src/panfrost/bifrost/bifrost_compile.c

   1 /*
   2  * Copyright (C) 2020 Collabora Ltd.
   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, sublicense,
   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 next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * 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 NONINFRINGEMENT.  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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors (Collabora):
  24  *      Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26
  27 #include "main/mtypes.h"
  28 #include "compiler/glsl/glsl_to_nir.h"
  29 #include "compiler/nir_types.h"
  30 #include "main/imports.h"
  31 #include "compiler/nir/nir_builder.h"
  32
  33 #include "disassemble.h"
  34 #include "bifrost_compile.h"
  35 #include "bifrost_nir.h"
  36 #include "compiler.h"
  37 #include "bi_quirks.h"
  38 #include "bi_print.h"
  39
  40 static bi_block *emit_cf_list(bi_context *ctx, struct exec_list *list);
  41 static bi_instruction *bi_emit_branch(bi_context *ctx);
  42 static void bi_block_add_successor(bi_block *block, bi_block *successor);
  43 static void bi_schedule_barrier(bi_context *ctx);
  44
  45 static void
  46 emit_jump(bi_context *ctx, nir_jump_instr *instr)
  47 {
  48         bi_instruction *branch = bi_emit_branch(ctx);
  49
  50         switch (instr->type) {
  51         case nir_jump_break:
  52                 branch->branch.target = ctx->break_block;
  53                 break;
  54         case nir_jump_continue:
  55                 branch->branch.target = ctx->continue_block;
  56                 break;
  57         default:
  58                 unreachable("Unhandled jump type");
  59         }
  60
  61         bi_block_add_successor(ctx->current_block, branch->branch.target);
  62 }
  63
  64 /* Gets a bytemask for a complete vecN write */
  65 static unsigned
  66 bi_mask_for_channels_32(unsigned i)
  67 {
  68         return (1 << (4 * i)) - 1;
  69 }
  70
  71 static bi_instruction
  72 bi_load(enum bi_class T, nir_intrinsic_instr *instr)
  73 {
  74         bi_instruction load = {
  75                 .type = T,
  76                 .writemask = bi_mask_for_channels_32(instr->num_components),
  77                 .src = { BIR_INDEX_CONSTANT },
  78                 .constant = { .u64 = nir_intrinsic_base(instr) },
  79         };
  80
  81         const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
  82
  83         if (info->has_dest)
  84                 load.dest = bir_dest_index(&instr->dest);
  85
  86         if (info->has_dest && info->index_map[NIR_INTRINSIC_TYPE] > 0)
  87                 load.dest_type = nir_intrinsic_type(instr);
  88
  89         nir_src *offset = nir_get_io_offset_src(instr);
  90
  91         if (nir_src_is_const(*offset))
  92                 load.constant.u64 += nir_src_as_uint(*offset);
  93         else
  94                 load.src[0] = bir_src_index(offset);
  95
  96         return load;
  97 }
  98
  99 static void
 100 bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 101 {
 102         bi_instruction ins = bi_load(BI_LOAD_VAR, instr);
 103         ins.load_vary.interp_mode = BIFROST_INTERP_DEFAULT; /* TODO */
 104         ins.load_vary.reuse = false; /* TODO */
 105         ins.load_vary.flat = instr->intrinsic != nir_intrinsic_load_interpolated_input;
 106         ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
 107         bi_emit(ctx, ins);
 108 }
 109
 110 static void
 111 bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr)
 112 {
 113         if (!ctx->emitted_atest) {
 114                 bi_instruction ins = {
 115                         .type = BI_ATEST
 116                 };
 117
 118                 bi_emit(ctx, ins);
 119                 bi_schedule_barrier(ctx);
 120                 ctx->emitted_atest = true;
 121         }
 122
 123         bi_instruction blend = {
 124                 .type = BI_BLEND,
 125                 .blend_location = nir_intrinsic_base(instr),
 126                 .src = {
 127                         bir_src_index(&instr->src[0])
 128                 },
 129                 .swizzle = {
 130                         { 0, 1, 2, 3 }
 131                 }
 132         };
 133
 134         bi_emit(ctx, blend);
 135         bi_schedule_barrier(ctx);
 136 }
 137
 138 static void
 139 bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 140 {
 141         bi_instruction address = bi_load(BI_LOAD_VAR_ADDRESS, instr);
 142         address.dest = bi_make_temp(ctx);
 143         address.dest_type = nir_type_uint64;
 144         address.writemask = (1 << 8) - 1;
 145
 146         bi_instruction st = {
 147                 .type = BI_STORE_VAR,
 148                 .src = {
