src/gallium/drivers/vc4/vc4_qpu_schedule.c

   1 /*
   2  * Copyright © 2010 Intel Corporation
   3  * Copyright © 2014 Broadcom
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  */
  24
  25 /**
  26  * @file vc4_qpu_schedule.c
  27  *
  28  * The basic model of the list scheduler is to take a basic block, compute a
  29  * DAG of the dependencies, and make a list of the DAG heads.  Heuristically
  30  * pick a DAG head, then put all the children that are now DAG heads into the
  31  * list of things to schedule.
  32  *
  33  * The goal of scheduling here is to pack pairs of operations together in a
  34  * single QPU instruction.
  35  */
  36
  37 #include "vc4_qir.h"
  38 #include "vc4_qpu.h"
  39 #include "util/ralloc.h"
  40
  41 static bool debug;
  42
  43 struct schedule_node_child;
  44
  45 struct schedule_node {
  46         struct list_head link;
  47         struct queued_qpu_inst *inst;
  48         struct schedule_node_child *children;
  49         uint32_t child_count;
  50         uint32_t child_array_size;
  51         uint32_t parent_count;
  52
  53         /* Longest cycles + n->latency of any parent of this node. */
  54         uint32_t unblocked_time;
  55
  56         /**
  57          * Minimum number of cycles from scheduling this instruction until the
  58          * end of the program, based on the slowest dependency chain through
  59          * the children.
  60          */
  61         uint32_t delay;
  62
  63         /**
  64          * cycles between this instruction being scheduled and when its result
  65          * can be consumed.
  66          */
  67         uint32_t latency;
  68
  69         /**
  70          * Which uniform from uniform_data[] this instruction read, or -1 if
  71          * not reading a uniform.
  72          */
  73         int uniform;
  74 };
  75
  76 struct schedule_node_child {
  77         struct schedule_node *node;
  78         bool write_after_read;
  79 };
  80
  81 /* When walking the instructions in reverse, we need to swap before/after in
  82  * add_dep().
  83  */
  84 enum direction { F, R };
  85
  86 struct schedule_state {
  87         struct schedule_node *last_r[6];
  88         struct schedule_node *last_ra[32];
  89         struct schedule_node *last_rb[32];
  90         struct schedule_node *last_sf;
  91         struct schedule_node *last_vpm_read;
  92         struct schedule_node *last_tmu_write;
  93         struct schedule_node *last_tlb;
  94         struct schedule_node *last_vpm;
  95         enum direction dir;
  96         /* Estimated cycle when the current instruction would start. */
  97         uint32_t time;
  98 };
  99
 100 static void
 101 add_dep(struct schedule_state *state,
 102         struct schedule_node *before,
 103         struct schedule_node *after,
 104         bool write)
 105 {
 106         bool write_after_read = !write && state->dir == R;
 107
 108         if (!before || !after)
 109                 return;
 110
 111         assert(before != after);
 112
 113         if (state->dir == R) {
 114                 struct schedule_node *t = before;
 115                 before = after;
 116                 after = t;
 117         }
 118
 119         for (int i = 0; i < before->child_count; i++) {
 120                 if (before->children[i].node == after &&
 121                     (before->children[i].write_after_read == write_after_read)) {
 122                         return;
 123                 }
 124         }
 125
 126         if (before->child_array_size <= before->child_count) {
 127                 before->child_array_size = MAX2(before->child_array_size * 2, 16);
 128                 before->children = reralloc(before, before->children,
