
 229
 230       if (counter == counter_vm) {
 231          imm.vm = wait_imm::unset_counter;
 232          events &= ~event_vmem;
 233       }
 234
 235       if (counter == counter_exp) {
 236          imm.exp = wait_imm::unset_counter;
 237          events &= ~(event_exp_pos | event_exp_param | event_exp_mrt_null | event_gds_gpr_lock | event_vmem_gpr_lock);
 238       }
 239
 240       if (counter == counter_vs) {
 241          imm.vs = wait_imm::unset_counter;
 242          events &= ~event_vmem_store;
 243       }
 244
 245       if (!(counters & counter_lgkm) && !(counters & counter_vm))
 246          events &= ~event_flat;
 247    }
 248 };
 249
 250 struct wait_ctx {
 251    Program *program;
 252    enum chip_class chip_class;
 253    uint16_t max_vm_cnt;
 254    uint16_t max_exp_cnt;
 255    uint16_t max_lgkm_cnt;
 256    uint16_t max_vs_cnt;
 257    uint16_t unordered_events = event_smem | event_flat;
 258
 259    uint8_t vm_cnt = 0;
 260    uint8_t exp_cnt = 0;
 261    uint8_t lgkm_cnt = 0;
 262    uint8_t vs_cnt = 0;
 263    bool pending_flat_lgkm = false;
 264    bool pending_flat_vm = false;
 265    bool pending_s_buffer_store = false; /* GFX10 workaround */
 266
 267    wait_imm barrier_imm[barrier_count];
 268    uint16_t barrier_events[barrier_count] = {}; /* use wait_event notion */
 269
 270    std::map<PhysReg,wait_entry> gpr_map;
 271
 272    /* used for vmem/smem scores */
 273    bool collect_statistics;
 274    Instruction *gen_instr;
 275    std::map<Instruction *, unsigned> unwaited_instrs[num_counters];
 276    std::map<PhysReg,std::set<Instruction *>> reg_instrs[num_counters];
 277    std::vector<unsigned> wait_distances[num_events];
 278
 279    wait_ctx() {}
 280    wait_ctx(Program *program_)
 281            : program(program_),
 282              chip_class(program_->chip_class),
 283              max_vm_cnt(program_->chip_class >= GFX9 ? 62 : 14),
 284              max_exp_cnt(6),
 285              max_lgkm_cnt(program_->chip_class >= GFX10 ? 62 : 14),
 286              max_vs_cnt(program_->chip_class >= GFX10 ? 62 : 0),
 287              unordered_events(event_smem | (program_->chip_class < GFX10 ? event_flat : 0)) {}
 288
 289    bool join(const wait_ctx* other, bool logical)
 290    {
 291       bool changed = other->exp_cnt > exp_cnt ||
 292                      other->vm_cnt > vm_cnt ||
 293                      other->lgkm_cnt > lgkm_cnt ||
 294                      other->vs_cnt > vs_cnt ||
 295                      (other->pending_flat_lgkm && !pending_flat_lgkm) ||
 296                      (other->pending_flat_vm && !pending_flat_vm);
 297
 298       exp_cnt = std::max(exp_cnt, other->exp_cnt);
 299       vm_cnt = std::max(vm_cnt, other->vm_cnt);
 300       lgkm_cnt = std::max(lgkm_cnt, other->lgkm_cnt);
 301       vs_cnt = std::max(vs_cnt, other->vs_cnt);
 302       pending_flat_lgkm |= other->pending_flat_lgkm;
 303       pending_flat_vm |= other->pending_flat_vm;
 304       pending_s_buffer_store |= other->pending_s_buffer_store;
 305
 306       for (std::pair<PhysReg,wait_entry> entry : other->gpr_map)
 307       {
 308          std::map<PhysReg,wait_entry>::iterator it = gpr_map.find(entry.first);
 309          if (entry.second.logical != logical)
 310             continue;
 311
 312          if (it != gpr_map.end()) {
 313             changed |= it->second.join(entry.second);
 314          } else {
 315             gpr_map.insert(entry);
 316             changed = true;
 317          }
 318       }
 319
 320       for (unsigned i = 0; i < barrier_count; i++) {
 321          changed |= barrier_imm[i].combine(other->barrier_imm[i]);
 322          changed |= other->barrier_events[i] & ~barrier_events[i];
