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