1 /* 2 * Copyright 2010 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 * Authors: 24 * Eric Anholt <eric (at) anholt.net> 25 * 26 */ 27 28 #include "brw_fs.h" 29 #include "brw_fs_live_variables.h" 30 #include "brw_vec4.h" 31 #include "brw_cfg.h" 32 #include "brw_shader.h" 33 34 using namespace brw; 35 36 /** @file brw_fs_schedule_instructions.cpp 37 * 38 * List scheduling of FS instructions. 39 * 40 * The basic model of the list scheduler is to take a basic block, 41 * compute a DAG of the dependencies (RAW ordering with latency, WAW 42 * ordering with latency, WAR ordering), and make a list of the DAG heads. 43 * Heuristically pick a DAG head, then put all the children that are 44 * now DAG heads into the list of things to schedule. 45 * 46 * The heuristic is the important part. We're trying to be cheap, 47 * since actually computing the optimal scheduling is NP complete. 48 * What we do is track a "current clock". When we schedule a node, we 49 * update the earliest-unblocked clock time of its children, and 50 * increment the clock. Then, when trying to schedule, we just pick 51 * the earliest-unblocked instruction to schedule. 52 * 53 * Note that often there will be many things which could execute 54 * immediately, and there are a range of heuristic options to choose 55 * from in picking among those. 56 */ 57 58 static bool debug = false; 59 60 class instruction_scheduler; 61 62 class schedule_node : public exec_node 63 { 64 public: 65 schedule_node(backend_instruction *inst, instruction_scheduler *sched); 66 void set_latency_gen4(); 67 void set_latency_gen7(bool is_haswell); 68 69 backend_instruction *inst; 70 schedule_node **children; 71 int *child_latency; 72 int child_count; 73 int parent_count; 74 int child_array_size; 75 int unblocked_time; 76 int latency; 77 78 /** 79 * Which iteration of pushing groups of children onto the candidates list 80 * this node was a part of. 81 */ 82 unsigned cand_generation; 83 84 /** 85 * This is the sum of the instruction's latency plus the maximum delay of 86 * its children, or just the issue_time if it's a leaf node. 87 */ 88 int delay; 89 90 /** 91 * Preferred exit node among the (direct or indirect) successors of this 92 * node. Among the scheduler nodes blocked by this node, this will be the 93 * one that may cause earliest program termination, or NULL if none of the 94 * successors is an exit node. 95 */ 96 schedule_node *exit; 97 98 bool is_barrier; 99 }; 100 101 /** 102 * Lower bound of the scheduling time after which one of the instructions 103 * blocked by this node may lead to program termination. 104 * 105 * exit_unblocked_time() determines a strict partial ordering relation '' on 106 * the set of scheduler nodes as follows: 107 * 108 * n m <-> exit_unblocked_time(n) < exit_unblocked_time(m) 109 * 110 * which can be used to heuristically order nodes according to how early they 111 * can unblock an exit node and lead to program termination. 112 */ 113 static inline int 114 exit_unblocked_time(const schedule_node *n) 115 { 116 return n->exit ? n->exit->unblocked_time : INT_MAX; 117 } 118 119 void 120 schedule_node::set_latency_gen4() 121 { 122 int chans = 8; 123 int math_latency = 22; 124 125 switch (inst->opcode) { 126 case SHADER_OPCODE_RCP: 127 this->latency = 1 * chans * math_latency; 128 break; 129 case SHADER_OPCODE_RSQ: 130 this->latency = 2 * chans * math_latency; 131 break; 132 case SHADER_OPCODE_INT_QUOTIENT: 133 case SHADER_OPCODE_SQRT: 134 case SHADER_OPCODE_LOG2: 135 /* full precision log. partial is 2. */ 136 this->latency = 3 * chans * math_latency; 137 break; 138 case SHADER_OPCODE_INT_REMAINDER: 139 case SHADER_OPCODE_EXP2: 140 /* full precision. partial is 3, same throughput. */ 141 this->latency = 4 * chans * math_latency; 142 break; 143 case SHADER_OPCODE_POW: 144 this->latency = 8 * chans * math_latency; 145 break; 146 case SHADER_OPCODE_SIN: 147 case SHADER_OPCODE_COS: 148 /* minimum latency, max is 12 rounds. */ 149 this->latency = 5 * chans * math_latency; 150 break; 151 default: 152 this->latency = 2; 153 break; 154 } 155 } 156 157 void 158 schedule_node::set_latency_gen7(bool is_haswell) 159 { 160 switch (inst->opcode) { 161 case BRW_OPCODE_MAD: 162 /* 2 cycles 163 * (since the last two src operands are in different register banks): 164 * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; 165 * 166 * 3 cycles on IVB, 4 on HSW 167 * (since the last two src operands are in the same register bank): 168 * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; 169 * 170 * 18 cycles on IVB, 16 on HSW 171 * (since the last two src operands are in different register banks): 172 * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; 173 * mov(8) null g4<4,5,1>F { align16 WE_normal 1Q }; 174 * 175 * 20 cycles on IVB, 18 on HSW 176 * (since the last two src operands are in the same register bank): 177 * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; 178 * mov(8) null g4<4,4,1>F { align16 WE_normal 1Q }; 179 */ 180 181 /* Our register allocator doesn't know about register banks, so use the 182 * higher latency. 183 */ 184 latency = is_haswell ? 16 : 18; 185 break; 186 187 case BRW_OPCODE_LRP: 188 /* 2 cycles 189 * (since the last two src operands are in different register banks): 190 * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; 191 * 192 * 3 cycles on IVB, 4 on HSW 193 * (since the last two src operands are in the same register bank): 194 * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; 195 * 196 * 16 cycles on IVB, 14 on HSW 197 * (since the last two src operands are in different register banks): 198 * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; 199 * mov(8) null g4<4,4,1>F { align16 WE_normal 1Q }; 200 * 201 * 16 cycles 202 * (since the last two src operands are in the same register bank): 203 * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; 204 * mov(8) null g4<4,4,1>F { align16 WE_normal 1Q }; 205 */ 206 207 /* Our register allocator doesn't know about register banks, so use the 208 * higher latency. 209 */ 210 latency = 14; 211 break; 212 213 case SHADER_OPCODE_RCP: 214 case SHADER_OPCODE_RSQ: 215 case SHADER_OPCODE_SQRT: 216 case SHADER_OPCODE_LOG2: 217 case SHADER_OPCODE_EXP2: 218 case SHADER_OPCODE_SIN: 219 case SHADER_OPCODE_COS: 220 /* 2 cycles: 221 * math inv(8) g4<1>F g2<0,1,0>F null { align1 WE_normal 1Q }; 222 * 223 * 18 cycles: 224 * math inv(8) g4<1>F g2<0,1,0>F null { align1 WE_normal 1Q }; 225 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 226 * 227 * Same for exp2, log2, rsq, sqrt, sin, cos. 228 */ 229 latency = is_haswell ? 14 : 16; 230 break; 231 232 case SHADER_OPCODE_POW: 233 /* 2 cycles: 234 * math pow(8) g4<1>F g2<0,1,0>F g2.1<0,1,0>F { align1 WE_normal 1Q }; 235 * 236 * 26 cycles: 237 * math pow(8) g4<1>F g2<0,1,0>F g2.1<0,1,0>F { align1 WE_normal 1Q }; 238 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 239 */ 240 latency = is_haswell ? 