 149                         address.dest,
 150                         bir_src_index(&instr->src[0])
 151                 },
 152                 .swizzle = {
 153                         { 0, 1, 2, 3 }
 154                 }
 155         };
 156
 157         bi_emit(ctx, address);
 158         bi_emit(ctx, st);
 159 }
 160
 161 static void
 162 bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr)
 163 {
 164         bi_instruction ld = bi_load(BI_LOAD_UNIFORM, instr);
 165         ld.src[1] = BIR_INDEX_ZERO; /* TODO: UBO index */
 166         bi_emit(ctx, ld);
 167 }
 168
 169 static void
 170 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
 171 {
 172
 173         switch (instr->intrinsic) {
 174         case nir_intrinsic_load_barycentric_pixel:
 175                 /* stub */
 176                 break;
 177         case nir_intrinsic_load_interpolated_input:
 178         case nir_intrinsic_load_input:
 179                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 180                         bi_emit_ld_vary(ctx, instr);
 181                 else if (ctx->stage == MESA_SHADER_VERTEX)
 182                         bi_emit(ctx, bi_load(BI_LOAD_ATTR, instr));
 183                 else {
 184                         unreachable("Unsupported shader stage");
 185                 }
 186                 break;
 187
 188         case nir_intrinsic_store_output:
 189                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 190                         bi_emit_frag_out(ctx, instr);
 191                 else if (ctx->stage == MESA_SHADER_VERTEX)
 192                         bi_emit_st_vary(ctx, instr);
 193                 else
 194                         unreachable("Unsupported shader stage");
 195                 break;
 196
 197         case nir_intrinsic_load_uniform:
 198                 bi_emit_ld_uniform(ctx, instr);
 199                 break;
 200
 201         default:
 202                 /* todo */
 203                 break;
 204         }
 205 }
 206
 207 static void
 208 emit_load_const(bi_context *ctx, nir_load_const_instr *instr)
 209 {
 210         /* Make sure we've been lowered */
 211         assert(instr->def.num_components == 1);
 212
 213         bi_instruction move = {
 214                 .type = BI_MOV,
 215                 .dest = bir_ssa_index(&instr->def),
 216                 .dest_type = instr->def.bit_size | nir_type_uint,
 217                 .writemask = (1 << (instr->def.bit_size / 8)) - 1,
 218                 .src = {
 219                         BIR_INDEX_CONSTANT
 220                 },
 221                 .constant = {
 222                         .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
 223                 }
 224         };
 225
 226         bi_emit(ctx, move);
 227 }
 228
 229 static enum bi_class
 230 bi_class_for_nir_alu(nir_op op)
 231 {
 232         switch (op) {
 233         case nir_op_iadd:
 234         case nir_op_fadd:
 235         case nir_op_fsub:
 236                 return BI_ADD;
 237         case nir_op_isub:
 238                 return BI_ISUB;
 239
 240         case nir_op_flt:
 241         case nir_op_fge:
 242         case nir_op_feq:
 243         case nir_op_fne:
 244         case nir_op_ilt:
 245         case nir_op_ige:
 246         case nir_op_ieq:
 247         case nir_op_ine:
 248                 return BI_CMP;
 249
 250         case nir_op_bcsel:
 251                 return BI_CSEL;
 252
 253         case nir_op_i2i8:
 254         case nir_op_i2i16:
 255         case nir_op_i2i32:
 256         case nir_op_i2i64:
 257         case nir_op_u2u8:
 258         case nir_op_u2u16:
 259         case nir_op_u2u32:
 260         case nir_op_u2u64:
 261         case nir_op_f2i16:
 262         case nir_op_f2i32:
 263         case nir_op_f2i64:
 264         case nir_op_f2u16:
 265         case nir_op_f2u32:
 266         case nir_op_f2u64:
 267         case nir_op_i2f16:
 268         case nir_op_i2f32:
 269         case nir_op_i2f64:
 270         case nir_op_u2f16:
 271         case nir_op_u2f32:
 272         case nir_op_u2f64:
 273                 return BI_CONVERT;
 274
 275         case nir_op_fmul:
 276                 return BI_FMA;
 277
 278         case nir_op_imin:
 279         case nir_op_imax:
 280         case nir_op_umin:
 281         case nir_op_umax:
 282         case nir_op_fmin:
 283         case nir_op_fmax:
 284                 return BI_MINMAX;
 285
 286         case nir_op_fsat:
 287         case nir_op_fneg:
 288         case nir_op_fabs:
 289         case nir_op_mov:
 290                 return BI_MOV;
 291
 292         case nir_op_frcp:
 293         case nir_op_frsq:
 294         case nir_op_fsin:
 295         case nir_op_fcos:
 296                 return BI_SPECIAL;