 323          barrier_events[i] |= other->barrier_events[i];
 324       }
 325
 326       /* these are used for statistics, so don't update "changed" */
 327       for (unsigned i = 0; i < num_counters; i++) {
 328          for (std::pair<Instruction *, unsigned> instr : other->unwaited_instrs[i]) {
 329             auto pos = unwaited_instrs[i].find(instr.first);
 330             if (pos == unwaited_instrs[i].end())
 331                unwaited_instrs[i].insert(instr);
 332             else
 333                pos->second = std::min(pos->second, instr.second);
 334          }
 335          /* don't use a foreach loop to avoid copies */
 336          for (auto it = other->reg_instrs[i].begin(); it != other->reg_instrs[i].end(); ++it)
 337             reg_instrs[i][it->first].insert(it->second.begin(), it->second.end());
 338       }
 339
 340       return changed;
 341    }
 342
 343    void wait_and_remove_from_entry(PhysReg reg, wait_entry& entry, counter_type counter) {
 344       if (collect_statistics && (entry.counters & counter)) {
 345          unsigned counter_idx = ffs(counter) - 1;
 346          for (Instruction *instr : reg_instrs[counter_idx][reg]) {
 347             auto pos = unwaited_instrs[counter_idx].find(instr);
 348             if (pos == unwaited_instrs[counter_idx].end())
 349                continue;
 350
 351             unsigned distance = pos->second;
 352             unsigned events = entry.events & get_events_for_counter(counter);
 353             while (events) {
 354                unsigned event_idx = u_bit_scan(&events);
 355                wait_distances[event_idx].push_back(distance);
 356             }
 357
 358             unwaited_instrs[counter_idx].erase(instr);
 359          }
 360          reg_instrs[counter_idx][reg].clear();
 361       }
 362
 363       entry.remove_counter(counter);
 364    }
 365
 366    void advance_unwaited_instrs()
 367    {
 368       for (unsigned i = 0; i < num_counters; i++) {
 369          for (auto it = unwaited_instrs[i].begin(); it != unwaited_instrs[i].end(); ++it)
 370             it->second++;
 371       }
 372    }
 373 };
 374
 375 wait_imm check_instr(Instruction* instr, wait_ctx& ctx)
 376 {
 377    wait_imm wait;
 378
 379    for (const Operand op : instr->operands) {
 380       if (op.isConstant() || op.isUndefined())
 381          continue;
 382
 383       /* check consecutively read gprs */
 384       for (unsigned j = 0; j < op.size(); j++) {
 385          PhysReg reg{op.physReg() + j};
 386          std::map<PhysReg,wait_entry>::iterator it = ctx.gpr_map.find(reg);
 387          if (it == ctx.gpr_map.end() || !it->second.wait_on_read)
 388             continue;
 389
 390          wait.combine(it->second.imm);
 391       }
 392    }
 393
 394    for (const Definition& def : instr->definitions) {
 395       /* check consecutively written gprs */
 396       for (unsigned j = 0; j < def.getTemp().size(); j++)
 397       {
 398          PhysReg reg{def.physReg() + j};
 399
 400          std::map<PhysReg,wait_entry>::iterator it = ctx.gpr_map.find(reg);
 401          if (it == ctx.gpr_map.end())
 402             continue;
 403
 404          /* Vector Memory reads and writes return in the order they were issued */
 405          if (instr->isVMEM() && ((it->second.events & vm_events) == event_vmem)) {
 406             it->second.remove_counter(counter_vm);
 407             if (!it->second.counters)
 408                it = ctx.gpr_map.erase(it);
 409             continue;
 410          }
 411
 412          /* LDS reads and writes return in the order they were issued. same for GDS */
 413          if (instr->format == Format::DS) {
