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 /* Gets a bytemask for a complete vecN write */
  64 static unsigned
  65 bi_mask_for_channels_32(unsigned i)
  66 {
  67         return (1 << (4 * i)) - 1;
  68 }
  69
  70 static void
  71 bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
  72 {
  73         bi_instruction ins = {
  74                 .type = BI_LOAD_VAR,
  75                 .load_vary = {
  76                         .load = {
  77                                 .location = nir_intrinsic_base(instr),
  78                                 .channels = instr->num_components,
  79                         },
  80                         .interp_mode = BIFROST_INTERP_DEFAULT, /* TODO */
  81                         .reuse = false, /* TODO */
  82                         .flat = instr->intrinsic != nir_intrinsic_load_interpolated_input
  83                 },
  84                 .dest = bir_dest_index(&instr->dest),
  85                 .dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
  86                 .writemask = bi_mask_for_channels_32(instr->num_components)
  87         };
  88
  89         nir_src *offset = nir_get_io_offset_src(instr);
  90
  91         if (nir_src_is_const(*offset))
  92                 ins.load_vary.load.location += nir_src_as_uint(*offset);
  93         else
  94                 ins.src[0] = bir_src_index(offset);
  95
  96         bi_emit(ctx, ins);
  97 }
  98
  99 static void
 100 bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr)
 101 {
 102         if (!ctx->emitted_atest) {
 103                 bi_instruction ins = {
 104                         .type = BI_ATEST
 105                 };
 106
 107                 bi_emit(ctx, ins);
 108                 bi_schedule_barrier(ctx);
 109                 ctx->emitted_atest = true;
 110         }
 111
 112         bi_instruction blend = {
 113                 .type = BI_BLEND,
 114                 .blend_location = nir_intrinsic_base(instr),
 115                 .src = {
 116                         bir_src_index(&instr->src[0])
 117                 },
 118                 .swizzle = {
 119                         { 0, 1, 2, 3 }
 120                 }
 121         };
 122
 123         bi_emit(ctx, blend);
 124         bi_schedule_barrier(ctx);
 125 }
 126
 127 static struct bi_load
 128 bi_direct_load_for_instr(nir_intrinsic_instr *instr)
 129 {
 130         nir_src *offset = nir_get_io_offset_src(instr);
 131         assert(nir_src_is_const(*offset)); /* no indirects */
 132
 133         struct bi_load load = {
 134                 .location = nir_intrinsic_base(instr) + nir_src_as_uint(*offset),
 135                 .channels = instr->num_components
 136         };
 137
 138         return load;
 139 }
 140
 141 static void
 142 bi_emit_ld_attr(bi_context *ctx, nir_intrinsic_instr *instr)
 143 {
 144         bi_instruction load = {
 145                 .type = BI_LOAD_ATTR,
 146                 .load = bi_direct_load_for_instr(instr),
 147                 .dest = bir_dest_index(&instr->dest),
 148                 .dest_type = nir_intrinsic_type(instr),
 149                 .writemask = bi_mask_for_channels_32(instr->num_components)
 150         };
 151
 152         bi_emit(ctx, load);
 153 }
 154
 155 static void
 156 bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 157 {
 158         nir_src *offset = nir_get_io_offset_src(instr);
 159         assert(nir_src_is_const(*offset)); /* no indirects */
 160
 161         bi_instruction address = {
 162                 .type = BI_LOAD_VAR_ADDRESS,
 163                 .load = bi_direct_load_for_instr(instr),
 164                 .dest_type = nir_intrinsic_type(instr),
 165                 .dest = bi_make_temp(ctx),
 166                 .writemask = bi_mask_for_channels_32(instr->num_components)
 167         };
 168
 169         bi_instruction st = {
 170                 .type = BI_STORE_VAR,
 171                 .src = {
 172                         address.dest,
 173                         bir_src_index(&instr->src[0])
 174                 },
 175                 .swizzle = {
 176                         { 0, 1, 2, 3 }
 177                 }
 178         };
 179
 180         bi_emit(ctx, address);
 181         bi_emit(ctx, st);
 182 }
 183
 184 static void
 185 bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr)
 186 {
 187         /* TODO: Indirect access */
 188
 189         bi_instruction ld = {
 190                 .type = BI_LOAD_UNIFORM,
 191                 .load = bi_direct_load_for_instr(instr),
 192                 .dest = bir_dest_index(&instr->dest),
 193                 .dest_type = nir_intrinsic_type(instr),
 194                 .writemask = bi_mask_for_channels_32(instr->num_components),
 195                 .src = {
 196                         BIR_INDEX_ZERO /* TODO: UBOs */
 197                 }
 198         };
 199
 200         bi_emit(ctx, ld);
 201 }
 202
 203 static void
 204 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
 205 {
 206
 207         switch (instr->intrinsic) {
 208         case nir_intrinsic_load_barycentric_pixel:
 209                 /* stub */
 210                 break;
