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 "compiler.h"
  36 #include "bi_quirks.h"
  37 #include "bi_print.h"
  38
  39 static bi_block *emit_cf_list(bi_context *ctx, struct exec_list *list);
  40 static bi_instruction *bi_emit_branch(bi_context *ctx);
  41 static void bi_block_add_successor(bi_block *block, bi_block *successor);
  42 static void bi_schedule_barrier(bi_context *ctx);
  43
  44 static void
  45 emit_jump(bi_context *ctx, nir_jump_instr *instr)
  46 {
  47         bi_instruction *branch = bi_emit_branch(ctx);
  48
  49         switch (instr->type) {
  50         case nir_jump_break:
  51                 branch->branch.target = ctx->break_block;
  52                 break;
  53         case nir_jump_continue:
  54                 branch->branch.target = ctx->continue_block;
  55                 break;
  56         default:
  57                 unreachable("Unhandled jump type");
  58         }
  59
  60         bi_block_add_successor(ctx->current_block, branch->branch.target);
  61 }
  62
  63 static void
  64 bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
  65 {
  66         bi_instruction ins = {
  67                 .type = BI_LOAD_VAR,
  68                 .load_vary = {
  69                         .load = {
  70                                 .location = nir_intrinsic_base(instr),
  71                                 .channels = instr->num_components,
  72                         },
  73                         .interp_mode = BIFROST_INTERP_DEFAULT, /* TODO */
  74                         .reuse = false, /* TODO */
  75                         .flat = instr->intrinsic != nir_intrinsic_load_interpolated_input
  76                 },
  77                 .dest = bir_dest_index(&instr->dest),
  78                 .dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
  79         };
  80
  81         nir_src *offset = nir_get_io_offset_src(instr);
  82
  83         if (nir_src_is_const(*offset))
  84                 ins.load_vary.load.location += nir_src_as_uint(*offset);
  85         else
  86                 ins.src[0] = bir_src_index(offset);
  87
  88         bi_emit(ctx, ins);
  89 }
  90
  91 static void
  92 bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr)
  93 {
  94         if (!ctx->emitted_atest) {
  95                 bi_instruction ins = {
  96                         .type = BI_ATEST
  97                 };
  98
  99                 bi_emit(ctx, ins);
 100                 bi_schedule_barrier(ctx);
 101                 ctx->emitted_atest = true;
 102         }
 103
 104         bi_instruction blend = {
 105                 .type = BI_BLEND,
 106                 .blend_location = nir_intrinsic_base(instr),
 107                 .src = {
 108                         bir_src_index(&instr->src[0])
 109                 },
 110                 .swizzle = {
 111                         { 0, 1, 2, 3 }
 112                 }
 113         };
 114
 115         bi_emit(ctx, blend);
 116         bi_schedule_barrier(ctx);
 117 }
 118
 119 static struct bi_load
 120 bi_direct_load_for_instr(nir_intrinsic_instr *instr)
 121 {
 122         nir_src *offset = nir_get_io_offset_src(instr);
 123         assert(nir_src_is_const(*offset)); /* no indirects */
 124
 125         struct bi_load load = {
 126                 .location = nir_intrinsic_base(instr) + nir_src_as_uint(*offset),
 127                 .channels = instr->num_components
 128         };
 129
 130         return load;
 131 }
 132
 133 static void
 134 bi_emit_ld_attr(bi_context *ctx, nir_intrinsic_instr *instr)
 135 {
 136         bi_instruction load = {
 137                 .type = BI_LOAD_ATTR,
 138                 .load = bi_direct_load_for_instr(instr),
 139                 .dest = bir_dest_index(&instr->dest),
 140                 .dest_type = nir_intrinsic_type(instr)
 141         };
 142
 143         bi_emit(ctx, load);
 144 }
 145
 146 static void
 147 bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 148 {
 149         nir_src *offset = nir_get_io_offset_src(instr);
 150         assert(nir_src_is_const(*offset)); /* no indirects */
 151
 152         bi_instruction address = {
 153                 .type = BI_LOAD_VAR_ADDRESS,
 154                 .load = bi_direct_load_for_instr(instr),
 155                 .dest_type = nir_intrinsic_type(instr),
 156                 .dest = bi_make_temp(ctx)
 157         };
 158
 159         bi_instruction st = {
 160                 .type = BI_STORE_VAR,
 161                 .src = {
 162                         address.dest,
 163                         bir_src_index(&instr->src[0])
 164                 },
 165                 .swizzle = {
 166                         { 0, 1, 2, 3 }
 167                 }
 168         };
 169
 170         bi_emit(ctx, address);
 171         bi_emit(ctx, st);
 172 }
 173
 174 static void
 175 bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr)
 176 {
 177         /* TODO: Indirect access */
 178
 179         bi_instruction ld = {
 180                 .type = BI_LOAD_UNIFORM,
 181                 .load = bi_direct_load_for_instr(instr),
 182                 .dest = bir_dest_index(&instr->dest),
 183                 .dest_type = nir_intrinsic_type(instr),
 184                 .src = {
 185                         BIR_INDEX_ZERO /* TODO: UBOs */