 414             bool gds = static_cast<DS_instruction*>(instr)->gds;
 415             if ((it->second.events & lgkm_events) == (gds ? event_gds : event_lds)) {
 416                it->second.remove_counter(counter_lgkm);
 417                if (!it->second.counters)
 418                   it = ctx.gpr_map.erase(it);
 419                continue;
 420             }
 421          }
 422
 423          wait.combine(it->second.imm);
 424       }
 425    }
 426
 427    return wait;
 428 }
 429
 430 wait_imm parse_wait_instr(wait_ctx& ctx, Instruction *instr)
 431 {
 432    if (instr->opcode == aco_opcode::s_waitcnt_vscnt &&
 433        instr->definitions[0].physReg() == sgpr_null) {
 434       wait_imm imm;
 435       imm.vs = std::min<uint8_t>(imm.vs, static_cast<SOPK_instruction*>(instr)->imm);
 436       return imm;
 437    } else if (instr->opcode == aco_opcode::s_waitcnt) {
 438       return wait_imm(ctx.chip_class, static_cast<SOPP_instruction*>(instr)->imm);
 439    }
 440    return wait_imm();
 441 }
 442
 443 wait_imm kill(Instruction* instr, wait_ctx& ctx)
 444 {
 445    wait_imm imm;
 446    if (ctx.exp_cnt || ctx.vm_cnt || ctx.lgkm_cnt)
 447       imm.combine(check_instr(instr, ctx));
 448
 449    imm.combine(parse_wait_instr(ctx, instr));
 450
 451
 452    /* It's required to wait for scalar stores before "writing back" data.
 453     * It shouldn't cost anything anyways since we're about to do s_endpgm.
 454     */
 455    if (ctx.lgkm_cnt && instr->opcode == aco_opcode::s_dcache_wb) {
 456       assert(ctx.chip_class >= GFX8);
 457       imm.lgkm = 0;
 458    }
 459
 460    if (ctx.chip_class >= GFX10) {
 461       /* GFX10: A store followed by a load at the same address causes a problem because
 462        * the load doesn't load the correct values unless we wait for the store first.
 463        * This is NOT mitigated by an s_nop.
 464        *
 465        * TODO: Refine this when we have proper alias analysis.
 466        */
 467       SMEM_instruction *smem = static_cast<SMEM_instruction *>(instr);
 468       if (ctx.pending_s_buffer_store &&
 469           !smem->definitions.empty() &&
 470           !smem->can_reorder && smem->barrier == barrier_buffer) {
 471          imm.lgkm = 0;
 472       }
 473    }
 474
 475    if (instr->format == Format::PSEUDO_BARRIER) {
 476       switch (instr->opcode) {
 477       case aco_opcode::p_memory_barrier_common:
 478          imm.combine(ctx.barrier_imm[ffs(barrier_atomic) - 1]);
 479          imm.combine(ctx.barrier_imm[ffs(barrier_buffer) - 1]);
 480          imm.combine(ctx.barrier_imm[ffs(barrier_image) - 1]);
 481          if (ctx.program->workgroup_size > ctx.program->wave_size)
 482             imm.combine(ctx.barrier_imm[ffs(barrier_shared) - 1]);
 483          break;
 484       case aco_opcode::p_memory_barrier_atomic:
 485          imm.combine(ctx.barrier_imm[ffs(barrier_atomic) - 1]);
 486          break;
 487       /* see comment in aco_scheduler.cpp's can_move_instr() on why these barriers are merged */
 488       case aco_opcode::p_memory_barrier_buffer:
 489       case aco_opcode::p_memory_barrier_image:
 490          imm.combine(ctx.barrier_imm[ffs(barrier_buffer) - 1]);
 491          imm.combine(ctx.barrier_imm[ffs(barrier_image) - 1]);
 492          break;
 493       case aco_opcode::p_memory_barrier_shared:
 494          if (ctx.program->workgroup_size > ctx.program->wave_size)
 495             imm.combine(ctx.barrier_imm[ffs(barrier_shared) - 1]);
 496          break;
 497       case aco_opcode::p_memory_barrier_gs_data:
 498          imm.combine(ctx.barrier_imm[ffs(barrier_gs_data) - 1]);
 499          break;
 500       case aco_opcode::p_memory_barrier_gs_sendmsg:
 501          imm.combine(ctx.barrier_imm[ffs(barrier_gs_sendmsg) - 1]);
 502          break;
 503       default:
 504          assert(false);
 505          break;
 506       }
 507    }
 508
 509    if (!imm.empty()) {
 510       if (ctx.pending_flat_vm && imm.vm != wait_imm::unset_counter)
 511          imm.vm = 0;
 512       if (ctx.pending_flat_lgkm && imm.lgkm != wait_imm::unset_counter)
 513          imm.lgkm = 0;
 514
 515       /* reset counters */
 516       ctx.exp_cnt = std::min(ctx.exp_cnt, imm.exp);
 517       ctx.vm_cnt = std::min(ctx.vm_cnt, imm.vm);
 518       ctx.lgkm_cnt = std::min(ctx.lgkm_cnt, imm.lgkm);
 519       ctx.vs_cnt = std::min(ctx.vs_cnt, imm.vs);
 520
 521       /* update barrier wait imms */
 522       for (unsigned i = 0; i < barrier_count; i++) {
 523          wait_imm& bar = ctx.barrier_imm[i];
 524          uint16_t& bar_ev = ctx.barrier_events[i];
 525          if (bar.exp != wait_imm::unset_counter && imm.exp <= bar.exp) {
 526             bar.exp = wait_imm::unset_counter;
 527             bar_ev &= ~exp_events;
 528          }
 529          if (bar.vm != wait_imm::unset_counter && imm.vm <= bar.vm) {
 530             bar.vm = wait_imm::unset_counter;
 531             bar_ev &= ~(vm_events & ~event_flat);
 532          }
 533          if (bar.lgkm != wait_imm::unset_counter && imm.lgkm <= bar.lgkm) {
 534             bar.lgkm = wait_imm::unset_counter;
 535             bar_ev &= ~(lgkm_events & ~event_flat);
 536          }
 537          if (bar.vs != wait_imm::unset_counter && imm.vs <= bar.vs) {
 538             bar.vs = wait_imm::unset_counter;
 539             bar_ev &= ~vs_events;
 540          }
 541          if (bar.vm == wait_imm::unset_counter && bar.lgkm == wait_imm::unset_counter)
 542             bar_ev &= ~event_flat;
 543       }
 544
 545       /* remove all gprs with higher counter from map */
 546       std::map<PhysReg,wait_entry>::iterator it = ctx.gpr_map.begin();
 547       while (it != ctx.gpr_map.end())
 548       {
 549          if (imm.exp != wait_imm::unset_counter && imm.exp <= it->second.imm.exp)
 550             ctx.wait_and_remove_from_entry(it->first, it->second, counter_exp);
 551          if (imm.vm != wait_imm::unset_counter && imm.vm <= it->second.imm.vm)
 552             ctx.wait_and_remove_from_entry(it->first, it->second, counter_vm);
 553          if (imm.lgkm != wait_imm::unset_counter && imm.lgkm <= it->second.imm.lgkm)
 554             ctx.wait_and_remove_from_entry(it->first, it->second, counter_lgkm);
 555          if (imm.lgkm != wait_imm::unset_counter && imm.vs <= it->second.imm.vs)
 556             ctx.wait_and_remove_from_entry(it->first, it->second, counter_vs);
 557          if (!it->second.counters)
 558             it = ctx.gpr_map.erase(it);
 559          else
 560             it++;
 561       }
 562    }
 563
 564    if (imm.vm == 0)
 565       ctx.pending_flat_vm = false;
 566    if (imm.lgkm == 0) {
 567       ctx.pending_flat_lgkm = false;
 568       ctx.pending_s_buffer_store = false;
 569    }
 570
 571    return imm;
 572 }
 573
 574 void update_barrier_counter(uint8_t *ctr, unsigned max)
 575 {
 576    if (*ctr != wait_imm::unset_counter && *ctr < max)
 577       (*ctr)++;
 578 }
 579
 580 void update_barrier_imm(wait_ctx& ctx, uint8_t counters, wait_event event, barrier_interaction barrier)
 581 {
 582    for (unsigned i = 0; i < barrier_count; i++) {
 583       wait_imm& bar = ctx.barrier_imm[i];
 584       uint16_t& bar_ev = ctx.barrier_events[i];
 585       if (barrier & (1 << i)) {
 586          bar_ev |= event;
 587          if (counters & counter_lgkm)
 588             bar.lgkm = 0;
 589          if (counters & counter_vm)