 211         case nir_intrinsic_load_interpolated_input:
 212         case nir_intrinsic_load_input:
 213                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 214                         bi_emit_ld_vary(ctx, instr);
 215                 else if (ctx->stage == MESA_SHADER_VERTEX)
 216                         bi_emit_ld_attr(ctx, instr);
 217                 else {
 218                         unreachable("Unsupported shader stage");
 219                 }
 220                 break;
 221
 222         case nir_intrinsic_store_output:
 223                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 224                         bi_emit_frag_out(ctx, instr);
 225                 else if (ctx->stage == MESA_SHADER_VERTEX)
 226                         bi_emit_st_vary(ctx, instr);
 227                 else
 228                         unreachable("Unsupported shader stage");
 229                 break;
 230
 231         case nir_intrinsic_load_uniform:
 232                 bi_emit_ld_uniform(ctx, instr);
 233                 break;
 234
 235         default:
 236                 /* todo */
 237                 break;
 238         }
 239 }
 240
 241 static void
 242 emit_load_const(bi_context *ctx, nir_load_const_instr *instr)
 243 {
 244         /* Make sure we've been lowered */
 245         assert(instr->def.num_components == 1);
 246
 247         bi_instruction move = {
 248                 .type = BI_MOV,
 249                 .dest = bir_ssa_index(&instr->def),
 250                 .dest_type = instr->def.bit_size | nir_type_uint,
 251                 .writemask = (1 << (instr->def.bit_size / 8)) - 1,
 252                 .src = {
 253                         BIR_INDEX_CONSTANT
 254                 },
 255                 .constant = {
 256                         .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
 257                 }
 258         };
 259
 260         bi_emit(ctx, move);
 261 }
 262
 263 static void
 264 emit_instr(bi_context *ctx, struct nir_instr *instr)
 265 {
 266         switch (instr->type) {
 267         case nir_instr_type_load_const:
 268                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 269                 break;
 270
 271         case nir_instr_type_intrinsic:
 272                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 273                 break;
 274
 275 #if 0
 276         case nir_instr_type_alu:
 277                 emit_alu(ctx, nir_instr_as_alu(instr));
 278                 break;
 279
 280         case nir_instr_type_tex:
 281                 emit_tex(ctx, nir_instr_as_tex(instr));
 282                 break;
 283 #endif
 284
 285         case nir_instr_type_jump:
 286                 emit_jump(ctx, nir_instr_as_jump(instr));
 287                 break;
 288
 289         case nir_instr_type_ssa_undef:
 290                 /* Spurious */
 291                 break;
 292
 293         default:
 294                 //unreachable("Unhandled instruction type");
 295                 break;
 296         }
 297 }
 298
 299
 300
 301 static bi_block *
 302 create_empty_block(bi_context *ctx)
 303 {
 304         bi_block *blk = rzalloc(ctx, bi_block);
 305
 306         blk->predecessors = _mesa_set_create(blk,
 307                         _mesa_hash_pointer,
 308                         _mesa_key_pointer_equal);
 309
 310         blk->name = ctx->block_name_count++;
 311
 312         return blk;
 313 }
 314
 315 static void
 316 bi_block_add_successor(bi_block *block, bi_block *successor)
 317 {
 318         assert(block);
 319         assert(successor);
 320
 321         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 322                 if (block->successors[i]) {
 323                        if (block->successors[i] == successor)
 324                                return;
 325                        else
 326                                continue;
 327                 }
 328
 329                 block->successors[i] = successor;
 330                 _mesa_set_add(successor->predecessors, block);
 331                 return;
 332         }
 333
 334         unreachable("Too many successors");
 335 }
 336
 337 static void
 338 bi_schedule_barrier(bi_context *ctx)
 339 {
 340         bi_block *temp = ctx->after_block;
 341         ctx->after_block = create_empty_block(ctx);
 342         list_addtail(&ctx->after_block->link, &ctx->blocks);
 343         list_inithead(&ctx->after_block->instructions);
 344         bi_block_add_successor(ctx->current_block, ctx->after_block);
 345         ctx->current_block = ctx->after_block;
 346         ctx->after_block = temp;
 347 }
 348
 349 static bi_block *
 350 emit_block(bi_context *ctx, nir_block *block)
 351 {
 352         if (ctx->after_block) {
 353                 ctx->current_block = ctx->after_block;
 354                 ctx->after_block = NULL;
 355         } else {
 356                 ctx->current_block = create_empty_block(ctx);
 357         }
 358
 359         list_addtail(&ctx->current_block->link, &ctx->blocks);
 360         list_inithead(&ctx->current_block->instructions);
 361
 362         nir_foreach_instr(instr, block) {
 363                 emit_instr(ctx, instr);
