src/gallium/drivers/lima/ir/pp/nir.c

   1 /*
   2  * Copyright (c) 2017 Lima Project
   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, sub license,
   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
  12  * next paragraph) shall be included in all copies or substantial portions
  13  * of the 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 NON-INFRINGEMENT. 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
  21  * DEALINGS IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <string.h>
  26
  27 #include "util/ralloc.h"
  28 #include "util/bitscan.h"
  29 #include "compiler/nir/nir.h"
  30
  31 #include "ppir.h"
  32
  33 static void *ppir_node_create_ssa(ppir_block *block, ppir_op op, nir_ssa_def *ssa)
  34 {
  35    ppir_node *node = ppir_node_create(block, op, ssa->index, 0);
  36    if (!node)
  37       return NULL;
  38
  39    ppir_dest *dest = ppir_node_get_dest(node);
  40    dest->type = ppir_target_ssa;
  41    dest->ssa.num_components = ssa->num_components;
  42    dest->ssa.live_in = INT_MAX;
  43    dest->ssa.live_out = 0;
  44    dest->write_mask = u_bit_consecutive(0, ssa->num_components);
  45
  46    if (node->type == ppir_node_type_load ||
  47        node->type == ppir_node_type_store)
  48       dest->ssa.is_head = true;
  49
  50    return node;
  51 }
  52
  53 static void *ppir_node_create_reg(ppir_block *block, ppir_op op,
  54                                   nir_reg_dest *reg, unsigned mask)
  55 {
  56    ppir_node *node = ppir_node_create(block, op, reg->reg->index, mask);
  57    if (!node)
  58       return NULL;
  59
  60    ppir_dest *dest = ppir_node_get_dest(node);
  61
  62    list_for_each_entry(ppir_reg, r, &block->comp->reg_list, list) {
  63       if (r->index == reg->reg->index) {
  64          dest->reg = r;
  65          break;
  66       }
  67    }
  68
  69    dest->type = ppir_target_register;
  70    dest->write_mask = mask;
  71
  72    if (node->type == ppir_node_type_load ||
  73        node->type == ppir_node_type_store)
  74       dest->reg->is_head = true;
  75
  76    return node;
  77 }
  78
  79 static void *ppir_node_create_dest(ppir_block *block, ppir_op op,
  80                                    nir_dest *dest, unsigned mask)
  81 {
  82    unsigned index = -1;
  83
  84    if (dest) {
  85       if (dest->is_ssa)
  86          return ppir_node_create_ssa(block, op, &dest->ssa);
  87       else
  88          return ppir_node_create_reg(block, op, &dest->reg, mask);
  89    }
  90
  91    return ppir_node_create(block, op, index, 0);
  92 }
  93
  94 static void ppir_node_add_src(ppir_compiler *comp, ppir_node *node,
  95                               ppir_src *ps, nir_src *ns, unsigned mask)
  96 {
  97    ppir_node *child = NULL;
  98
  99    if (ns->is_ssa) {
 100       child = comp->var_nodes[ns->ssa->index];
 101       ppir_node_add_dep(node, child);
 102    }
 103    else {
 104       nir_register *reg = ns->reg.reg;
 105       while (mask) {
 106          int swizzle = ps->swizzle[u_bit_scan(&mask)];
 107          child = comp->var_nodes[(reg->index << 2) + comp->reg_base + swizzle];
 108          ppir_node_add_dep(node, child);
 109       }
 110    }
 111
 112    ppir_dest *dest = ppir_node_get_dest(child);
 113    ppir_node_target_assign(ps, dest);
 114 }
 115
 116 static int nir_to_ppir_opcodes[nir_num_opcodes] = {
 117    /* not supported */
 118    [0 ... nir_last_opcode] = -1,
 119
 120    [nir_op_mov] = ppir_op_mov,
 121    [nir_op_fmul] = ppir_op_mul,
 122    [nir_op_fabs] = ppir_op_abs,
 123    [nir_op_fneg] = ppir_op_neg,
 124    [nir_op_fadd] = ppir_op_add,
 125    [nir_op_fdot2] = ppir_op_dot2,
 126    [nir_op_fdot3] = ppir_op_dot3,
 127    [nir_op_fdot4] = ppir_op_dot4,
 128    [nir_op_frsq] = ppir_op_rsqrt,
 129    [nir_op_flog2] = ppir_op_log2,
 130    [nir_op_fexp2] = ppir_op_exp2,
