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_fand] = ppir_op_and,
 141    [nir_op_for] = ppir_op_or,
 142    [nir_op_fxor] = ppir_op_xor,
 143    [nir_op_sge] = ppir_op_ge,
 144    [nir_op_fge] = ppir_op_ge,
 145    [nir_op_slt] = ppir_op_lt,
 146    [nir_op_flt] = ppir_op_lt,
 147    [nir_op_seq] = ppir_op_eq,
 148    [nir_op_feq] = ppir_op_eq,
 149    [nir_op_sne] = ppir_op_ne,
 150    [nir_op_fne] = ppir_op_ne,
 151    [nir_op_fnot] = ppir_op_not,
 152    [nir_op_fcsel] = ppir_op_select,
 153    [nir_op_inot] = ppir_op_not,
 154    [nir_op_ftrunc] = ppir_op_trunc,
 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_dot2:
 180       src_mask = 0b0011;
 181       break;
 182    case ppir_op_dot3:
 183       src_mask = 0b0111;
 184       break;
 185    case ppir_op_dot4:
 186       src_mask = 0b1111;
 187       break;
 188    default:
 189       src_mask = pd->write_mask;
 190       break;
 191    }
 192
 193    unsigned num_child = nir_op_infos[instr->op].num_inputs;
 194    node->num_src = num_child;
 195
 196    for (int i = 0; i < num_child; i++) {
 197       nir_alu_src *ns = instr->src + i;
 198       ppir_src *ps = node->src + i;
 199       memcpy(ps->swizzle, ns->swizzle, sizeof(ps->swizzle));
 200       ppir_node_add_src(block->comp, &node->node, ps, &ns->src, src_mask);
 201
 202       ps->absolute = ns->abs;
 203       ps->negate = ns->negate;
 204    }
 205
 206    return &node->node;
 207 }
 208
 209 static ppir_block *ppir_block_create(ppir_compiler *comp);
 210
 211 static bool ppir_emit_discard_block(ppir_compiler *comp)
 212 {
 213    ppir_block *block = ppir_block_create(comp);
 214    ppir_discard_node *discard;
 215    if (!block)
 216       return false;
 217
 218    comp->discard_block = block;
 219    block->comp  = comp;
 220
 221    discard = ppir_node_create(block, ppir_op_discard, -1, 0);
 222    if (discard)
 223       list_addtail(&discard->node.list, &block->node_list);
 224    else
 225       return false;
 226
 227    return true;
 228 }
 229
 230 static ppir_node *ppir_emit_discard_if(ppir_block *block, nir_instr *ni)
 231 {
 232    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 233    ppir_node *node;
 234    ppir_compiler *comp = block->comp;
 235    ppir_branch_node *branch;
 236
 237    if (!comp->discard_block && !ppir_emit_discard_block(comp))
 238       return NULL;
 239
 240    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 241    if (!node)
 242       return NULL;
 243    branch = ppir_node_to_branch(node);
 244
 245    /* second src and condition will be updated during lowering */
 246    ppir_node_add_src(block->comp, node, &branch->src[0],
 247                      &instr->src[0], u_bit_consecutive(0, instr->num_components));
 248    branch->target = comp->discard_block;
 249
 250    return node;
 251 }
 252
 253 static ppir_node *ppir_emit_discard(ppir_block *block, nir_instr *ni)
 254 {
 255    ppir_node *node = ppir_node_create(block, ppir_op_discard, -1, 0);
 256
 257    return node;
 258 }
 259
 260 static ppir_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 261 {
 262    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 263    unsigned mask = 0;
 264    ppir_load_node *lnode;
 265    ppir_store_node *snode;
 266
 267    switch (instr->intrinsic) {
 268    case nir_intrinsic_load_input:
 269       if (!instr->dest.is_ssa)
 270          mask = u_bit_consecutive(0, instr->num_components);
 271
 272       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 273       if (!lnode)
 274          return NULL;
 275
 276       lnode->num_components = instr->num_components;
 277       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 278       return &lnode->node;
 279
 280    case nir_intrinsic_load_frag_coord:
 281       if (!instr->dest.is_ssa)
 282          mask = u_bit_consecutive(0, instr->num_components);
 283
 284       lnode = ppir_node_create_dest(block, ppir_op_load_fragcoord, &instr->dest, mask);
 285       if (!lnode)
 286          return NULL;
 287
 288       lnode->num_components = instr->num_components;
 289       return &lnode->node;
 290
 291    case nir_intrinsic_load_uniform:
 292       if (!instr->dest.is_ssa)
 293          mask = u_bit_consecutive(0, instr->num_components);
 294
 295       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 296       if (!lnode)
 297          return NULL;
 298
 299       lnode->num_components = instr->num_components;
 300       lnode->index = nir_intrinsic_base(instr);
 301       lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 302
 303       return &lnode->node;
 304
 305    case nir_intrinsic_store_output:
 306       snode = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 307       if (!snode)
 308          return NULL;
 309
 310       snode->index = nir_intrinsic_base(instr);
 311
 312       for (int i = 0; i < instr->num_components; i++)
 313          snode->src.swizzle[i] = i;
 314
 315       ppir_node_add_src(block->comp, &snode->node, &snode->src, instr->src,
 316                         u_bit_consecutive(0, instr->num_components));
 317
 318       return &snode->node;
 319
 320    case nir_intrinsic_discard:
 321       return ppir_emit_discard(block, ni);
 322
 323    case nir_intrinsic_discard_if:
 324       return ppir_emit_discard_if(block, ni);
 325
 326    default:
 327       ppir_error("unsupported nir_intrinsic_instr %s\n",
 328                  nir_intrinsic_infos[instr->intrinsic].name);
 329       return NULL;
 330    }
 331 }
 332
 333 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 334 {
 335    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 336    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 337    if (!node)
 338       return NULL;
 339
 340    assert(instr->def.bit_size == 32);
 341
 342    for (int i = 0; i < instr->def.num_components; i++)
 343       node->constant.value[i].i = instr->value[i].i32;
 344    node->constant.num = instr->def.num_components;
 345
 346    return &node->node;
 347 }
 348
 349 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 350 {
 351    ppir_error("nir_ssa_undef_instr not support\n");
 352    return NULL;
 353 }
 354
 355 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 356 {
 357    nir_tex_instr *instr = nir_instr_as_tex(ni);
 358    ppir_load_texture_node *node;
 359
 360    if (instr->op != nir_texop_tex) {
 361       ppir_error("unsupported texop %d\n", instr->op);
 362       return NULL;
 363    }
 364
 365    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, 0);
 366    if (!node)
 367       return NULL;
 368
 369    node->sampler = instr->texture_index;
 370
 371    switch (instr->sampler_dim) {
 372    case GLSL_SAMPLER_DIM_2D:
 373    case GLSL_SAMPLER_DIM_RECT:
 374    case GLSL_SAMPLER_DIM_EXTERNAL:
 375       break;
 376    default:
 377       ppir_debug("unsupported sampler dim: %d\n", instr->sampler_dim);
 378       return NULL;
 379    }
 380
 381    node->sampler_dim = instr->sampler_dim;
 382
 383    for (int i = 0; i < instr->coord_components; i++)
 384          node->src_coords.swizzle[i] = i;
 385
 386    assert(instr->num_srcs == 1);
 387    for (int i = 0; i < instr->num_srcs; i++) {
 388       switch (instr->src[i].src_type) {
 389       case nir_tex_src_coord:
 390          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 391                            u_bit_consecutive(0, instr->coord_components));
 392          break;
 393       default:
 394          ppir_debug("unknown texture source");
 395          return NULL;
 396       }
 397    }
 398
 399    return &node->node;
 400 }
 401
 402 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 403 {
 404    ppir_error("nir_jump_instr not support\n");
 405    return NULL;
 406 }
 407
 408 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 409    [nir_instr_type_alu]        = ppir_emit_alu,
 410    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 411    [nir_instr_type_load_const] = ppir_emit_load_const,
 412    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 413    [nir_instr_type_tex]        = ppir_emit_tex,
 414    [nir_instr_type_jump]       = ppir_emit_jump,
 415 };
 416
 417 static ppir_block *ppir_block_create(ppir_compiler *comp)
 418 {
 419    ppir_block *block = rzalloc(comp, ppir_block);
 420    if (!block)
 421       return NULL;
 422
 423    list_inithead(&block->node_list);
 424    list_inithead(&block->instr_list);
 425
 426    return block;
 427 }
 428
 429 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 430 {
 431    ppir_block *block = ppir_block_create(comp);
 432    if (!block)
 433       return false;
 434
 435    list_addtail(&block->list, &comp->block_list);
 436    block->comp = comp;
 437
 438    nir_foreach_instr(instr, nblock) {
 439       assert(instr->type < nir_instr_type_phi);
 440       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 441       if (!node)
 442          return false;
 443
 444       list_addtail(&node->list, &block->node_list);
 445    }
 446
 447    return true;
 448 }
 449
 450 static bool ppir_emit_if(ppir_compiler *comp, nir_if *nif)
 451 {
 452    ppir_error("if nir_cf_node not support\n");
 453    return false;
 454 }
 455
 456 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 457 {
 458    ppir_error("loop nir_cf_node not support\n");
 459    return false;
 460 }
 461