 590             bar.vm = 0;
 591          if (counters & counter_exp)
 592             bar.exp = 0;
 593          if (counters & counter_vs)
 594             bar.vs = 0;
 595       } else if (!(bar_ev & ctx.unordered_events) && !(ctx.unordered_events & event)) {
 596          if (counters & counter_lgkm && (bar_ev & lgkm_events) == event)
 597             update_barrier_counter(&bar.lgkm, ctx.max_lgkm_cnt);
 598          if (counters & counter_vm && (bar_ev & vm_events) == event)
 599             update_barrier_counter(&bar.vm, ctx.max_vm_cnt);
 600          if (counters & counter_exp && (bar_ev & exp_events) == event)
 601             update_barrier_counter(&bar.exp, ctx.max_exp_cnt);
 602          if (counters & counter_vs && (bar_ev & vs_events) == event)
 603             update_barrier_counter(&bar.vs, ctx.max_vs_cnt);
 604       }
 605    }
 606 }
 607
 608 void update_counters(wait_ctx& ctx, wait_event event, barrier_interaction barrier=barrier_none)
 609 {
 610    uint8_t counters = get_counters_for_event(event);
 611
 612    if (counters & counter_lgkm && ctx.lgkm_cnt <= ctx.max_lgkm_cnt)
 613       ctx.lgkm_cnt++;
 614    if (counters & counter_vm && ctx.vm_cnt <= ctx.max_vm_cnt)
 615       ctx.vm_cnt++;
 616    if (counters & counter_exp && ctx.exp_cnt <= ctx.max_exp_cnt)
 617       ctx.exp_cnt++;
 618    if (counters & counter_vs && ctx.vs_cnt <= ctx.max_vs_cnt)
 619       ctx.vs_cnt++;
 620
 621    update_barrier_imm(ctx, counters, event, barrier);
 622
 623    if (ctx.unordered_events & event)
 624       return;
 625
 626    if (ctx.pending_flat_lgkm)
 627       counters &= ~counter_lgkm;
 628    if (ctx.pending_flat_vm)
 629       counters &= ~counter_vm;
 630
 631    for (std::pair<const PhysReg,wait_entry>& e : ctx.gpr_map) {
 632       wait_entry& entry = e.second;
 633
 634       if (entry.events & ctx.unordered_events)
 635          continue;
 636
 637       assert(entry.events);
 638
 639       if ((counters & counter_exp) && (entry.events & exp_events) == event && entry.imm.exp < ctx.max_exp_cnt)
 640          entry.imm.exp++;
 641       if ((counters & counter_lgkm) && (entry.events & lgkm_events) == event && entry.imm.lgkm < ctx.max_lgkm_cnt)
 642          entry.imm.lgkm++;
 643       if ((counters & counter_vm) && (entry.events & vm_events) == event && entry.imm.vm < ctx.max_vm_cnt)
 644          entry.imm.vm++;
 645       if ((counters & counter_vs) && (entry.events & vs_events) == event && entry.imm.vs < ctx.max_vs_cnt)
 646          entry.imm.vs++;
 647    }
 648 }
 649
 650 void update_counters_for_flat_load(wait_ctx& ctx, barrier_interaction barrier=barrier_none)
 651 {
 652    assert(ctx.chip_class < GFX10);
 653
 654    if (ctx.lgkm_cnt <= ctx.max_lgkm_cnt)
 655       ctx.lgkm_cnt++;
 656    if (ctx.vm_cnt <= ctx.max_vm_cnt)
 657       ctx.vm_cnt++;
 658
 659    update_barrier_imm(ctx, counter_vm | counter_lgkm, event_flat, barrier);
 660
 661    for (std::pair<PhysReg,wait_entry> e : ctx.gpr_map)
 662    {
 663       if (e.second.counters & counter_vm)
 664          e.second.imm.vm = 0;
 665       if (e.second.counters & counter_lgkm)
 666          e.second.imm.lgkm = 0;
 667    }
 668    ctx.pending_flat_lgkm = true;
 669    ctx.pending_flat_vm = true;
 670 }
 671
 672 void insert_wait_entry(wait_ctx& ctx, PhysReg reg, RegClass rc, wait_event event, bool wait_on_read)
 673 {
 674    uint16_t counters = get_counters_for_event(event);
 675    wait_imm imm;
 676    if (counters & counter_lgkm)
 677       imm.lgkm = 0;
 678    if (counters & counter_vm)
 679       imm.vm = 0;
 680    if (counters & counter_exp)
 681       imm.exp = 0;
 682    if (counters & counter_vs)
 683       imm.vs = 0;
 684