 186                 }
 187         };
 188
 189         bi_emit(ctx, ld);
 190 }
 191
 192 static void
 193 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
 194 {
 195
 196         switch (instr->intrinsic) {
 197         case nir_intrinsic_load_barycentric_pixel:
 198                 /* stub */
 199                 break;
 200         case nir_intrinsic_load_interpolated_input:
 201         case nir_intrinsic_load_input:
 202                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 203                         bi_emit_ld_vary(ctx, instr);
 204                 else if (ctx->stage == MESA_SHADER_VERTEX)
 205                         bi_emit_ld_attr(ctx, instr);
 206                 else {
 207                         unreachable("Unsupported shader stage");
 208                 }
 209                 break;
 210
 211         case nir_intrinsic_store_output:
 212                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 213                         bi_emit_frag_out(ctx, instr);
 214                 else if (ctx->stage == MESA_SHADER_VERTEX)
 215                         bi_emit_st_vary(ctx, instr);
 216                 else
 217                         unreachable("Unsupported shader stage");
 218                 break;
 219
 220         case nir_intrinsic_load_uniform:
 221                 bi_emit_ld_uniform(ctx, instr);
 222                 break;
 223
 224         default:
 225                 /* todo */
 226                 break;
 227         }
 228 }
 229
 230 static void
 231 emit_load_const(bi_context *ctx, nir_load_const_instr *instr)
 232 {
 233         /* Make sure we've been lowered */
 234         assert(instr->def.num_components == 1);
 235
 236         bi_instruction move = {
 237                 .type = BI_MOV,
 238                 .dest = bir_ssa_index(&instr->def),
 239                 .dest_type = instr->def.bit_size | nir_type_uint,
 240                 .src = {
 241                         BIR_INDEX_CONSTANT
 242                 },
 243                 .constant = {
 244                         .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
 245                 }
 246         };
 247
 248         bi_emit(ctx, move);
 249 }
 250
 251 static void
 252 emit_instr(bi_context *ctx, struct nir_instr *instr)
 253 {
 254         switch (instr->type) {
 255         case nir_instr_type_load_const:
 256                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 257                 break;
 258
 259         case nir_instr_type_intrinsic:
 260                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 261                 break;
 262
 263 #if 0
 264         case nir_instr_type_alu:
 265                 emit_alu(ctx, nir_instr_as_alu(instr));
 266                 break;
 267
 268         case nir_instr_type_tex:
 269                 emit_tex(ctx, nir_instr_as_tex(instr));
 270                 break;
 271 #endif
 272
 273         case nir_instr_type_jump:
 274                 emit_jump(ctx, nir_instr_as_jump(instr));
 275                 break;
 276
 277         case nir_instr_type_ssa_undef:
 278                 /* Spurious */
 279                 break;
 280
 281         default:
 282                 //unreachable("Unhandled instruction type");
 283                 break;
 284         }
 285 }
 286
 287
 288
 289 static bi_block *
 290 create_empty_block(bi_context *ctx)
 291 {
 292         bi_block *blk = rzalloc(ctx, bi_block);
 293
 294         blk->predecessors = _mesa_set_create(blk,
 295                         _mesa_hash_pointer,
 296                         _mesa_key_pointer_equal);
 297
 298         blk->name = ctx->block_name_count++;
 299
 300         return blk;
 301 }
 302
 303 static void
 304 bi_block_add_successor(bi_block *block, bi_block *successor)
 305 {
 306         assert(block);
 307         assert(successor);
 308
 309         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 310                 if (block->successors[i]) {
 311                        if (block->successors[i] == successor)
 312                                return;
 313                        else
 314                                continue;
 315                 }
 316
 317                 block->successors[i] = successor;
 318                 _mesa_set_add(successor->predecessors, block);
 319                 return;
 320         }
 321
 322         unreachable("Too many successors");
 323 }
 324
 325 static void
 326 bi_schedule_barrier(bi_context *ctx)
 327 {
 328         bi_block *temp = ctx->after_block;
 329         ctx->after_block = create_empty_block(ctx);
 330         list_addtail(&ctx->after_block->link, &ctx->blocks);
 331         list_inithead(&ctx->after_block->instructions);
 332         bi_block_add_successor(ctx->current_block, ctx->after_block);
 333         ctx->current_block = ctx->after_block;
 334         ctx->after_block = temp;
 335 }
 336
 337 static bi_block *
 338 emit_block(bi_context *ctx, nir_block *block)
 339 {
 340         if (ctx->after_block) {
 341                 ctx->current_block = ctx->after_block;
 342                 ctx->after_block = NULL;
 343         } else {
 344                 ctx->current_block = create_empty_block(ctx);
 345         }
 346
 347         list_addtail(&ctx->current_block->link, &ctx->blocks);