22 : 24; 241 break; 242 243 case SHADER_OPCODE_TEX: 244 case SHADER_OPCODE_TXD: 245 case SHADER_OPCODE_TXF: 246 case SHADER_OPCODE_TXF_LZ: 247 case SHADER_OPCODE_TXL: 248 case SHADER_OPCODE_TXL_LZ: 249 /* 18 cycles: 250 * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; 251 * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; 252 * send(8) g4<1>UW g114<8,8,1>F 253 * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; 254 * 255 * 697 +/-49 cycles (min 610, n=26): 256 * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; 257 * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; 258 * send(8) g4<1>UW g114<8,8,1>F 259 * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; 260 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 261 * 262 * So the latency on our first texture load of the batchbuffer takes 263 * ~700 cycles, since the caches are cold at that point. 264 * 265 * 840 +/- 92 cycles (min 720, n=25): 266 * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; 267 * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; 268 * send(8) g4<1>UW g114<8,8,1>F 269 * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; 270 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 271 * send(8) g4<1>UW g114<8,8,1>F 272 * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; 273 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 274 * 275 * On the second load, it takes just an extra ~140 cycles, and after 276 * accounting for the 14 cycles of the MOV's latency, that makes ~130. 277 * 278 * 683 +/- 49 cycles (min = 602, n=47): 279 * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; 280 * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; 281 * send(8) g4<1>UW g114<8,8,1>F 282 * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; 283 * send(8) g50<1>UW g114<8,8,1>F 284 * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; 285 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 286 * 287 * The unit appears to be pipelined, since this matches up with the 288 * cache-cold case, despite there being two loads here. If you replace 289 * the g4 in the MOV to null with g50, it's still 693 +/- 52 (n=39). 290 * 291 * So, take some number between the cache-hot 140 cycles and the 292 * cache-cold 700 cycles. No particular tuning was done on this. 293 * 294 * I haven't done significant testing of the non-TEX opcodes. TXL at 295 * least looked about the same as TEX. 296 */ 297 latency = 200; 298 break; 299 300 case SHADER_OPCODE_TXS: 301 /* Testing textureSize(sampler2D, 0), one load was 420 +/- 41 302 * cycles (n=15): 303 * mov(8) g114<1>UD 0D { align1 WE_normal 1Q }; 304 * send(8) g6<1>UW g114<8,8,1>F 305 * sampler (10, 0, 10, 1) mlen 1 rlen 4 { align1 WE_normal 1Q }; 306 * mov(16) g6<1>F g6<8,8,1>D { align1 WE_normal 1Q }; 307 * 308 * 309 * Two loads was 535 +/- 30 cycles (n=19): 310 * mov(16) g114<1>UD 0D { align1 WE_normal 1H }; 311 * send(16) g6<1>UW g114<8,8,1>F 312 * sampler (10, 0, 10, 2) mlen 2 rlen 8 { align1 WE_normal 1H }; 313 * mov(16) g114<1>UD 0D { align1 WE_normal 1H }; 314 * mov(16) g6<1>F g6<8,8,1>D { align1 WE_normal 1H }; 315 * send(16) g8<1>UW g114<8,8,1>F 316 * sampler (10, 0, 10, 2) mlen 2 rlen 8 { align1 WE_normal 1H }; 317 * mov(16) g8<1>F g8<8,8,1>D { align1 WE_normal 1H }; 318 * add(16) g6<1>F g6<8,8,1>F g8<8,8,1>F { align1 WE_normal 1H }; 319 * 320 * Since the only caches that should matter are just the 321 * instruction/state cache containing the surface state, assume that we 322 * always have hot caches. 323 */ 324 latency = 100; 325 break; 326 327 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4: 328 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7: 329 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD: 330 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7: 331 case VS_OPCODE_PULL_CONSTANT_LOAD: 332 /* testing using varying-index pull constants: 333 * 334 * 16 cycles: 335 * mov(8) g4<1>D g2.1<0,1,0>F { align1 WE_normal 1Q }; 336 * send(8) g4<1>F g4<8,8,1>D 337 * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; 338 * 339 * ~480 cycles: 340 * mov(8) g4<1>D g2.1<0,1,0>F { align1 WE_normal 1Q }; 341 * send(8) g4<1>F g4<8,8,1>D 342 * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; 343 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 344 * 345 * ~620 cycles: 346 * mov(8) g4<1>D g2.1<0,1,0>F { align1 WE_normal 1Q }; 347 * send(8) g4<1>F g4<8,8,1>D 348 * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; 349 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 350 * send(8) g4<1>F g4<8,8,1>D 351 * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; 352 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 353 * 354 * So, if it's cache-hot, it's about 140. If it's cache cold, it's 355 * about 460. We expect to mostly be cache hot, so pick something more 356 * in that direction. 357 */ 358 latency = 200; 359 break; 360 361 case SHADER_OPCODE_GEN7_SCRATCH_READ: 362 /* Testing a load from offset 0, that had been previously written: 363 * 364 * send(8) g114<1>UW g0<8,8,1>F data (0, 0, 0) mlen 1 rlen 1 { align1 WE_normal 1Q }; 365 * mov(8) null g114<8,8,1>F { align1 WE_normal 1Q }; 366 * 367 * The cycles spent seemed to be grouped around 40-50 (as low as 38), 368 * then around 140. Presumably this is cache hit vs miss. 369 */ 370 latency = 50; 371 break; 372 373 case SHADER_OPCODE_UNTYPED_ATOMIC: 374 case SHADER_OPCODE_TYPED_ATOMIC: 375 /* Test code: 376 * mov(8) g112<1>ud 0x00000000ud { align1 WE_all 1Q }; 377 * mov(1) g112.7<1>ud g1.7<0,1,0>ud { align1 WE_all }; 378 * mov(8) g113<1>ud 0x00000000ud { align1 WE_normal 1Q }; 379 * send(8) g4<1>ud g112<8,8,1>ud 380 * data (38, 5, 6) mlen 2 rlen 1 { align1 WE_normal 1Q }; 381 * 382 * Running it 100 times as fragment shader on a 128x128 quad 383 * gives an average latency of 13867 cycles per atomic op, 384 * standard deviation 3%. Note that this is a rather 385 * pessimistic estimate, the actual latency in cases with few 386 * collisions between threads and favorable pipelining has been 387 * seen to be reduced by a factor of 100. 388 */ 389 latency = 14000; 390 break; 391 392 case SHADER_OPCODE_UNTYPED_SURFACE_READ: 393 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE: 394 case SHADER_OPCODE_TYPED_SURFACE_READ: 395 case SHADER_OPCODE_TYPED_SURFACE_WRITE: 396 /* Test code: 397 * mov(8) g112<1>UD 0x00000000UD { align1 WE_all 1Q }; 398 * mov(1) g112.7<1>UD g1.7<0,1,0>UD { align1 WE_all }; 399 * mov(8) g113<1>UD 0x00000000UD { align1 WE_normal 1Q }; 400 * send(8) g4<1>UD g112<8,8,1>UD 401 * data (38, 6, 5) mlen 2 rlen 1 { align1 WE_normal 1Q }; 402 * . 403 * . [repeats 8 times] 404 * . 