src/compiler/nir/nir_opt_vectorize.c

   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) _mesa_fnv32_1a_accumulate((hash), (data))
  31
  32 static uint32_t
  33 hash_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
  41 static uint32_t
  42 hash_alu_src(uint32_t hash, const nir_alu_src *src)
  43 {
  44    assert(!src->abs && !src->negate);
  45
  46    /* intentionally don't hash swizzle */
  47
  48    return hash_src(hash, &src->src);
  49 }
  50
  51 static uint32_t
  52 hash_alu(uint32_t hash, const nir_alu_instr *instr)
  53 {
  54    hash = HASH(hash, instr->op);
  55
  56    hash = HASH(hash, instr->dest.dest.ssa.bit_size);
  57
  58    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
  59       hash = hash_alu_src(hash, &instr->src[i]);
  60
  61    return hash;
  62 }
  63
  64 static uint32_t
  65 hash_instr(const nir_instr *instr)
  66 {
  67    uint32_t hash = _mesa_fnv32_1a_offset_bias;
  68
  69    switch (instr->type) {
  70    case nir_instr_type_alu:
  71       return hash_alu(hash, nir_instr_as_alu(instr));
  72    default:
  73       unreachable("bad instruction type");
  74    }
  75 }
  76
  77 static bool
  78 srcs_equal(const nir_src *src1, const nir_src *src2)
  79 {
  80    assert(src1->is_ssa);
  81    assert(src2->is_ssa);
  82
  83    return src1->ssa == src2->ssa ||
  84       nir_src_is_const(*src1) == nir_src_is_const(*src2);
  85 }
  86
  87 static bool
  88 alu_srcs_equal(const nir_alu_src *src1, const nir_alu_src *src2)
  89 {
  90    assert(!src1->abs);
  91    assert(!src1->negate);
  92    assert(!src2->abs);
  93    assert(!src2->negate);
  94
  95    return srcs_equal(&src1->src, &src2->src);
  96 }
  97
  98 static bool
  99 instrs_equal(const nir_instr *instr1, const nir_instr *instr2)
 100 {
 101    switch (instr1->type) {
 102    case nir_instr_type_alu: {
 103       nir_alu_instr *alu1 = nir_instr_as_alu(instr1);
 104       nir_alu_instr *alu2 = nir_instr_as_alu(instr2);
 105
 106       if (alu1->op != alu2->op)
 107          return false;
 108
 109       if (alu1->dest.dest.ssa.bit_size != alu2->dest.dest.ssa.bit_size)
 110          return false;
 111
 112       for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
 113          if (!alu_srcs_equal(&alu1->src[i], &alu2->src[i]))
 114             return false;
 115       }
 116
 117       return true;
 118    }
 119
 120    default:
 121       unreachable("bad instruction type");
 122    }
 123 }
 124
 125 static bool
 126 instr_can_rewrite(nir_instr *instr)
 127 {
 128    switch (instr->type) {
 129    case nir_instr_type_alu: {
 130       nir_alu_instr *alu = nir_instr_as_alu(instr);
 131
 132       /* Don't try and vectorize mov's. Either they'll be handled by copy
 133        * prop, or they're actually necessary and trying to vectorize them
 134        * would result in fighting with copy prop.
 135        */
 136       if (alu->op == nir_op_mov)
 137          return false;
 138
 139       if (nir_op_infos[alu->op].output_size != 0)
 140          return false;
 141
 142       for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 143          if (nir_op_infos[alu->op].input_sizes[i] != 0)
 144             return false;
 145       }
 146
 147       return true;
 148    }
 149
 150    /* TODO support phi nodes */
 151    default:
 152       break;
 153    }
 154
 155    return false;
 156 }
 157
 158 /*
 159  * Tries to combine two instructions whose sources are different components of
 160  * the same instructions into one vectorized instruction. Note that instr1
 161  * should dominate instr2.
