src/compiler/nir/nir_lower_bool_to_bitsize.c

   1 /*
   2  * Copyright © 2018 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
  24 #include "nir.h"
  25 #include "nir_builder.h"
  26
  27 static bool
  28 assert_ssa_def_is_not_1bit(nir_ssa_def *def, UNUSED void *unused)
  29 {
  30    assert(def->bit_size > 1);
  31    return true;
  32 }
  33
  34 static bool
  35 rewrite_1bit_ssa_def_to_32bit(nir_ssa_def *def, void *_progress)
  36 {
  37    bool *progress = _progress;
  38    if (def->bit_size == 1) {
  39       def->bit_size = 32;
  40       *progress = true;
  41    }
  42    return true;
  43 }
  44
  45 static uint32_t
  46 get_bool_convert_opcode(uint32_t dst_bit_size)
  47 {
  48    switch (dst_bit_size) {
  49    case 32: return nir_op_i2i32;
  50    case 16: return nir_op_i2i16;
  51    case 8:  return nir_op_i2i8;
  52    default:
  53       unreachable("invalid boolean bit-size");
  54    }
  55 }
  56
  57 static void
  58 make_sources_canonical(nir_builder *b, nir_alu_instr *alu, uint32_t start_idx)
  59 {
  60    /* TODO: for now we take the bit-size of the first source as the canonical
  61     * form but we could try to be smarter.
  62     */
  63    const nir_op_info *op_info = &nir_op_infos[alu->op];
  64    uint32_t bit_size = nir_src_bit_size(alu->src[start_idx].src);
  65    for (uint32_t i = start_idx + 1; i < op_info->num_inputs; i++) {
  66       if (nir_src_bit_size(alu->src[i].src) != bit_size) {
  67          b->cursor = nir_before_instr(&alu->instr);
  68          nir_op convert_op = get_bool_convert_opcode(bit_size);
  69          nir_ssa_def *new_src =
  70             nir_build_alu(b, convert_op, alu->src[i].src.ssa, NULL, NULL, NULL);
  71          /* Retain the write mask and swizzle of the original instruction so
  72           * that we don’t unnecessarily create a vectorized instruction.
  73           */
  74          nir_alu_instr *conv_instr =
  75             nir_instr_as_alu(nir_builder_last_instr(b));
  76          conv_instr->dest.write_mask = alu->dest.write_mask;
  77          conv_instr->dest.dest.ssa.num_components =
  78             alu->dest.dest.ssa.num_components;
  79          memcpy(conv_instr->src[0].swizzle,
  80                 alu->src[i].swizzle,
  81                 sizeof(conv_instr->src[0].swizzle));
  82          nir_instr_rewrite_src(&alu->instr,
  83                                &alu->src[i].src, nir_src_for_ssa(new_src));
  84          /* The swizzle will have been handled by the conversion instruction
  85           * so we can reset it back to the default
  86           */
  87          for (unsigned j = 0; j < NIR_MAX_VEC_COMPONENTS; j++)
  88             alu->src[i].swizzle[j] = j;
  89       }
  90    }
  91 }
  92
  93 static bool
  94 lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
  95 {
  96    const nir_op_info *op_info = &nir_op_infos[alu->op];
  97
  98    /* For operations that can take multiple boolean sources we need to ensure
  99     * that all booleans have the same bit-size
 100     */
 101    switch (alu->op) {
 102    case nir_op_mov:
 103    case nir_op_vec2:
 104    case nir_op_vec3:
 105    case nir_op_vec4:
 106    case nir_op_vec8:
 107    case nir_op_vec16:
 108    case nir_op_inot:
 109    case nir_op_iand:
 110    case nir_op_ior:
 111    case nir_op_ixor:
 112       if (nir_dest_bit_size(alu->dest.dest) > 1)
 113          return false; /* Not a boolean instruction */
 114       /* Fallthrough */
 115
 116    case nir_op_ball_fequal2:
 117    case nir_op_ball_fequal3:
 118    case nir_op_ball_fequal4:
 119    case nir_op_bany_fnequal2:
 120    case nir_op_bany_fnequal3:
 121    case nir_op_bany_fnequal4:
 122    case nir_op_ball_iequal2:
 123    case nir_op_ball_iequal3:
 124    case nir_op_ball_iequal4:
 125    case nir_op_bany_inequal2:
 126    case nir_op_bany_inequal3:
 127    case nir_op_bany_inequal4:
 128    case nir_op_ieq:
 129    case nir_op_ine:
 130       make_sources_canonical(b, alu, 0);
 131       break;
 132
 133    case nir_op_bcsel:
 134       /* bcsel may be choosing between boolean sources too */
 135       if (nir_dest_bit_size(alu->dest.dest) == 1)
 136          make_sources_canonical(b, alu, 1);
 137       break;
 138
 139    default:
 140       break;
 141    }
 142
 143    /* Now that we have a canonical boolean bit-size, go on and rewrite the
 144     * instruction to match the canonical bit-size.
