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    [nir_op_fddx] = ppir_op_ddx,
 154    [nir_op_fddy] = ppir_op_ddy,
 155 };
 156
 157 static ppir_node *ppir_emit_alu(ppir_block *block, nir_instr *ni)
 158 {
 159    nir_alu_instr *instr = nir_instr_as_alu(ni);
 160    int op = nir_to_ppir_opcodes[instr->op];
 161
 162    if (op < 0) {
 163       ppir_error("unsupported nir_op: %s\n", nir_op_infos[instr->op].name);
 164       return NULL;
 165    }
 166
 167    ppir_alu_node *node = ppir_node_create_dest(block, op, &instr->dest.dest,
 168                                                instr->dest.write_mask);
 169    if (!node)
 170       return NULL;
 171
 172    ppir_dest *pd = &node->dest;
 173    nir_alu_dest *nd = &instr->dest;
 174    if (nd->saturate)
 175       pd->modifier = ppir_outmod_clamp_fraction;
 176
 177    unsigned src_mask;
 178    switch (op) {
 179    case ppir_op_sum3:
 180       src_mask = 0b0111;
 181       break;
 182    case ppir_op_sum4:
 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    case nir_intrinsic_load_point_coord:
 279    case nir_intrinsic_load_front_face:
 280       if (!instr->dest.is_ssa)
 281          mask = u_bit_consecutive(0, instr->num_components);
 282
 283       ppir_op op;
 284       switch (instr->intrinsic) {
 285       case nir_intrinsic_load_frag_coord:
 286          op = ppir_op_load_fragcoord;
 287          break;
 288       case nir_intrinsic_load_point_coord:
 289          op = ppir_op_load_pointcoord;
 290          break;
 291       case nir_intrinsic_load_front_face:
 292          op = ppir_op_load_frontface;
 293          break;
 294       default:
 295          assert(0);
 296          break;
 297       }
 298
 299       lnode = ppir_node_create_dest(block, op, &instr->dest, mask);
 300       if (!lnode)
 301          return NULL;
 302
 303       lnode->num_components = instr->num_components;
 304       return &lnode->node;
 305
 306    case nir_intrinsic_load_uniform:
 307       if (!instr->dest.is_ssa)
 308          mask = u_bit_consecutive(0, instr->num_components);
 309
 310       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 311       if (!lnode)
 312          return NULL;
 313
 314       lnode->num_components = instr->num_components;
 315       lnode->index = nir_intrinsic_base(instr);
 316       lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 317
 318       return &lnode->node;
 319
 320    case nir_intrinsic_store_output:
 321       snode = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 322       if (!snode)
 323          return NULL;
 324
 325       snode->index = nir_intrinsic_base(instr);
 326
 327       for (int i = 0; i < instr->num_components; i++)
 328          snode->src.swizzle[i] = i;
 329
 330       ppir_node_add_src(block->comp, &snode->node, &snode->src, instr->src,
 331                         u_bit_consecutive(0, instr->num_components));
 332
 333       return &snode->node;
 334
 335    case nir_intrinsic_discard:
 336       return ppir_emit_discard(block, ni);
 337
 338    case nir_intrinsic_discard_if:
 339       return ppir_emit_discard_if(block, ni);
 340
 341    default:
 342       ppir_error("unsupported nir_intrinsic_instr %s\n",
 343                  nir_intrinsic_infos[instr->intrinsic].name);
 344       return NULL;
 345    }
 346 }
 347
 348 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 349 {
 350    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 351    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 352    if (!node)
 353       return NULL;
 354
 355    assert(instr->def.bit_size == 32);
 356
 357    for (int i = 0; i < instr->def.num_components; i++)
 358       node->constant.value[i].i = instr->value[i].i32;
 359    node->constant.num = instr->def.num_components;
 360
 361    return &node->node;
 362 }
 363
 364 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 365 {
 366    ppir_error("nir_ssa_undef_instr not support\n");
 367    return NULL;
 368 }
 369
 370 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 371 {
 372    nir_tex_instr *instr = nir_instr_as_tex(ni);
 373    ppir_load_texture_node *node;
 374
 375    if (instr->op != nir_texop_tex) {
 376       ppir_error("unsupported texop %d\n", instr->op);
 377       return NULL;
 378    }
 379
 380    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, 0);
 381    if (!node)
 382       return NULL;
 383
 384    node->sampler = instr->texture_index;
 385
 386    switch (instr->sampler_dim) {
 387    case GLSL_SAMPLER_DIM_2D:
 388    case GLSL_SAMPLER_DIM_RECT:
 389    case GLSL_SAMPLER_DIM_EXTERNAL:
 390       break;
 391    default:
 392       ppir_error("unsupported sampler dim: %d\n", instr->sampler_dim);
 393       return NULL;
 394    }
 395
 396    node->sampler_dim = instr->sampler_dim;
 397
 398    for (int i = 0; i < instr->coord_components; i++)
 399          node->src_coords.swizzle[i] = i;
 400
 401    for (int i = 0; i < instr->num_srcs; i++) {
 402       switch (instr->src[i].src_type) {
 403       case nir_tex_src_coord:
 404          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 405                            u_bit_consecutive(0, instr->coord_components));
 406          break;
 407       default:
 408          ppir_error("unsupported texture source type\n");
 409          assert(0);
 410          return NULL;
 411       }
 412    }
 413
 414    return &node->node;
 415 }
 416
 417 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 418 {
 419    ppir_error("nir_jump_instr not support\n");
 420    return NULL;
 421 }
 422
 423 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 424    [nir_instr_type_alu]        = ppir_emit_alu,
 425    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 426    [nir_instr_type_load_const] = ppir_emit_load_const,
 427    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 428    [nir_instr_type_tex]        = ppir_emit_tex,
 429    [nir_instr_type_jump]       = ppir_emit_jump,
 430 };
 431
 432 static ppir_block *ppir_block_create(ppir_compiler *comp)
 433 {
 434    ppir_block *block = rzalloc(comp, ppir_block);
 435    if (!block)
 436       return NULL;
 437
 438    list_inithead(&block->node_list);
 439    list_inithead(&block->instr_list);
 440
 441    return block;
 442 }
 443
 444 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 445 {
 446    ppir_block *block = ppir_block_create(comp);
 447    if (!block)
 448       return false;
 449
 450    list_addtail(&block->list, &comp->block_list);
 451    block->comp = comp;
 452
 453    nir_foreach_instr(instr, nblock) {
 454       assert(instr->type < nir_instr_type_phi);
 455       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 456       if (!node)
 457          return false;
 458
 459       list_addtail(&node->list, &block->node_list);
 460    }
 461
 462    return true;
 463 }
 464
 465 static bool ppir_emit_if(ppir_compiler *comp, nir_if *nif)
 466 {
 467    ppir_error("if nir_cf_node not support\n");
 468    return false;
 469 }
 470
 471 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 472 {
 473    ppir_error("loop nir_cf_node not support\n");
 474    return false;
 475 }
 476
 477 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 478 {
 479    ppir_error("function nir_cf_node not support\n");
 480    return false;
 481 }
 482
 483 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 484 {
 485    foreach_list_typed(nir_cf_node, node, node, list) {
 486       bool ret;
 487
 488       switch (node->type) {
 489       case nir_cf_node_block:
 490          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 491          break;
 492       case nir_cf_node_if:
 493          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 494          break;
 495       case nir_cf_node_loop:
 496          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 497          break;
 498       case nir_cf_node_function:
 499          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 500          break;
 501       default:
 502          ppir_error("unknown NIR node type %d\n", node->type);
 503          return false;
 504       }
 505
 506       if (!ret)
 507          return false;
 508    }
 509
 510    return true;
 511 }
 512
 513 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 514 {
 515    ppir_compiler *comp = rzalloc_size(
 516       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 517    if (!comp)
 518       return NULL;
 519
 520    list_inithead(&comp->block_list);
 521    list_inithead(&comp->reg_list);
 522
 523    comp->var_nodes = (ppir_node **)(comp + 1);
 524    comp->reg_base = num_ssa;
 525    comp->prog = prog;
 526    return comp;
 527 }
 528
 529 static void ppir_add_ordering_deps(ppir_compiler *comp)
 530 {
 531    /* Some intrinsics do not have explicit dependencies and thus depend
 532     * on instructions order. Consider discard_if and store_ouput as
 533     * example. If we don't add fake dependency of discard_if to store_output
 534     * scheduler may put store_output first and since store_output terminates
 535     * shader on Utgard PP, rest of it will never be executed.