 162  */
 163
 164 static nir_instr *
 165 instr_try_combine(nir_instr *instr1, nir_instr *instr2)
 166 {
 167    assert(instr1->type == nir_instr_type_alu);
 168    assert(instr2->type == nir_instr_type_alu);
 169    nir_alu_instr *alu1 = nir_instr_as_alu(instr1);
 170    nir_alu_instr *alu2 = nir_instr_as_alu(instr2);
 171
 172    assert(alu1->dest.dest.ssa.bit_size == alu2->dest.dest.ssa.bit_size);
 173    unsigned alu1_components = alu1->dest.dest.ssa.num_components;
 174    unsigned alu2_components = alu2->dest.dest.ssa.num_components;
 175    unsigned total_components = alu1_components + alu2_components;
 176
 177    if (total_components > 4)
 178       return NULL;
 179
 180    nir_builder b;
 181    nir_builder_init(&b, nir_cf_node_get_function(&instr1->block->cf_node));
 182    b.cursor = nir_after_instr(instr1);
 183
 184    nir_alu_instr *new_alu = nir_alu_instr_create(b.shader, alu1->op);
 185    nir_ssa_dest_init(&new_alu->instr, &new_alu->dest.dest,
 186                      total_components, alu1->dest.dest.ssa.bit_size, NULL);
 187    new_alu->dest.write_mask = (1 << total_components) - 1;
 188
 189    for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
 190       /* handle constant merging case */
 191       if (alu1->src[i].src.ssa != alu2->src[i].src.ssa) {
 192          nir_const_value *c1 = nir_src_as_const_value(alu1->src[i].src);
 193          nir_const_value *c2 = nir_src_as_const_value(alu2->src[i].src);
 194          assert(c1 && c2);
 195          nir_const_value value[4];
 196          unsigned bit_size = alu1->src[i].src.ssa->bit_size;
 197
 198          for (unsigned j = 0; j < total_components; j++) {
 199             value[j].u64 = j < alu1_components ?
 200                               c1[alu1->src[i].swizzle[j]].u64 :
 201                               c2[alu2->src[i].swizzle[j - alu1_components]].u64;
 202          }
 203          nir_ssa_def *def = nir_build_imm(&b, total_components, bit_size, value);
 204
 205          new_alu->src[i].src = nir_src_for_ssa(def);
 206          for (unsigned j = 0; j < total_components; j++)
 207             new_alu->src[i].swizzle[j] = j;
 208          continue;
 209       }
 210
 211       new_alu->src[i].src = alu1->src[i].src;
 212
 213       for (unsigned j = 0; j < alu1_components; j++)
 214          new_alu->src[i].swizzle[j] = alu1->src[i].swizzle[j];
 215
 216       for (unsigned j = 0; j < alu2_components; j++) {
 217          new_alu->src[i].swizzle[j + alu1_components] =
 218             alu2->src[i].swizzle[j];
 219       }
 220    }
 221
 222    nir_builder_instr_insert(&b, &new_alu->instr);
 223
 224    unsigned swiz[4] = {0, 1, 2, 3};
 225    nir_ssa_def *new_alu1 = nir_swizzle(&b, &new_alu->dest.dest.ssa, swiz,
 226                                        alu1_components);
 227
 228    for (unsigned i = 0; i < alu2_components; i++)
 229       swiz[i] += alu1_components;
 230    nir_ssa_def *new_alu2 = nir_swizzle(&b, &new_alu->dest.dest.ssa, swiz,
 231                                        alu2_components);
 232
 233    nir_foreach_use_safe(src, &alu1->dest.dest.ssa) {
 234       if (src->parent_instr->type == nir_instr_type_alu) {
 235          /* For ALU instructions, rewrite the source directly to avoid a
 236           * round-trip through copy propagation.
 237           */
 238
 239          nir_instr_rewrite_src(src->parent_instr, src,
 240                                nir_src_for_ssa(&new_alu->dest.dest.ssa));
 241       } else {
 242          nir_instr_rewrite_src(src->parent_instr, src,
 243                                nir_src_for_ssa(new_alu1));
 244       }
 245    }
 246
 247    nir_foreach_if_use_safe(src, &alu1->dest.dest.ssa) {
 248       nir_if_rewrite_condition(src->parent_if, nir_src_for_ssa(new_alu1));
 249    }
 250
 251    assert(list_is_empty(&alu1->dest.dest.ssa.uses));
 252    assert(list_is_empty(&alu1->dest.dest.ssa.if_uses));
 253
 254    nir_foreach_use_safe(src, &alu2->dest.dest.ssa) {
 255       if (src->parent_instr->type == nir_instr_type_alu) {
 256          /* For ALU instructions, rewrite the source directly to avoid a
 257           * round-trip through copy propagation.
