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 bi_instruction
  71 bi_load(enum bi_class T, nir_intrinsic_instr *instr)
  72 {
  73         bi_instruction load = {
  74                 .type = T,
  75                 .writemask = bi_mask_for_channels_32(instr->num_components),
  76                 .src = { BIR_INDEX_CONSTANT },
  77                 .constant = { .u64 = nir_intrinsic_base(instr) },
  78         };
  79
  80         const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
  81
  82         if (info->has_dest)
  83                 load.dest = bir_dest_index(&instr->dest);
  84
  85         if (info->has_dest && info->index_map[NIR_INTRINSIC_TYPE] > 0)
  86                 load.dest_type = nir_intrinsic_type(instr);
  87
  88         nir_src *offset = nir_get_io_offset_src(instr);
  89
  90         if (nir_src_is_const(*offset))
  91                 load.constant.u64 += nir_src_as_uint(*offset);
  92         else
  93                 load.src[0] = bir_src_index(offset);
  94
  95         return load;
  96 }
  97
  98 static void
  99 bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 100 {
 101         bi_instruction ins = bi_load(BI_LOAD_VAR, instr);
 102         ins.load_vary.interp_mode = BIFROST_INTERP_DEFAULT; /* TODO */
 103         ins.load_vary.reuse = false; /* TODO */
 104         ins.load_vary.flat = instr->intrinsic != nir_intrinsic_load_interpolated_input;
 105         ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
 106         bi_emit(ctx, ins);
 107 }
 108
 109 static void
 110 bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr)
 111 {
 112         if (!ctx->emitted_atest) {
 113                 bi_instruction ins = {
 114                         .type = BI_ATEST
 115                 };
 116
 117                 bi_emit(ctx, ins);
 118                 bi_schedule_barrier(ctx);
 119                 ctx->emitted_atest = true;
 120         }
 121
 122         bi_instruction blend = {
 123                 .type = BI_BLEND,
 124                 .blend_location = nir_intrinsic_base(instr),
 125                 .src = {
 126                         bir_src_index(&instr->src[0])
 127                 },
 128                 .swizzle = {
 129                         { 0, 1, 2, 3 }
 130                 }
 131         };
 132
 133         bi_emit(ctx, blend);
 134         bi_schedule_barrier(ctx);
 135 }
 136
 137 static void
 138 bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 139 {
 140         bi_instruction address = bi_load(BI_LOAD_VAR_ADDRESS, instr);
 141         address.dest = bi_make_temp(ctx);
 142         address.dest_type = nir_type_uint64;
 143         address.writemask = (1 << 8) - 1;
 144
 145         bi_instruction st = {
 146                 .type = BI_STORE_VAR,
 147                 .src = {
 148                         address.dest,
 149                         bir_src_index(&instr->src[0])
 150                 },
 151                 .swizzle = {
 152                         { 0, 1, 2, 3 }
 153                 }
 154         };
 155
 156         bi_emit(ctx, address);
 157         bi_emit(ctx, st);
 158 }
 159
 160 static void
 161 bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr)
 162 {
 163         bi_instruction ld = bi_load(BI_LOAD_UNIFORM, instr);
 164         ld.src[1] = BIR_INDEX_ZERO; /* TODO: UBO index */
 165         bi_emit(ctx, ld);
 166 }
 167
 168 static void
 169 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
 170 {
 171
 172         switch (instr->intrinsic) {
 173         case nir_intrinsic_load_barycentric_pixel:
 174                 /* stub */
 175                 break;
 176         case nir_intrinsic_load_interpolated_input:
 177         case nir_intrinsic_load_input:
 178                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 179                         bi_emit_ld_vary(ctx, instr);
 180                 else if (ctx->stage == MESA_SHADER_VERTEX)
 181                         bi_emit(ctx, bi_load(BI_LOAD_ATTR, instr));
 182                 else {
 183                         unreachable("Unsupported shader stage");
 184                 }
 185                 break;
 186
 187         case nir_intrinsic_store_output:
 188                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 189                         bi_emit_frag_out(ctx, instr);
 190                 else if (ctx->stage == MESA_SHADER_VERTEX)
 191                         bi_emit_st_vary(ctx, instr);
 192                 else
 193                         unreachable("Unsupported shader stage");
 194                 break;
 195
 196         case nir_intrinsic_load_uniform:
 197                 bi_emit_ld_uniform(ctx, instr);
 198                 break;
 199
 200         default:
 201                 /* todo */
 202                 break;
 203         }
 204 }
 205
 206 static void
 207 emit_load_const(bi_context *ctx, nir_load_const_instr *instr)
