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 #include "pipe/p_state.h"
  31
  32
  33 #include "ppir.h"
  34
  35 static void *ppir_node_create_ssa(ppir_block *block, ppir_op op, nir_ssa_def *ssa)
  36 {
  37    ppir_node *node = ppir_node_create(block, op, ssa->index, 0);
  38    if (!node)
  39       return NULL;
  40
  41    ppir_dest *dest = ppir_node_get_dest(node);
  42    dest->type = ppir_target_ssa;
  43    dest->ssa.num_components = ssa->num_components;
  44    dest->ssa.live_in = INT_MAX;
  45    dest->ssa.live_out = 0;
  46    dest->write_mask = u_bit_consecutive(0, ssa->num_components);
  47
  48    if (node->type == ppir_node_type_load ||
  49        node->type == ppir_node_type_store)
  50       dest->ssa.is_head = true;
  51
  52    return node;
  53 }
  54
  55 static void *ppir_node_create_reg(ppir_block *block, ppir_op op,
  56                                   nir_reg_dest *reg, unsigned mask)
  57 {
  58    ppir_node *node = ppir_node_create(block, op, reg->reg->index, mask);
  59    if (!node)
  60       return NULL;
  61
  62    ppir_dest *dest = ppir_node_get_dest(node);
  63
  64    list_for_each_entry(ppir_reg, r, &block->comp->reg_list, list) {
  65       if (r->index == reg->reg->index) {
  66          dest->reg = r;
  67          break;
  68       }
  69    }
  70
  71    dest->type = ppir_target_register;
  72    dest->write_mask = mask;
  73
  74    if (node->type == ppir_node_type_load ||
  75        node->type == ppir_node_type_store)
  76       dest->reg->is_head = true;
  77
  78    return node;
  79 }
  80
  81 static void *ppir_node_create_dest(ppir_block *block, ppir_op op,
  82                                    nir_dest *dest, unsigned mask)
  83 {
  84    unsigned index = -1;
  85
  86    if (dest) {
  87       if (dest->is_ssa)
  88          return ppir_node_create_ssa(block, op, &dest->ssa);
  89       else
  90          return ppir_node_create_reg(block, op, &dest->reg, mask);
  91    }
  92
  93    return ppir_node_create(block, op, index, 0);
  94 }
  95
  96 static void ppir_node_add_src(ppir_compiler *comp, ppir_node *node,
  97                               ppir_src *ps, nir_src *ns, unsigned mask)
  98 {
  99    ppir_node *child = NULL;
 100
 101    if (ns->is_ssa) {
 102       child = comp->var_nodes[ns->ssa->index];
 103       ppir_node_add_dep(node, child);
 104    }
 105    else {
 106       nir_register *reg = ns->reg.reg;
 107       while (mask) {
 108          int swizzle = ps->swizzle[u_bit_scan(&mask)];
 109          child = comp->var_nodes[(reg->index << 2) + comp->reg_base + swizzle];
 110          ppir_node_add_dep(node, child);
 111       }
 112    }
 113
 114    ppir_node_target_assign(ps, child);
 115 }
 116
 117 static int nir_to_ppir_opcodes[nir_num_opcodes] = {
 118    /* not supported */
 119    [0 ... nir_last_opcode] = -1,
 120
 121    [nir_op_mov] = ppir_op_mov,
 122    [nir_op_fmul] = ppir_op_mul,
 123    [nir_op_fabs] = ppir_op_abs,
 124    [nir_op_fneg] = ppir_op_neg,
 125    [nir_op_fadd] = ppir_op_add,
 126    [nir_op_fsum3] = ppir_op_sum3,
 127    [nir_op_fsum4] = ppir_op_sum4,
 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    [nir_op_fddx] = ppir_op_ddx,
 153    [nir_op_fddy] = ppir_op_ddy,
 154 };
 155
 156 static ppir_node *ppir_emit_alu(ppir_block *block, nir_instr *ni)
 157 {
 158    nir_alu_instr *instr = nir_instr_as_alu(ni);
 159    int op = nir_to_ppir_opcodes[instr->op];
 160
 161    if (op < 0) {
 162       ppir_error("unsupported nir_op: %s\n", nir_op_infos[instr->op].name);
 163       return NULL;
 164    }
 165
 166    ppir_alu_node *node = ppir_node_create_dest(block, op, &instr->dest.dest,
 167                                                instr->dest.write_mask);
 168    if (!node)
 169       return NULL;
 170
 171    ppir_dest *pd = &node->dest;
 172    nir_alu_dest *nd = &instr->dest;
 173    if (nd->saturate)
 174       pd->modifier = ppir_outmod_clamp_fraction;
 175
 176    unsigned src_mask;
 177    switch (op) {
 178    case ppir_op_sum3:
 179       src_mask = 0b0111;
 180       break;
 181    case ppir_op_sum4:
 182       src_mask = 0b1111;
 183       break;
 184    default:
 185       src_mask = pd->write_mask;
 186       break;
 187    }
 188
 189    unsigned num_child = nir_op_infos[instr->op].num_inputs;
 190    node->num_src = num_child;
 191
 192    for (int i = 0; i < num_child; i++) {
 193       nir_alu_src *ns = instr->src + i;
 194       ppir_src *ps = node->src + i;
 195       memcpy(ps->swizzle, ns->swizzle, sizeof(ps->swizzle));
 196       ppir_node_add_src(block->comp, &node->node, ps, &ns->src, src_mask);
 197
 198       ps->absolute = ns->abs;
 199       ps->negate = ns->negate;
 200    }
 201
 202    return &node->node;
 203 }
 204
 205 static ppir_block *ppir_block_create(ppir_compiler *comp);
 206
 207 static bool ppir_emit_discard_block(ppir_compiler *comp)
 208 {
 209    ppir_block *block = ppir_block_create(comp);
 210    ppir_discard_node *discard;
 211    if (!block)
 212       return false;
 213
 214    comp->discard_block = block;