 129                                             struct schedule_node_child,
 130                                             before->child_array_size);
 131         }
 132
 133         before->children[before->child_count].node = after;
 134         before->children[before->child_count].write_after_read =
 135                 write_after_read;
 136         before->child_count++;
 137         after->parent_count++;
 138 }
 139
 140 static void
 141 add_read_dep(struct schedule_state *state,
 142               struct schedule_node *before,
 143               struct schedule_node *after)
 144 {
 145         add_dep(state, before, after, false);
 146 }
 147
 148 static void
 149 add_write_dep(struct schedule_state *state,
 150               struct schedule_node **before,
 151               struct schedule_node *after)
 152 {
 153         add_dep(state, *before, after, true);
 154         *before = after;
 155 }
 156
 157 static bool
 158 qpu_writes_r4(uint64_t inst)
 159 {
 160         uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
 161
 162         switch(sig) {
 163         case QPU_SIG_COLOR_LOAD:
 164         case QPU_SIG_LOAD_TMU0:
 165         case QPU_SIG_LOAD_TMU1:
 166         case QPU_SIG_ALPHA_MASK_LOAD:
 167                 return true;
 168         default:
 169                 return false;
 170         }
 171 }
 172
 173 static void
 174 process_raddr_deps(struct schedule_state *state, struct schedule_node *n,
 175                    uint32_t raddr, bool is_a)
 176 {
 177         switch (raddr) {
 178         case QPU_R_VARY:
 179                 add_write_dep(state, &state->last_r[5], n);
 180                 break;
 181
 182         case QPU_R_VPM:
 183                 add_write_dep(state, &state->last_vpm_read, n);
 184                 break;
 185
 186         case QPU_R_UNIF:
 187         case QPU_R_NOP:
 188         case QPU_R_ELEM_QPU:
 189         case QPU_R_XY_PIXEL_COORD:
 190         case QPU_R_MS_REV_FLAGS:
 191                 break;
 192
 193         default:
 194                 if (raddr < 32) {
 195                         if (is_a)
 196                                 add_read_dep(state, state->last_ra[raddr], n);
 197                         else
 198                                 add_read_dep(state, state->last_rb[raddr], n);
 199                 } else {
 200                         fprintf(stderr, "unknown raddr %d\n", raddr);
 201                         abort();
 202                 }
 203                 break;
 204         }
 205 }
 206
 207 static bool
 208 is_tmu_write(uint32_t waddr)
 209 {
 210         switch (waddr) {
 211         case QPU_W_TMU0_S:
 212         case QPU_W_TMU0_T:
 213         case QPU_W_TMU0_R:
 214         case QPU_W_TMU0_B:
 215         case QPU_W_TMU1_S:
 216         case QPU_W_TMU1_T:
 217         case QPU_W_TMU1_R:
 218         case QPU_W_TMU1_B:
 219                 return true;
 220         default:
 221                 return false;
 222         }
 223 }
 224
 225 static bool
 226 reads_uniform(uint64_t inst)
 227 {
 228         if (QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_LOAD_IMM)
 229                 return false;
 230
 231         return (QPU_GET_FIELD(inst, QPU_RADDR_A) == QPU_R_UNIF ||
 232                 (QPU_GET_FIELD(inst, QPU_RADDR_B) == QPU_R_UNIF &&
 233                  QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_SMALL_IMM) ||
 234                 is_tmu_write(QPU_GET_FIELD(inst, QPU_WADDR_ADD)) ||
 235                 is_tmu_write(QPU_GET_FIELD(inst, QPU_WADDR_MUL)));
 236 }
 237
 238 static void
 239 process_mux_deps(struct schedule_state *state, struct schedule_node *n,
 240                  uint32_t mux)
 241 {