 297
 298         default:
 299                 unreachable("Unknown ALU op");
 300         }
 301 }
 302
 303 static enum bi_cond
 304 bi_cond_for_nir(nir_op op)
 305 {
 306         switch (op) {
 307         case nir_op_flt:
 308         case nir_op_ilt:
 309                 return BI_COND_LT;
 310         case nir_op_fge:
 311         case nir_op_ige:
 312                 return BI_COND_GE;
 313         case nir_op_feq:
 314         case nir_op_ieq:
 315                 return BI_COND_EQ;
 316         case nir_op_fne:
 317         case nir_op_ine:
 318                 return BI_COND_NE;
 319         default:
 320                 unreachable("Invalid compare");
 321         }
 322 }
 323
 324 static void
 325 emit_alu(bi_context *ctx, nir_alu_instr *instr)
 326 {
 327         /* Assume it's something we can handle normally */
 328         bi_instruction alu = {
 329                 .type = bi_class_for_nir_alu(instr->op),
 330                 .dest = bir_dest_index(&instr->dest.dest),
 331                 .dest_type = nir_op_infos[instr->op].output_type
 332                         | nir_dest_bit_size(instr->dest.dest),
 333         };
 334
 335         /* TODO: Implement lowering of special functions for older Bifrost */
 336         assert((alu.type != BI_SPECIAL) || !(ctx->quirks & BIFROST_NO_FAST_OP));
 337
 338         if (instr->dest.dest.is_ssa) {
 339                 /* Construct a writemask */
 340                 unsigned bits_per_comp = instr->dest.dest.ssa.bit_size;
 341                 unsigned comps = instr->dest.dest.ssa.num_components;
 342                 assert(comps == 1);
 343                 unsigned bits = bits_per_comp * comps;
 344                 unsigned bytes = MAX2(bits / 8, 1);
 345                 alu.writemask = (1 << bytes) - 1;
 346         } else {
 347                 unsigned comp_mask = instr->dest.write_mask;
 348
 349                 alu.writemask = pan_to_bytemask(nir_dest_bit_size(instr->dest.dest),
 350                                 comp_mask);
 351         }
 352
 353         /* We inline constants as we go. This tracks how many constants have
 354          * been inlined, since we're limited to 64-bits of constants per
 355          * instruction */
 356
 357         unsigned dest_bits = nir_dest_bit_size(instr->dest.dest);
 358         unsigned constants_left = (64 / dest_bits);
 359         unsigned constant_shift = 0;
 360
 361         /* Copy sources */
 362
 363         unsigned num_inputs = nir_op_infos[instr->op].num_inputs;
 364         assert(num_inputs <= ARRAY_SIZE(alu.src));
 365
 366         for (unsigned i = 0; i < num_inputs; ++i) {
 367                 unsigned bits = nir_src_bit_size(instr->src[i].src);
 368                 alu.src_types[i] = nir_op_infos[instr->op].input_types[i]
 369                         | bits;
 370
 371                 /* Try to inline a constant */
 372                 if (nir_src_is_const(instr->src[i].src) && constants_left && (dest_bits == bits)) {
 373                         alu.constant.u64 |=
 374                                 (nir_src_as_uint(instr->src[i].src)) << constant_shift;
 375
 376                         alu.src[i] = BIR_INDEX_CONSTANT | constant_shift;
 377                         --constants_left;
 378                         constant_shift += dest_bits;
 379                         continue;
 380                 }
 381
 382                 alu.src[i] = bir_src_index(&instr->src[i].src);
 383
 384                 /* We assert scalarization above */
 385                 alu.swizzle[i][0] = instr->src[i].swizzle[0];
 386         }
 387
 388         /* Op-specific fixup */
 389         switch (instr->op) {
 390         case nir_op_fmul:
 391                 alu.src[2] = BIR_INDEX_ZERO; /* FMA */
 392                 break;
 393         case nir_op_fsat:
 394                 alu.outmod = BIFROST_SAT; /* MOV */
 395                 break;
 396         case nir_op_fneg:
 397                 alu.src_neg[0] = true; /* MOV */
 398                 break;
 399         case nir_op_fabs:
 400                 alu.src_abs[0] = true; /* MOV */
 401                 break;
 402         case nir_op_fsub:
 403                 alu.src_neg[1] = true; /* ADD */
 404                 break;
 405         case nir_op_fmax:
 406         case nir_op_imax:
 407         case nir_op_umax:
 408                 alu.op.minmax = BI_MINMAX_MAX; /* MINMAX */
 409                 break;
 410         case nir_op_frcp:
 411                 alu.op.special = BI_SPECIAL_FRCP;
 412                 break;
 413         case nir_op_frsq:
 414                 alu.op.special = BI_SPECIAL_FRSQ;
 415                 break;
 416         case nir_op_fsin:
 417                 alu.op.special = BI_SPECIAL_FSIN;