405 * mov(8) g112<1>UD 0x00000000UD { align1 WE_all 1Q }; 406 * mov(1) g112.7<1>UD g1.7<0,1,0>UD { align1 WE_all }; 407 * mov(8) g113<1>UD 0x00000000UD { align1 WE_normal 1Q }; 408 * send(8) g4<1>UD g112<8,8,1>UD 409 * data (38, 6, 5) mlen 2 rlen 1 { align1 WE_normal 1Q }; 410 * 411 * Running it 100 times as fragment shader on a 128x128 quad 412 * gives an average latency of 583 cycles per surface read, 413 * standard deviation 0.9%. 414 */ 415 latency = is_haswell ? 300 : 600; 416 break; 417 418 default: 419 /* 2 cycles: 420 * mul(8) g4<1>F g2<0,1,0>F 0.5F { align1 WE_normal 1Q }; 421 * 422 * 16 cycles: 423 * mul(8) g4<1>F g2<0,1,0>F 0.5F { align1 WE_normal 1Q }; 424 * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; 425 */ 426 latency = 14; 427 break; 428 } 429 } 430 431 class instruction_scheduler { 432 public: 433 instruction_scheduler(backend_shader *s, int grf_count, 434 int hw_reg_count, int block_count, 435 instruction_scheduler_mode mode) 436 { 437 this->bs = s; 438 this->mem_ctx = ralloc_context(NULL); 439 this->grf_count = grf_count; 440 this->hw_reg_count = hw_reg_count; 441 this->instructions.make_empty(); 442 this->instructions_to_schedule = 0; 443 this->post_reg_alloc = (mode == SCHEDULE_POST); 444 this->mode = mode; 445 if (!post_reg_alloc) { 446 this->reg_pressure_in = rzalloc_array(mem_ctx, int, block_count); 447 448 this->livein = ralloc_array(mem_ctx, BITSET_WORD *, block_count); 449 for (int i = 0; i < block_count; i++) 450 this->livein[i] = rzalloc_array(mem_ctx, BITSET_WORD, 451 BITSET_WORDS(grf_count)); 452 453 this->liveout = ralloc_array(mem_ctx, BITSET_WORD *, block_count); 454 for (int i = 0; i < block_count; i++) 455 this->liveout[i] = rzalloc_array(mem_ctx, BITSET_WORD, 456 BITSET_WORDS(grf_count)); 457 458 this->hw_liveout = ralloc_array(mem_ctx, BITSET_WORD *, block_count); 459 for (int i = 0; i < block_count; i++) 460 this->hw_liveout[i] = rzalloc_array(mem_ctx, BITSET_WORD, 461 BITSET_WORDS(hw_reg_count)); 462 463 this->written = rzalloc_array(mem_ctx, bool, grf_count); 464 465 this->reads_remaining = rzalloc_array(mem_ctx, int, grf_count); 466 467 this->hw_reads_remaining = rzalloc_array(mem_ctx, int, hw_reg_count); 468 } else { 469 this->reg_pressure_in = NULL; 470 this->livein = NULL; 471 this->liveout = NULL; 472 this->hw_liveout = NULL; 473 this->written = NULL; 474 this->reads_remaining = NULL; 475 this->hw_reads_remaining = NULL; 476 } 477 } 478 479 ~instruction_scheduler() 480 { 481 ralloc_free(this->mem_ctx); 482 } 483 void add_barrier_deps(schedule_node *n); 484 void add_dep(schedule_node *before, schedule_node *after, int latency); 485 void add_dep(schedule_node *before, schedule_node *after); 486 487 void run(cfg_t *cfg); 488 void add_insts_from_block(bblock_t *block); 489 void compute_delays(); 490 void compute_exits(); 491 virtual void calculate_deps() = 0; 492 virtual schedule_node *choose_instruction_to_schedule() = 0; 493 494 /** 495 * Returns how many cycles it takes the instruction to issue. 496 * 497 * Instructions in gen hardware are handled one simd4 vector at a time, 498 * with 1 cycle per vector dispatched. Thus SIMD8 pixel shaders take 2 499 * cycles to dispatch and SIMD16 (compressed) instructions take 4. 500 */ 501 virtual int issue_time(backend_instruction *inst) = 0; 502 503 virtual void count_reads_remaining(backend_instruction *inst) = 0; 504 virtual void setup_liveness(cfg_t *cfg) = 0; 505 virtual void update_register_pressure(backend_instruction *inst) = 0; 506 virtual int get_register_pressure_benefit(backend_instruction *inst) = 0; 507 508 void schedule_instructions(bblock_t *block); 509 510 void *mem_ctx; 511 512 bool post_reg_alloc; 513 int instructions_to_schedule; 514 int grf_count; 515 int hw_reg_count; 516 int reg_pressure; 517 int block_idx; 518 exec_list instructions; 519 backend_shader *bs; 520 521 instruction_scheduler_mode mode; 522 523 /* 524 * The register pressure at the beginning of each basic block. 525 */ 526 527 int *reg_pressure_in; 528 529 /* 530 * The virtual GRF's whose range overlaps the beginning of each basic block. 531 */ 532 533 BITSET_WORD **livein; 534 535 /* 536 * The virtual GRF's whose range overlaps the end of each basic block. 537 */ 538 539 BITSET_WORD **liveout; 540 541 /* 542 * The hardware GRF's whose range overlaps the end of each basic block. 543 */ 544 545 BITSET_WORD **hw_liveout; 546 547 /* 548 * Whether we've scheduled a write for this virtual GRF yet. 549 */ 550 551 bool *written; 552 553 /* 554 * How many reads we haven't scheduled for this virtual GRF yet. 555 */ 556 557 int *reads_remaining; 558 559 /* 560 * How many reads we haven't scheduled for this hardware GRF yet. 561 */ 562 563 int *hw_reads_remaining; 564 }; 565 566 class fs_instruction_scheduler : public instruction_scheduler 567 { 568 public: 569 fs_instruction_scheduler(fs_visitor *v, int grf_count, int hw_reg_count, 570 int block_count, 571 instruction_scheduler_mode mode); 572 void calculate_deps(); 573 bool is_compressed(fs_inst *inst); 574 schedule_node *choose_instruction_to_schedule(); 575 int issue_time(backend_instruction *inst); 576 fs_visitor *v; 577 578 void count_reads_remaining(backend_instruction *inst); 579 void setup_liveness(cfg_t *cfg); 580 void update_register_pressure(backend_instruction *inst); 581 int get_register_pressure_benefit(backend_instruction *inst); 582 }; 583 584 fs_instruction_scheduler::fs_instruction_scheduler(fs_visitor *v, 585 int grf_count, int hw_reg_count, 586 int block_count, 587 instruction_scheduler_mode mode) 588 : instruction_scheduler(v, grf_count, hw_reg_count, block_count, mode), 589 v(v) 590 { 591 } 592 593 static bool 594 is_src_duplicate(fs_inst *inst, int src) 595 { 596 for (int i = 0; i < src; i++) 597 if (inst->src[i].equals(inst->src[src])) 598 return true; 599 600 return false; 601 } 602 603 void 604 fs_instruction_scheduler::count_reads_remaining(backend_instruction *be) 605 { 606 fs_inst *inst = (fs_inst *)be; 607 608 if (!reads_remaining) 609 return; 610 611 for (int i = 0; i < inst->sources; i++) { 612 if (is_src_duplicate(inst, i)) 613 continue; 614 615 if (inst->src[i].file == VGRF) { 616 reads_remaining[inst->src[i].nr]++; 617 } else if (inst->src[i].file == FIXED_GRF) { 618 if (inst->src[i].nr >= hw_reg_count) 619 continue; 620 621 for (unsigned j = 0; j < regs_read(inst, i); j++) 622 hw_reads_remaining[inst->src[i].nr + j]++; 623 } 624 } 625 } 626 627 void 628 fs_instruction_scheduler::setup_liveness(cfg_t *cfg) 629 { 630 /* First, compute liveness on a per-GRF level using the in/out sets from 631 * liveness calculation. 