 208 {
 209         /* Make sure we've been lowered */
 210         assert(instr->def.num_components == 1);
 211
 212         bi_instruction move = {
 213                 .type = BI_MOV,
 214                 .dest = bir_ssa_index(&instr->def),
 215                 .dest_type = instr->def.bit_size | nir_type_uint,
 216                 .writemask = (1 << (instr->def.bit_size / 8)) - 1,
 217                 .src = {
 218                         BIR_INDEX_CONSTANT
 219                 },
 220                 .constant = {
 221                         .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
 222                 }
 223         };
 224
 225         bi_emit(ctx, move);
 226 }
 227
 228 static void
 229 emit_instr(bi_context *ctx, struct nir_instr *instr)
 230 {
 231         switch (instr->type) {
 232         case nir_instr_type_load_const:
 233                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 234                 break;
 235
 236         case nir_instr_type_intrinsic:
 237                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 238                 break;
 239
 240 #if 0
 241         case nir_instr_type_alu:
 242                 emit_alu(ctx, nir_instr_as_alu(instr));
 243                 break;
 244
 245         case nir_instr_type_tex:
 246                 emit_tex(ctx, nir_instr_as_tex(instr));
 247                 break;
 248 #endif
 249
 250         case nir_instr_type_jump:
 251                 emit_jump(ctx, nir_instr_as_jump(instr));
 252                 break;
 253
 254         case nir_instr_type_ssa_undef:
 255                 /* Spurious */
 256                 break;
 257
 258         default:
 259                 //unreachable("Unhandled instruction type");
 260                 break;
 261         }
 262 }
 263
 264
 265
 266 static bi_block *
 267 create_empty_block(bi_context *ctx)
 268 {
 269         bi_block *blk = rzalloc(ctx, bi_block);
 270
 271         blk->predecessors = _mesa_set_create(blk,
 272                         _mesa_hash_pointer,
 273                         _mesa_key_pointer_equal);
 274
 275         blk->name = ctx->block_name_count++;
 276
 277         return blk;
 278 }
 279
 280 static void
 281 bi_block_add_successor(bi_block *block, bi_block *successor)
 282 {
 283         assert(block);
 284         assert(successor);
 285
 286         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 287                 if (block->successors[i]) {
 288                        if (block->successors[i] == successor)
 289                                return;
 290                        else
 291                                continue;
 292                 }
 293
 294                 block->successors[i] = successor;
 295                 _mesa_set_add(successor->predecessors, block);
 296                 return;
 297         }
 298
 299         unreachable("Too many successors");
 300 }
 301
 302 static void
 303 bi_schedule_barrier(bi_context *ctx)
 304 {
 305         bi_block *temp = ctx->after_block;
 306         ctx->after_block = create_empty_block(ctx);
 307         list_addtail(&ctx->after_block->link, &ctx->blocks);
 308         list_inithead(&ctx->after_block->instructions);
 309         bi_block_add_successor(ctx->current_block, ctx->after_block);
 310         ctx->current_block = ctx->after_block;
 311         ctx->after_block = temp;
 312 }
 313
 314 static bi_block *
 315 emit_block(bi_context *ctx, nir_block *block)
 316 {
 317         if (ctx->after_block) {
 318                 ctx->current_block = ctx->after_block;
 319                 ctx->after_block = NULL;
 320         } else {
 321                 ctx->current_block = create_empty_block(ctx);
 322         }
 323
 324         list_addtail(&ctx->current_block->link, &ctx->blocks);
 325         list_inithead(&ctx->current_block->instructions);
 326
 327         nir_foreach_instr(instr, block) {
 328                 emit_instr(ctx, instr);
 329                 ++ctx->instruction_count;
 330         }
 331
 332         return ctx->current_block;
 333 }
 334
 335 /* Emits an unconditional branch to the end of the current block, returning a
 336  * pointer so the user can fill in details */
 337
 338 static bi_instruction *
 339 bi_emit_branch(bi_context *ctx)
 340 {
 341         bi_instruction branch = {
 342                 .type = BI_BRANCH,
 343                 .branch = {
 344                         .cond = BI_COND_ALWAYS
 345                 }
 346         };
 347
 348         return bi_emit(ctx, branch);
 349 }
 350
 351 /* Sets a condition for a branch by examing the NIR condition. If we're
 352  * familiar with the condition, we unwrap it to fold it into the branch
 353  * instruction. Otherwise, we consume the condition directly. We
 354  * generally use 1-bit booleans which allows us to use small types for
 355  * the conditions.