 348         list_inithead(&ctx->current_block->instructions);
 349
 350         nir_foreach_instr(instr, block) {
 351                 emit_instr(ctx, instr);
 352                 ++ctx->instruction_count;
 353         }
 354
 355         return ctx->current_block;
 356 }
 357
 358 /* Emits an unconditional branch to the end of the current block, returning a
 359  * pointer so the user can fill in details */
 360
 361 static bi_instruction *
 362 bi_emit_branch(bi_context *ctx)
 363 {
 364         bi_instruction branch = {
 365                 .type = BI_BRANCH,
 366                 .branch = {
 367                         .cond = BI_COND_ALWAYS
 368                 }
 369         };
 370
 371         return bi_emit(ctx, branch);
 372 }
 373
 374 /* Sets a condition for a branch by examing the NIR condition. If we're
 375  * familiar with the condition, we unwrap it to fold it into the branch
 376  * instruction. Otherwise, we consume the condition directly. We
 377  * generally use 1-bit booleans which allows us to use small types for
 378  * the conditions.
 379  */
 380
 381 static void
 382 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 383 {
 384         /* TODO: Try to unwrap instead of always bailing */
 385         branch->src[0] = bir_src_index(cond);
 386         branch->src[1] = BIR_INDEX_ZERO;
 387         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 388         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 389 }
 390
 391 static void
 392 emit_if(bi_context *ctx, nir_if *nif)
 393 {
 394         bi_block *before_block = ctx->current_block;
 395
 396         /* Speculatively emit the branch, but we can't fill it in until later */
 397         bi_instruction *then_branch = bi_emit_branch(ctx);
 398         bi_set_branch_cond(then_branch, &nif->condition, true);
 399
 400         /* Emit the two subblocks. */
 401         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 402         bi_block *end_then_block = ctx->current_block;
 403
 404         /* Emit a jump from the end of the then block to the end of the else */
 405         bi_instruction *then_exit = bi_emit_branch(ctx);
 406
 407         /* Emit second block, and check if it's empty */
 408
 409         int count_in = ctx->instruction_count;
 410         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 411         bi_block *end_else_block = ctx->current_block;
 412         ctx->after_block = create_empty_block(ctx);
 413
 414         /* Now that we have the subblocks emitted, fix up the branches */
 415
 416         assert(then_block);
 417         assert(else_block);
 418
 419         if (ctx->instruction_count == count_in) {
 420                 /* The else block is empty, so don't emit an exit jump */
 421                 bi_remove_instruction(then_exit);
 422                 then_branch->branch.target = ctx->after_block;
 423         } else {
 424                 then_branch->branch.target = else_block;
 425                 then_exit->branch.target = ctx->after_block;
 426                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 427         }
 428
 429         /* Wire up the successors */
 430
 431         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 432
 433         bi_block_add_successor(before_block, then_block); /* fallthrough */
 434         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 435 }
 436
 437 static void
 438 emit_loop(bi_context *ctx, nir_loop *nloop)
 439 {
 440         /* Remember where we are */
 441         bi_block *start_block = ctx->current_block;
 442
 443         bi_block *saved_break = ctx->break_block;
 444         bi_block *saved_continue = ctx->continue_block;
 445
 446         ctx->continue_block = create_empty_block(ctx);
 447         ctx->break_block = create_empty_block(ctx);
 448         ctx->after_block = ctx->continue_block;
 449
 450         /* Emit the body itself */
 451         emit_cf_list(ctx, &nloop->body);
 452
 453         /* Branch back to loop back */
 454         bi_instruction *br_back = bi_emit_branch(ctx);
 455         br_back->branch.target = ctx->continue_block;
 456         bi_block_add_successor(start_block, ctx->continue_block);
 457         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 458
 459         ctx->after_block = ctx->break_block;
 460
 461         /* Pop off */
 462         ctx->break_block = saved_break;
 463         ctx->continue_block = saved_continue;
 464         ++ctx->loop_count;
 465 }
 466
 467 static bi_block *
 468 emit_cf_list(bi_context *ctx, struct exec_list *list)
 469 {
 470         bi_block *start_block = NULL;
 471
 472         foreach_list_typed(nir_cf_node, node, node, list) {
 473                 switch (node->type) {
 474                 case nir_cf_node_block: {
 475                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 476
 477                         if (!start_block)
 478                                 start_block = block;
 479
 480                         break;
 481                 }
 482
 483                 case nir_cf_node_if:
 484                         emit_if(ctx, nir_cf_node_as_if(node));
 485                         break;
 486
 487                 case nir_cf_node_loop:
 488                         emit_loop(ctx, nir_cf_node_as_loop(node));