632 */ 633 for (int block = 0; block < cfg->num_blocks; block++) { 634 for (int i = 0; i < v->live_intervals->num_vars; i++) { 635 if (BITSET_TEST(v->live_intervals->block_data[block].livein, i)) { 636 int vgrf = v->live_intervals->vgrf_from_var[i]; 637 if (!BITSET_TEST(livein[block], vgrf)) { 638 reg_pressure_in[block] += v->alloc.sizes[vgrf]; 639 BITSET_SET(livein[block], vgrf); 640 } 641 } 642 643 if (BITSET_TEST(v->live_intervals->block_data[block].liveout, i)) 644 BITSET_SET(liveout[block], v->live_intervals->vgrf_from_var[i]); 645 } 646 } 647 648 /* Now, extend the live in/live out sets for when a range crosses a block 649 * boundary, which matches what our register allocator/interference code 650 * does to account for force_writemask_all and incompatible exec_mask's. 651 */ 652 for (int block = 0; block < cfg->num_blocks - 1; block++) { 653 for (int i = 0; i < grf_count; i++) { 654 if (v->virtual_grf_start[i] <= cfg->blocks[block]->end_ip && 655 v->virtual_grf_end[i] >= cfg->blocks[block + 1]->start_ip) { 656 if (!BITSET_TEST(livein[block + 1], i)) { 657 reg_pressure_in[block + 1] += v->alloc.sizes[i]; 658 BITSET_SET(livein[block + 1], i); 659 } 660 661 BITSET_SET(liveout[block], i); 662 } 663 } 664 } 665 666 int payload_last_use_ip[hw_reg_count]; 667 v->calculate_payload_ranges(hw_reg_count, payload_last_use_ip); 668 669 for (int i = 0; i < hw_reg_count; i++) { 670 if (payload_last_use_ip[i] == -1) 671 continue; 672 673 for (int block = 0; block < cfg->num_blocks; block++) { 674 if (cfg->blocks[block]->start_ip <= payload_last_use_ip[i]) 675 reg_pressure_in[block]++; 676 677 if (cfg->blocks[block]->end_ip <= payload_last_use_ip[i]) 678 BITSET_SET(hw_liveout[block], i); 679 } 680 } 681 } 682 683 void 684 fs_instruction_scheduler::update_register_pressure(backend_instruction *be) 685 { 686 fs_inst *inst = (fs_inst *)be; 687 688 if (!reads_remaining) 689 return; 690 691 if (inst->dst.file == VGRF) { 692 written[inst->dst.nr] = true; 693 } 694 695 for (int i = 0; i < inst->sources; i++) { 696 if (is_src_duplicate(inst, i)) 697 continue; 698 699 if (inst->src[i].file == VGRF) { 700 reads_remaining[inst->src[i].nr]--; 701 } else if (inst->src[i].file == FIXED_GRF && 702 inst->src[i].nr < hw_reg_count) { 703 for (unsigned off = 0; off < regs_read(inst, i); off++) 704 hw_reads_remaining[inst->src[i].nr + off]--; 705 } 706 } 707 } 708 709 int 710 fs_instruction_scheduler::get_register_pressure_benefit(backend_instruction *be) 711 { 712 fs_inst *inst = (fs_inst *)be; 713 int benefit = 0; 714 715 if (inst->dst.file == VGRF) { 716 if (!BITSET_TEST(livein[block_idx], inst->dst.nr) && 717 !written[inst->dst.nr]) 718 benefit -= v->alloc.sizes[inst->dst.nr]; 719 } 720 721 for (int i = 0; i < inst->sources; i++) { 722 if (is_src_duplicate(inst, i)) 723 continue; 724 725 if (inst->src[i].file == VGRF && 726 !BITSET_TEST(liveout[block_idx], inst->src[i].nr) && 727 reads_remaining[inst->src[i].nr] == 1) 728 benefit += v->alloc.sizes[inst->src[i].nr]; 729 730 if (inst->src[i].file == FIXED_GRF && 731 inst->src[i].nr < hw_reg_count) { 732 for (unsigned off = 0; off < regs_read(inst, i); off++) { 733 int reg = inst->src[i].nr + off; 734 if (!BITSET_TEST(hw_liveout[block_idx], reg) && 735 hw_reads_remaining[reg] == 1) { 736 benefit++; 737 } 738 } 739 } 740 } 741 742 return benefit; 743 } 744 745 class vec4_instruction_scheduler : public instruction_scheduler 746 { 747 public: 748 vec4_instruction_scheduler(vec4_visitor *v, int grf_count); 749 void calculate_deps(); 750 schedule_node *choose_instruction_to_schedule(); 751 int issue_time(backend_instruction *inst); 752 vec4_visitor *v; 753 754 void count_reads_remaining(backend_instruction *inst); 755 void setup_liveness(cfg_t *cfg); 756 void update_register_pressure(backend_instruction *inst); 757 int get_register_pressure_benefit(backend_instruction *inst); 758 }; 759 760 vec4_instruction_scheduler::vec4_instruction_scheduler(vec4_visitor *v, 761 int grf_count) 762 : instruction_scheduler(v, grf_count, 0, 0, SCHEDULE_POST), 763 v(v) 764 { 765 } 766 767 void 768 vec4_instruction_scheduler::count_reads_remaining(backend_instruction *be) 769 { 770 } 771 772 void 773 vec4_instruction_scheduler::setup_liveness(cfg_t *cfg) 774 { 775 } 776 777 void 778 vec4_instruction_scheduler::update_register_pressure(backend_instruction *be) 779 { 780 } 781 782 int 783 vec4_instruction_scheduler::get_register_pressure_benefit(backend_instruction *be) 784 { 785 return 0; 786 } 787 788 schedule_node::schedule_node(backend_instruction *inst, 789 instruction_scheduler *sched) 790 { 791 const struct gen_device_info *devinfo = sched->bs->devinfo; 792 793 this->inst = inst; 794 this->child_array_size = 0; 795 this->children = NULL; 796 this->child_latency = NULL; 797 this->child_count = 0; 798 this->parent_count = 0; 799 this->unblocked_time = 0; 800 this->cand_generation = 0; 801 this->delay = 0; 802 this->exit = NULL; 803 this->is_barrier = false; 804 805 /* We can't measure Gen6 timings directly but expect them to be much 806 * closer to Gen7 than Gen4. 807 */ 808 if (!sched->post_reg_alloc) 809 this->latency = 1; 810 else if (devinfo->gen >= 6) 811 set_latency_gen7(devinfo->is_haswell); 812 else 813 set_latency_gen4(); 814 } 815 816 void 817 instruction_scheduler::add_insts_from_block(bblock_t *block) 818 { 819 foreach_inst_in_block(backend_instruction, inst, block) { 820 schedule_node *n = new(mem_ctx) schedule_node(inst, this); 821 822 instructions.push_tail(n); 823 } 824 825 this->instructions_to_schedule = block->end_ip - block->start_ip + 1; 826 } 827 828 /** Computation of the delay member of each node. */ 829 void 830 instruction_scheduler::compute_delays() 831 { 832 foreach_in_list_reverse(schedule_node, n, &instructions) { 833 if (!n->child_count) { 834 n->delay = issue_time(n->inst); 835 } else { 836 for (int i = 0; i < n->child_count; i++) { 837 assert(n->children[i]->delay); 838 n->delay = MAX2(n->delay, n->latency + n->children[i]->delay); 839 } 840 } 841 } 842 } 843 844 void 845 instruction_scheduler::compute_exits() 846 { 847 /* Calculate a lower bound of the scheduling time of each node in the 848 * graph. This is analogous to the node's critical path but calculated 849 * from the top instead of from the bottom of the block. 850 */ 851 foreach_in_list(schedule_node, n, &instructions) { 852 for (int i = 0; i < n->child_count; i++) { 853 n->children[i]->unblocked_time = 854 MAX2(n->children[i]->unblocked_time, 855 n->unblocked_time + issue_time(n->inst) + n->child_latency[i]); 856 } 857 } 858 859 /* Calculate the exit of each node by induction based on the exit nodes of 860 * its children. The preferred exit of a node is the one among the exit 861 * nodes of its children which can be unblocked first according to the 862 * optimistic unblocked time estimate calculated above. 