 242         if (mux != QPU_MUX_A && mux != QPU_MUX_B)
 243                 add_read_dep(state, state->last_r[mux], n);
 244 }
 245
 246
 247 static void
 248 process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
 249                    uint32_t waddr, bool is_add)
 250 {
 251         uint64_t inst = n->inst->inst;
 252         bool is_a = is_add ^ ((inst & QPU_WS) != 0);
 253
 254         if (waddr < 32) {
 255                 if (is_a) {
 256                         add_write_dep(state, &state->last_ra[waddr], n);
 257                 } else {
 258                         add_write_dep(state, &state->last_rb[waddr], n);
 259                 }
 260         } else if (is_tmu_write(waddr)) {
 261                 add_write_dep(state, &state->last_tmu_write, n);
 262         } else if (qpu_waddr_is_tlb(waddr) ||
 263                    waddr == QPU_W_MS_FLAGS) {
 264                 add_write_dep(state, &state->last_tlb, n);
 265         } else {
 266                 switch (waddr) {
 267                 case QPU_W_ACC0:
 268                 case QPU_W_ACC1:
 269                 case QPU_W_ACC2:
 270                 case QPU_W_ACC3:
 271                 case QPU_W_ACC5:
 272                         add_write_dep(state, &state->last_r[waddr - QPU_W_ACC0],
 273                                       n);
 274                         break;
 275
 276                 case QPU_W_VPM:
 277                         add_write_dep(state, &state->last_vpm, n);
 278                         break;
 279
 280                 case QPU_W_VPMVCD_SETUP:
 281                         if (is_a)
 282                                 add_write_dep(state, &state->last_vpm_read, n);
 283                         else
 284                                 add_write_dep(state, &state->last_vpm, n);
 285                         break;
 286
 287                 case QPU_W_SFU_RECIP:
 288                 case QPU_W_SFU_RECIPSQRT:
 289                 case QPU_W_SFU_EXP:
 290                 case QPU_W_SFU_LOG:
 291                         add_write_dep(state, &state->last_r[4], n);
 292                         break;
 293
 294                 case QPU_W_TLB_STENCIL_SETUP:
 295                         /* This isn't a TLB operation that does things like
 296                          * implicitly lock the scoreboard, but it does have to
 297                          * appear before TLB_Z, and each of the TLB_STENCILs
 298                          * have to schedule in the same order relative to each
 299                          * other.
 300                          */
 301                         add_write_dep(state, &state->last_tlb, n);
 302                         break;
 303
 304                 case QPU_W_MS_FLAGS:
 305                         add_write_dep(state, &state->last_tlb, n);
 306                         break;
 307
 308                 case QPU_W_NOP:
 309                         break;
 310
 311                 default:
 312                         fprintf(stderr, "Unknown waddr %d\n", waddr);
 313                         abort();
 314                 }
 315         }
 316 }
 317
 318 static void
 319 process_cond_deps(struct schedule_state *state, struct schedule_node *n,
 320                   uint32_t cond)
 321 {
 322         switch (cond) {
 323         case QPU_COND_NEVER:
 324         case QPU_COND_ALWAYS:
 325                 break;
 326         default:
 327                 add_read_dep(state, state->last_sf, n);
 328                 break;
 329         }
 330 }
 331
 332 /**
 333  * Common code for dependencies that need to be tracked both forward and
 334  * backward.
 335  *
 336  * This is for things like "all reads of r4 have to happen between the r4
 337  * writes that surround them".