 462 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 463 {
 464    ppir_error("function nir_cf_node not support\n");
 465    return false;
 466 }
 467
 468 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 469 {
 470    foreach_list_typed(nir_cf_node, node, node, list) {
 471       bool ret;
 472
 473       switch (node->type) {
 474       case nir_cf_node_block:
 475          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 476          break;
 477       case nir_cf_node_if:
 478          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 479          break;
 480       case nir_cf_node_loop:
 481          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 482          break;
 483       case nir_cf_node_function:
 484          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 485          break;
 486       default:
 487          ppir_error("unknown NIR node type %d\n", node->type);
 488          return false;
 489       }
 490
 491       if (!ret)
 492          return false;
 493    }
 494
 495    return true;
 496 }
 497
 498 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 499 {
 500    ppir_compiler *comp = rzalloc_size(
 501       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 502    if (!comp)
 503       return NULL;
 504
 505    list_inithead(&comp->block_list);
 506    list_inithead(&comp->reg_list);
 507
 508    comp->var_nodes = (ppir_node **)(comp + 1);
 509    comp->reg_base = num_ssa;
 510    comp->prog = prog;
 511    return comp;
 512 }
 513
 514 static void ppir_add_ordering_deps(ppir_compiler *comp)
 515 {
 516    /* Some intrinsics do not have explicit dependencies and thus depend
 517     * on instructions order. Consider discard_if and store_ouput as
 518     * example. If we don't add fake dependency of discard_if to store_output
 519     * scheduler may put store_output first and since store_output terminates
 520     * shader on Utgard PP, rest of it will never be executed.
 521     * Add fake dependencies for discard/branch/store to preserve
 522     * instruction order.
 523     *
 524     * TODO: scheduler should schedule discard_if as early as possible otherwise
 525     * we may end up with suboptimal code for cases like this:
 526     *
 527     * s3 = s1 < s2
 528     * discard_if s3
 529     * s4 = s1 + s2
 530     * store s4
 531     *
 532     * In this case store depends on discard_if and s4, but since dependencies can
 533     * be scheduled in any order it can result in code like this:
 534     *
 535     * instr1: s3 = s1 < s3
 536     * instr2: s4 = s1 + s2
 537     * instr3: discard_if s3
 538     * instr4: store s4
 539     */
 540    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 541       ppir_node *prev_node = NULL;
 542       list_for_each_entry(ppir_node, node, &block->node_list, list) {
 543          if (node->type == ppir_node_type_discard ||
 544              node->type == ppir_node_type_store ||
 545              node->type == ppir_node_type_branch) {
 546             if (prev_node)
 547                ppir_node_add_dep(node, prev_node);
 548             prev_node = node;
 549          }
 550       }
 551    }
 552 }
 553
 554 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 555                       struct ra_regs *ra)
 556 {
 557    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 558    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 559    if (!comp)
 560       return false;
 561
 562    comp->ra = ra;
 563
 564    foreach_list_typed(nir_register, reg, node, &func->registers) {
 565       ppir_reg *r = rzalloc(comp, ppir_reg);
 566       if (!r)
 567          return false;
 568
 569       r->index = reg->index;
 570       r->num_components = reg->num_components;
 571       r->live_in = INT_MAX;
 572       r->live_out = 0;
 573       r->is_head = false;
 574       list_addtail(&r->list, &comp->reg_list);
 575    }
 576
 577    if (!ppir_emit_cf_list(comp, &func->body))
 578       goto err_out0;
 579
 580    /* If we have discard block add it to the very end */
 581    if (comp->discard_block)
 582       list_addtail(&comp->discard_block->list, &comp->block_list);
 583
 584    ppir_add_ordering_deps(comp);
 585
 586    ppir_node_print_prog(comp);
 587
 588    if (!ppir_lower_prog(comp))
 589       goto err_out0;
 590
 591    if (!ppir_node_to_instr(comp))
 592       goto err_out0;
 593
 594    if (!ppir_schedule_prog(comp))
 595       goto err_out0;
 596
 597    if (!ppir_regalloc_prog(comp))
 598       goto err_out0;
 599
 600    if (!ppir_codegen_prog(comp))
 601       goto err_out0;
 602
 603    ralloc_free(comp);
 604    return true;
 605
 606 err_out0:
 607    ralloc_free(comp);
 608    return false;
 609 }
 610