 418                 break;
 419         case nir_op_fcos:
 420                 alu.op.special = BI_SPECIAL_FCOS;
 421                 break;
 422         case nir_op_flt:
 423         case nir_op_ilt:
 424         case nir_op_fge:
 425         case nir_op_ige:
 426         case nir_op_feq:
 427         case nir_op_ieq:
 428         case nir_op_fne:
 429         case nir_op_ine:
 430                 alu.op.compare = bi_cond_for_nir(instr->op);
 431                 break;
 432         default:
 433                 break;
 434         }
 435
 436         bi_emit(ctx, alu);
 437 }
 438
 439 static void
 440 emit_instr(bi_context *ctx, struct nir_instr *instr)
 441 {
 442         switch (instr->type) {
 443         case nir_instr_type_load_const:
 444                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 445                 break;
 446
 447         case nir_instr_type_intrinsic:
 448                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 449                 break;
 450
 451         case nir_instr_type_alu:
 452                 emit_alu(ctx, nir_instr_as_alu(instr));
 453                 break;
 454
 455 #if 0
 456         case nir_instr_type_tex:
 457                 emit_tex(ctx, nir_instr_as_tex(instr));
 458                 break;
 459 #endif
 460
 461         case nir_instr_type_jump:
 462                 emit_jump(ctx, nir_instr_as_jump(instr));
 463                 break;
 464
 465         case nir_instr_type_ssa_undef:
 466                 /* Spurious */
 467                 break;
 468
 469         default:
 470                 //unreachable("Unhandled instruction type");
 471                 break;
 472         }
 473 }
 474
 475
 476
 477 static bi_block *
 478 create_empty_block(bi_context *ctx)
 479 {
 480         bi_block *blk = rzalloc(ctx, bi_block);
 481
 482         blk->predecessors = _mesa_set_create(blk,
 483                         _mesa_hash_pointer,
 484                         _mesa_key_pointer_equal);
 485
 486         blk->name = ctx->block_name_count++;
 487
 488         return blk;
 489 }
 490
 491 static void
 492 bi_block_add_successor(bi_block *block, bi_block *successor)
 493 {
 494         assert(block);
 495         assert(successor);
 496
 497         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 498                 if (block->successors[i]) {
 499                        if (block->successors[i] == successor)
 500                                return;
 501                        else
 502                                continue;
 503                 }
 504
 505                 block->successors[i] = successor;
 506                 _mesa_set_add(successor->predecessors, block);
 507                 return;
 508         }
 509
 510         unreachable("Too many successors");
 511 }
 512
 513 static void
 514 bi_schedule_barrier(bi_context *ctx)
 515 {
 516         bi_block *temp = ctx->after_block;
 517         ctx->after_block = create_empty_block(ctx);
 518         list_addtail(&ctx->after_block->link, &ctx->blocks);
 519         list_inithead(&ctx->after_block->instructions);
 520         bi_block_add_successor(ctx->current_block, ctx->after_block);
 521         ctx->current_block = ctx->after_block;
 522         ctx->after_block = temp;
 523 }
 524
 525 static bi_block *
 526 emit_block(bi_context *ctx, nir_block *block)
 527 {
 528         if (ctx->after_block) {
 529                 ctx->current_block = ctx->after_block;
 530                 ctx->after_block = NULL;
 531         } else {
 532                 ctx->current_block = create_empty_block(ctx);
 533         }
 534
 535         list_addtail(&ctx->current_block->link, &ctx->blocks);
 536         list_inithead(&ctx->current_block->instructions);
 537
 538         nir_foreach_instr(instr, block) {
 539                 emit_instr(ctx, instr);
 540                 ++ctx->instruction_count;
 541         }
 542
 543         return ctx->current_block;
 544 }
 545
 546 /* Emits an unconditional branch to the end of the current block, returning a
 547  * pointer so the user can fill in details */
 548
 549 static bi_instruction *
 550 bi_emit_branch(bi_context *ctx)
 551 {
 552         bi_instruction branch = {
 553                 .type = BI_BRANCH,
 554                 .branch = {
 555                         .cond = BI_COND_ALWAYS
 556                 }
 557         };
 558
 559         return bi_emit(ctx, branch);
 560 }
 561
 562 /* Sets a condition for a branch by examing the NIR condition. If we're
 563  * familiar with the condition, we unwrap it to fold it into the branch
 564  * instruction. Otherwise, we consume the condition directly. We
 565  * generally use 1-bit booleans which allows us to use small types for
 566  * the conditions.