 338  */
 339 static void
 340 calculate_deps(struct schedule_state *state, struct schedule_node *n)
 341 {
 342         uint64_t inst = n->inst->inst;
 343         uint32_t add_op = QPU_GET_FIELD(inst, QPU_OP_ADD);
 344         uint32_t mul_op = QPU_GET_FIELD(inst, QPU_OP_MUL);
 345         uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
 346         uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
 347         uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
 348         uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
 349         uint32_t add_a = QPU_GET_FIELD(inst, QPU_ADD_A);
 350         uint32_t add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
 351         uint32_t mul_a = QPU_GET_FIELD(inst, QPU_MUL_A);
 352         uint32_t mul_b = QPU_GET_FIELD(inst, QPU_MUL_B);
 353         uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
 354
 355         if (sig != QPU_SIG_LOAD_IMM) {
 356                 process_raddr_deps(state, n, raddr_a, true);
 357                 if (sig != QPU_SIG_SMALL_IMM)
 358                         process_raddr_deps(state, n, raddr_b, false);
 359         }
 360
 361         if (add_op != QPU_A_NOP) {
 362                 process_mux_deps(state, n, add_a);
 363                 process_mux_deps(state, n, add_b);
 364         }
 365         if (mul_op != QPU_M_NOP) {
 366                 process_mux_deps(state, n, mul_a);
 367                 process_mux_deps(state, n, mul_b);
 368         }
 369
 370         process_waddr_deps(state, n, waddr_add, true);
 371         process_waddr_deps(state, n, waddr_mul, false);
 372         if (qpu_writes_r4(inst))
 373                 add_write_dep(state, &state->last_r[4], n);
 374
 375         switch (sig) {
 376         case QPU_SIG_SW_BREAKPOINT:
 377         case QPU_SIG_NONE:
 378         case QPU_SIG_THREAD_SWITCH:
 379         case QPU_SIG_LAST_THREAD_SWITCH:
 380         case QPU_SIG_SMALL_IMM:
 381         case QPU_SIG_LOAD_IMM:
 382                 break;
 383
 384         case QPU_SIG_LOAD_TMU0:
 385         case QPU_SIG_LOAD_TMU1:
 386                 /* TMU loads are coming from a FIFO, so ordering is important.
 387                  */
 388                 add_write_dep(state, &state->last_tmu_write, n);
 389                 break;
 390
 391         case QPU_SIG_COLOR_LOAD:
 392                 add_read_dep(state, state->last_tlb, n);
 393                 break;
 394
 395         case QPU_SIG_PROG_END:
 396         case QPU_SIG_WAIT_FOR_SCOREBOARD:
 397         case QPU_SIG_SCOREBOARD_UNLOCK:
 398         case QPU_SIG_COVERAGE_LOAD:
 399         case QPU_SIG_COLOR_LOAD_END:
 400         case QPU_SIG_ALPHA_MASK_LOAD:
 401         case QPU_SIG_BRANCH:
 402                 fprintf(stderr, "Unhandled signal bits %d\n", sig);
 403                 abort();
 404         }
 405
 406         process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_ADD));
 407         process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_ADD));
 408         if (inst & QPU_SF)
 409                 add_write_dep(state, &state->last_sf, n);
 410 }
 411
 412 static void
 413 calculate_forward_deps(struct vc4_compile *c, struct list_head *schedule_list)
 414 {
 415         struct schedule_state state;
 416
 417         memset(&state, 0, sizeof(state));
 418         state.dir = F;
 419
 420         list_for_each_entry(struct schedule_node, node, schedule_list, link)
 421                 calculate_deps(&state, node);
 422 }
 423
 424 static void
 425 calculate_reverse_deps(struct vc4_compile *c, struct list_head *schedule_list)
 426 {
 427         struct list_head *node;
 428         struct schedule_state state;
 429
 430         memset(&state, 0, sizeof(state));
 431         state.dir = R;
 432
 433         for (node = schedule_list->prev; schedule_list != node; node = node->prev) {
 434                 calculate_deps(&state, (struct schedule_node *)node);
 435         }
 436 }
 437
 438 struct choose_scoreboard {
 439         int tick;
 440         int last_sfu_write_tick;
 441         uint32_t last_waddr_a, last_waddr_b;
 442 };
 443
 444 static bool
 445 reads_too_soon_after_write(struct choose_scoreboard *scoreboard, uint64_t inst)
 446 {
 447         uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
 448         uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
 449         uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
 450         uint32_t src_muxes[] = {