863 */ 864 foreach_in_list_reverse(schedule_node, n, &instructions) { 865 n->exit = (n->inst->opcode == FS_OPCODE_DISCARD_JUMP ? n : NULL); 866 867 for (int i = 0; i < n->child_count; i++) { 868 if (exit_unblocked_time(n->children[i]) < exit_unblocked_time(n)) 869 n->exit = n->children[i]->exit; 870 } 871 } 872 } 873 874 /** 875 * Add a dependency between two instruction nodes. 876 * 877 * The @after node will be scheduled after @before. We will try to 878 * schedule it @latency cycles after @before, but no guarantees there. 879 */ 880 void 881 instruction_scheduler::add_dep(schedule_node *before, schedule_node *after, 882 int latency) 883 { 884 if (!before || !after) 885 return; 886 887 assert(before != after); 888 889 for (int i = 0; i < before->child_count; i++) { 890 if (before->children[i] == after) { 891 before->child_latency[i] = MAX2(before->child_latency[i], latency); 892 return; 893 } 894 } 895 896 if (before->child_array_size <= before->child_count) { 897 if (before->child_array_size < 16) 898 before->child_array_size = 16; 899 else 900 before->child_array_size *= 2; 901 902 before->children = reralloc(mem_ctx, before->children, 903 schedule_node *, 904 before->child_array_size); 905 before->child_latency = reralloc(mem_ctx, before->child_latency, 906 int, before->child_array_size); 907 } 908 909 before->children[before->child_count] = after; 910 before->child_latency[before->child_count] = latency; 911 before->child_count++; 912 after->parent_count++; 913 } 914 915 void 916 instruction_scheduler::add_dep(schedule_node *before, schedule_node *after) 917 { 918 if (!before) 919 return; 920 921 add_dep(before, after, before->latency); 922 } 923 924 /** 925 * Sometimes we really want this node to execute after everything that 926 * was before it and before everything that followed it. This adds 927 * the deps to do so. 928 */ 929 void 930 instruction_scheduler::add_barrier_deps(schedule_node *n) 931 { 932 schedule_node *prev = (schedule_node *)n->prev; 933 schedule_node *next = (schedule_node *)n->next; 934 935 n->is_barrier = true; 936 937 if (prev) { 938 while (!prev->is_head_sentinel()) { 939 add_dep(prev, n, 0); 940 if (prev->is_barrier) 941 break; 942 prev = (schedule_node *)prev->prev; 943 } 944 } 945 946 if (next) { 947 while (!next->is_tail_sentinel()) { 948 add_dep(n, next, 0); 949 if (next->is_barrier) 950 break; 951 next = (schedule_node *)next->next; 952 } 953 } 954 } 955 956 /* instruction scheduling needs to be aware of when an MRF write 957 * actually writes 2 MRFs. 958 */ 959 bool 960 fs_instruction_scheduler::is_compressed(fs_inst *inst) 961 { 962 return inst->exec_size == 16; 963 } 964 965 static bool 966 is_scheduling_barrier(const fs_inst *inst) 967 { 968 return inst->opcode == FS_OPCODE_PLACEHOLDER_HALT || 969 inst->is_control_flow() || 970 inst->has_side_effects(); 971 } 972 973 void 974 fs_instruction_scheduler::calculate_deps() 975 { 976 /* Pre-register-allocation, this tracks the last write per VGRF offset. 977 * After register allocation, reg_offsets are gone and we track individual 978 * GRF registers. 979 */ 980 schedule_node *last_grf_write[grf_count * 16]; 981 schedule_node *last_mrf_write[BRW_MAX_MRF(v->devinfo->gen)]; 982 schedule_node *last_conditional_mod[4] = {}; 983 schedule_node *last_accumulator_write = NULL; 984 /* Fixed HW registers are assumed to be separate from the virtual 985 * GRFs, so they can be tracked separately. We don't really write 986 * to fixed GRFs much, so don't bother tracking them on a more 987 * granular level. 988 */ 989 schedule_node *last_fixed_grf_write = NULL; 990 991 memset(last_grf_write, 0, sizeof(last_grf_write)); 992 memset(last_mrf_write, 0, sizeof(last_mrf_write)); 993 994 /* top-to-bottom dependencies: RAW and WAW. */ 995 foreach_in_list(schedule_node, n, &instructions) { 996 fs_inst *inst = (fs_inst *)n->inst; 997 998 if (is_scheduling_barrier(inst)) 999 add_barrier_deps(n); 1000 1001 /* read-after-write deps. */ 1002 for (int i = 0; i < inst->sources; i++) { 1003 if (inst->src[i].file == VGRF) { 1004 if (post_reg_alloc) { 1005 for (unsigned r = 0; r < regs_read(inst, i); r++) 1006 add_dep(last_grf_write[inst->src[i].nr + r], n); 1007 } else { 1008 for (unsigned r = 0; r < regs_read(inst, i); r++) { 1009 add_dep(last_grf_write[inst->src[i].nr * 16 + 1010 inst->src[i].offset / REG_SIZE + r], n); 1011 } 1012 } 1013 } else if (inst->src[i].file == FIXED_GRF) { 1014 if (post_reg_alloc) { 1015 for (unsigned r = 0; r < regs_read(inst, i); r++) 1016 add_dep(last_grf_write[inst->src[i].nr + r], n); 1017 } else { 1018 add_dep(last_fixed_grf_write, n); 1019 } 1020 } else if (inst->src[i].is_accumulator()) { 1021 add_dep(last_accumulator_write, n); 1022 } else if (inst->src[i].file == ARF) { 1023 add_barrier_deps(n); 1024 } 1025 } 1026 1027 if (inst->base_mrf != -1) { 1028 for (int i = 0; i < inst->mlen; i++) { 1029 /* It looks like the MRF regs are released in the send 1030 * instruction once it's sent, not when the result comes 1031 * back. 1032 */ 1033 add_dep(last_mrf_write[inst->base_mrf + i], n); 1034 } 1035 } 1036 1037 if (const unsigned mask = inst->flags_read(v->devinfo)) { 1038 assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); 1039 1040 for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { 1041 if (mask & (1 << i)) 1042 add_dep(last_conditional_mod[i], n); 1043 } 1044 } 1045 1046 if (inst->reads_accumulator_implicitly()) { 1047 add_dep(last_accumulator_write, n); 1048 } 1049 1050 /* write-after-write deps. */ 1051 if (inst->dst.file == VGRF) { 1052 if (post_reg_alloc) { 1053 for (unsigned r = 0; r < regs_written(inst); r++) { 1054 add_dep(last_grf_write[inst->dst.nr + r], n); 1055 last_grf_write[inst->dst.nr + r] = n; 1056 } 1057 } else { 1058 for (unsigned r = 0; r < regs_written(inst); r++) { 1059 add_dep(last_grf_write[inst->dst.nr * 16 + 1060 inst->dst.offset / REG_SIZE + r], n); 1061 last_grf_write[inst->dst.nr * 16 + 1062 inst->dst.offset / REG_SIZE + r] = n; 1063 } 1064 } 1065 } else if (inst->dst.file == MRF) { 1066 int reg = inst->dst.nr & ~BRW_MRF_COMPR4; 1067 1068 add_dep(last_mrf_write[reg], n); 1069 last_mrf_write[reg] = n; 1070 if (is_compressed(inst)) { 1071 if (inst->dst.nr & BRW_MRF_COMPR4) 1072 reg += 4; 1073 else 1074 reg++; 1075 add_dep(last_mrf_write[reg], n); 1076 last_mrf_write[reg] = n; 1077 } 1078 } else if (inst->dst.file == FIXED_GRF) { 1079 if (post_reg_alloc) { 1080 for (unsigned r = 0; r < regs_written(inst); r++) 1081 last_grf_write[inst->dst.nr + r] = n; 1082 } else { 1083 last_fixed_grf_write = n; 1084 } 1085 } else if (inst->dst.is_accumulator()) { 1086 add_dep(last_accumulator_write, n); 1087 last_accumulator_write = n; 1088 } else if (inst->dst.file == ARF && !inst->dst.is_null()) { 1089 add_barrier_deps(n); 1090 } 1091 1092 if (inst->mlen > 0 && inst->base_mrf != -1) { 1093 for (int i = 0; i < v->implied_mrf_writes(inst); i++) { 1094 add_dep(last_mrf_write[inst->base_mrf + i], n); 1095 