 685    wait_entry new_entry(event, imm, !rc.is_linear(), wait_on_read);
 686
 687    for (unsigned i = 0; i < rc.size(); i++) {
 688       auto it = ctx.gpr_map.emplace(PhysReg{reg.reg()+i}, new_entry);
 689       if (!it.second)
 690          it.first->second.join(new_entry);
 691    }
 692
 693    if (ctx.collect_statistics) {
 694       unsigned counters_todo = counters;
 695       while (counters_todo) {
 696          unsigned i = u_bit_scan(&counters_todo);
 697          ctx.unwaited_instrs[i].insert(std::make_pair(ctx.gen_instr, 0u));
 698          for (unsigned j = 0; j < rc.size(); j++)
 699             ctx.reg_instrs[i][PhysReg{reg.reg()+j}].insert(ctx.gen_instr);
 700       }
 701    }
 702 }
 703
 704 void insert_wait_entry(wait_ctx& ctx, Operand op, wait_event event)
 705 {
 706    if (!op.isConstant() && !op.isUndefined())
 707       insert_wait_entry(ctx, op.physReg(), op.regClass(), event, false);
 708 }
 709
 710 void insert_wait_entry(wait_ctx& ctx, Definition def, wait_event event)
 711 {
 712    insert_wait_entry(ctx, def.physReg(), def.regClass(), event, true);
 713 }
 714
 715 void gen(Instruction* instr, wait_ctx& ctx)
 716 {
 717    switch (instr->format) {
 718    case Format::EXP: {
 719       Export_instruction* exp_instr = static_cast<Export_instruction*>(instr);
 720
 721       wait_event ev;
 722       if (exp_instr->dest <= 9)
 723          ev = event_exp_mrt_null;
 724       else if (exp_instr->dest <= 15)
 725          ev = event_exp_pos;
 726       else
 727          ev = event_exp_param;
 728       update_counters(ctx, ev);
 729
 730       /* insert new entries for exported vgprs */
 731       for (unsigned i = 0; i < 4; i++)
 732       {
 733          if (exp_instr->enabled_mask & (1 << i)) {
 734             unsigned idx = exp_instr->compressed ? i >> 1 : i;
 735             assert(idx < exp_instr->operands.size());
 736             insert_wait_entry(ctx, exp_instr->operands[idx], ev);
 737
 738          }
 739       }
 740       insert_wait_entry(ctx, exec, s2, ev, false);
 741       break;
 742    }
 743    case Format::FLAT: {
 744       if (ctx.chip_class < GFX10 && !instr->definitions.empty())
 745          update_counters_for_flat_load(ctx, barrier_buffer);
 746       else
 747          update_counters(ctx, event_flat, barrier_buffer);
 748
 749       if (!instr->definitions.empty())
 750          insert_wait_entry(ctx, instr->definitions[0], event_flat);
 751       break;
 752    }
 753    case Format::SMEM: {
 754       SMEM_instruction *smem = static_cast<SMEM_instruction*>(instr);
 755       update_counters(ctx, event_smem, static_cast<SMEM_instruction*>(instr)->barrier);
 756
 757       if (!instr->definitions.empty())
 758          insert_wait_entry(ctx, instr->definitions[0], event_smem);
 759       else if (ctx.chip_class >= GFX10 &&
 760                !smem->can_reorder &&
 761                smem->barrier == barrier_buffer)
 762          ctx.pending_s_buffer_store = true;
 763
 764       break;
 765    }
 766    case Format::DS: {
 767       bool gds = static_cast<DS_instruction*>(instr)->gds;
 768       update_counters(ctx, gds ? event_gds : event_lds, gds ? barrier_none : barrier_shared);
 769       if (gds)
 770          update_counters(ctx, event_gds_gpr_lock);
 771
 772       if (!instr->definitions.empty())
 773          insert_wait_entry(ctx, instr->definitions[0], gds ? event_gds : event_lds);
 774
 775       if (gds) {
 776          for (const Operand& op : instr->operands)
 777             insert_wait_entry(ctx, op, event_gds_gpr_lock);
 778          insert_wait_entry(ctx, exec, s2, event_gds_gpr_lock, false);
 779       }
 780       break;
 781    }
 782    case Format::MUBUF:
 783    case Format::MTBUF:
 784    case Format::MIMG:
 785    case Format::GLOBAL: {
 786       wait_event ev = !instr->definitions.empty() || ctx.chip_class < GFX10 ? event_vmem : event_vmem_store;
 787       update_counters(ctx, ev, get_barrier_interaction(instr));
 788
 789       if (!instr->definitions.empty())
 790          insert_wait_entry(ctx, instr->definitions[0], ev);
 791
 792       if (ctx.chip_class == GFX6 &&
 793           instr->format != Format::MIMG &&
 794           instr->operands.size() == 4) {
 795          ctx.exp_cnt++;
 796          update_counters(ctx, event_vmem_gpr_lock);
 797          insert_wait_entry(ctx, instr->operands[3], event_vmem_gpr_lock);
 798       } else if (ctx.chip_class == GFX6 &&
 799                  instr->format == Format::MIMG &&
 800                  instr->operands[1].regClass().type() == RegType::vgpr) {
 801          ctx.exp_cnt++;
 802          update_counters(ctx, event_vmem_gpr_lock);
 803          insert_wait_entry(ctx, instr->operands[1], event_vmem_gpr_lock);
 804       }
 805
 806       break;
 807    }
 808    case Format::SOPP: {
 809       if (instr->opcode == aco_opcode::s_sendmsg ||
 810           instr->opcode == aco_opcode::s_sendmsghalt)
 811          update_counters(ctx, event_sendmsg, get_barrier_interaction(instr));
 812    }
 813    default:
 814       break;
 815    }
 816 }
 817
 818 void emit_waitcnt(wait_ctx& ctx, std::vector<aco_ptr<Instruction>>& instructions, wait_imm imm)
 819 {
 820    if (imm.vs != wait_imm::unset_counter) {
 821       assert(ctx.chip_class >= GFX10);
 822       SOPK_instruction* waitcnt_vs = create_instruction<SOPK_instruction>(aco_opcode::s_waitcnt_vscnt, Format::SOPK, 0, 1);
 823       waitcnt_vs->definitions[0] = Definition(sgpr_null, s1);
 824       waitcnt_vs->imm = imm.vs;
 825       instructions.emplace_back(waitcnt_vs);
 826       imm.vs = wait_imm::unset_counter;
 827    }
 828    if (!imm.empty()) {
 829       SOPP_instruction* waitcnt = create_instruction<SOPP_instruction>(aco_opcode::s_waitcnt, Format::SOPP, 0, 0);
 830       waitcnt->imm = imm.pack(ctx.chip_class);
 831       waitcnt->block = -1;
 832       instructions.emplace_back(waitcnt);
 833    }
 834 }
 835
 836 void handle_block(Program *program, Block& block, wait_ctx& ctx)
 837 {
 838    std::vector<aco_ptr<Instruction>> new_instructions;
 839
 840    wait_imm queued_imm;
 841
 842    ctx.collect_statistics = program->collect_statistics;
 843
 844    for (aco_ptr<Instruction>& instr : block.instructions) {
 845       bool is_wait = !parse_wait_instr(ctx, instr.get()).empty();
 846
 847       queued_imm.combine(kill(instr.get(), ctx));
 848
 849       ctx.gen_instr = instr.get();
 850       gen(instr.get(), ctx);
 851
 852       if (instr->format != Format::PSEUDO_BARRIER && !is_wait) {
 853          if (!queued_imm.empty()) {
 854             emit_waitcnt(ctx, new_instructions, queued_imm);
 855             queued_imm = wait_imm();
 856          }
 857          new_instructions.emplace_back(std::move(instr));
 858
 859          if (ctx.collect_statistics)
 860             ctx.advance_unwaited_instrs();
 861       }
 862    }
 863
 864    if (!queued_imm.empty())
 865       emit_waitcnt(ctx, new_instructions, queued_imm);
 866
 867    block.instructions.swap(new_instructions);
 868 }
 869
 870 } /* end namespace */
 871
 872 static uint32_t calculate_score(std::vector<wait_ctx> &ctx_vec, uint32_t event_mask)
 873 {
 874    double result = 0.0;
 875    unsigned num_waits = 0;
 876    while (event_mask) {
 877       unsigned event_index = u_bit_scan(&event_mask);
 878       for (const wait_ctx &ctx : ctx_vec) {
 879          for (unsigned dist : ctx.wait_distances[event_index]) {