 567  */
 568
 569 static void
 570 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 571 {
 572         /* TODO: Try to unwrap instead of always bailing */
 573         branch->src[0] = bir_src_index(cond);
 574         branch->src[1] = BIR_INDEX_ZERO;
 575         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 576         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 577 }
 578
 579 static void
 580 emit_if(bi_context *ctx, nir_if *nif)
 581 {
 582         bi_block *before_block = ctx->current_block;
 583
 584         /* Speculatively emit the branch, but we can't fill it in until later */
 585         bi_instruction *then_branch = bi_emit_branch(ctx);
 586         bi_set_branch_cond(then_branch, &nif->condition, true);
 587
 588         /* Emit the two subblocks. */
 589         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 590         bi_block *end_then_block = ctx->current_block;
 591
 592         /* Emit a jump from the end of the then block to the end of the else */
 593         bi_instruction *then_exit = bi_emit_branch(ctx);
 594
 595         /* Emit second block, and check if it's empty */
 596
 597         int count_in = ctx->instruction_count;
 598         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 599         bi_block *end_else_block = ctx->current_block;
 600         ctx->after_block = create_empty_block(ctx);
 601
 602         /* Now that we have the subblocks emitted, fix up the branches */
 603
 604         assert(then_block);
 605         assert(else_block);
 606
 607         if (ctx->instruction_count == count_in) {
 608                 /* The else block is empty, so don't emit an exit jump */
 609                 bi_remove_instruction(then_exit);
 610                 then_branch->branch.target = ctx->after_block;
 611         } else {
 612                 then_branch->branch.target = else_block;
 613                 then_exit->branch.target = ctx->after_block;
 614                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 615         }
 616
 617         /* Wire up the successors */
 618
 619         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 620
 621         bi_block_add_successor(before_block, then_block); /* fallthrough */
 622         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 623 }
 624
 625 static void
 626 emit_loop(bi_context *ctx, nir_loop *nloop)
 627 {
 628         /* Remember where we are */
 629         bi_block *start_block = ctx->current_block;
 630
 631         bi_block *saved_break = ctx->break_block;
 632         bi_block *saved_continue = ctx->continue_block;
 633
 634         ctx->continue_block = create_empty_block(ctx);
 635         ctx->break_block = create_empty_block(ctx);
 636         ctx->after_block = ctx->continue_block;
 637
 638         /* Emit the body itself */
 639         emit_cf_list(ctx, &nloop->body);
 640
 641         /* Branch back to loop back */
 642         bi_instruction *br_back = bi_emit_branch(ctx);
 643         br_back->branch.target = ctx->continue_block;
 644         bi_block_add_successor(start_block, ctx->continue_block);
 645         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 646
 647         ctx->after_block = ctx->break_block;
 648
 649         /* Pop off */
 650         ctx->break_block = saved_break;
 651         ctx->continue_block = saved_continue;
 652         ++ctx->loop_count;
 653 }
 654
 655 static bi_block *
 656 emit_cf_list(bi_context *ctx, struct exec_list *list)
 657 {
 658         bi_block *start_block = NULL;
 659
 660         foreach_list_typed(nir_cf_node, node, node, list) {
 661                 switch (node->type) {
 662                 case nir_cf_node_block: {
 663                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 664
 665                         if (!start_block)
 666                                 start_block = block;
 667
 668                         break;
 669                 }
 670
 671                 case nir_cf_node_if:
 672                         emit_if(ctx, nir_cf_node_as_if(node));
 673                         break;
 674
 675                 case nir_cf_node_loop:
 676                         emit_loop(ctx, nir_cf_node_as_loop(node));
 677                         break;
 678
 679                 default:
 680                         unreachable("Unknown control flow");
 681                 }
 682         }
 683
 684         return start_block;
 685 }
 686
 687 static int
 688 glsl_type_size(const struct glsl_type *type, bool bindless)
 689 {
 690         return glsl_count_attribute_slots(type, false);
 691 }
 692
 693 static void
 694 bi_optimize_nir(nir_shader *nir)
 695 {
 696         bool progress;
 697         unsigned lower_flrp = 16 | 32 | 64;
 698
 699         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 700         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 701
 702         nir_lower_tex_options lower_tex_options = {
 703                 .lower_txs_lod = true,