 131    [nir_op_fsqrt] = ppir_op_sqrt,
 132    [nir_op_fsin] = ppir_op_sin,
 133    [nir_op_fcos] = ppir_op_cos,
 134    [nir_op_fmax] = ppir_op_max,
 135    [nir_op_fmin] = ppir_op_min,
 136    [nir_op_frcp] = ppir_op_rcp,
 137    [nir_op_ffloor] = ppir_op_floor,
 138    [nir_op_fceil] = ppir_op_ceil,
 139    [nir_op_ffract] = ppir_op_fract,
 140    [nir_op_sge] = ppir_op_ge,
 141    [nir_op_fge] = ppir_op_ge,
 142    [nir_op_slt] = ppir_op_lt,
 143    [nir_op_flt] = ppir_op_lt,
 144    [nir_op_seq] = ppir_op_eq,
 145    [nir_op_feq] = ppir_op_eq,
 146    [nir_op_sne] = ppir_op_ne,
 147    [nir_op_fne] = ppir_op_ne,
 148    [nir_op_fcsel] = ppir_op_select,
 149    [nir_op_inot] = ppir_op_not,
 150    [nir_op_ftrunc] = ppir_op_trunc,
 151    [nir_op_fsat] = ppir_op_sat,
 152 };
 153
 154 static ppir_node *ppir_emit_alu(ppir_block *block, nir_instr *ni)
 155 {
 156    nir_alu_instr *instr = nir_instr_as_alu(ni);
 157    int op = nir_to_ppir_opcodes[instr->op];
 158
 159    if (op < 0) {
 160       ppir_error("unsupported nir_op: %s\n", nir_op_infos[instr->op].name);
 161       return NULL;
 162    }
 163
 164    ppir_alu_node *node = ppir_node_create_dest(block, op, &instr->dest.dest,
 165                                                instr->dest.write_mask);
 166    if (!node)
 167       return NULL;
 168
 169    ppir_dest *pd = &node->dest;
 170    nir_alu_dest *nd = &instr->dest;
 171    if (nd->saturate)
 172       pd->modifier = ppir_outmod_clamp_fraction;
 173
 174    unsigned src_mask;
 175    switch (op) {
 176    case ppir_op_dot2:
 177       src_mask = 0b0011;
 178       break;
 179    case ppir_op_dot3:
 180       src_mask = 0b0111;
 181       break;
 182    case ppir_op_dot4:
 183       src_mask = 0b1111;
 184       break;
 185    default:
 186       src_mask = pd->write_mask;
 187       break;
 188    }
 189
 190    unsigned num_child = nir_op_infos[instr->op].num_inputs;
 191    node->num_src = num_child;
 192
 193    for (int i = 0; i < num_child; i++) {
 194       nir_alu_src *ns = instr->src + i;
 195       ppir_src *ps = node->src + i;
 196       memcpy(ps->swizzle, ns->swizzle, sizeof(ps->swizzle));
 197       ppir_node_add_src(block->comp, &node->node, ps, &ns->src, src_mask);
 198
 199       ps->absolute = ns->abs;
 200       ps->negate = ns->negate;
 201    }
 202
 203    return &node->node;
 204 }
 205
 206 static ppir_block *ppir_block_create(ppir_compiler *comp);
 207
 208 static bool ppir_emit_discard_block(ppir_compiler *comp)
 209 {
 210    ppir_block *block = ppir_block_create(comp);
 211    ppir_discard_node *discard;
 212    if (!block)
 213       return false;
 214
 215    comp->discard_block = block;
 216    block->comp  = comp;
 217
 218    discard = ppir_node_create(block, ppir_op_discard, -1, 0);
 219    if (discard)
 220       list_addtail(&discard->node.list, &block->node_list);
 221    else
 222       return false;
 223
 224    return true;
 225 }
 226
 227 static ppir_node *ppir_emit_discard_if(ppir_block *block, nir_instr *ni)
 228 {
 229    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 230    ppir_node *node;
 231    ppir_compiler *comp = block->comp;
 232    ppir_branch_node *branch;
 233
 234    if (!comp->discard_block && !ppir_emit_discard_block(comp))
 235       return NULL;
 236
 237    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 238    if (!node)
 239       return NULL;
 240    branch = ppir_node_to_branch(node);
 241
 242    /* second src and condition will be updated during lowering */
 243    ppir_node_add_src(block->comp, node, &branch->src[0],
 244                      &instr->src[0], u_bit_consecutive(0, instr->num_components));
 245    branch->target = comp->discard_block;
 246
 247    return node;
 248 }
 249
 250 static ppir_node *ppir_emit_discard(ppir_block *block, nir_instr *ni)
 251 {
