1/* 2 * Copyright © 2015 Connor Abbott 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 "nir.h" 26#include "nir_vla.h" 27#include "nir_builder.h" 28#include "util/u_dynarray.h" 29 30#define HASH(hash, data) XXH32(&data, sizeof(data), hash) 31 32static uint32_t 33hash_src(uint32_t hash, const nir_src *src) 34{ 35 assert(src->is_ssa); 36 void *hash_data = nir_src_is_const(*src) ? NULL : src->ssa; 37 38 return HASH(hash, hash_data); 39} 40 41static uint32_t 42hash_alu_src(uint32_t hash, const nir_alu_src *src, 43 uint32_t num_components, uint32_t max_vec) 44{ 45 assert(!src->abs && !src->negate); 46 47 /* hash whether a swizzle accesses elements beyond the maximum 48 * vectorization factor: 49 * For example accesses to .x and .y are considered different variables 50 * compared to accesses to .z and .w for 16-bit vec2. 51 */ 52 uint32_t swizzle = (src->swizzle[0] & ~(max_vec - 1)); 53 hash = HASH(hash, swizzle); 54 55 return hash_src(hash, &src->src); 56} 57 58static uint32_t 59hash_instr(const void *data) 60{ 61 const nir_instr *instr = (nir_instr *) data; 62 assert(instr->type == nir_instr_type_alu); 63 nir_alu_instr *alu = nir_instr_as_alu(instr); 64 65 uint32_t hash = HASH(0, alu->op); 66 hash = HASH(hash, alu->dest.dest.ssa.bit_size); 67 68 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) 69 hash = hash_alu_src(hash, &alu->src[i], 70 alu->dest.dest.ssa.num_components, 71 instr->pass_flags); 72 73 return hash; 74} 75 76static bool 77srcs_equal(const nir_src *src1, const nir_src *src2) 78{ 79 assert(src1->is_ssa); 80 assert(src2->is_ssa); 81 82 return src1->ssa == src2->ssa || 83 (nir_src_is_const(*src1) && nir_src_is_const(*src2)); 84} 85 86static bool 87alu_srcs_equal(const nir_alu_src *src1, const nir_alu_src *src2, 88 uint32_t max_vec) 89{ 90 assert(!src1->abs); 91 assert(!src1->negate); 92 assert(!src2->abs); 93 assert(!src2->negate); 94 95 uint32_t mask = ~(max_vec - 1); 96 if ((src1->swizzle[0] & mask) != (src2->swizzle[0] & mask)) 97 return false; 98 99 return srcs_equal(&src1->src, &src2->src); 100} 101 102static bool 103instrs_equal(const void *data1, const void *data2) 104{ 105 const nir_instr *instr1 = (nir_instr *) data1; 106 const nir_instr *instr2 = (nir_instr *) data2; 107 assert(instr1->type == nir_instr_type_alu); 108 assert(instr2->type == nir_instr_type_alu); 109 110 nir_alu_instr *alu1 = nir_instr_as_alu(instr1); 111 nir_alu_instr *alu2 = nir_instr_as_alu(instr2); 112 113 if (alu1->op != alu2->op) 114 return false; 115 116 if (alu1->dest.dest.ssa.bit_size != alu2->dest.dest.ssa.bit_size) 117 return false; 118 119 for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) { 120 if (!alu_srcs_equal(&alu1->src[i], &alu2->src[i], instr1->pass_flags)) 121 return false; 122 } 123 124 return true; 125} 126 127static bool 128instr_can_rewrite(nir_instr *instr, bool vectorize_16bit) 129{ 130 switch (instr->type) { 131 case nir_instr_type_alu: { 132 nir_alu_instr *alu = nir_instr_as_alu(instr); 133 134 /* Don't try and vectorize mov's. Either they'll be handled by copy 135 * prop, or they're actually necessary and trying to vectorize them 136 * would result in fighting with copy prop. 137 */ 138 if (alu->op == nir_op_mov) 