last_mrf_write[inst->base_mrf + i] = n; 1096 } 1097 } 1098 1099 if (const unsigned mask = inst->flags_written()) { 1100 assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); 1101 1102 for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { 1103 if (mask & (1 << i)) { 1104 add_dep(last_conditional_mod[i], n, 0); 1105 last_conditional_mod[i] = n; 1106 } 1107 } 1108 } 1109 1110 if (inst->writes_accumulator_implicitly(v->devinfo) && 1111 !inst->dst.is_accumulator()) { 1112 add_dep(last_accumulator_write, n); 1113 last_accumulator_write = n; 1114 } 1115 } 1116 1117 /* bottom-to-top dependencies: WAR */ 1118 memset(last_grf_write, 0, sizeof(last_grf_write)); 1119 memset(last_mrf_write, 0, sizeof(last_mrf_write)); 1120 memset(last_conditional_mod, 0, sizeof(last_conditional_mod)); 1121 last_accumulator_write = NULL; 1122 last_fixed_grf_write = NULL; 1123 1124 foreach_in_list_reverse_safe(schedule_node, n, &instructions) { 1125 fs_inst *inst = (fs_inst *)n->inst; 1126 1127 /* write-after-read deps. */ 1128 for (int i = 0; i < inst->sources; i++) { 1129 if (inst->src[i].file == VGRF) { 1130 if (post_reg_alloc) { 1131 for (unsigned r = 0; r < regs_read(inst, i); r++) 1132 add_dep(n, last_grf_write[inst->src[i].nr + r], 0); 1133 } else { 1134 for (unsigned r = 0; r < regs_read(inst, i); r++) { 1135 add_dep(n, last_grf_write[inst->src[i].nr * 16 + 1136 inst->src[i].offset / REG_SIZE + r], 0); 1137 } 1138 } 1139 } else if (inst->src[i].file == FIXED_GRF) { 1140 if (post_reg_alloc) { 1141 for (unsigned r = 0; r < regs_read(inst, i); r++) 1142 add_dep(n, last_grf_write[inst->src[i].nr + r], 0); 1143 } else { 1144 add_dep(n, last_fixed_grf_write, 0); 1145 } 1146 } else if (inst->src[i].is_accumulator()) { 1147 add_dep(n, last_accumulator_write, 0); 1148 } else if (inst->src[i].file == ARF) { 1149 add_barrier_deps(n); 1150 } 1151 } 1152 1153 if (inst->base_mrf != -1) { 1154 for (int i = 0; i < inst->mlen; i++) { 1155 /* It looks like the MRF regs are released in the send 1156 * instruction once it's sent, not when the result comes 1157 * back. 1158 */ 1159 add_dep(n, last_mrf_write[inst->base_mrf + i], 2); 1160 } 1161 } 1162 1163 if (const unsigned mask = inst->flags_read(v->devinfo)) { 1164 assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); 1165 1166 for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { 1167 if (mask & (1 << i)) 1168 add_dep(n, last_conditional_mod[i]); 1169 } 1170 } 1171 1172 if (inst->reads_accumulator_implicitly()) { 1173 add_dep(n, last_accumulator_write); 1174 } 1175 1176 /* Update the things this instruction wrote, so earlier reads 1177 * can mark this as WAR dependency. 1178 */ 1179 if (inst->dst.file == VGRF) { 1180 if (post_reg_alloc) { 1181 for (unsigned r = 0; r < regs_written(inst); r++) 1182 last_grf_write[inst->dst.nr + r] = n; 1183 } else { 1184 for (unsigned r = 0; r < regs_written(inst); r++) { 1185 last_grf_write[inst->dst.nr * 16 + 1186 inst->dst.offset / REG_SIZE + r] = n; 1187 } 1188 } 1189 } else if (inst->dst.file == MRF) { 1190 int reg = inst->dst.nr & ~BRW_MRF_COMPR4; 1191 1192 last_mrf_write[reg] = n; 1193 1194 if (is_compressed(inst)) { 1195 if (inst->dst.nr & BRW_MRF_COMPR4) 1196 reg += 4; 1197 else 1198 reg++; 1199 1200 last_mrf_write[reg] = n; 1201 } 1202 } else if (inst->dst.file == FIXED_GRF) { 1203 if (post_reg_alloc) { 1204 for (unsigned r = 0; r < regs_written(inst); r++) 1205 last_grf_write[inst->dst.nr + r] = n; 1206 } else { 1207 last_fixed_grf_write = n; 1208 } 1209 } else if (inst->dst.is_accumulator()) { 1210 last_accumulator_write = n; 1211 } else if (inst->dst.file == ARF && !inst->dst.is_null()) { 1212 add_barrier_deps(n); 1213 } 1214 1215 if (inst->mlen > 0 && inst->base_mrf != -1) { 1216 for (int i = 0; i < v->implied_mrf_writes(inst); i++) { 1217 last_mrf_write[inst->base_mrf + i] = n; 1218 } 1219 } 1220 1221 if (const unsigned mask = inst->flags_written()) { 1222 assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); 1223 1224 for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { 1225 if (mask & (1 << i)) 1226 last_conditional_mod[i] = n; 1227 } 1228 } 1229 1230 if (inst->writes_accumulator_implicitly(v->devinfo)) { 1231 last_accumulator_write = n; 1232 } 1233 } 1234 } 1235 1236 static bool 1237 is_scheduling_barrier(const vec4_instruction *inst) 1238 { 1239 return inst->is_control_flow() || 1240 inst->has_side_effects(); 1241 } 1242 1243 void 1244 vec4_instruction_scheduler::calculate_deps() 1245 { 1246 schedule_node *last_grf_write[grf_count]; 1247 schedule_node *last_mrf_write[BRW_MAX_MRF(v->devinfo->gen)]; 1248 schedule_node *last_conditional_mod = NULL; 1249 schedule_node *last_accumulator_write = NULL; 1250 /* Fixed HW registers are assumed to be separate from the virtual 1251 * GRFs, so they can be tracked separately. We don't really write 1252 * to fixed GRFs much, so don't bother tracking them on a more 1253 * granular level. 1254 */ 1255 schedule_node *last_fixed_grf_write = NULL; 1256 1257 memset(last_grf_write, 0, sizeof(last_grf_write)); 1258 memset(last_mrf_write, 0, sizeof(last_mrf_write)); 1259 1260 /* top-to-bottom dependencies: RAW and WAW. */ 1261 foreach_in_list(schedule_node, n, &instructions) { 1262 vec4_instruction *inst = (vec4_instruction *)n->inst; 1263 1264 if (is_scheduling_barrier(inst)) 1265 add_barrier_deps(n); 1266 1267 /* read-after-write deps. */ 1268 for (int i = 0; i < 3; i++) { 1269 if (inst->src[i].file == VGRF) { 1270 for (unsigned j = 0; j < regs_read(inst, i); ++j) 1271 add_dep(last_grf_write[inst->src[i].nr + j], n); 1272 } else if (inst->src[i].file == FIXED_GRF) { 1273 add_dep(last_fixed_grf_write, n); 1274 } else if (inst->src[i].is_accumulator()) { 1275 assert(last_accumulator_write); 1276 add_dep(last_accumulator_write, n); 1277 } else if (inst->src[i].file == ARF) { 1278 add_barrier_deps(n); 1279 } 1280 } 1281 1282 if (!inst->is_send_from_grf()) { 1283 for (int i = 0; i < inst->mlen; i++) { 1284 /* It looks like the MRF regs are released in the send 1285 * instruction once it's sent, not when the result comes 1286 * back. 1287 */ 1288 add_dep(last_mrf_write[inst->base_mrf + i], n); 1289 } 1290 } 1291 1292 if (inst->reads_flag()) { 1293 assert(last_conditional_mod); 1294 add_dep(last_conditional_mod, n); 1295 } 1296 1297 if (inst->reads_accumulator_implicitly()) { 1298 assert(last_accumulator_write); 1299 add_dep(last_accumulator_write, n); 1300 } 1301 1302 /* write-after-write deps. */ 1303 if (inst->dst.file == VGRF) { 1304 for (unsigned j = 0; j < regs_written(inst); ++j) { 1305 add_dep(last_grf_write[inst->dst.nr + j], n); 1306 last_grf_write[inst->dst.nr + j] = n; 1307 } 1308 } else if (inst->dst.file == MRF) { 1309 add_dep(last_mrf_write[inst->dst.nr], n); 1310 last_mrf_write[inst->dst.nr] = n; 1311 } else if (inst->dst.file == FIXED_GRF) { 1312 last_fixed_grf_write = n; 1313 } else if (inst->dst.is_accumulator()) { 1314 add_dep(last_accumulator_write, n); 1315 last_accumulator_write = n; 1316 } else if (inst->dst.file == ARF && !inst->dst.is_null()) { 1317 add_barrier_deps(n); 1318 } 1319 1320 if (inst->mlen > 0 && !inst->is_send_from_grf()) { 1321 for (int i = 0; i < v->implied_mrf_writes(inst); i++) { 1322 add_dep(last_mrf_write[inst->base_mrf + i], n); 1323 last_mrf_write[inst->base_mrf + i] = n; 1324 } 1325 } 1326 1327 if (inst->writes_flag()) { 1328 add_dep(last_conditional_mod, n, 0); 1329 last_conditional_mod = n; 1330 } 1331 1332 if (inst->writes_accumulator_implicitly(v->devinfo) && 1333 !inst->dst.is_accumulator()) { 1334 add_dep(last_accumulator_write, n); 1335 last_accumulator_write = n; 1336 } 1337 } 1338 1339 /* bottom-to-top dependencies: WAR */ 1340 memset(last_grf_write, 0, sizeof(last_grf_write)); 1341 memset(last_mrf_write, 0, sizeof(last_mrf_write)); 1342 last_conditional_mod = NULL; 1343 last_accumulator_write = NULL; 1344 last_fixed_grf_write = NULL; 1345 1346 foreach_in_list_reverse_safe(schedule_node, n, &instructions) { 1347 vec4_instruction *inst = (vec4_instruction *)n->inst; 1348 1349 /* write-after-read deps. */ 1350 for (int i = 0; i < 3; i++) { 1351 if (inst->src[i].file == VGRF) { 1352 for (unsigned j = 0; j < regs_read(inst, i); ++j) 1353 add_dep(n, last_grf_write[inst->src[i].nr + j]); 1354 } else if (inst->src[i].file == FIXED_GRF) { 1355 add_dep(n, last_fixed_grf_write); 1356 } else if (inst->src[i].is_accumulator()) { 1357 add_dep(n, last_accumulator_write); 1358 } else if (inst->src[i].file == ARF) { 1359 add_barrier_deps(n); 1360 } 1361 } 1362 1363 if (!inst->is_send_from_grf()) { 1364 for (int i = 0; i < inst->mlen; i++) { 1365 /* It looks like the MRF regs are released in the send 1366 * instruction once it's sent, not when the result comes 1367 * back. 1368 */ 1369 add_dep(n, last_mrf_write[inst->base_mrf + i], 2); 1370 } 1371 } 1372 1373 if (inst->reads_flag()) { 1374 add_dep(n, last_conditional_mod); 1375 } 1376 1377 if (inst->reads_accumulator_implicitly()) { 1378 add_dep(n, last_accumulator_write); 1379 } 1380 1381 /* Update the things this instruction wrote, so earlier reads 1382 * can mark this as WAR dependency. 1383 */ 1384 if (inst->dst.file == VGRF) { 1385 for (unsigned j = 0; j < regs_written(inst); ++j) 1386 last_grf_write[inst->dst.nr + j] = n; 1387 } else if (inst->dst.file == MRF) { 1388 last_mrf_write[inst->dst.nr] = n; 1389 } else if (inst->dst.file == FIXED_GRF) { 1390 last_fixed_grf_write = n; 1391 } else if (inst->dst.is_accumulator()) { 1392 last_accumulator_write = n; 1393 } else if (inst->dst.file == ARF && !inst->dst.is_null()) { 1394 add_barrier_deps(n); 1395 } 1396 1397 if (inst->mlen > 0 && !inst->is_send_from_grf()) { 1398 for (int i = 0; i < v->implied_mrf_writes(inst); i++) { 1399 last_mrf_write[inst->base_mrf + i] = n; 1400 } 1401 } 1402 1403 if (inst->writes_flag()) { 1404 last_conditional_mod = n; 1405 } 1406 1407 if (inst->writes_accumulator_implicitly(v->devinfo)) { 1408 last_accumulator_write = n; 1409 } 1410 } 1411 } 1412 1413 schedule_node * 1414 fs_instruction_scheduler::choose_instruction_to_schedule() 1415 { 1416 schedule_node *chosen = NULL; 1417 1418 if (mode == SCHEDULE_PRE || mode == SCHEDULE_POST) { 1419 int chosen_time = 0; 1420 1421 /* Of the instructions ready to execute or the closest to being ready, 1422 * choose the one most likely to unblock an early program exit, or 1423 * otherwise the oldest one. 1424 */ 1425 foreach_in_list(schedule_node, n, &instructions) { 1426 if (!chosen || 1427 exit_unblocked_time(n) < exit_unblocked_time(chosen) || 1428 (exit_unblocked_time(n) == exit_unblocked_time(chosen) && 1429 n->unblocked_time < chosen_time)) { 1430 chosen = n; 1431 chosen_time = n->unblocked_time; 1432 } 1433 } 1434 } else { 1435 /* Before register allocation, we don't care about the latencies of 1436 * instructions. All we care about is reducing live intervals of 1437 * variables so that we can avoid register spilling, or get SIMD16 1438 * shaders which naturally do a better job of hiding instruction 1439 * latency. 1440 */ 1441 foreach_in_list(schedule_node, n, &instructions) { 1442 fs_inst *inst = (fs_inst *)n->inst; 1443 1444 if (!chosen) { 1445 chosen = n; 1446 continue; 1447 } 1448 1449 /* Most important: If we can definitely reduce register pressure, do 1450 * so immediately. 1451 */ 1452 int register_pressure_benefit = get_register_pressure_benefit(n->inst); 1453 int chosen_register_pressure_benefit = 1454 get_register_pressure_benefit(chosen->inst); 1455 1456 if (register_pressure_benefit > 0 && 1457 register_pressure_benefit > chosen_register_pressure_benefit) { 1458 chosen = n; 1459 continue; 1460 } else if (chosen_register_pressure_benefit > 0 && 1461 (register_pressure_benefit < 1462 chosen_register_pressure_benefit)) { 1463 continue; 1464 } 1465 1466 if (mode == SCHEDULE_PRE_LIFO) { 1467 /* Prefer instructions that recently became available for 1468 * scheduling. These are the things that are most likely to 1469 * (eventually) make a variable dead and reduce register pressure. 1470 * Typical register pressure estimates don't work for us because 1471 * most of our pressure comes from texturing, where no single 1472 * instruction to schedule will make a vec4 value dead. 1473 */ 1474 if (n->cand_generation > chosen->cand_generation) { 1475 chosen = n; 1476 continue; 1477 } else if (n->cand_generation < chosen->cand_generation) { 1478 continue; 1479 } 1480 1481 /* On MRF-using chips, prefer non-SEND instructions. If we don't 1482 * do this, then because we prefer instructions that just became 1483 * candidates, we'll end up in a pattern of scheduling a SEND, 1484 * then the MRFs for the next SEND, then the next SEND, then the 1485 * MRFs, etc., without ever consuming the results of a send. 1486 */ 1487 if (v->devinfo->gen < 7) { 1488 fs_inst *chosen_inst = (fs_inst *)chosen->inst; 1489 1490 /* We use size_written > 4 * exec_size as our test for the kind 1491 * of send instruction to avoid -- only sends generate many 1492 * regs, and a single-result send is probably actually reducing 1493 * register pressure. 1494 */ 1495 if (inst->size_written <= 4 * inst->exec_size && 1496 chosen_inst->size_written > 4 * chosen_inst->exec_size) { 1497 chosen = n; 1498 continue; 1499 } else if (inst->size_written > chosen_inst->size_written) { 1500 continue; 1501 } 1502 } 1503 } 1504 1505 /* For instructions pushed on the cands list at the same time, prefer 1506 * the one with the highest delay to the end of the program. This is 1507 * most likely to have its values able to be consumed first (such as 1508 * for a large tree of lowered ubo loads, which appear reversed in 1509 * the instruction stream with respect to when they can be consumed). 