 489                         break;
 490
 491                 default:
 492                         unreachable("Unknown control flow");
 493                 }
 494         }
 495
 496         return start_block;
 497 }
 498
 499 static int
 500 glsl_type_size(const struct glsl_type *type, bool bindless)
 501 {
 502         return glsl_count_attribute_slots(type, false);
 503 }
 504
 505 static void
 506 bi_optimize_nir(nir_shader *nir)
 507 {
 508         bool progress;
 509         unsigned lower_flrp = 16 | 32 | 64;
 510
 511         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 512         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 513
 514         nir_lower_tex_options lower_tex_options = {
 515                 .lower_txs_lod = true,
 516                 .lower_txp = ~0,
 517                 .lower_tex_without_implicit_lod = true,
 518                 .lower_txd = true,
 519         };
 520
 521         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 522         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 523         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 524
 525         do {
 526                 progress = false;
 527
 528                 NIR_PASS(progress, nir, nir_lower_var_copies);
 529                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 530
 531                 NIR_PASS(progress, nir, nir_copy_prop);
 532                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 533                 NIR_PASS(progress, nir, nir_opt_dce);
 534                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 535                 NIR_PASS(progress, nir, nir_opt_cse);
 536                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 537                 NIR_PASS(progress, nir, nir_opt_algebraic);
 538                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 539
 540                 if (lower_flrp != 0) {
 541                         bool lower_flrp_progress = false;
 542                         NIR_PASS(lower_flrp_progress,
 543                                  nir,
 544                                  nir_lower_flrp,
 545                                  lower_flrp,
 546                                  false /* always_precise */,
 547                                  nir->options->lower_ffma);
 548                         if (lower_flrp_progress) {
 549                                 NIR_PASS(progress, nir,
 550                                          nir_opt_constant_folding);
 551                                 progress = true;
 552                         }
 553
 554                         /* Nothing should rematerialize any flrps, so we only
 555                          * need to do this lowering once.
 556                          */
 557                         lower_flrp = 0;
 558                 }
 559
 560                 NIR_PASS(progress, nir, nir_opt_undef);
 561                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 562                          nir_var_shader_in |
 563                          nir_var_shader_out |
 564                          nir_var_function_temp);
 565         } while (progress);
 566
 567         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 568         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 569         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 570
 571         /* Take us out of SSA */
 572         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 573         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 574 }
 575
 576 void
 577 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 578 {
 579         bi_context *ctx = rzalloc(NULL, bi_context);
 580         ctx->nir = nir;
 581         ctx->stage = nir->info.stage;
 582         ctx->quirks = bifrost_get_quirks(product_id);
 583         list_inithead(&ctx->blocks);
 584
 585         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 586          * (so we don't accidentally duplicate the epilogue since mesa/st has
 587          * messed with our I/O quite a bit already) */
 588
 589         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 590
 591         if (ctx->stage == MESA_SHADER_VERTEX) {
 592                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 593                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 594         }
 595
 596         NIR_PASS_V(nir, nir_split_var_copies);
 597         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 598         NIR_PASS_V(nir, nir_lower_var_copies);
 599         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 600         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 601         NIR_PASS_V(nir, nir_lower_ssbo);
 602
 603         bi_optimize_nir(nir);
 604         nir_print_shader(nir, stdout);
 605
 606         nir_foreach_function(func, nir) {
 607                 if (!func->impl)
 608                         continue;
 609
 610                 ctx->impl = func->impl;
 611                 emit_cf_list(ctx, &func->impl->body);
 612                 break; /* TODO: Multi-function shaders */
 613         }
 614
 615         bi_print_shader(ctx, stdout);
 616         bi_schedule(ctx);
 617         bi_print_shader(ctx, stdout);
 618
 619         ralloc_free(ctx);
 620 }