 258           */
 259
 260          nir_alu_instr *use = nir_instr_as_alu(src->parent_instr);
 261
 262          unsigned src_index = 5;
 263          for (unsigned i = 0; i < nir_op_infos[use->op].num_inputs; i++) {
 264             if (&use->src[i].src == src) {
 265                src_index = i;
 266                break;
 267             }
 268          }
 269          assert(src_index != 5);
 270
 271          nir_instr_rewrite_src(src->parent_instr, src,
 272                                nir_src_for_ssa(&new_alu->dest.dest.ssa));
 273
 274          for (unsigned i = 0;
 275               i < nir_ssa_alu_instr_src_components(use, src_index); i++) {
 276             use->src[src_index].swizzle[i] += alu1_components;
 277          }
 278       } else {
 279          nir_instr_rewrite_src(src->parent_instr, src,
 280                                nir_src_for_ssa(new_alu2));
 281       }
 282    }
 283
 284    nir_foreach_if_use_safe(src, &alu2->dest.dest.ssa) {
 285       nir_if_rewrite_condition(src->parent_if, nir_src_for_ssa(new_alu2));
 286    }
 287
 288    assert(list_is_empty(&alu2->dest.dest.ssa.uses));
 289    assert(list_is_empty(&alu2->dest.dest.ssa.if_uses));
 290
 291    nir_instr_remove(instr1);
 292    nir_instr_remove(instr2);
 293
 294    return &new_alu->instr;
 295 }
 296
 297 /*
 298  * Use an array to represent a stack of instructions that are equivalent.
 299  *
 300  * We push and pop instructions off the stack in dominance order. The first
 301  * element dominates the second element which dominates the third, etc. When
 302  * trying to add to the stack, first we try and combine the instruction with
 303  * each of the instructions on the stack and, if successful, replace the
 304  * instruction on the stack with the newly-combined instruction.
 305  */
 306
 307 static struct util_dynarray *
 308 vec_instr_stack_create(void *mem_ctx)
 309 {
 310    struct util_dynarray *stack = ralloc(mem_ctx, struct util_dynarray);
 311    util_dynarray_init(stack, mem_ctx);
 312    return stack;
 313 }
 314
 315 /* returns true if we were able to successfully replace the instruction */
 316
 317 static bool
 318 vec_instr_stack_push(struct util_dynarray *stack, nir_instr *instr)
 319 {
 320    /* Walk the stack from child to parent to make live ranges shorter by
 321     * matching the closest thing we can
 322     */
 323    util_dynarray_foreach_reverse(stack, nir_instr *, stack_instr) {
 324       nir_instr *new_instr = instr_try_combine(*stack_instr, instr);
 325       if (new_instr) {
 326          *stack_instr = new_instr;
 327          return true;
 328       }
 329    }
 330
 331    util_dynarray_append(stack, nir_instr *, instr);
 332    return false;
 333 }
 334
 335 static void
 336 vec_instr_stack_pop(struct util_dynarray *stack, nir_instr *instr)
 337 {
 338    ASSERTED nir_instr *last = util_dynarray_pop(stack, nir_instr *);
 339    assert(last == instr);
 340 }
 341
 342 static bool
 343 cmp_func(const void *data1, const void *data2)
 344 {
 345    const struct util_dynarray *arr1 = data1;
 346    const struct util_dynarray *arr2 = data2;
 347
 348    const nir_instr *instr1 = *(nir_instr **)util_dynarray_begin(arr1);
 349    const nir_instr *instr2 = *(nir_instr **)util_dynarray_begin(arr2);
 350
 351    return instrs_equal(instr1, instr2);
 352 }
 353
 354 static uint32_t
 355 hash_stack(const void *data)
 356 {
 357    const struct util_dynarray *stack = data;
 358    const nir_instr *first = *(nir_instr **)util_dynarray_begin(stack);
 359    return hash_instr(first);
 360 }
 361
 362 static struct set *
 363 vec_instr_set_create(void)
 364 {