 451                 QPU_GET_FIELD(inst, QPU_ADD_A),
 452                 QPU_GET_FIELD(inst, QPU_ADD_B),
 453                 QPU_GET_FIELD(inst, QPU_MUL_A),
 454                 QPU_GET_FIELD(inst, QPU_MUL_B),
 455         };
 456         for (int i = 0; i < ARRAY_SIZE(src_muxes); i++) {
 457                 if ((src_muxes[i] == QPU_MUX_A &&
 458                      raddr_a < 32 &&
 459                      scoreboard->last_waddr_a == raddr_a) ||
 460                     (src_muxes[i] == QPU_MUX_B &&
 461                      sig != QPU_SIG_SMALL_IMM &&
 462                      raddr_b < 32 &&
 463                      scoreboard->last_waddr_b == raddr_b)) {
 464                         return true;
 465                 }
 466
 467                 if (src_muxes[i] == QPU_MUX_R4) {
 468                         if (scoreboard->tick -
 469                             scoreboard->last_sfu_write_tick <= 2) {
 470                                 return true;
 471                         }
 472                 }
 473         }
 474
 475         return false;
 476 }
 477
 478 static bool
 479 pixel_scoreboard_too_soon(struct choose_scoreboard *scoreboard, uint64_t inst)
 480 {
 481         return (scoreboard->tick < 2 && qpu_inst_is_tlb(inst));
 482 }
 483
 484 static int
 485 get_instruction_priority(uint64_t inst)
 486 {
 487         uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
 488         uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
 489         uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
 490         uint32_t baseline_score;
 491         uint32_t next_score = 0;
 492
 493         /* Schedule TLB operations as late as possible, to get more
 494          * parallelism between shaders.
 495          */
 496         if (qpu_inst_is_tlb(inst))
 497                 return next_score;
 498         next_score++;
 499
 500         /* Schedule texture read results collection late to hide latency. */
 501         if (sig == QPU_SIG_LOAD_TMU0 || sig == QPU_SIG_LOAD_TMU1)
 502                 return next_score;
 503         next_score++;
 504
 505         /* Default score for things that aren't otherwise special. */
 506         baseline_score = next_score;
 507         next_score++;
 508
 509         /* Schedule texture read setup early to hide their latency better. */
 510         if (is_tmu_write(waddr_add) || is_tmu_write(waddr_mul))
 511                 return next_score;
 512         next_score++;
 513
 514         return baseline_score;
 515 }
 516
 517 static struct schedule_node *
 518 choose_instruction_to_schedule(struct choose_scoreboard *scoreboard,
 519                                struct list_head *schedule_list,
 520                                struct schedule_node *prev_inst)
 521 {
 522         struct schedule_node *chosen = NULL;
 523         int chosen_prio = 0;
 524
 525         list_for_each_entry(struct schedule_node, n, schedule_list, link) {
 526                 uint64_t inst = n->inst->inst;
 527
 528                 /* "An instruction must not read from a location in physical
 529                  *  regfile A or B that was written to by the previous
 530                  *  instruction."
 531                  */
 532                 if (reads_too_soon_after_write(scoreboard, inst))
 533                         continue;
 534
 535                 /* "A scoreboard wait must not occur in the first two
 536                  *  instructions of a fragment shader. This is either the
 537                  *  explicit Wait for Scoreboard signal or an implicit wait
 538                  *  with the first tile-buffer read or write instruction."
 539                  */
 540                 if (pixel_scoreboard_too_soon(scoreboard, inst))
 541                         continue;
 542
 543                 /* If we're trying to pair with another instruction, check
 544                  * that they're compatible.
 545                  */
 546                 if (prev_inst) {
 547                         if (prev_inst->uniform != -1 && n->uniform != -1)
 548                                 continue;
 549
 550                         inst = qpu_merge_inst(prev_inst->inst->inst, inst);
 551                         if (!inst)
 552                                 continue;
 553                 }
 554
 555                 int prio = get_instruction_priority(inst);
 556
 557                 /* Found a valid instruction.  If nothing better comes along,
 558                  * this one works.