 252    ppir_node *node = ppir_node_create(block, ppir_op_discard, -1, 0);
 253
 254    return node;
 255 }
 256
 257 static ppir_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 258 {
 259    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 260    unsigned mask = 0;
 261    ppir_load_node *lnode;
 262    ppir_store_node *snode;
 263
 264    switch (instr->intrinsic) {
 265    case nir_intrinsic_load_input:
 266       if (!instr->dest.is_ssa)
 267          mask = u_bit_consecutive(0, instr->num_components);
 268
 269       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 270       if (!lnode)
 271          return NULL;
 272
 273       lnode->num_components = instr->num_components;
 274       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 275       return &lnode->node;
 276
 277    case nir_intrinsic_load_frag_coord:
 278       if (!instr->dest.is_ssa)
 279          mask = u_bit_consecutive(0, instr->num_components);
 280
 281       lnode = ppir_node_create_dest(block, ppir_op_load_fragcoord, &instr->dest, mask);
 282       if (!lnode)
 283          return NULL;
 284
 285       lnode->num_components = instr->num_components;
 286       return &lnode->node;
 287
 288    case nir_intrinsic_load_point_coord:
 289       if (!instr->dest.is_ssa)
 290          mask = u_bit_consecutive(0, instr->num_components);
 291
 292       lnode = ppir_node_create_dest(block, ppir_op_load_pointcoord, &instr->dest, mask);
 293       if (!lnode)
 294          return NULL;
 295
 296       lnode->num_components = instr->num_components;
 297       return &lnode->node;
 298
 299    case nir_intrinsic_load_uniform:
 300       if (!instr->dest.is_ssa)
 301          mask = u_bit_consecutive(0, instr->num_components);
 302
 303       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 304       if (!lnode)
 305          return NULL;
 306
 307       lnode->num_components = instr->num_components;
 308       lnode->index = nir_intrinsic_base(instr);
 309       lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 310
 311       return &lnode->node;
 312
 313    case nir_intrinsic_store_output:
 314       snode = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 315       if (!snode)
 316          return NULL;
 317
 318       snode->index = nir_intrinsic_base(instr);
 319
 320       for (int i = 0; i < instr->num_components; i++)
 321          snode->src.swizzle[i] = i;
 322
 323       ppir_node_add_src(block->comp, &snode->node, &snode->src, instr->src,
 324                         u_bit_consecutive(0, instr->num_components));
 325
 326       return &snode->node;
 327
 328    case nir_intrinsic_discard:
 329       return ppir_emit_discard(block, ni);
 330
 331    case nir_intrinsic_discard_if:
 332       return ppir_emit_discard_if(block, ni);
 333
 334    default:
 335       ppir_error("unsupported nir_intrinsic_instr %s\n",
 336                  nir_intrinsic_infos[instr->intrinsic].name);
 337       return NULL;
 338    }
 339 }
 340
 341 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 342 {
 343    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 344    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 345    if (!node)
 346       return NULL;
 347
 348    assert(instr->def.bit_size == 32);
 349
 350    for (int i = 0; i < instr->def.num_components; i++)
 351       node->constant.value[i].i = instr->value[i].i32;
 352    node->constant.num = instr->def.num_components;
 353
 354    return &node->node;
 355 }
 356
 357 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 358 {
 359    ppir_error("nir_ssa_undef_instr not support\n");
 360    return NULL;
 361 }
 362
 363 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 364 {
 365    nir_tex_instr *instr = nir_instr_as_tex(ni);
 366    ppir_load_texture_node *node;
 367
 368    if (instr->op != nir_texop_tex) {
