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_dest *dest = ppir_node_get_dest(child);
 115    ppir_node_target_assign(ps, dest);
 116 }
 117
 118 static int nir_to_ppir_opcodes[nir_num_opcodes] = {
 119    /* not supported */
 120    [0 ... nir_last_opcode] = -1,
 121
 122    [nir_op_mov] = ppir_op_mov,
 123    [nir_op_fmul] = ppir_op_mul,
 124    [nir_op_fabs] = ppir_op_abs,
 125    [nir_op_fneg] = ppir_op_neg,
 126    [nir_op_fadd] = ppir_op_add,
 127    [nir_op_fsum3] = ppir_op_sum3,
 128    [nir_op_fsum4] = ppir_op_sum4,
 129    [nir_op_frsq] = ppir_op_rsqrt,
 130    [nir_op_flog2] = ppir_op_log2,
 131    [nir_op_fexp2] = ppir_op_exp2,
 132    [nir_op_fsqrt] = ppir_op_sqrt,
 133    [nir_op_fsin] = ppir_op_sin,
 134    [nir_op_fcos] = ppir_op_cos,
 135    [nir_op_fmax] = ppir_op_max,
 136    [nir_op_fmin] = ppir_op_min,
 137    [nir_op_frcp] = ppir_op_rcp,
 138    [nir_op_ffloor] = ppir_op_floor,
 139    [nir_op_fceil] = ppir_op_ceil,
 140    [nir_op_ffract] = ppir_op_fract,
 141    [nir_op_sge] = ppir_op_ge,
 142    [nir_op_fge] = ppir_op_ge,
 143    [nir_op_slt] = ppir_op_lt,
 144    [nir_op_flt] = ppir_op_lt,
 145    [nir_op_seq] = ppir_op_eq,
 146    [nir_op_feq] = ppir_op_eq,
 147    [nir_op_sne] = ppir_op_ne,
 148    [nir_op_fne] = ppir_op_ne,
 149    [nir_op_fcsel] = ppir_op_select,
 150    [nir_op_inot] = ppir_op_not,
 151    [nir_op_ftrunc] = ppir_op_trunc,
 152    [nir_op_fsat] = ppir_op_sat,
 153 };
 154
 155 static ppir_node *ppir_emit_alu(ppir_block *block, nir_instr *ni)
 156 {
 157    nir_alu_instr *instr = nir_instr_as_alu(ni);
 158    int op = nir_to_ppir_opcodes[instr->op];
 159
 160    if (op < 0) {
 161       ppir_error("unsupported nir_op: %s\n", nir_op_infos[instr->op].name);
 162       return NULL;
 163    }
 164
 165    ppir_alu_node *node = ppir_node_create_dest(block, op, &instr->dest.dest,
 166                                                instr->dest.write_mask);
 167    if (!node)
 168       return NULL;
 169
 170    ppir_dest *pd = &node->dest;
 171    nir_alu_dest *nd = &instr->dest;
 172    if (nd->saturate)
 173       pd->modifier = ppir_outmod_clamp_fraction;
 174
 175    unsigned src_mask;
 176    switch (op) {
 177    case ppir_op_sum3:
 178       src_mask = 0b0111;
 179       break;
 180    case ppir_op_sum4:
 181       src_mask = 0b1111;
 182       break;
 183    default:
 184       src_mask = pd->write_mask;
 185       break;
 186    }
 187
 188    unsigned num_child = nir_op_infos[instr->op].num_inputs;
 189    node->num_src = num_child;
 190
 191    for (int i = 0; i < num_child; i++) {
 192       nir_alu_src *ns = instr->src + i;
 193       ppir_src *ps = node->src + i;
 194       memcpy(ps->swizzle, ns->swizzle, sizeof(ps->swizzle));
 195       ppir_node_add_src(block->comp, &node->node, ps, &ns->src, src_mask);
 196
 197       ps->absolute = ns->abs;
 198       ps->negate = ns->negate;
 199    }
 200
 201    return &node->node;
 202 }
 203
 204 static ppir_block *ppir_block_create(ppir_compiler *comp);
 205
 206 static bool ppir_emit_discard_block(ppir_compiler *comp)
 207 {
 208    ppir_block *block = ppir_block_create(comp);
 209    ppir_discard_node *discard;
 210    if (!block)
 211       return false;
 212
 213    comp->discard_block = block;
 214    block->comp  = comp;
 215
 216    discard = ppir_node_create(block, ppir_op_discard, -1, 0);
 217    if (discard)
 218       list_addtail(&discard->node.list, &block->node_list);
 219    else
 220       return false;
 221
 222    return true;
 223 }
 224
 225 static ppir_node *ppir_emit_discard_if(ppir_block *block, nir_instr *ni)
 226 {
 227    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 228    ppir_node *node;
 229    ppir_compiler *comp = block->comp;
 230    ppir_branch_node *branch;
 231
 232    if (!comp->discard_block && !ppir_emit_discard_block(comp))
 233       return NULL;
 234
 235    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 236    if (!node)
 237       return NULL;
 238    branch = ppir_node_to_branch(node);
 239
 240    /* second src and condition will be updated during lowering */
 241    ppir_node_add_src(block->comp, node, &branch->src[0],
 242                      &instr->src[0], u_bit_consecutive(0, instr->num_components));
 243    branch->target = comp->discard_block;
 244
 245    return node;
 246 }
 247
 248 static ppir_node *ppir_emit_discard(ppir_block *block, nir_instr *ni)
 249 {
 250    ppir_node *node = ppir_node_create(block, ppir_op_discard, -1, 0);