 880             double score = dist;
 881             /* for many events, excessive distances provide little benefit, so
 882              * decrease the score in that case. */
 883             double threshold = INFINITY;
 884             double inv_strength = 0.000001;
 885             switch (1 << event_index) {
 886             case event_smem:
 887                threshold = 70.0;
 888                inv_strength = 75.0;
 889                break;
 890             case event_vmem:
 891             case event_vmem_store:
 892             case event_flat:
 893                threshold = 230.0;
 894                inv_strength = 150.0;
 895                break;
 896             case event_lds:
 897                threshold = 16.0;
 898                break;
 899             default:
 900                break;
 901             }
 902             if (score > threshold) {
 903                score -= threshold;
 904                score = threshold + score / (1.0 + score / inv_strength);
 905             }
 906
 907             /* we don't want increases in high scores to hide decreases in low scores,
 908              * so raise to the power of 0.1 before averaging. */
 909             result += pow(score, 0.1);
 910             num_waits++;
 911          }
 912       }
 913    }
 914    return round(pow(result / num_waits, 10.0) * 10.0);
 915 }
 916
 917 void insert_wait_states(Program* program)
 918 {
 919    /* per BB ctx */
 920    std::vector<bool> done(program->blocks.size());
 921    std::vector<wait_ctx> in_ctx(program->blocks.size(), wait_ctx(program));
 922    std::vector<wait_ctx> out_ctx(program->blocks.size(), wait_ctx(program));
 923
 924    std::stack<unsigned> loop_header_indices;
 925    unsigned loop_progress = 0;
 926
 927    for (unsigned i = 0; i < program->blocks.size();) {
 928       Block& current = program->blocks[i++];
 929       wait_ctx ctx = in_ctx[current.index];
 930
 931       if (current.kind & block_kind_loop_header) {
 932          loop_header_indices.push(current.index);
 933       } else if (current.kind & block_kind_loop_exit) {
 934          bool repeat = false;
 935          if (loop_progress == loop_header_indices.size()) {
 936             i = loop_header_indices.top();
 937             repeat = true;
 938          }
 939          loop_header_indices.pop();
 940          loop_progress = std::min<unsigned>(loop_progress, loop_header_indices.size());
 941          if (repeat)
 942             continue;
 943       }
 944
 945       bool changed = false;
 946       for (unsigned b : current.linear_preds)
 947          changed |= ctx.join(&out_ctx[b], false);
 948       for (unsigned b : current.logical_preds)
 949          changed |= ctx.join(&out_ctx[b], true);
 950
 951       if (done[current.index] && !changed) {
 952          in_ctx[current.index] = std::move(ctx);
 953          continue;
 954       } else {
 955          in_ctx[current.index] = ctx;
 956       }
 957
 958       if (current.instructions.empty()) {
 959          out_ctx[current.index] = std::move(ctx);
 960          continue;
 961       }
 962
 963       loop_progress = std::max<unsigned>(loop_progress, current.loop_nest_depth);
 964       done[current.index] = true;
 965
 966       handle_block(program, current, ctx);
 967
 968       out_ctx[current.index] = std::move(ctx);
 969    }
 970
 971    if (program->collect_statistics) {
 972       program->statistics[statistic_vmem_score] =
 973          calculate_score(out_ctx, event_vmem | event_flat | event_vmem_store);
 974       program->statistics[statistic_smem_score] =
 975          calculate_score(out_ctx, event_smem);
 976    }
 977 }
 978
 979 }
 980