 145     */
 146    uint32_t bit_size = nir_src_bit_size(alu->src[0].src);
 147    assert(bit_size > 1);
 148
 149    nir_op opcode = alu->op;
 150    switch (opcode) {
 151    case nir_op_mov:
 152    case nir_op_vec2:
 153    case nir_op_vec3:
 154    case nir_op_vec4:
 155    case nir_op_vec8:
 156    case nir_op_vec16:
 157    case nir_op_inot:
 158    case nir_op_iand:
 159    case nir_op_ior:
 160    case nir_op_ixor:
 161       /* Nothing to do here, we do not specialize these opcodes by bit-size */
 162       break;
 163
 164    case nir_op_f2b1:
 165       opcode = bit_size == 8 ? nir_op_f2b8 :
 166                                bit_size == 16 ? nir_op_f2b16 : nir_op_f2b32;
 167       break;
 168
 169    case nir_op_i2b1:
 170       opcode = bit_size == 8 ? nir_op_i2b8 :
 171                                bit_size == 16 ? nir_op_i2b16 : nir_op_i2b32;
 172       break;
 173
 174    case nir_op_b2b1:
 175       /* Since the canonical bit size is the size of the src, it's a no-op */
 176       opcode = nir_op_mov;
 177       break;
 178
 179    case nir_op_b2b32:
 180       /* For up-converting booleans, sign-extend */
 181       opcode = nir_op_i2i32;
 182       break;
 183
 184    case nir_op_flt:
 185       opcode = bit_size == 8 ? nir_op_flt8 :
 186                                bit_size == 16 ? nir_op_flt16 : nir_op_flt32;
 187       break;
 188
 189    case nir_op_fge:
 190       opcode = bit_size == 8 ? nir_op_fge8 :
 191                                bit_size == 16 ? nir_op_fge16 : nir_op_fge32;
 192       break;
 193
 194    case nir_op_feq:
 195       opcode = bit_size == 8 ? nir_op_feq8 :
 196                                bit_size == 16 ? nir_op_feq16 : nir_op_feq32;
 197       break;
 198
 199    case nir_op_fneu:
 200       opcode = bit_size == 8 ? nir_op_fneu8 :
 201                                bit_size == 16 ? nir_op_fneu16 : nir_op_fneu32;
 202       break;
 203
 204    case nir_op_ilt:
 205       opcode = bit_size == 8 ? nir_op_ilt8 :
 206                                bit_size == 16 ? nir_op_ilt16 : nir_op_ilt32;
 207       break;
 208
 209    case nir_op_ige:
 210       opcode = bit_size == 8 ? nir_op_ige8 :
 211                                bit_size == 16 ? nir_op_ige16 : nir_op_ige32;
 212       break;
 213
 214    case nir_op_ieq:
 215       opcode = bit_size == 8 ? nir_op_ieq8 :
 216                                bit_size == 16 ? nir_op_ieq16 : nir_op_ieq32;
 217       break;
 218
 219    case nir_op_ine:
 220       opcode = bit_size == 8 ? nir_op_ine8 :
 221                                bit_size == 16 ? nir_op_ine16 : nir_op_ine32;
 222       break;
 223
 224    case nir_op_ult:
 225       opcode = bit_size == 8 ? nir_op_ult8 :
 226                                bit_size == 16 ? nir_op_ult16 : nir_op_ult32;
 227       break;
 228
 229    case nir_op_uge:
 230       opcode = bit_size == 8 ? nir_op_uge8 :
 231                                bit_size == 16 ? nir_op_uge16 : nir_op_uge32;
 232       break;
 233
 234    case nir_op_ball_fequal2:
 235       opcode = bit_size == 8 ? nir_op_b8all_fequal2 :
 236                                bit_size == 16 ? nir_op_b16all_fequal2 :
 237                                                 nir_op_b32all_fequal2;
 238       break;
 239
 240    case nir_op_ball_fequal3:
 241       opcode = bit_size == 8 ? nir_op_b8all_fequal3 :
 242                                bit_size == 16 ? nir_op_b16all_fequal3 :
 243                                                 nir_op_b32all_fequal3;
 244       break;
 245
 246    case nir_op_ball_fequal4:
 247       opcode = bit_size == 8 ? nir_op_b8all_fequal4 :
 248                                bit_size == 16 ? nir_op_b16all_fequal4 :
