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_fmov] = ppir_op_mov,
 121    [nir_op_imov] = ppir_op_mov,
 122    [nir_op_fmul] = ppir_op_mul,
 123    [nir_op_fadd] = ppir_op_add,
 124    [nir_op_fdot2] = ppir_op_dot2,
 125    [nir_op_fdot3] = ppir_op_dot3,
 126    [nir_op_fdot4] = ppir_op_dot4,
 127    [nir_op_frsq] = ppir_op_rsqrt,
 128    [nir_op_flog2] = ppir_op_log2,
 129    [nir_op_fexp2] = ppir_op_exp2,
 130    [nir_op_fsqrt] = ppir_op_sqrt,
 131    [nir_op_fsin] = ppir_op_sin,
 132    [nir_op_fcos] = ppir_op_cos,
 133    [nir_op_fmax] = ppir_op_max,
 134    [nir_op_fmin] = ppir_op_min,
 135    [nir_op_frcp] = ppir_op_rcp,
 136    [nir_op_ffloor] = ppir_op_floor,
 137    [nir_op_ffract] = ppir_op_fract,
 138    [nir_op_fand] = ppir_op_and,
 139    [nir_op_for] = ppir_op_or,
 140    [nir_op_fxor] = ppir_op_xor,
 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_fnot] = ppir_op_not,
 150    [nir_op_fcsel] = ppir_op_select,
 151    [nir_op_inot] = ppir_op_not,
 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_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 207 {
 208    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 209    unsigned mask = 0;
 210    ppir_load_node *lnode;
 211    ppir_store_node *snode;
 212    nir_const_value *const_offset;
 213
 214    switch (instr->intrinsic) {
 215    case nir_intrinsic_load_input:
 216       if (!instr->dest.is_ssa)
 217          mask = u_bit_consecutive(0, instr->num_components);
 218
 219       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 220       if (!lnode)
 221          return NULL;
 222
 223       lnode->num_components = instr->num_components;
 224       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 225       return &lnode->node;
 226
 227    case nir_intrinsic_load_uniform:
 228       if (!instr->dest.is_ssa)
 229          mask = u_bit_consecutive(0, instr->num_components);
 230
 231       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 232       if (!lnode)
 233          return NULL;
 234
 235       lnode->num_components = instr->num_components;
 236       lnode->index = nir_intrinsic_base(instr);
 237
 238       const_offset = nir_src_as_const_value(instr->src[0]);
 239       assert(const_offset);
 240       lnode->index += (uint32_t)const_offset->f32[0];
 241
 242       return &lnode->node;
 243
 244    case nir_intrinsic_store_output:
 245       snode = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 246       if (!snode)
 247          return NULL;
 248
 249       snode->index = nir_intrinsic_base(instr);
 250
 251       for (int i = 0; i < instr->num_components; i++)
 252          snode->src.swizzle[i] = i;
 253
 254       ppir_node_add_src(block->comp, &snode->node, &snode->src, instr->src,
 255                         u_bit_consecutive(0, instr->num_components));
 256
 257       return &snode->node;
 258
 259    default:
 260       ppir_error("unsupported nir_intrinsic_instr %d\n", instr->intrinsic);
 261       return NULL;
 262    }
 263 }
 264
 265 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 266 {
 267    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 268    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 269    if (!node)
 270       return NULL;
 271
 272    assert(instr->def.bit_size == 32);
 273
 274    for (int i = 0; i < instr->def.num_components; i++)
 275       node->constant.value[i].i = instr->value.i32[i];
 276    node->constant.num = instr->def.num_components;
 277
 278    return &node->node;
 279 }
 280
 281 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 282 {
 283    ppir_error("nir_ssa_undef_instr not support\n");
 284    return NULL;
 285 }
 286
 287 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 288 {
 289    nir_tex_instr *instr = nir_instr_as_tex(ni);
 290    ppir_load_texture_node *node;
 291
 292    if (instr->op != nir_texop_tex) {
 293       ppir_error("unsupported texop %d\n", instr->op);
 294       return NULL;
 295    }
 296
 297    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, 0);
 298    if (!node)
 299       return NULL;
 300
 301    node->sampler = instr->texture_index;
 302
 303    switch (instr->sampler_dim) {
 304    case GLSL_SAMPLER_DIM_2D:
 305    case GLSL_SAMPLER_DIM_RECT:
 306    case GLSL_SAMPLER_DIM_EXTERNAL:
 307       break;
 308    default:
 309       ppir_debug("unsupported sampler dim: %d\n", instr->sampler_dim);
 310       return NULL;
 311    }
 312
 313    node->sampler_dim = instr->sampler_dim;
 314
 315    for (int i = 0; i < instr->coord_components; i++)
 316          node->src_coords.swizzle[i] = i;
 317
 318    assert(instr->num_srcs == 1);
 319    for (int i = 0; i < instr->num_srcs; i++) {
 320       switch (instr->src[i].src_type) {
 321       case nir_tex_src_coord:
 322          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 323                            u_bit_consecutive(0, instr->coord_components));
 324          break;
 325       default:
 326          ppir_debug("unknown texture source");
 327          return NULL;
 328       }
 329    }
 330
 331    return &node->node;
 332 }
 333
 334 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 335 {
 336    ppir_error("nir_jump_instr not support\n");
 337    return NULL;
 338 }
 339
 340 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 341    [nir_instr_type_alu]        = ppir_emit_alu,
 342    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 343    [nir_instr_type_load_const] = ppir_emit_load_const,
 344    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 345    [nir_instr_type_tex]        = ppir_emit_tex,
 346    [nir_instr_type_jump]       = ppir_emit_jump,
 347 };
 348
 349 static ppir_block *ppir_block_create(ppir_compiler *comp)
 350 {
 351    ppir_block *block = rzalloc(comp, ppir_block);
 352    if (!block)
 353       return NULL;
 354
 355    list_inithead(&block->node_list);
 356    list_inithead(&block->instr_list);
 357
 358    return block;
 359 }
 360
 361 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 362 {
 363    ppir_block *block = ppir_block_create(comp);
 364    if (!block)
 365       return false;
 366
 367    list_addtail(&block->list, &comp->block_list);
 368    block->comp = comp;
 369
 370    nir_foreach_instr(instr, nblock) {
 371       assert(instr->type < nir_instr_type_phi);
 372       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 373       if (node)
 374          list_addtail(&node->list, &block->node_list);
 375    }
 376
 377    return true;
 378 }
 379
 380 static bool ppir_emit_if(ppir_compiler *comp, nir_if *nif)
 381 {
 382    ppir_error("if nir_cf_node not support\n");
 383    return false;
 384 }
 385
 386 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 387 {
 388    ppir_error("loop nir_cf_node not support\n");
 389    return false;
 390 }
 391
 392 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 393 {
 394    ppir_error("function nir_cf_node not support\n");
 395    return false;
 396 }
 397
 398 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 399 {
 400    foreach_list_typed(nir_cf_node, node, node, list) {
 401       bool ret;
 402
 403       switch (node->type) {
 404       case nir_cf_node_block:
 405          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 406          break;
 407       case nir_cf_node_if:
 408          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 409          break;
 410       case nir_cf_node_loop:
 411          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 412          break;
 413       case nir_cf_node_function:
 414          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 415          break;
 416       default:
 417          ppir_error("unknown NIR node type %d\n", node->type);
 418          return false;
 419       }
 420
 421       if (!ret)
 422          return false;
 423    }
 424
 425    return true;
 426 }
 427
 428 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 429 {
 430    ppir_compiler *comp = rzalloc_size(
 431       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 432    if (!comp)
 433       return NULL;
 434
 435    list_inithead(&comp->block_list);
 436    list_inithead(&comp->reg_list);
 437
 438    comp->var_nodes = (ppir_node **)(comp + 1);
 439    comp->reg_base = num_ssa;
 440    comp->prog = prog;
 441    return comp;
 442 }
 443
 444 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 445                       struct ra_regs *ra)
 446 {
 447    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 448    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 449    if (!comp)
 450       return false;
 451
 452    comp->ra = ra;
 453
 454    foreach_list_typed(nir_register, reg, node, &func->registers) {
 455       ppir_reg *r = rzalloc(comp, ppir_reg);
 456       if (!r)
 457          return false;
 458
 459       r->index = reg->index;
 460       r->num_components = reg->num_components;
 461       r->live_in = INT_MAX;
 462       r->live_out = 0;
 463       r->is_head = false;
 464       list_addtail(&r->list, &comp->reg_list);
 465    }
 466
 467    if (!ppir_emit_cf_list(comp, &func->body))
 468       goto err_out0;
 469    ppir_node_print_prog(comp);
 470
 471    if (!ppir_lower_prog(comp))
 472       goto err_out0;
 473
 474    if (!ppir_node_to_instr(comp))
 475       goto err_out0;
 476
 477    if (!ppir_schedule_prog(comp))
 478       goto err_out0;
 479
 480    if (!ppir_regalloc_prog(comp))
 481       goto err_out0;
 482
 483    if (!ppir_codegen_prog(comp))
 484       goto err_out0;
 485
 486    ralloc_free(comp);
 487    return true;
 488
 489 err_out0:
 490    ralloc_free(comp);
 491    return false;
 492 }
 493