 215    block->comp  = comp;
 216
 217    discard = ppir_node_create(block, ppir_op_discard, -1, 0);
 218    if (discard)
 219       list_addtail(&discard->node.list, &block->node_list);
 220    else
 221       return false;
 222
 223    return true;
 224 }
 225
 226 static ppir_node *ppir_emit_discard_if(ppir_block *block, nir_instr *ni)
 227 {
 228    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 229    ppir_node *node;
 230    ppir_compiler *comp = block->comp;
 231    ppir_branch_node *branch;
 232
 233    if (!comp->discard_block && !ppir_emit_discard_block(comp))
 234       return NULL;
 235
 236    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 237    if (!node)
 238       return NULL;
 239    branch = ppir_node_to_branch(node);
 240
 241    /* second src and condition will be updated during lowering */
 242    ppir_node_add_src(block->comp, node, &branch->src[0],
 243                      &instr->src[0], u_bit_consecutive(0, instr->num_components));
 244    branch->target = comp->discard_block;
 245
 246    return node;
 247 }
 248
 249 static ppir_node *ppir_emit_discard(ppir_block *block, nir_instr *ni)
 250 {
 251    ppir_node *node = ppir_node_create(block, ppir_op_discard, -1, 0);
 252
 253    return node;
 254 }
 255
 256 static ppir_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 257 {
 258    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 259    unsigned mask = 0;
 260    ppir_load_node *lnode;
 261    ppir_store_node *snode;
 262
 263    switch (instr->intrinsic) {
 264    case nir_intrinsic_load_input:
 265       if (!instr->dest.is_ssa)
 266          mask = u_bit_consecutive(0, instr->num_components);
 267
 268       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 269       if (!lnode)
 270          return NULL;
 271
 272       lnode->num_components = instr->num_components;
 273       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 274       return &lnode->node;
 275
 276    case nir_intrinsic_load_frag_coord:
 277    case nir_intrinsic_load_point_coord:
 278    case nir_intrinsic_load_front_face:
 279       if (!instr->dest.is_ssa)
 280          mask = u_bit_consecutive(0, instr->num_components);
 281
 282       ppir_op op;
 283       switch (instr->intrinsic) {
 284       case nir_intrinsic_load_frag_coord:
 285          op = ppir_op_load_fragcoord;
 286          break;
 287       case nir_intrinsic_load_point_coord:
 288          op = ppir_op_load_pointcoord;
 289          break;
 290       case nir_intrinsic_load_front_face:
 291          op = ppir_op_load_frontface;
 292          break;
 293       default:
 294          assert(0);
 295          break;
 296       }
 297
 298       lnode = ppir_node_create_dest(block, op, &instr->dest, mask);
 299       if (!lnode)
 300          return NULL;
 301
 302       lnode->num_components = instr->num_components;
 303       return &lnode->node;
 304
 305    case nir_intrinsic_load_uniform:
 306       if (!instr->dest.is_ssa)
 307          mask = u_bit_consecutive(0, instr->num_components);
 308
 309       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 310       if (!lnode)
 311          return NULL;
 312
 313       lnode->num_components = instr->num_components;
 314       lnode->index = nir_intrinsic_base(instr);
 315       lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 316
 317       return &lnode->node;
 318
 319    case nir_intrinsic_store_output:
 320       snode = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 321       if (!snode)
 322          return NULL;
 323
 324       snode->index = nir_intrinsic_base(instr);
 325
 326       for (int i = 0; i < instr->num_components; i++)
 327          snode->src.swizzle[i] = i;
 328
 329       ppir_node_add_src(block->comp, &snode->node, &snode->src, instr->src,
 330                         u_bit_consecutive(0, instr->num_components));
 331
 332       return &snode->node;
 333
 334    case nir_intrinsic_discard:
 335       return ppir_emit_discard(block, ni);
 336
 337    case nir_intrinsic_discard_if:
 338       return ppir_emit_discard_if(block, ni);
 339
 340    default:
 341       ppir_error("unsupported nir_intrinsic_instr %s\n",
 342                  nir_intrinsic_infos[instr->intrinsic].name);
 343       return NULL;
 344    }
 345 }
 346
 347 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 348 {
 349    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 350    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 351    if (!node)
 352       return NULL;
 353
 354    assert(instr->def.bit_size == 32);
 355
 356    for (int i = 0; i < instr->def.num_components; i++)
 357       node->constant.value[i].i = instr->value[i].i32;
 358    node->constant.num = instr->def.num_components;
 359
 360    return &node->node;
 361 }
 362
 363 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 364 {
 365    ppir_error("nir_ssa_undef_instr not support\n");
 366    return NULL;
 367 }
 368
 369 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 370 {