 251
 252    return node;
 253 }
 254
 255 static ppir_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 256 {
 257    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 258    unsigned mask = 0;
 259    ppir_load_node *lnode;
 260    ppir_store_node *snode;
 261
 262    switch (instr->intrinsic) {
 263    case nir_intrinsic_load_input:
 264       if (!instr->dest.is_ssa)
 265          mask = u_bit_consecutive(0, instr->num_components);
 266
 267       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 268       if (!lnode)
 269          return NULL;
 270
 271       lnode->num_components = instr->num_components;
 272       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 273       return &lnode->node;
 274
 275    case nir_intrinsic_load_frag_coord:
 276    case nir_intrinsic_load_point_coord:
 277    case nir_intrinsic_load_front_face:
 278       if (!instr->dest.is_ssa)
 279          mask = u_bit_consecutive(0, instr->num_components);
 280
 281       ppir_op op;
 282       switch (instr->intrinsic) {
 283       case nir_intrinsic_load_frag_coord:
 284          op = ppir_op_load_fragcoord;
 285          break;
 286       case nir_intrinsic_load_point_coord:
 287          op = ppir_op_load_pointcoord;
 288          break;
 289       case nir_intrinsic_load_front_face:
 290          op = ppir_op_load_frontface;
 291          break;
 292       default:
 293          assert(0);
 294          break;
 295       }
 296
 297       lnode = ppir_node_create_dest(block, op, &instr->dest, mask);
 298       if (!lnode)
 299          return NULL;
 300
 301       lnode->num_components = instr->num_components;
 302       return &lnode->node;
 303
 304    case nir_intrinsic_load_uniform:
 305       if (!instr->dest.is_ssa)
 306          mask = u_bit_consecutive(0, instr->num_components);
 307
 308       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 309       if (!lnode)
 310          return NULL;
 311
 312       lnode->num_components = instr->num_components;
 313       lnode->index = nir_intrinsic_base(instr);
 314       lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 315
 316       return &lnode->node;
 317
 318    case nir_intrinsic_store_output:
 319       snode = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 320       if (!snode)
 321          return NULL;
 322
 323       snode->index = nir_intrinsic_base(instr);
 324
 325       for (int i = 0; i < instr->num_components; i++)
 326          snode->src.swizzle[i] = i;
 327
 328       ppir_node_add_src(block->comp, &snode->node, &snode->src, instr->src,
 329                         u_bit_consecutive(0, instr->num_components));
 330
 331       return &snode->node;
 332
 333    case nir_intrinsic_discard:
 334       return ppir_emit_discard(block, ni);
 335
 336    case nir_intrinsic_discard_if:
 337       return ppir_emit_discard_if(block, ni);
 338
 339    default:
 340       ppir_error("unsupported nir_intrinsic_instr %s\n",
 341                  nir_intrinsic_infos[instr->intrinsic].name);
 342       return NULL;
 343    }
 344 }
 345
 346 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 347 {
 348    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 349    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 350    if (!node)
 351       return NULL;
 352
 353    assert(instr->def.bit_size == 32);
 354
 355    for (int i = 0; i < instr->def.num_components; i++)
 356       node->constant.value[i].i = instr->value[i].i32;
 357    node->constant.num = instr->def.num_components;
 358
 359    return &node->node;
 360 }
 361
 362 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 363 {
 364    ppir_error("nir_ssa_undef_instr not support\n");
 365    return NULL;
 366 }
 367
 368 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 369 {
 370    nir_tex_instr *instr = nir_instr_as_tex(ni);
 371    ppir_load_texture_node *node;
 372
 373    if (instr->op != nir_texop_tex) {
 374       ppir_error("unsupported texop %d\n", instr->op);
 375       return NULL;
 376    }
 377
 378    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, 0);
 379    if (!node)
 380       return NULL;
 381
 382    node->sampler = instr->texture_index;
 383
 384    switch (instr->sampler_dim) {
 385    case GLSL_SAMPLER_DIM_2D:
 386    case GLSL_SAMPLER_DIM_RECT:
 387    case GLSL_SAMPLER_DIM_EXTERNAL:
 388       break;
 389    default:
 390       ppir_error("unsupported sampler dim: %d\n", instr->sampler_dim);