 559                  */
 560                 if (!chosen) {
 561                         chosen = n;
 562                         chosen_prio = prio;
 563                         continue;
 564                 }
 565
 566                 if (prio > chosen_prio) {
 567                         chosen = n;
 568                         chosen_prio = prio;
 569                 } else if (prio < chosen_prio) {
 570                         continue;
 571                 }
 572
 573                 if (n->delay > chosen->delay) {
 574                         chosen = n;
 575                         chosen_prio = prio;
 576                 } else if (n->delay < chosen->delay) {
 577                         continue;
 578                 }
 579         }
 580
 581         return chosen;
 582 }
 583
 584 static void
 585 update_scoreboard_for_chosen(struct choose_scoreboard *scoreboard,
 586                              uint64_t inst)
 587 {
 588         uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
 589         uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
 590
 591         if (!(inst & QPU_WS)) {
 592                 scoreboard->last_waddr_a = waddr_add;
 593                 scoreboard->last_waddr_b = waddr_mul;
 594         } else {
 595                 scoreboard->last_waddr_b = waddr_add;
 596                 scoreboard->last_waddr_a = waddr_mul;
 597         }
 598
 599         if ((waddr_add >= QPU_W_SFU_RECIP && waddr_add <= QPU_W_SFU_LOG) ||
 600             (waddr_mul >= QPU_W_SFU_RECIP && waddr_mul <= QPU_W_SFU_LOG)) {
 601                 scoreboard->last_sfu_write_tick = scoreboard->tick;
 602         }
 603 }
 604
 605 static void
 606 dump_state(struct list_head *schedule_list)
 607 {
 608         list_for_each_entry(struct schedule_node, n, schedule_list, link) {
 609                 fprintf(stderr, "         t=%4d: ", n->unblocked_time);
 610                 vc4_qpu_disasm(&n->inst->inst, 1);
 611                 fprintf(stderr, "\n");
 612
 613                 for (int i = 0; i < n->child_count; i++) {
 614                         struct schedule_node *child = n->children[i].node;
 615                         if (!child)
 616                                 continue;
 617
 618                         fprintf(stderr, "                 - ");
 619                         vc4_qpu_disasm(&child->inst->inst, 1);
 620                         fprintf(stderr, " (%d parents, %c)\n",
 621                                 child->parent_count,
 622                                 n->children[i].write_after_read ? 'w' : 'r');
 623                 }
 624         }
 625 }
 626
 627 /** Recursive computation of the delay member of a node. */
 628 static void
 629 compute_delay(struct schedule_node *n)
 630 {
 631         if (!n->child_count) {
 632                 n->delay = 1;
 633         } else {
 634                 for (int i = 0; i < n->child_count; i++) {
 635                         if (!n->children[i].node->delay)
 636                                 compute_delay(n->children[i].node);
 637                         n->delay = MAX2(n->delay,
 638                                         n->children[i].node->delay + n->latency);
 639                 }
 640         }
 641 }
 642
 643 static void
 644 mark_instruction_scheduled(struct list_head *schedule_list,
 645                            uint32_t time,
 646                            struct schedule_node *node,
 647                            bool war_only)
 648 {
 649         if (!node)
 650                 return;
 651
 652         for (int i = node->child_count - 1; i >= 0; i--) {
 653                 struct schedule_node *child =
 654                         node->children[i].node;
 655
 656                 if (!child)
 657                         continue;
 658
 659                 if (war_only && !node->children[i].write_after_read)
 660                         continue;
 661
 662                 /* If the requirement is only that the node not appear before
 663                  * the last read of its destination, then it can be scheduled
 664                  * immediately after (or paired with!) the thing reading the
 665                  * destination.
 666                  */
 667                 int latency_from_previous = war_only ? 0 : node->latency;
 668                 child->unblocked_time = MAX2(child->unblocked_time,
 669                                              time + latency_from_previous);
 670                 child->parent_count--;
 671                 if (child->parent_count == 0)
 672                         list_add(&child->link, schedule_list);
 673
 674                 node->children[i].node = NULL;
 675         }
 676 }
 677
 678 static uint32_t
 679 schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list)
 680 {
 681         struct choose_scoreboard scoreboard;
 682         uint32_t time = 0;
 683
 684         /* We reorder the uniforms as we schedule instructions, so save the
 685          * old data off and replace it.