1510 */ 1511 if (n->delay > chosen->delay) { 1512 chosen = n; 1513 continue; 1514 } else if (n->delay < chosen->delay) { 1515 continue; 1516 } 1517 1518 /* Prefer the node most likely to unblock an early program exit. 1519 */ 1520 if (exit_unblocked_time(n) < exit_unblocked_time(chosen)) { 1521 chosen = n; 1522 continue; 1523 } else if (exit_unblocked_time(n) > exit_unblocked_time(chosen)) { 1524 continue; 1525 } 1526 1527 /* If all other metrics are equal, we prefer the first instruction in 1528 * the list (program execution). 1529 */ 1530 } 1531 } 1532 1533 return chosen; 1534 } 1535 1536 schedule_node * 1537 vec4_instruction_scheduler::choose_instruction_to_schedule() 1538 { 1539 schedule_node *chosen = NULL; 1540 int chosen_time = 0; 1541 1542 /* Of the instructions ready to execute or the closest to being ready, 1543 * choose the oldest one. 1544 */ 1545 foreach_in_list(schedule_node, n, &instructions) { 1546 if (!chosen || n->unblocked_time < chosen_time) { 1547 chosen = n; 1548 chosen_time = n->unblocked_time; 1549 } 1550 } 1551 1552 return chosen; 1553 } 1554 1555 int 1556 fs_instruction_scheduler::issue_time(backend_instruction *inst) 1557 { 1558 if (is_compressed((fs_inst *)inst)) 1559 return 4; 1560 else 1561 return 2; 1562 } 1563 1564 int 1565 vec4_instruction_scheduler::issue_time(backend_instruction *inst) 1566 { 1567 /* We always execute as two vec4s in parallel. */ 1568 return 2; 1569 } 1570 1571 void 1572 instruction_scheduler::schedule_instructions(bblock_t *block) 1573 { 1574 const struct gen_device_info *devinfo = bs->devinfo; 1575 int time = 0; 1576 if (!post_reg_alloc) 1577 reg_pressure = reg_pressure_in[block->num]; 1578 block_idx = block->num; 1579 1580 /* Remove non-DAG heads from the list. */ 1581 foreach_in_list_safe(schedule_node, n, &instructions) { 1582 if (n->parent_count != 0) 1583 n->remove(); 1584 } 1585 1586 unsigned cand_generation = 1; 1587 while (!instructions.is_empty()) { 1588 schedule_node *chosen = choose_instruction_to_schedule(); 1589 1590 /* Schedule this instruction. */ 1591 assert(chosen); 1592 chosen->remove(); 1593 chosen->inst->exec_node::remove(); 1594 block->instructions.push_tail(chosen->inst); 1595 instructions_to_schedule--; 1596 1597 if (!post_reg_alloc) { 1598 reg_pressure -= get_register_pressure_benefit(chosen->inst); 1599 update_register_pressure(chosen->inst); 1600 } 1601 1602 /* If we expected a delay for scheduling, then bump the clock to reflect 1603 * that. In reality, the hardware will switch to another hyperthread 1604 * and may not return to dispatching our thread for a while even after 1605 * we're unblocked. After this, we have the time when the chosen 1606 * instruction will start executing. 1607 */ 1608 time = MAX2(time, chosen->unblocked_time); 1609 1610 /* Update the clock for how soon an instruction could start after the 1611 * chosen one. 1612 */ 1613 time += issue_time(chosen->inst); 1614 1615 if (debug) { 1616 fprintf(stderr, "clock %4d, scheduled: ", time); 1617 bs->dump_instruction(chosen->inst); 1618 if (!post_reg_alloc) 1619 fprintf(stderr, "(register pressure %d)\n", reg_pressure); 1620 } 1621 1622 /* Now that we've scheduled a new instruction, some of its 1623 * children can be promoted to the list of instructions ready to 1624 * be scheduled. Update the children's unblocked time for this 1625 * DAG edge as we do so. 1626 */ 1627 for (int i = chosen->child_count - 1; i >= 0; i--) { 1628 schedule_node *child = chosen->children[i]; 1629 1630 child->unblocked_time = MAX2(child->unblocked_time, 1631 time + chosen->child_latency[i]); 1632 1633 if (debug) { 1634 fprintf(stderr, "\tchild %d, %d parents: ", i, child->parent_count); 1635 bs->dump_instruction(child->inst); 1636 } 1637 1638 child->cand_generation = cand_generation; 1639 child->parent_count--; 1640 if (child->parent_count == 0) { 1641 if (debug) { 1642 fprintf(stderr, "\t\tnow available\n"); 1643 } 1644 instructions.push_head(child); 1645 } 1646 } 1647 cand_generation++; 1648 1649 /* Shared resource: the mathbox. There's one mathbox per EU on Gen6+ 1650 * but it's more limited pre-gen6, so if we send something off to it then 1651 * the next math instruction isn't going to make progress until the first 1652 * is done. 1653 */ 1654 if (devinfo->gen < 6 && chosen->inst->is_math()) { 1655 foreach_in_list(schedule_node, n, &instructions) { 1656 if (n->inst->is_math()) 1657 n->unblocked_time = MAX2(n->unblocked_time, 1658 time + chosen->latency); 1659 } 1660 } 1661 } 1662 1663 assert(instructions_to_schedule == 0); 1664 1665 block->cycle_count = time; 1666 } 1667 1668 static unsigned get_cycle_count(cfg_t *cfg) 1669 { 1670 unsigned count = 0, multiplier = 1; 1671 foreach_block(block, cfg) { 1672 if (block->start()->opcode == BRW_OPCODE_DO) 1673 multiplier *= 10; /* assume that loops execute ~10 times */ 1674 1675 count += block->cycle_count * multiplier; 1676 1677 if (block->end()->opcode == BRW_OPCODE_WHILE) 1678 multiplier /= 10; 1679 } 1680 1681 return count; 1682 } 1683 1684 void 1685 instruction_scheduler::run(cfg_t *cfg) 1686 { 1687 if (debug && !post_reg_alloc) { 1688 fprintf(stderr, "\nInstructions before scheduling (reg_alloc %d)\n", 1689 post_reg_alloc); 1690 bs->dump_instructions(); 1691 } 1692 1693 if (!post_reg_alloc) 1694 setup_liveness(cfg); 1695 1696 foreach_block(block, cfg) { 1697 if (reads_remaining) { 1698 memset(reads_remaining, 0, 1699 grf_count * sizeof(*reads_remaining)); 1700 memset(hw_reads_remaining, 0, 1701 hw_reg_count * sizeof(*hw_reads_remaining)); 1702 memset(written, 0, grf_count * sizeof(*written)); 1703 1704 foreach_inst_in_block(fs_inst, inst, block) 1705 count_reads_remaining(inst); 1706 } 1707 1708 add_insts_from_block(block); 1709 1710 calculate_deps(); 1711 1712 compute_delays(); 1713 compute_exits(); 1714 1715 schedule_instructions(block); 1716 } 1717 1718 if (debug && !post_reg_alloc) { 1719 fprintf(stderr, "\nInstructions after scheduling (reg_alloc %d)\n", 1720 post_reg_alloc); 1721 bs->dump_instructions(); 1722 } 1723 1724 cfg->cycle_count = get_cycle_count(cfg); 1725 } 1726 1727 void 1728 fs_visitor::schedule_instructions(instruction_scheduler_mode mode) 1729 { 1730 if (mode != SCHEDULE_POST) 1731 calculate_live_intervals(); 1732 1733 int grf_count; 1734 if (mode == SCHEDULE_POST) 1735 grf_count = grf_used; 1736 else 1737 grf_count = alloc.count; 1738 1739 fs_instruction_scheduler sched(this, grf_count, first_non_payload_grf, 1740 cfg->num_blocks, mode); 1741 sched.run(cfg); 1742 1743 invalidate_live_intervals(); 1744 } 1745 1746 void 1747 vec4_visitor::opt_schedule_instructions() 1748 { 1749 vec4_instruction_scheduler sched(this, prog_data->total_grf); 1750 sched.run(cfg); 1751 1752 invalidate_live_intervals(); 1753 } 1754