 704                 .lower_txp = ~0,
 705                 .lower_tex_without_implicit_lod = true,
 706                 .lower_txd = true,
 707         };
 708
 709         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 710         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 711         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 712
 713         do {
 714                 progress = false;
 715
 716                 NIR_PASS(progress, nir, nir_lower_var_copies);
 717                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 718
 719                 NIR_PASS(progress, nir, nir_copy_prop);
 720                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 721                 NIR_PASS(progress, nir, nir_opt_dce);
 722                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 723                 NIR_PASS(progress, nir, nir_opt_cse);
 724                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 725                 NIR_PASS(progress, nir, nir_opt_algebraic);
 726                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 727
 728                 if (lower_flrp != 0) {
 729                         bool lower_flrp_progress = false;
 730                         NIR_PASS(lower_flrp_progress,
 731                                  nir,
 732                                  nir_lower_flrp,
 733                                  lower_flrp,
 734                                  false /* always_precise */,
 735                                  nir->options->lower_ffma);
 736                         if (lower_flrp_progress) {
 737                                 NIR_PASS(progress, nir,
 738                                          nir_opt_constant_folding);
 739                                 progress = true;
 740                         }
 741
 742                         /* Nothing should rematerialize any flrps, so we only
 743                          * need to do this lowering once.
 744                          */
 745                         lower_flrp = 0;
 746                 }
 747
 748                 NIR_PASS(progress, nir, nir_opt_undef);
 749                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 750                          nir_var_shader_in |
 751                          nir_var_shader_out |
 752                          nir_var_function_temp);
 753         } while (progress);
 754
 755         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 756         NIR_PASS(progress, nir, bifrost_nir_lower_algebraic_late);
 757         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 758         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 759
 760         /* Take us out of SSA */
 761         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 762         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 763
 764         /* We're a primary scalar architecture but there's enough vector that
 765          * we use a vector IR so let's not also deal with scalar hacks on top
 766          * of the vector hacks */
 767
 768         NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
 769         NIR_PASS(progress, nir, nir_lower_vec_to_movs);
 770         NIR_PASS(progress, nir, nir_opt_dce);
 771 }
 772
 773 void
 774 bifrost_compile_shader_nir(nir_shader *nir, panfrost_program *program, unsigned product_id)
 775 {
 776         bi_context *ctx = rzalloc(NULL, bi_context);
 777         ctx->nir = nir;
 778         ctx->stage = nir->info.stage;
 779         ctx->quirks = bifrost_get_quirks(product_id);
 780         list_inithead(&ctx->blocks);
 781
 782         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 783          * (so we don't accidentally duplicate the epilogue since mesa/st has
 784          * messed with our I/O quite a bit already) */
 785
 786         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 787
 788         if (ctx->stage == MESA_SHADER_VERTEX) {
 789                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 790                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 791         }
 792
 793         NIR_PASS_V(nir, nir_split_var_copies);
 794         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 795         NIR_PASS_V(nir, nir_lower_var_copies);
 796         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 797         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 798         NIR_PASS_V(nir, nir_lower_ssbo);
 799
 800         bi_optimize_nir(nir);
 801         nir_print_shader(nir, stdout);
 802
 803         nir_foreach_function(func, nir) {
 804                 if (!func->impl)
 805                         continue;
 806
 807                 ctx->impl = func->impl;
 808                 emit_cf_list(ctx, &func->impl->body);
 809                 break; /* TODO: Multi-function shaders */
 810         }
 811
 812         bi_print_shader(ctx, stdout);
 813         bi_schedule(ctx);
 814
 815         ralloc_free(ctx);
 816 }