139 return false; 140 141 /* no need to hash instructions which are already vectorized */ 142 if (alu->dest.dest.ssa.num_components >= 4) 143 return false; 144 145 if (vectorize_16bit && 146 (alu->dest.dest.ssa.num_components >= 2 || 147 alu->dest.dest.ssa.bit_size != 16)) 148 return false; 149 150 if (nir_op_infos[alu->op].output_size != 0) 151 return false; 152 153 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) { 154 if (nir_op_infos[alu->op].input_sizes[i] != 0) 155 return false; 156 157 /* don't hash instructions which are already swizzled 158 * outside of max_components: these should better be scalarized */ 159 uint32_t mask = vectorize_16bit ? ~1 : ~3; 160 for (unsigned j = 0; j < alu->dest.dest.ssa.num_components; j++) { 161 if ((alu->src[i].swizzle[0] & mask) != (alu->src[i].swizzle[j] & mask)) 162 return false; 163 } 164 } 165 166 return true; 167 } 168 169 /* TODO support phi nodes */ 170 default: 171 break; 172 } 173 174 return false; 175} 176 177/* 178 * Tries to combine two instructions whose sources are different components of 179 * the same instructions into one vectorized instruction. Note that instr1 180 * should dominate instr2. 181 */ 182 183static nir_instr * 184instr_try_combine(struct nir_shader *nir, struct set *instr_set, 185 nir_instr *instr1, nir_instr *instr2) 186{ 187 assert(instr1->type == nir_instr_type_alu); 188 assert(instr2->type == nir_instr_type_alu); 189 nir_alu_instr *alu1 = nir_instr_as_alu(instr1); 190 nir_alu_instr *alu2 = nir_instr_as_alu(instr2); 191 192 assert(alu1->dest.dest.ssa.bit_size == alu2->dest.dest.ssa.bit_size); 193 unsigned alu1_components = alu1->dest.dest.ssa.num_components; 194 unsigned alu2_components = alu2->dest.dest.ssa.num_components; 195 unsigned total_components = alu1_components + alu2_components; 196 197 if (total_components > 4) 198 return NULL; 199 200 if (nir->options->vectorize_vec2_16bit) { 201 assert(total_components == 2); 202 assert(alu1->dest.dest.ssa.bit_size == 16); 203 } 204 205 nir_builder b; 206 nir_builder_init(&b, nir_cf_node_get_function(&instr1->block->cf_node)); 207 b.cursor = nir_after_instr(instr1); 208 209 nir_alu_instr *new_alu = nir_alu_instr_create(b.shader, alu1->op); 210 nir_ssa_dest_init(&new_alu->instr, &new_alu->dest.dest, 211 total_components, alu1->dest.dest.ssa.bit_size, NULL); 212 new_alu->dest.write_mask = (1 << total_components) - 1; 213 new_alu->instr.pass_flags = alu1->instr.pass_flags; 214 215 /* If either channel is exact, we have to preserve it even if it's 216 * not optimal for other channels. 217 */ 218 new_alu->exact = alu1->exact || alu2->exact; 219 220 /* If all channels don't wrap, we can say that the whole vector doesn't 221 * wrap. 222 */ 223 new_alu->no_signed_wrap = alu1->no_signed_wrap && alu2->no_signed_wrap; 224 new_alu->no_unsigned_wrap = alu1->no_unsigned_wrap && alu2->no_unsigned_wrap; 225 226 for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) { 227 /* handle constant merging case */ 228 if (alu1->src[i].src.ssa != alu2->src[i].src.ssa) { 229 nir_const_value *c1 = nir_src_as_const_value(alu1->src[i].src); 230 nir_const_value *c2 = nir_src_as_const_value(alu2->src[i].src); 231 assert(c1 && c2); 232 nir_const_value value[NIR_MAX_VEC_COMPONENTS]; 233 unsigned bit_size = alu1->src[i].src.ssa->bit_size; 234 235 for (unsigned j = 0; j < total_components; j++) { 236 value[j].u64 = j < alu1_components ? 