 686          */
 687         uint32_t *uniform_data = c->uniform_data;
 688         enum quniform_contents *uniform_contents = c->uniform_contents;
 689         c->uniform_contents = ralloc_array(c, enum quniform_contents,
 690                                            c->num_uniforms);
 691         c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms);
 692         c->uniform_array_size = c->num_uniforms;
 693         uint32_t next_uniform = 0;
 694
 695         memset(&scoreboard, 0, sizeof(scoreboard));
 696         scoreboard.last_waddr_a = ~0;
 697         scoreboard.last_waddr_b = ~0;
 698         scoreboard.last_sfu_write_tick = -10;
 699
 700         if (debug) {
 701                 fprintf(stderr, "initial deps:\n");
 702                 dump_state(schedule_list);
 703                 fprintf(stderr, "\n");
 704         }
 705
 706         /* Remove non-DAG heads from the list. */
 707         list_for_each_entry_safe(struct schedule_node, n, schedule_list, link) {
 708                 if (n->parent_count != 0)
 709                         list_del(&n->link);
 710         }
 711
 712         while (!list_empty(schedule_list)) {
 713                 struct schedule_node *chosen =
 714                         choose_instruction_to_schedule(&scoreboard,
 715                                                        schedule_list,
 716                                                        NULL);
 717                 struct schedule_node *merge = NULL;
 718
 719                 /* If there are no valid instructions to schedule, drop a NOP
 720                  * in.
 721                  */
 722                 uint64_t inst = chosen ? chosen->inst->inst : qpu_NOP();
 723
 724                 if (debug) {
 725                         fprintf(stderr, "t=%4d: current list:\n",
 726                                 time);
 727                         dump_state(schedule_list);
 728                         fprintf(stderr, "t=%4d: chose: ", time);
 729                         vc4_qpu_disasm(&inst, 1);
 730                         fprintf(stderr, "\n");
 731                 }
 732
 733                 /* Schedule this instruction onto the QPU list. Also try to
 734                  * find an instruction to pair with it.
 735                  */
 736                 if (chosen) {
 737                         time = MAX2(chosen->unblocked_time, time);
 738                         list_del(&chosen->link);
 739                         mark_instruction_scheduled(schedule_list, time,
 740                                                    chosen, true);
 741                         if (chosen->uniform != -1) {
 742                                 c->uniform_data[next_uniform] =
 743                                         uniform_data[chosen->uniform];
 744                                 c->uniform_contents[next_uniform] =
 745                                         uniform_contents[chosen->uniform];
 746                                 next_uniform++;
 747                         }
 748
 749                         merge = choose_instruction_to_schedule(&scoreboard,
 750                                                                schedule_list,
 751                                                                chosen);
 752                         if (merge) {
 753                                 time = MAX2(merge->unblocked_time, time);
 754                                 list_del(&merge->link);
 755                                 inst = qpu_merge_inst(inst, merge->inst->inst);
 756                                 assert(inst != 0);
 757                                 if (merge->uniform != -1) {
 758                                         c->uniform_data[next_uniform] =
 759                                                 uniform_data[merge->uniform];
 760                                         c->uniform_contents[next_uniform] =
 761                                                 uniform_contents[merge->uniform];
 762                                         next_uniform++;
 763                                 }
 764
 765                                 if (debug) {
 766                                         fprintf(stderr, "t=%4d: merging: ",
 767                                                 time);
 768                                         vc4_qpu_disasm(&merge->inst->inst, 1);
 769                                         fprintf(stderr, "\n");
 770                                         fprintf(stderr, "            resulting in: ");
 771                                         vc4_qpu_disasm(&inst, 1);
 772                                         fprintf(stderr, "\n");
 773                                 }
 774                         }
 775                 }
 776
 777                 if (debug) {
 778                         fprintf(stderr, "\n");
 779                 }
 780
 781                 qpu_serialize_one_inst(c, inst);
 782
 783                 update_scoreboard_for_chosen(&scoreboard, inst);
 784
 785                 /* Now that we've scheduled a new instruction, some of its
 786                  * children can be promoted to the list of instructions ready to
 787                  * be scheduled.  Update the children's unblocked time for this
 788                  * DAG edge as we do so.