 356  */
 357
 358 static void
 359 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 360 {
 361         /* TODO: Try to unwrap instead of always bailing */
 362         branch->src[0] = bir_src_index(cond);
 363         branch->src[1] = BIR_INDEX_ZERO;
 364         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 365         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 366 }
 367
 368 static void
 369 emit_if(bi_context *ctx, nir_if *nif)
 370 {
 371         bi_block *before_block = ctx->current_block;
 372
 373         /* Speculatively emit the branch, but we can't fill it in until later */
 374         bi_instruction *then_branch = bi_emit_branch(ctx);
 375         bi_set_branch_cond(then_branch, &nif->condition, true);
 376
 377         /* Emit the two subblocks. */
 378         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 379         bi_block *end_then_block = ctx->current_block;
 380
 381         /* Emit a jump from the end of the then block to the end of the else */
 382         bi_instruction *then_exit = bi_emit_branch(ctx);
 383
 384         /* Emit second block, and check if it's empty */
 385
 386         int count_in = ctx->instruction_count;
 387         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 388         bi_block *end_else_block = ctx->current_block;
 389         ctx->after_block = create_empty_block(ctx);
 390
 391         /* Now that we have the subblocks emitted, fix up the branches */
 392
 393         assert(then_block);
 394         assert(else_block);
 395
 396         if (ctx->instruction_count == count_in) {
 397                 /* The else block is empty, so don't emit an exit jump */
 398                 bi_remove_instruction(then_exit);
 399                 then_branch->branch.target = ctx->after_block;
 400         } else {
 401                 then_branch->branch.target = else_block;
 402                 then_exit->branch.target = ctx->after_block;
 403                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 404         }
 405
 406         /* Wire up the successors */
 407
 408         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 409
 410         bi_block_add_successor(before_block, then_block); /* fallthrough */
 411         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 412 }
 413
 414 static void
 415 emit_loop(bi_context *ctx, nir_loop *nloop)
 416 {
 417         /* Remember where we are */
 418         bi_block *start_block = ctx->current_block;
 419
 420         bi_block *saved_break = ctx->break_block;
 421         bi_block *saved_continue = ctx->continue_block;
 422
 423         ctx->continue_block = create_empty_block(ctx);
 424         ctx->break_block = create_empty_block(ctx);
 425         ctx->after_block = ctx->continue_block;
 426
 427         /* Emit the body itself */
 428         emit_cf_list(ctx, &nloop->body);
 429
 430         /* Branch back to loop back */
 431         bi_instruction *br_back = bi_emit_branch(ctx);
 432         br_back->branch.target = ctx->continue_block;
 433         bi_block_add_successor(start_block, ctx->continue_block);
 434         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 435
 436         ctx->after_block = ctx->break_block;
 437
 438         /* Pop off */
 439         ctx->break_block = saved_break;
 440         ctx->continue_block = saved_continue;
 441         ++ctx->loop_count;
 442 }
 443
 444 static bi_block *
 445 emit_cf_list(bi_context *ctx, struct exec_list *list)
 446 {
 447         bi_block *start_block = NULL;
 448
 449         foreach_list_typed(nir_cf_node, node, node, list) {
 450                 switch (node->type) {
 451                 case nir_cf_node_block: {
 452                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 453
 454                         if (!start_block)
 455                                 start_block = block;
 456
 457                         break;
 458                 }
 459
 460                 case nir_cf_node_if:
 461                         emit_if(ctx, nir_cf_node_as_if(node));
 462                         break;
 463
 464                 case nir_cf_node_loop:
 465                         emit_loop(ctx, nir_cf_node_as_loop(node));
 466                         break;
 467
 468                 default:
 469                         unreachable("Unknown control flow");
 470                 }
 471         }
 472
 473         return start_block;
 474 }
 475
 476 static int
 477 glsl_type_size(const struct glsl_type *type, bool bindless)
 478 {
 479         return glsl_count_attribute_slots(type, false);
 480 }
 481
 482 static void
 483 bi_optimize_nir(nir_shader *nir)
 484 {
 485         bool progress;
 486         unsigned lower_flrp = 16 | 32 | 64;