 391       return NULL;
 392    }
 393
 394    node->sampler_dim = instr->sampler_dim;
 395
 396    for (int i = 0; i < instr->coord_components; i++)
 397          node->src_coords.swizzle[i] = i;
 398
 399    for (int i = 0; i < instr->num_srcs; i++) {
 400       switch (instr->src[i].src_type) {
 401       case nir_tex_src_coord:
 402          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 403                            u_bit_consecutive(0, instr->coord_components));
 404          break;
 405       default:
 406          ppir_error("unsupported texture source type\n");
 407          assert(0);
 408          return NULL;
 409       }
 410    }
 411
 412    return &node->node;
 413 }
 414
 415 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 416 {
 417    ppir_error("nir_jump_instr not support\n");
 418    return NULL;
 419 }
 420
 421 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 422    [nir_instr_type_alu]        = ppir_emit_alu,
 423    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 424    [nir_instr_type_load_const] = ppir_emit_load_const,
 425    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 426    [nir_instr_type_tex]        = ppir_emit_tex,
 427    [nir_instr_type_jump]       = ppir_emit_jump,
 428 };
 429
 430 static ppir_block *ppir_block_create(ppir_compiler *comp)
 431 {
 432    ppir_block *block = rzalloc(comp, ppir_block);
 433    if (!block)
 434       return NULL;
 435
 436    list_inithead(&block->node_list);
 437    list_inithead(&block->instr_list);
 438
 439    return block;
 440 }
 441
 442 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 443 {
 444    ppir_block *block = ppir_block_create(comp);
 445    if (!block)
 446       return false;
 447
 448    list_addtail(&block->list, &comp->block_list);
 449    block->comp = comp;
 450
 451    nir_foreach_instr(instr, nblock) {
 452       assert(instr->type < nir_instr_type_phi);
 453       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 454       if (!node)
 455          return false;
 456
 457       list_addtail(&node->list, &block->node_list);
 458    }
 459
 460    return true;
 461 }
 462
 463 static bool ppir_emit_if(ppir_compiler *comp, nir_if *nif)
 464 {
 465    ppir_error("if nir_cf_node not support\n");
 466    return false;
 467 }
 468
 469 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 470 {
 471    ppir_error("loop nir_cf_node not support\n");
 472    return false;
 473 }
 474
 475 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 476 {
 477    ppir_error("function nir_cf_node not support\n");
 478    return false;
 479 }
 480
 481 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 482 {
 483    foreach_list_typed(nir_cf_node, node, node, list) {
 484       bool ret;
 485
 486       switch (node->type) {
 487       case nir_cf_node_block:
 488          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 489          break;
 490       case nir_cf_node_if:
 491          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 492          break;
 493       case nir_cf_node_loop:
 494          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 495          break;
 496       case nir_cf_node_function:
 497          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 498          break;
 499       default:
 500          ppir_error("unknown NIR node type %d\n", node->type);
 501          return false;
 502       }
 503
 504       if (!ret)
 505          return false;
 506    }
 507
 508    return true;
 509 }
 510
 511 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 512 {
 513    ppir_compiler *comp = rzalloc_size(
 514       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 515    if (!comp)
 516       return NULL;
 517
 518    list_inithead(&comp->block_list);
 519    list_inithead(&comp->reg_list);
 520
 521    comp->var_nodes = (ppir_node **)(comp + 1);
 522    comp->reg_base = num_ssa;
 523    comp->prog = prog;
 524    return comp;
 525 }
 526
 527 static void ppir_add_ordering_deps(ppir_compiler *comp)
 528 {
 529    /* Some intrinsics do not have explicit dependencies and thus depend
 530     * on instructions order. Consider discard_if and store_ouput as
 531     * example. If we don't add fake dependency of discard_if to store_output
 532     * scheduler may put store_output first and since store_output terminates
 533     * shader on Utgard PP, rest of it will never be executed.