 371    nir_tex_instr *instr = nir_instr_as_tex(ni);
 372    ppir_load_texture_node *node;
 373
 374    if (instr->op != nir_texop_tex) {
 375       ppir_error("unsupported texop %d\n", instr->op);
 376       return NULL;
 377    }
 378
 379    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, 0);
 380    if (!node)
 381       return NULL;
 382
 383    node->sampler = instr->texture_index;
 384
 385    switch (instr->sampler_dim) {
 386    case GLSL_SAMPLER_DIM_2D:
 387    case GLSL_SAMPLER_DIM_RECT:
 388    case GLSL_SAMPLER_DIM_EXTERNAL:
 389       break;
 390    default:
 391       ppir_error("unsupported sampler dim: %d\n", instr->sampler_dim);
 392       return NULL;
 393    }
 394
 395    node->sampler_dim = instr->sampler_dim;
 396
 397    for (int i = 0; i < instr->coord_components; i++)
 398          node->src_coords.swizzle[i] = i;
 399
 400    for (int i = 0; i < instr->num_srcs; i++) {
 401       switch (instr->src[i].src_type) {
 402       case nir_tex_src_coord:
 403          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 404                            u_bit_consecutive(0, instr->coord_components));
 405          break;
 406       default:
 407          ppir_error("unsupported texture source type\n");
 408          assert(0);
 409          return NULL;
 410       }
 411    }
 412
 413    return &node->node;
 414 }
 415
 416 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 417 {
 418    ppir_error("nir_jump_instr not support\n");
 419    return NULL;
 420 }
 421
 422 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 423    [nir_instr_type_alu]        = ppir_emit_alu,
 424    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 425    [nir_instr_type_load_const] = ppir_emit_load_const,
 426    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 427    [nir_instr_type_tex]        = ppir_emit_tex,
 428    [nir_instr_type_jump]       = ppir_emit_jump,
 429 };
 430
 431 static ppir_block *ppir_block_create(ppir_compiler *comp)
 432 {
 433    ppir_block *block = rzalloc(comp, ppir_block);
 434    if (!block)
 435       return NULL;
 436
 437    list_inithead(&block->node_list);
 438    list_inithead(&block->instr_list);
 439
 440    return block;
 441 }
 442
 443 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 444 {
 445    ppir_block *block = ppir_block_create(comp);
 446    if (!block)
 447       return false;
 448
 449    list_addtail(&block->list, &comp->block_list);
 450    block->comp = comp;
 451
 452    nir_foreach_instr(instr, nblock) {
 453       assert(instr->type < nir_instr_type_phi);
 454       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 455       if (!node)
 456          return false;
 457
 458       list_addtail(&node->list, &block->node_list);
 459    }
 460
 461    return true;
 462 }
 463
 464 static bool ppir_emit_if(ppir_compiler *comp, nir_if *nif)
 465 {
 466    ppir_error("if nir_cf_node not support\n");
 467    return false;
 468 }
 469
 470 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 471 {
 472    ppir_error("loop nir_cf_node not support\n");
 473    return false;
 474 }
 475
 476 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 477 {
 478    ppir_error("function nir_cf_node not support\n");
 479    return false;
 480 }
 481
 482 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 483 {
 484    foreach_list_typed(nir_cf_node, node, node, list) {
 485       bool ret;
 486
 487       switch (node->type) {
 488       case nir_cf_node_block:
 489          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 490          break;
 491       case nir_cf_node_if:
 492          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 493          break;
 494       case nir_cf_node_loop:
 495          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 496          break;
 497       case nir_cf_node_function:
 498          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 499          break;
 500       default:
 501          ppir_error("unknown NIR node type %d\n", node->type);
 502          return false;
 503       }
 504
 505       if (!ret)
 506          return false;
 507    }
 508
 509    return true;
 510 }
 511
 512 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 513 {
 514    ppir_compiler *comp = rzalloc_size(
 515       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 516    if (!comp)
 517       return NULL;
 518
 519    list_inithead(&comp->block_list);
 520    list_inithead(&comp->reg_list);
 521
 522    comp->var_nodes = (ppir_node **)(comp + 1);
 523    comp->reg_base = num_ssa;
 524    comp->prog = prog;
 525    return comp;
 526 }
 527
 528 static void ppir_add_ordering_deps(ppir_compiler *comp)
 529 {
 530    /* Some intrinsics do not have explicit dependencies and thus depend
 531     * on instructions order. Consider discard_if and store_ouput as
 532     * example. If we don't add fake dependency of discard_if to store_output
 533     * scheduler may put store_output first and since store_output terminates
 534     * shader on Utgard PP, rest of it will never be executed.