 369       ppir_error("unsupported texop %d\n", instr->op);
 370       return NULL;
 371    }
 372
 373    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, 0);
 374    if (!node)
 375       return NULL;
 376
 377    node->sampler = instr->texture_index;
 378
 379    switch (instr->sampler_dim) {
 380    case GLSL_SAMPLER_DIM_2D:
 381    case GLSL_SAMPLER_DIM_RECT:
 382    case GLSL_SAMPLER_DIM_EXTERNAL:
 383       break;
 384    default:
 385       ppir_debug("unsupported sampler dim: %d\n", instr->sampler_dim);
 386       return NULL;
 387    }
 388
 389    node->sampler_dim = instr->sampler_dim;
 390
 391    for (int i = 0; i < instr->coord_components; i++)
 392          node->src_coords.swizzle[i] = i;
 393
 394    assert(instr->num_srcs == 1);
 395    for (int i = 0; i < instr->num_srcs; i++) {
 396       switch (instr->src[i].src_type) {
 397       case nir_tex_src_coord:
 398          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 399                            u_bit_consecutive(0, instr->coord_components));
 400          break;
 401       default:
 402          ppir_debug("unknown texture source");
 403          return NULL;
 404       }
 405    }
 406
 407    return &node->node;
 408 }
 409
 410 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 411 {
 412    ppir_error("nir_jump_instr not support\n");
 413    return NULL;
 414 }
 415
 416 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 417    [nir_instr_type_alu]        = ppir_emit_alu,
 418    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 419    [nir_instr_type_load_const] = ppir_emit_load_const,
 420    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 421    [nir_instr_type_tex]        = ppir_emit_tex,
 422    [nir_instr_type_jump]       = ppir_emit_jump,
 423 };
 424
 425 static ppir_block *ppir_block_create(ppir_compiler *comp)
 426 {
 427    ppir_block *block = rzalloc(comp, ppir_block);
 428    if (!block)
 429       return NULL;
 430
 431    list_inithead(&block->node_list);
 432    list_inithead(&block->instr_list);
 433
 434    return block;
 435 }
 436
 437 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 438 {
 439    ppir_block *block = ppir_block_create(comp);
 440    if (!block)
 441       return false;
 442
 443    list_addtail(&block->list, &comp->block_list);
 444    block->comp = comp;
 445
 446    nir_foreach_instr(instr, nblock) {
 447       assert(instr->type < nir_instr_type_phi);
 448       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 449       if (!node)
 450          return false;
 451
 452       list_addtail(&node->list, &block->node_list);
 453    }
 454
 455    return true;
 456 }
 457
 458 static bool ppir_emit_if(ppir_compiler *comp, nir_if *nif)
 459 {
 460    ppir_error("if nir_cf_node not support\n");
 461    return false;
 462 }
 463
 464 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 465 {
 466    ppir_error("loop nir_cf_node not support\n");
 467    return false;
 468 }
 469
 470 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 471 {
 472    ppir_error("function nir_cf_node not support\n");
 473    return false;
 474 }
 475
 476 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 477 {
 478    foreach_list_typed(nir_cf_node, node, node, list) {
 479       bool ret;
 480
 481       switch (node->type) {
 482       case nir_cf_node_block:
 483          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 484          break;
 485       case nir_cf_node_if:
 486          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 487          break;
 488       case nir_cf_node_loop:
 489          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 490          break;
 491       case nir_cf_node_function:
 492          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 493          break;
 494       default:
 495          ppir_error("unknown NIR node type %d\n", node->type);
 496          return false;
 497       }
 498
 499       if (!ret)
 500          return false;
 501    }
 502
 503    return true;
 504 }
 505
 506 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 507 {
 508    ppir_compiler *comp = rzalloc_size(
 509       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 510    if (!comp)
 511       return NULL;
 512
 513    list_inithead(&comp->block_list);
 514    list_inithead(&comp->reg_list);
 515
 516    comp->var_nodes = (ppir_node **)(comp + 1);
 517    comp->reg_base = num_ssa;
 518    comp->prog = prog;
 519    return comp;
 520 }
 521
 522 static void ppir_add_ordering_deps(ppir_compiler *comp)
 523 {
 524    /* Some intrinsics do not have explicit dependencies and thus depend
 525     * on instructions order. Consider discard_if and store_ouput as
 526     * example. If we don't add fake dependency of discard_if to store_output
 527     * scheduler may put store_output first and since store_output terminates
 528     * shader on Utgard PP, rest of it will never be executed.
 529     * Add fake dependencies for discard/branch/store to preserve
 530     * instruction order.
 531     *
 532     * TODO: scheduler should schedule discard_if as early as possible otherwise
 533     * we may end up with suboptimal code for cases like this:
 534     *
 535     * s3 = s1 < s2
 536     * discard_if s3
 537     * s4 = s1 + s2
 538     * store s4
 539     *
 540     * In this case store depends on discard_if and s4, but since dependencies can
 541     * be scheduled in any order it can result in code like this:
 542     *
 543     * instr1: s3 = s1 < s3
 544     * instr2: s4 = s1 + s2
 545     * instr3: discard_if s3
 546     * instr4: store s4
 547     */
 548    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 549       ppir_node *prev_node = NULL;
 550       list_for_each_entry(ppir_node, node, &block->node_list, list) {
 551          if (node->type == ppir_node_type_discard ||
 552              node->type == ppir_node_type_store ||
 553              node->type == ppir_node_type_branch) {
 554             if (prev_node)
 555                ppir_node_add_dep(node, prev_node);
 556             prev_node = node;
 557          }
 558       }
 559    }
 560 }
 561
 562 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 563                       struct ra_regs *ra)
 564 {
 565    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 566    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 567    if (!comp)
 568       return false;
 569
 570    comp->ra = ra;
 571
 572    foreach_list_typed(nir_register, reg, node, &func->registers) {
 573       ppir_reg *r = rzalloc(comp, ppir_reg);
 574       if (!r)
 575          return false;
 576
 577       r->index = reg->index;
 578       r->num_components = reg->num_components;
 579       r->live_in = INT_MAX;
 580       r->live_out = 0;
 581       r->is_head = false;
 582       list_addtail(&r->list, &comp->reg_list);
 583    }
 584
 585    if (!ppir_emit_cf_list(comp, &func->body))
 586       goto err_out0;
 587
 588    /* If we have discard block add it to the very end */
 589    if (comp->discard_block)
 590       list_addtail(&comp->discard_block->list, &comp->block_list);
 591
 592    ppir_add_ordering_deps(comp);
 593
 594    ppir_node_print_prog(comp);
 595
 596    if (!ppir_lower_prog(comp))
 597       goto err_out0;
 598
 599    if (!ppir_node_to_instr(comp))
 600       goto err_out0;
 601
 602    if (!ppir_schedule_prog(comp))
 603       goto err_out0;
 604
 605    if (!ppir_regalloc_prog(comp))
 606       goto err_out0;
 607
 608    if (!ppir_codegen_prog(comp))
 609       goto err_out0;
 610
 611    ralloc_free(comp);
 612    return true;
 613
 614 err_out0:
 615    ralloc_free(comp);
 616    return false;
 617 }
 618