 364                 ++ctx->instruction_count;
 365         }
 366
 367         return ctx->current_block;
 368 }
 369
 370 /* Emits an unconditional branch to the end of the current block, returning a
 371  * pointer so the user can fill in details */
 372
 373 static bi_instruction *
 374 bi_emit_branch(bi_context *ctx)
 375 {
 376         bi_instruction branch = {
 377                 .type = BI_BRANCH,
 378                 .branch = {
 379                         .cond = BI_COND_ALWAYS
 380                 }
 381         };
 382
 383         return bi_emit(ctx, branch);
 384 }
 385
 386 /* Sets a condition for a branch by examing the NIR condition. If we're
 387  * familiar with the condition, we unwrap it to fold it into the branch
 388  * instruction. Otherwise, we consume the condition directly. We
 389  * generally use 1-bit booleans which allows us to use small types for
 390  * the conditions.
 391  */
 392
 393 static void
 394 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 395 {
 396         /* TODO: Try to unwrap instead of always bailing */
 397         branch->src[0] = bir_src_index(cond);
 398         branch->src[1] = BIR_INDEX_ZERO;
 399         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 400         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 401 }
 402
 403 static void
 404 emit_if(bi_context *ctx, nir_if *nif)
 405 {
 406         bi_block *before_block = ctx->current_block;
 407
 408         /* Speculatively emit the branch, but we can't fill it in until later */
 409         bi_instruction *then_branch = bi_emit_branch(ctx);
 410         bi_set_branch_cond(then_branch, &nif->condition, true);
 411
 412         /* Emit the two subblocks. */
 413         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 414         bi_block *end_then_block = ctx->current_block;
 415
 416         /* Emit a jump from the end of the then block to the end of the else */
 417         bi_instruction *then_exit = bi_emit_branch(ctx);
 418
 419         /* Emit second block, and check if it's empty */
 420
 421         int count_in = ctx->instruction_count;
 422         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 423         bi_block *end_else_block = ctx->current_block;
 424         ctx->after_block = create_empty_block(ctx);
 425
 426         /* Now that we have the subblocks emitted, fix up the branches */
 427
 428         assert(then_block);
 429         assert(else_block);
 430
 431         if (ctx->instruction_count == count_in) {
 432                 /* The else block is empty, so don't emit an exit jump */
 433                 bi_remove_instruction(then_exit);
 434                 then_branch->branch.target = ctx->after_block;
 435         } else {
 436                 then_branch->branch.target = else_block;
 437                 then_exit->branch.target = ctx->after_block;
 438                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 439         }
 440
 441         /* Wire up the successors */
 442
 443         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 444
 445         bi_block_add_successor(before_block, then_block); /* fallthrough */
 446         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 447 }
 448
 449 static void
 450 emit_loop(bi_context *ctx, nir_loop *nloop)
 451 {
 452         /* Remember where we are */
 453         bi_block *start_block = ctx->current_block;
 454
 455         bi_block *saved_break = ctx->break_block;
 456         bi_block *saved_continue = ctx->continue_block;
 457
 458         ctx->continue_block = create_empty_block(ctx);
 459         ctx->break_block = create_empty_block(ctx);
 460         ctx->after_block = ctx->continue_block;
 461
 462         /* Emit the body itself */
 463         emit_cf_list(ctx, &nloop->body);
 464
 465         /* Branch back to loop back */
 466         bi_instruction *br_back = bi_emit_branch(ctx);
 467         br_back->branch.target = ctx->continue_block;
 468         bi_block_add_successor(start_block, ctx->continue_block);
 469         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 470
 471         ctx->after_block = ctx->break_block;
 472
 473         /* Pop off */
 474         ctx->break_block = saved_break;
 475         ctx->continue_block = saved_continue;
 476         ++ctx->loop_count;
 477 }
 478
 479 static bi_block *
 480 emit_cf_list(bi_context *ctx, struct exec_list *list)
 481 {
 482         bi_block *start_block = NULL;
 483
 484         foreach_list_typed(nir_cf_node, node, node, list) {
 485                 switch (node->type) {
 486                 case nir_cf_node_block: {
 487                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 488
 489                         if (!start_block)
 490                                 start_block = block;
 491
 492                         break;
 493                 }
 494
 495                 case nir_cf_node_if:
 496                         emit_if(ctx, nir_cf_node_as_if(node));
 497                         break;
 498
 499                 case nir_cf_node_loop:
 500                         emit_loop(ctx, nir_cf_node_as_loop(node));
 501                         break;
 502
 503                 default:
 504                         unreachable("Unknown control flow");
 505                 }
 506         }
 507
 508         return start_block;
 509 }
 510
 511 static int
 512 glsl_type_size(const struct glsl_type *type, bool bindless)
 513 {
 514         return glsl_count_attribute_slots(type, false);
 515 }
 516
 517 static void
 518 bi_optimize_nir(nir_shader *nir)
 519 {
 520         bool progress;
 521         unsigned lower_flrp = 16 | 32 | 64;
 522
 523         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 524         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 525
 526         nir_lower_tex_options lower_tex_options = {
 527                 .lower_txs_lod = true,
 528                 .lower_txp = ~0,
 529                 .lower_tex_without_implicit_lod = true,
 530                 .lower_txd = true,
 531         };
 532
 533         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 534         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 535         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 536
 537         do {
 538                 progress = false;
 539
 540                 NIR_PASS(progress, nir, nir_lower_var_copies);
 541                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 542
 543                 NIR_PASS(progress, nir, nir_copy_prop);
 544                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 545                 NIR_PASS(progress, nir, nir_opt_dce);
 546                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 547                 NIR_PASS(progress, nir, nir_opt_cse);
 548                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 549                 NIR_PASS(progress, nir, nir_opt_algebraic);
 550                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 551
 552                 if (lower_flrp != 0) {
 553                         bool lower_flrp_progress = false;
 554                         NIR_PASS(lower_flrp_progress,
 555                                  nir,
 556                                  nir_lower_flrp,
 557                                  lower_flrp,
 558                                  false /* always_precise */,
 559                                  nir->options->lower_ffma);
 560                         if (lower_flrp_progress) {
 561                                 NIR_PASS(progress, nir,
 562                                          nir_opt_constant_folding);
 563                                 progress = true;
 564                         }
 565
 566                         /* Nothing should rematerialize any flrps, so we only
 567                          * need to do this lowering once.
 568                          */
 569                         lower_flrp = 0;
 570                 }
 571
 572                 NIR_PASS(progress, nir, nir_opt_undef);
 573                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 574                          nir_var_shader_in |
 575                          nir_var_shader_out |
 576                          nir_var_function_temp);
 577         } while (progress);
 578
 579         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 580         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 581         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 582
 583         /* Take us out of SSA */
 584         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 585         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 586 }
 587
 588 void
 589 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 590 {
 591         bi_context *ctx = rzalloc(NULL, bi_context);
 592         ctx->nir = nir;
 593         ctx->stage = nir->info.stage;
 594         ctx->quirks = bifrost_get_quirks(product_id);
 595         list_inithead(&ctx->blocks);
 596
 597         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 598          * (so we don't accidentally duplicate the epilogue since mesa/st has
 599          * messed with our I/O quite a bit already) */
 600
 601         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 602
 603         if (ctx->stage == MESA_SHADER_VERTEX) {
 604                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 605                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 606         }
 607
 608         NIR_PASS_V(nir, nir_split_var_copies);
 609         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 610         NIR_PASS_V(nir, nir_lower_var_copies);
 611         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 612         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 613         NIR_PASS_V(nir, nir_lower_ssbo);
 614
 615         bi_optimize_nir(nir);
 616         nir_print_shader(nir, stdout);
 617
 618         nir_foreach_function(func, nir) {
 619                 if (!func->impl)
 620                         continue;
 621
 622                 ctx->impl = func->impl;
 623                 emit_cf_list(ctx, &func->impl->body);
 624                 break; /* TODO: Multi-function shaders */
 625         }
 626
 627         bi_print_shader(ctx, stdout);
 628         bi_schedule(ctx);
 629
 630         ralloc_free(ctx);
 631 }