 249                                                 nir_op_b32all_fequal4;
 250       break;
 251
 252    case nir_op_bany_fnequal2:
 253       opcode = bit_size == 8 ? nir_op_b8any_fnequal2 :
 254                                bit_size == 16 ? nir_op_b16any_fnequal2 :
 255                                                 nir_op_b32any_fnequal2;
 256       break;
 257
 258    case nir_op_bany_fnequal3:
 259       opcode = bit_size == 8 ? nir_op_b8any_fnequal3 :
 260                                bit_size == 16 ? nir_op_b16any_fnequal3 :
 261                                                 nir_op_b32any_fnequal3;
 262       break;
 263
 264    case nir_op_bany_fnequal4:
 265       opcode = bit_size == 8 ? nir_op_b8any_fnequal4 :
 266                                bit_size == 16 ? nir_op_b16any_fnequal4 :
 267                                                 nir_op_b32any_fnequal4;
 268       break;
 269
 270    case nir_op_ball_iequal2:
 271       opcode = bit_size == 8 ? nir_op_b8all_iequal2 :
 272                                bit_size == 16 ? nir_op_b16all_iequal2 :
 273                                                 nir_op_b32all_iequal2;
 274       break;
 275
 276    case nir_op_ball_iequal3:
 277       opcode = bit_size == 8 ? nir_op_b8all_iequal3 :
 278                                bit_size == 16 ? nir_op_b16all_iequal3 :
 279                                                 nir_op_b32all_iequal3;
 280       break;
 281
 282    case nir_op_ball_iequal4:
 283       opcode = bit_size == 8 ? nir_op_b8all_iequal4 :
 284                                bit_size == 16 ? nir_op_b16all_iequal4 :
 285                                                 nir_op_b32all_iequal4;
 286       break;
 287
 288    case nir_op_bany_inequal2:
 289       opcode = bit_size == 8 ? nir_op_b8any_inequal2 :
 290                                bit_size == 16 ? nir_op_b16any_inequal2 :
 291                                                 nir_op_b32any_inequal2;
 292       break;
 293
 294    case nir_op_bany_inequal3:
 295       opcode = bit_size == 8 ? nir_op_b8any_inequal3 :
 296                                bit_size == 16 ? nir_op_b16any_inequal3 :
 297                                                 nir_op_b32any_inequal3;
 298       break;
 299
 300    case nir_op_bany_inequal4:
 301       opcode = bit_size == 8 ? nir_op_b8any_inequal4 :
 302                                bit_size == 16 ? nir_op_b16any_inequal4 :
 303                                                 nir_op_b32any_inequal4;
 304       break;
 305
 306    case nir_op_bcsel:
 307       opcode = bit_size == 8 ? nir_op_b8csel :
 308                                bit_size == 16 ? nir_op_b16csel : nir_op_b32csel;
 309
 310       /* The destination of the selection may have a different bit-size from
 311        * the bcsel condition.
 312        */
 313       bit_size = nir_src_bit_size(alu->src[1].src);
 314       break;
 315
 316    default:
 317       assert(alu->dest.dest.ssa.bit_size > 1);
 318       for (unsigned i = 0; i < op_info->num_inputs; i++)
 319          assert(alu->src[i].src.ssa->bit_size > 1);
 320       return false;
 321    }
 322
 323    alu->op = opcode;
 324
 325    if (alu->dest.dest.ssa.bit_size == 1)
 326       alu->dest.dest.ssa.bit_size = bit_size;
 327
 328    return true;
 329 }
 330
 331 static bool
 332 lower_load_const_instr(nir_load_const_instr *load)
 333 {
 334    bool progress = false;
 335
 336    if (load->def.bit_size > 1)
 337       return progress;
 338
 339    /* TODO: It is not clear if there is any case in which we can ever hit
 340     * this path, so for now we just provide a 32-bit default.