237 c1[alu1->src[i].swizzle[j]].u64 : 238 c2[alu2->src[i].swizzle[j - alu1_components]].u64; 239 } 240 nir_ssa_def *def = nir_build_imm(&b, total_components, bit_size, value); 241 242 new_alu->src[i].src = nir_src_for_ssa(def); 243 for (unsigned j = 0; j < total_components; j++) 244 new_alu->src[i].swizzle[j] = j; 245 continue; 246 } 247 248 new_alu->src[i].src = alu1->src[i].src; 249 250 for (unsigned j = 0; j < alu1_components; j++) 251 new_alu->src[i].swizzle[j] = alu1->src[i].swizzle[j]; 252 253 for (unsigned j = 0; j < alu2_components; j++) { 254 new_alu->src[i].swizzle[j + alu1_components] = 255 alu2->src[i].swizzle[j]; 256 } 257 } 258 259 nir_builder_instr_insert(&b, &new_alu->instr); 260 261 unsigned swiz[NIR_MAX_VEC_COMPONENTS]; 262 for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) 263 swiz[i] = i; 264 nir_ssa_def *new_alu1 = nir_swizzle(&b, &new_alu->dest.dest.ssa, swiz, 265 alu1_components); 266 267 for (unsigned i = 0; i < alu2_components; i++) 268 swiz[i] += alu1_components; 269 nir_ssa_def *new_alu2 = nir_swizzle(&b, &new_alu->dest.dest.ssa, swiz, 270 alu2_components); 271 272 nir_foreach_use_safe(src, &alu1->dest.dest.ssa) { 273 nir_instr *user_instr = src->parent_instr; 274 if (user_instr->type == nir_instr_type_alu) { 275 /* Check if user is found in the hashset */ 276 struct set_entry *entry = _mesa_set_search(instr_set, user_instr); 277 278 /* For ALU instructions, rewrite the source directly to avoid a 279 * round-trip through copy propagation. 280 */ 281 nir_instr_rewrite_src(user_instr, src, 282 nir_src_for_ssa(&new_alu->dest.dest.ssa)); 283 284 /* Rehash user if it was found in the hashset */ 285 if (entry && entry->key == user_instr) { 286 _mesa_set_remove(instr_set, entry); 287 _mesa_set_add(instr_set, src->parent_instr); 288 } 289 } else { 290 nir_instr_rewrite_src(user_instr, src, nir_src_for_ssa(new_alu1)); 291 } 292 } 293 294 nir_foreach_if_use_safe(src, &alu1->dest.dest.ssa) { 295 nir_if_rewrite_condition(src->parent_if, nir_src_for_ssa(new_alu1)); 296 } 297 298 assert(nir_ssa_def_is_unused(&alu1->dest.dest.ssa)); 299 300 nir_foreach_use_safe(src, &alu2->dest.dest.ssa) { 301 if (src->parent_instr->type == nir_instr_type_alu) { 302 /* For ALU instructions, rewrite the source directly to avoid a 303 * round-trip through copy propagation. 304 */ 305 306 nir_alu_instr *use = nir_instr_as_alu(src->parent_instr); 307 308 unsigned src_index = 5; 309 for (unsigned i = 0; i < nir_op_infos[use->op].num_inputs; i++) { 310 if (&use->src[i].src == src) { 311 src_index = i; 312 break; 313 } 314 } 315 assert(src_index != 5); 316 317 nir_instr_rewrite_src(src->parent_instr, src, 318 nir_src_for_ssa(&new_alu->dest.dest.ssa)); 319 320 for (unsigned i = 0; 321 i < nir_ssa_alu_instr_src_components(use, src_index); i++) { 322 use->src[src_index].swizzle[i] += alu1_components; 323 } 324 } else { 325 nir_instr_rewrite_src(src->parent_instr, src, 326 nir_src_for_ssa(new_alu2)); 327 } 328 } 329 330 nir_foreach_if_use_safe(src, &alu2->dest.dest.ssa) { 