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