 534     * Add fake dependencies for discard/branch/store to preserve
 535     * instruction order.
 536     *
 537     * TODO: scheduler should schedule discard_if as early as possible otherwise
 538     * we may end up with suboptimal code for cases like this:
 539     *
 540     * s3 = s1 < s2
 541     * discard_if s3
 542     * s4 = s1 + s2
 543     * store s4
 544     *
 545     * In this case store depends on discard_if and s4, but since dependencies can
 546     * be scheduled in any order it can result in code like this:
 547     *
 548     * instr1: s3 = s1 < s3
 549     * instr2: s4 = s1 + s2
 550     * instr3: discard_if s3
 551     * instr4: store s4
 552     */
 553    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 554       ppir_node *prev_node = NULL;
 555       list_for_each_entry(ppir_node, node, &block->node_list, list) {
 556          if (node->type == ppir_node_type_discard ||
 557              node->type == ppir_node_type_store ||
 558              node->type == ppir_node_type_branch) {
 559             if (prev_node)
 560                ppir_node_add_dep(node, prev_node);
 561             prev_node = node;
 562          }
 563       }
 564    }
 565 }
 566
 567 static void ppir_print_shader_db(struct nir_shader *nir, ppir_compiler *comp,
 568                                  const struct pipe_debug_callback *debug)
 569 {
 570    const struct shader_info *info = &nir->info;
 571    char *shaderdb;
 572    int ret = asprintf(&shaderdb,
 573                       "%s shader: %d inst, %d loops, %d:%d spills:fills\n",
 574                       gl_shader_stage_name(info->stage),
 575                       comp->cur_instr_index,
 576                       comp->num_loops,
 577                       comp->num_spills,
 578                       comp->num_fills);
 579    assert(ret >= 0);
 580
 581    if (lima_debug & LIMA_DEBUG_SHADERDB)
 582       fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
 583
 584    pipe_debug_message(debug, SHADER_INFO, "%s", shaderdb);
 585    free(shaderdb);
 586 }
 587
 588 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 589                       struct ra_regs *ra,
 590                       const struct pipe_debug_callback *debug)
 591 {
 592    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 593    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 594    if (!comp)
 595       return false;
 596
 597    comp->ra = ra;
 598
 599    foreach_list_typed(nir_register, reg, node, &func->registers) {
 600       ppir_reg *r = rzalloc(comp, ppir_reg);
 601       if (!r)
 602          return false;
 603
 604       r->index = reg->index;
 605       r->num_components = reg->num_components;
 606       r->live_in = INT_MAX;
 607       r->live_out = 0;
 608       r->is_head = false;
 609       list_addtail(&r->list, &comp->reg_list);
 610    }
 611
 612    if (!ppir_emit_cf_list(comp, &func->body))
 613       goto err_out0;
 614
 615    /* If we have discard block add it to the very end */
 616    if (comp->discard_block)
 617       list_addtail(&comp->discard_block->list, &comp->block_list);
 618
 619    ppir_add_ordering_deps(comp);
 620
 621    ppir_node_print_prog(comp);
 622
 623    if (!ppir_lower_prog(comp))
 624       goto err_out0;
 625
 626    if (!ppir_node_to_instr(comp))
 627       goto err_out0;
 628
 629    if (!ppir_schedule_prog(comp))
 630       goto err_out0;
 631
 632    if (!ppir_regalloc_prog(comp))
 633       goto err_out0;
 634
 635    if (!ppir_codegen_prog(comp))
 636       goto err_out0;
 637
 638    ppir_print_shader_db(nir, comp, debug);
 639
 640    ralloc_free(comp);
 641    return true;
 642
 643 err_out0:
 644    ralloc_free(comp);
 645    return false;
 646 }
 647