 535     * Add fake dependencies for discard/branch/store to preserve
 536     * instruction order.
 537     *
 538     * TODO: scheduler should schedule discard_if as early as possible otherwise
 539     * we may end up with suboptimal code for cases like this:
 540     *
 541     * s3 = s1 < s2
 542     * discard_if s3
 543     * s4 = s1 + s2
 544     * store s4
 545     *
 546     * In this case store depends on discard_if and s4, but since dependencies can
 547     * be scheduled in any order it can result in code like this:
 548     *
 549     * instr1: s3 = s1 < s3
 550     * instr2: s4 = s1 + s2
 551     * instr3: discard_if s3
 552     * instr4: store s4
 553     */
 554    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 555       ppir_node *prev_node = NULL;
 556       list_for_each_entry(ppir_node, node, &block->node_list, list) {
 557          if (node->type == ppir_node_type_discard ||
 558              node->type == ppir_node_type_store ||
 559              node->type == ppir_node_type_branch) {
 560             if (prev_node)
 561                ppir_node_add_dep(node, prev_node);
 562             prev_node = node;
 563          }
 564       }
 565    }
 566 }
 567
 568 static void ppir_print_shader_db(struct nir_shader *nir, ppir_compiler *comp,
 569                                  struct pipe_debug_callback *debug)
 570 {
 571    const struct shader_info *info = &nir->info;
 572    char *shaderdb;
 573    int ret = asprintf(&shaderdb,
 574                       "%s shader: %d inst, %d loops, %d:%d spills:fills\n",
 575                       gl_shader_stage_name(info->stage),
 576                       comp->cur_instr_index,
 577                       comp->num_loops,
 578                       comp->num_spills,
 579                       comp->num_fills);
 580    assert(ret >= 0);
 581
 582    if (lima_debug & LIMA_DEBUG_SHADERDB)
 583       fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
 584
 585    pipe_debug_message(debug, SHADER_INFO, "%s", shaderdb);
 586    free(shaderdb);
 587 }
 588
 589 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 590                       struct ra_regs *ra,
 591                       struct pipe_debug_callback *debug)
 592 {
 593    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 594    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 595    if (!comp)
 596       return false;
 597
 598    comp->ra = ra;
 599
 600    foreach_list_typed(nir_register, reg, node, &func->registers) {
 601       ppir_reg *r = rzalloc(comp, ppir_reg);
 602       if (!r)
 603          return false;
 604
 605       r->index = reg->index;
 606       r->num_components = reg->num_components;
 607       r->live_in = INT_MAX;
 608       r->live_out = 0;
 609       r->is_head = false;
 610       list_addtail(&r->list, &comp->reg_list);
 611    }
 612
 613    if (!ppir_emit_cf_list(comp, &func->body))
 614       goto err_out0;
 615
 616    /* If we have discard block add it to the very end */
 617    if (comp->discard_block)
 618       list_addtail(&comp->discard_block->list, &comp->block_list);
 619
 620    ppir_add_ordering_deps(comp);
 621
 622    ppir_node_print_prog(comp);
 623
 624    if (!ppir_lower_prog(comp))
 625       goto err_out0;
 626
 627    if (!ppir_node_to_instr(comp))
 628       goto err_out0;
 629
 630    if (!ppir_schedule_prog(comp))
 631       goto err_out0;
 632
 633    if (!ppir_regalloc_prog(comp))
 634       goto err_out0;
 635
 636    if (!ppir_codegen_prog(comp))
 637       goto err_out0;
 638
 639    ppir_print_shader_db(nir, comp, debug);
 640
 641    ralloc_free(comp);
 642    return true;
 643
 644 err_out0:
 645    ralloc_free(comp);
 646    return false;
 647 }
 648