 365    return _mesa_set_create(NULL, hash_stack, cmp_func);
 366 }
 367
 368 static void
 369 vec_instr_set_destroy(struct set *instr_set)
 370 {
 371    _mesa_set_destroy(instr_set, NULL);
 372 }
 373
 374 static bool
 375 vec_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr)
 376 {
 377    if (!instr_can_rewrite(instr))
 378       return false;
 379
 380    struct util_dynarray *new_stack = vec_instr_stack_create(instr_set);
 381    vec_instr_stack_push(new_stack, instr);
 382
 383    struct set_entry *entry = _mesa_set_search(instr_set, new_stack);
 384
 385    if (entry) {
 386       ralloc_free(new_stack);
 387       struct util_dynarray *stack = (struct util_dynarray *) entry->key;
 388       return vec_instr_stack_push(stack, instr);
 389    }
 390
 391    _mesa_set_add(instr_set, new_stack);
 392    return false;
 393 }
 394
 395 static void
 396 vec_instr_set_remove(struct set *instr_set, nir_instr *instr)
 397 {
 398    if (!instr_can_rewrite(instr))
 399       return;
 400
 401    /*
 402     * It's pretty unfortunate that we have to do this, but it's a side effect
 403     * of the hash set interfaces. The hash set assumes that we're only
 404     * interested in storing one equivalent element at a time, and if we try to
 405     * insert a duplicate element it will remove the original. We could hack up
 406     * the comparison function to "know" which input is an instruction we
 407     * passed in and which is an array that's part of the entry, but that
 408     * wouldn't work because we need to pass an array to _mesa_set_add() in
 409     * vec_instr_add_or_rewrite() above, and _mesa_set_add() will call our
 410     * comparison function as well.
 411     */
 412    struct util_dynarray *temp = vec_instr_stack_create(instr_set);
 413    vec_instr_stack_push(temp, instr);
 414    struct set_entry *entry = _mesa_set_search(instr_set, temp);
 415    ralloc_free(temp);
 416
 417    if (entry) {
 418       struct util_dynarray *stack = (struct util_dynarray *) entry->key;
 419
 420       if (util_dynarray_num_elements(stack, nir_instr *) > 1)
 421          vec_instr_stack_pop(stack, instr);
 422       else
 423          _mesa_set_remove(instr_set, entry);
 424    }
 425 }
 426
 427 static bool
 428 vectorize_block(nir_block *block, struct set *instr_set)
 429 {
 430    bool progress = false;
 431
 432    nir_foreach_instr_safe(instr, block) {
 433       if (vec_instr_set_add_or_rewrite(instr_set, instr))
 434          progress = true;
 435    }
 436
 437    for (unsigned i = 0; i < block->num_dom_children; i++) {
 438       nir_block *child = block->dom_children[i];
 439       progress |= vectorize_block(child, instr_set);
 440    }
 441
 442    nir_foreach_instr_reverse(instr, block)
 443       vec_instr_set_remove(instr_set, instr);
 444
 445    return progress;
 446 }
 447
 448 static bool
 449 nir_opt_vectorize_impl(nir_function_impl *impl)
 450 {
 451    struct set *instr_set = vec_instr_set_create();
 452
 453    nir_metadata_require(impl, nir_metadata_dominance);
 454
 455    bool progress = vectorize_block(nir_start_block(impl), instr_set);
 456
 457    if (progress)
 458       nir_metadata_preserve(impl, nir_metadata_block_index |
 459                                   nir_metadata_dominance);
 460
 461    vec_instr_set_destroy(instr_set);
 462    return progress;
 463 }
 464
 465 bool
 466 nir_opt_vectorize(nir_shader *shader)
 467 {
 468    bool progress = false;
 469
 470    nir_foreach_function(function, shader) {
 471       if (function->impl)
 472          progress |= nir_opt_vectorize_impl(function->impl);
 473    }
 474
 475    return progress;
 476 }