 536     * Add fake dependencies for discard/branch/store to preserve
 537     * instruction order.
 538     *
 539     * TODO: scheduler should schedule discard_if as early as possible otherwise
 540     * we may end up with suboptimal code for cases like this:
 541     *
 542     * s3 = s1 < s2
 543     * discard_if s3
 544     * s4 = s1 + s2
 545     * store s4
 546     *
 547     * In this case store depends on discard_if and s4, but since dependencies can
 548     * be scheduled in any order it can result in code like this:
 549     *
 550     * instr1: s3 = s1 < s3
 551     * instr2: s4 = s1 + s2
 552     * instr3: discard_if s3
 553     * instr4: store s4
 554     */
 555    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 556       ppir_node *prev_node = NULL;
 557       list_for_each_entry(ppir_node, node, &block->node_list, list) {
 558          if (node->type == ppir_node_type_discard ||
 559              node->type == ppir_node_type_store ||
 560              node->type == ppir_node_type_branch) {
 561             if (prev_node)
 562                ppir_node_add_dep(node, prev_node);
 563             prev_node = node;
 564          }
 565       }
 566    }
 567 }
 568
 569 static void ppir_print_shader_db(struct nir_shader *nir, ppir_compiler *comp,
 570                                  struct pipe_debug_callback *debug)
 571 {
 572    const struct shader_info *info = &nir->info;
 573    char *shaderdb;
 574    int ret = asprintf(&shaderdb,
 575                       "%s shader: %d inst, %d loops, %d:%d spills:fills\n",
 576                       gl_shader_stage_name(info->stage),
 577                       comp->cur_instr_index,
 578                       comp->num_loops,
 579                       comp->num_spills,
 580                       comp->num_fills);
 581    assert(ret >= 0);
 582
 583    if (lima_debug & LIMA_DEBUG_SHADERDB)
 584       fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
 585
 586    pipe_debug_message(debug, SHADER_INFO, "%s", shaderdb);
 587    free(shaderdb);
 588 }
 589
 590 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 591                       struct ra_regs *ra,
 592                       struct pipe_debug_callback *debug)
 593 {
 594    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 595    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 596    if (!comp)
 597       return false;
 598
 599    comp->ra = ra;
 600
 601    foreach_list_typed(nir_register, reg, node, &func->registers) {
 602       ppir_reg *r = rzalloc(comp, ppir_reg);
 603       if (!r)
 604          return false;
 605
 606       r->index = reg->index;
 607       r->num_components = reg->num_components;
 608       r->live_in = INT_MAX;
 609       r->live_out = 0;
 610       r->is_head = false;
 611       list_addtail(&r->list, &comp->reg_list);
 612    }
 613
 614    if (!ppir_emit_cf_list(comp, &func->body))
 615       goto err_out0;
 616
 617    /* If we have discard block add it to the very end */
 618    if (comp->discard_block)
 619       list_addtail(&comp->discard_block->list, &comp->block_list);
 620
 621    ppir_add_ordering_deps(comp);
 622
 623    ppir_node_print_prog(comp);
 624
 625    if (!ppir_lower_prog(comp))
 626       goto err_out0;
 627
 628    if (!ppir_node_to_instr(comp))
 629       goto err_out0;
 630
 631    if (!ppir_schedule_prog(comp))
 632       goto err_out0;
 633
 634    if (!ppir_regalloc_prog(comp))
 635       goto err_out0;
 636
 637    if (!ppir_codegen_prog(comp))
 638       goto err_out0;
 639
 640    ppir_print_shader_db(nir, comp, debug);
 641
 642    ralloc_free(comp);
 643    return true;
 644
 645 err_out0:
 646    ralloc_free(comp);
 647    return false;
 648 }
 649