 487
 488         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 489         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 490
 491         nir_lower_tex_options lower_tex_options = {
 492                 .lower_txs_lod = true,
 493                 .lower_txp = ~0,
 494                 .lower_tex_without_implicit_lod = true,
 495                 .lower_txd = true,
 496         };
 497
 498         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 499         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 500         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 501
 502         do {
 503                 progress = false;
 504
 505                 NIR_PASS(progress, nir, nir_lower_var_copies);
 506                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 507
 508                 NIR_PASS(progress, nir, nir_copy_prop);
 509                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 510                 NIR_PASS(progress, nir, nir_opt_dce);
 511                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 512                 NIR_PASS(progress, nir, nir_opt_cse);
 513                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 514                 NIR_PASS(progress, nir, nir_opt_algebraic);
 515                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 516
 517                 if (lower_flrp != 0) {
 518                         bool lower_flrp_progress = false;
 519                         NIR_PASS(lower_flrp_progress,
 520                                  nir,
 521                                  nir_lower_flrp,
 522                                  lower_flrp,
 523                                  false /* always_precise */,
 524                                  nir->options->lower_ffma);
 525                         if (lower_flrp_progress) {
 526                                 NIR_PASS(progress, nir,
 527                                          nir_opt_constant_folding);
 528                                 progress = true;
 529                         }
 530
 531                         /* Nothing should rematerialize any flrps, so we only
 532                          * need to do this lowering once.
 533                          */
 534                         lower_flrp = 0;
 535                 }
 536
 537                 NIR_PASS(progress, nir, nir_opt_undef);
 538                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 539                          nir_var_shader_in |
 540                          nir_var_shader_out |
 541                          nir_var_function_temp);
 542         } while (progress);
 543
 544         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 545         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 546         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 547
 548         /* Take us out of SSA */
 549         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 550         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 551 }
 552
 553 void
 554 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 555 {
 556         bi_context *ctx = rzalloc(NULL, bi_context);
 557         ctx->nir = nir;
 558         ctx->stage = nir->info.stage;
 559         ctx->quirks = bifrost_get_quirks(product_id);
 560         list_inithead(&ctx->blocks);
 561
 562         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 563          * (so we don't accidentally duplicate the epilogue since mesa/st has
 564          * messed with our I/O quite a bit already) */
 565
 566         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 567
 568         if (ctx->stage == MESA_SHADER_VERTEX) {
 569                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 570                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 571         }
 572
 573         NIR_PASS_V(nir, nir_split_var_copies);
 574         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 575         NIR_PASS_V(nir, nir_lower_var_copies);
 576         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 577         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 578         NIR_PASS_V(nir, nir_lower_ssbo);
 579
 580         bi_optimize_nir(nir);
 581         nir_print_shader(nir, stdout);
 582
 583         nir_foreach_function(func, nir) {
 584                 if (!func->impl)
 585                         continue;
 586
 587                 ctx->impl = func->impl;
 588                 emit_cf_list(ctx, &func->impl->body);
 589                 break; /* TODO: Multi-function shaders */
 590         }
 591
 592         bi_print_shader(ctx, stdout);
 593         bi_schedule(ctx);
 594
 595         ralloc_free(ctx);
 596 }