331 nir_if_rewrite_condition(src->parent_if, nir_src_for_ssa(new_alu2)); 332 } 333 334 assert(nir_ssa_def_is_unused(&alu2->dest.dest.ssa)); 335 336 nir_instr_remove(instr1); 337 nir_instr_remove(instr2); 338 339 return &new_alu->instr; 340} 341 342static struct set * 343vec_instr_set_create(void) 344{ 345 return _mesa_set_create(NULL, hash_instr, instrs_equal); 346} 347 348static void 349vec_instr_set_destroy(struct set *instr_set) 350{ 351 _mesa_set_destroy(instr_set, NULL); 352} 353 354static bool 355vec_instr_set_add_or_rewrite(struct nir_shader *nir, struct set *instr_set, 356 nir_instr *instr, 357 nir_opt_vectorize_cb filter, void *data) 358{ 359 if (!instr_can_rewrite(instr, nir->options->vectorize_vec2_16bit)) 360 return false; 361 362 if (filter && !filter(instr, data)) 363 return false; 364 365 /* set max vector to instr pass flags: this is used to hash swizzles */ 366 instr->pass_flags = nir->options->vectorize_vec2_16bit ? 2 : 4; 367 368 struct set_entry *entry = _mesa_set_search(instr_set, instr); 369 if (entry) { 370 nir_instr *old_instr = (nir_instr *) entry->key; 371 _mesa_set_remove(instr_set, entry); 372 nir_instr *new_instr = instr_try_combine(nir, instr_set, 373 old_instr, instr); 374 if (new_instr) { 375 if (instr_can_rewrite(new_instr, nir->options->vectorize_vec2_16bit) && 376 (!filter || filter(new_instr, data))) 377 _mesa_set_add(instr_set, new_instr); 378 return true; 379 } 380 } 381 382 _mesa_set_add(instr_set, instr); 383 return false; 384} 385 386static bool 387vectorize_block(struct nir_shader *nir, nir_block *block, 388 struct set *instr_set, 389 nir_opt_vectorize_cb filter, void *data) 390{ 391 bool progress = false; 392 393 nir_foreach_instr_safe(instr, block) { 394 if (vec_instr_set_add_or_rewrite(nir, instr_set, instr, filter, data)) 395 progress = true; 396 } 397 398 for (unsigned i = 0; i < block->num_dom_children; i++) { 399 nir_block *child = block->dom_children[i]; 400 progress |= vectorize_block(nir, child, instr_set, filter, data); 401 } 402 403 nir_foreach_instr_reverse(instr, block) { 404 if (instr_can_rewrite(instr, nir->options->vectorize_vec2_16bit) && 405 (!filter || filter(instr, data))) 406 _mesa_set_remove_key(instr_set, instr); 407 } 408 409 return progress; 410} 411 412static bool 413nir_opt_vectorize_impl(struct nir_shader *nir, nir_function_impl *impl, 414 nir_opt_vectorize_cb filter, void *data) 415{ 416 struct set *instr_set = vec_instr_set_create(); 417 418 nir_metadata_require(impl, nir_metadata_dominance); 419 420 bool progress = vectorize_block(nir, nir_start_block(impl), instr_set, 421 filter, data); 422 423 if (progress) { 424 nir_metadata_preserve(impl, nir_metadata_block_index | 425 nir_metadata_dominance); 426 } else { 427 nir_metadata_preserve(impl, nir_metadata_all); 428 } 429 430 vec_instr_set_destroy(instr_set); 431 return progress; 432} 433 434bool 435nir_opt_vectorize(nir_shader *shader, nir_opt_vectorize_cb filter, 436 void *data) 437{ 438 bool progress = false; 439 440 nir_foreach_function(function, shader) { 441 if (function->impl) 442 progress |= nir_opt_vectorize_impl(shader, function->impl, filter, data); 443 } 444 445 return progress; 446} 447