 341     *
 342     * TODO2: after some changed on nir_const_value and other on upstream, we
 343     * removed the initialization of a general value like this:
 344     *   nir_const_value value = load->value
 345     *
 346     * to initialize per value component. Need to confirm if that is correct,
 347     * but look at the TOO before.
 348     */
 349    for (unsigned i = 0; i < load->def.num_components; i++) {
 350       load->value[i].u32 = load->value[i].b ? NIR_TRUE : NIR_FALSE;
 351       load->def.bit_size = 32;
 352       progress = true;
 353    }
 354
 355    return progress;
 356 }
 357
 358 static bool
 359 lower_phi_instr(nir_builder *b, nir_phi_instr *phi)
 360 {
 361    if (nir_dest_bit_size(phi->dest) != 1)
 362       return false;
 363
 364    /* Ensure all phi sources have a canonical bit-size. We choose the
 365     * bit-size of the first phi source as the canonical form.
 366     *
 367     * TODO: maybe we can be smarter about how we choose the canonical form.
 368     */
 369    uint32_t dst_bit_size = 0;
 370    nir_foreach_phi_src(phi_src, phi) {
 371       uint32_t src_bit_size = nir_src_bit_size(phi_src->src);
 372       if (dst_bit_size == 0) {
 373          dst_bit_size = src_bit_size;
 374       } else if (src_bit_size != dst_bit_size) {
 375          assert(phi_src->src.is_ssa);
 376          b->cursor = nir_before_src(&phi_src->src, false);
 377          nir_op convert_op = get_bool_convert_opcode(dst_bit_size);
 378          nir_ssa_def *new_src =
 379             nir_build_alu(b, convert_op, phi_src->src.ssa, NULL, NULL, NULL);
 380          nir_instr_rewrite_src(&phi->instr, &phi_src->src,
 381                                nir_src_for_ssa(new_src));
 382       }
 383    }
 384
 385    phi->dest.ssa.bit_size = dst_bit_size;
 386
 387    return true;
 388 }
 389
 390 static bool
 391 nir_lower_bool_to_bitsize_impl(nir_builder *b, nir_function_impl *impl)
 392 {
 393    bool progress = false;
 394
 395    nir_foreach_block(block, impl) {
 396       nir_foreach_instr_safe(instr, block) {
 397          switch (instr->type) {
 398          case nir_instr_type_alu:
 399             progress |= lower_alu_instr(b, nir_instr_as_alu(instr));
 400             break;
 401
 402          case nir_instr_type_load_const:
 403             progress |= lower_load_const_instr(nir_instr_as_load_const(instr));
 404             break;
 405
 406          case nir_instr_type_phi:
 407             progress |= lower_phi_instr(b, nir_instr_as_phi(instr));
 408             break;
 409
 410          case nir_instr_type_ssa_undef:
 411          case nir_instr_type_intrinsic:
 412          case nir_instr_type_tex:
 413             nir_foreach_ssa_def(instr, rewrite_1bit_ssa_def_to_32bit,
 414                                 &progress);
 415             break;
 416
 417          default:
 418             nir_foreach_ssa_def(instr, assert_ssa_def_is_not_1bit, NULL);
 419          }
 420       }
 421    }
 422
 423    if (progress) {
 424       nir_metadata_preserve(impl, nir_metadata_block_index |
 425                                   nir_metadata_dominance);
 426    }
 427
 428    return progress;
 429 }
 430
 431 bool
 432 nir_lower_bool_to_bitsize(nir_shader *shader)
 433 {
 434    nir_builder b;
 435    bool progress = false;
 436
 437    nir_foreach_function(function, shader) {
 438       if (function->impl) {
 439          nir_builder_init(&b, function->impl);
 440          progress = nir_lower_bool_to_bitsize_impl(&b, function->impl) || progress;
 441       }
 442    }
 443
 444    return progress;
 445 }