 789                  */
 790                 mark_instruction_scheduled(schedule_list, time, chosen, false);
 791                 mark_instruction_scheduled(schedule_list, time, merge, false);
 792
 793                 scoreboard.tick++;
 794                 time++;
 795         }
 796
 797         assert(next_uniform == c->num_uniforms);
 798
 799         return time;
 800 }
 801
 802 static uint32_t waddr_latency(uint32_t waddr)
 803 {
 804         if (waddr < 32)
 805                 return 2;
 806
 807         /* Some huge number, really. */
 808         if (waddr >= QPU_W_TMU0_S && waddr <= QPU_W_TMU1_B)
 809                 return 100;
 810
 811         switch(waddr) {
 812         case QPU_W_SFU_RECIP:
 813         case QPU_W_SFU_RECIPSQRT:
 814         case QPU_W_SFU_EXP:
 815         case QPU_W_SFU_LOG:
 816                 return 3;
 817         default:
 818                 return 1;
 819         }
 820 }
 821
 822 static uint32_t
 823 instruction_latency(uint64_t inst)
 824 {
 825         return MAX2(waddr_latency(QPU_GET_FIELD(inst, QPU_WADDR_ADD)),
 826                     waddr_latency(QPU_GET_FIELD(inst, QPU_WADDR_MUL)));
 827 }
 828
 829 uint32_t
 830 qpu_schedule_instructions(struct vc4_compile *c)
 831 {
 832         void *mem_ctx = ralloc_context(NULL);
 833         struct list_head schedule_list;
 834
 835         list_inithead(&schedule_list);
 836
 837         if (debug) {
 838                 fprintf(stderr, "Pre-schedule instructions\n");
 839                 list_for_each_entry(struct queued_qpu_inst, q,
 840                                     &c->qpu_inst_list, link) {
 841                         vc4_qpu_disasm(&q->inst, 1);
 842                         fprintf(stderr, "\n");
 843                 }
 844                 fprintf(stderr, "\n");
 845         }
 846
 847         /* Wrap each instruction in a scheduler structure. */
 848         uint32_t next_uniform = 0;
 849         while (!list_empty(&c->qpu_inst_list)) {
 850                 struct queued_qpu_inst *inst =
 851                         (struct queued_qpu_inst *)c->qpu_inst_list.next;
 852                 struct schedule_node *n = rzalloc(mem_ctx, struct schedule_node);
 853
 854                 n->inst = inst;
 855                 n->latency = instruction_latency(inst->inst);
 856
 857                 if (reads_uniform(inst->inst)) {
 858                         n->uniform = next_uniform++;
 859                 } else {
 860                         n->uniform = -1;
 861                 }
 862                 list_del(&inst->link);
 863                 list_addtail(&n->link, &schedule_list);
 864         }
 865         assert(next_uniform == c->num_uniforms);
 866
 867         calculate_forward_deps(c, &schedule_list);
 868         calculate_reverse_deps(c, &schedule_list);
 869
 870         list_for_each_entry(struct schedule_node, n, &schedule_list, link) {
 871                 compute_delay(n);
 872         }
 873
 874         uint32_t cycles = schedule_instructions(c, &schedule_list);
 875
 876         if (debug) {
 877                 fprintf(stderr, "Post-schedule instructions\n");
 878                 vc4_qpu_disasm(c->qpu_insts, c->qpu_inst_count);
 879                 fprintf(stderr, "\n");
 880         }
 881
 882         ralloc_free(mem_ctx);
 883
 884         return cycles;
 885 }