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 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
  93 {
  94
  95         switch (instr->intrinsic) {
  96         case nir_intrinsic_load_barycentric_pixel:
  97                 /* stub */
  98                 break;
  99         case nir_intrinsic_load_interpolated_input:
 100                 bi_emit_ld_vary(ctx, instr);
 101                 break;
 102         default:
 103                 /* todo */
 104                 break;
 105         }
 106 }
 107
 108 static void
 109 emit_instr(bi_context *ctx, struct nir_instr *instr)
 110 {
 111         switch (instr->type) {
 112 #if 0
 113         case nir_instr_type_load_const:
 114                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 115                 break;
 116 #endif
 117
 118         case nir_instr_type_intrinsic:
 119                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 120                 break;
 121
 122 #if 0
 123         case nir_instr_type_alu:
 124                 emit_alu(ctx, nir_instr_as_alu(instr));
 125                 break;
 126
 127         case nir_instr_type_tex:
 128                 emit_tex(ctx, nir_instr_as_tex(instr));
 129                 break;
 130 #endif
 131
 132         case nir_instr_type_jump:
 133                 emit_jump(ctx, nir_instr_as_jump(instr));
 134                 break;
 135
 136         case nir_instr_type_ssa_undef:
 137                 /* Spurious */
 138                 break;
 139
 140         default:
 141                 //unreachable("Unhandled instruction type");
 142                 break;
 143         }
 144 }
 145
 146
 147
 148 static bi_block *
 149 create_empty_block(bi_context *ctx)
 150 {
 151         bi_block *blk = rzalloc(ctx, bi_block);
 152
 153         blk->predecessors = _mesa_set_create(blk,
 154                         _mesa_hash_pointer,
 155                         _mesa_key_pointer_equal);
 156
 157         blk->name = ctx->block_name_count++;
 158
 159         return blk;
 160 }
 161
 162 static void
 163 bi_block_add_successor(bi_block *block, bi_block *successor)
 164 {
 165         assert(block);
 166         assert(successor);
 167
 168         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 169                 if (block->successors[i]) {
 170                        if (block->successors[i] == successor)
 171                                return;
 172                        else
 173                                continue;
 174                 }
 175
 176                 block->successors[i] = successor;
 177                 _mesa_set_add(successor->predecessors, block);
 178                 return;
 179         }
 180
 181         unreachable("Too many successors");
 182 }
 183
 184 static void
 185 bi_schedule_barrier(bi_context *ctx)
 186 {
 187         bi_block *temp = ctx->after_block;
 188         ctx->after_block = create_empty_block(ctx);
 189         list_addtail(&ctx->after_block->link, &ctx->blocks);
 190         list_inithead(&ctx->after_block->instructions);
 191         bi_block_add_successor(ctx->current_block, ctx->after_block);
 192         ctx->current_block = ctx->after_block;
 193         ctx->after_block = temp;
 194 }
 195
 196 static bi_block *
 197 emit_block(bi_context *ctx, nir_block *block)
 198 {
 199         if (ctx->after_block) {
 200                 ctx->current_block = ctx->after_block;
 201                 ctx->after_block = NULL;
 202         } else {
 203                 ctx->current_block = create_empty_block(ctx);
 204         }
 205
 206         list_addtail(&ctx->current_block->link, &ctx->blocks);
 207         list_inithead(&ctx->current_block->instructions);
 208
 209         nir_foreach_instr(instr, block) {
 210                 emit_instr(ctx, instr);
 211                 ++ctx->instruction_count;
 212         }
 213
 214         return ctx->current_block;
 215 }
 216
 217 /* Emits an unconditional branch to the end of the current block, returning a
 218  * pointer so the user can fill in details */
 219
 220 static bi_instruction *
 221 bi_emit_branch(bi_context *ctx)
 222 {
 223         bi_instruction branch = {
 224                 .type = BI_BRANCH,
 225                 .branch = {
 226                         .cond = BI_COND_ALWAYS
 227                 }
 228         };
 229
 230         return bi_emit(ctx, branch);
 231 }
 232
 233 /* Sets a condition for a branch by examing the NIR condition. If we're
 234  * familiar with the condition, we unwrap it to fold it into the branch
 235  * instruction. Otherwise, we consume the condition directly. We
 236  * generally use 1-bit booleans which allows us to use small types for
 237  * the conditions.
 238  */
 239
 240 static void
 241 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 242 {
 243         /* TODO: Try to unwrap instead of always bailing */
 244         branch->src[0] = bir_src_index(cond);
 245         branch->src[1] = BIR_INDEX_ZERO;
 246         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 247         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 248 }
 249
 250 static void
 251 emit_if(bi_context *ctx, nir_if *nif)
 252 {
 253         bi_block *before_block = ctx->current_block;
 254
 255         /* Speculatively emit the branch, but we can't fill it in until later */
 256         bi_instruction *then_branch = bi_emit_branch(ctx);
 257         bi_set_branch_cond(then_branch, &nif->condition, true);
 258
 259         /* Emit the two subblocks. */
 260         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 261         bi_block *end_then_block = ctx->current_block;
 262
 263         /* Emit a jump from the end of the then block to the end of the else */
 264         bi_instruction *then_exit = bi_emit_branch(ctx);
 265
 266         /* Emit second block, and check if it's empty */
 267
 268         int count_in = ctx->instruction_count;
 269         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 270         bi_block *end_else_block = ctx->current_block;
 271         ctx->after_block = create_empty_block(ctx);
 272
 273         /* Now that we have the subblocks emitted, fix up the branches */
 274
 275         assert(then_block);
 276         assert(else_block);
 277
 278         if (ctx->instruction_count == count_in) {
 279                 /* The else block is empty, so don't emit an exit jump */
 280                 bi_remove_instruction(then_exit);
 281                 then_branch->branch.target = ctx->after_block;
 282         } else {
 283                 then_branch->branch.target = else_block;
 284                 then_exit->branch.target = ctx->after_block;
 285                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 286         }
 287
 288         /* Wire up the successors */
 289
 290         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 291
 292         bi_block_add_successor(before_block, then_block); /* fallthrough */
 293         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 294 }
 295
 296 static void
 297 emit_loop(bi_context *ctx, nir_loop *nloop)
 298 {
 299         /* Remember where we are */
 300         bi_block *start_block = ctx->current_block;
 301
 302         bi_block *saved_break = ctx->break_block;
 303         bi_block *saved_continue = ctx->continue_block;
 304
 305         ctx->continue_block = create_empty_block(ctx);
 306         ctx->break_block = create_empty_block(ctx);
 307         ctx->after_block = ctx->continue_block;
 308
 309         /* Emit the body itself */
 310         emit_cf_list(ctx, &nloop->body);
 311
 312         /* Branch back to loop back */
 313         bi_instruction *br_back = bi_emit_branch(ctx);
 314         br_back->branch.target = ctx->continue_block;
 315         bi_block_add_successor(start_block, ctx->continue_block);
 316         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 317
 318         ctx->after_block = ctx->break_block;
 319
 320         /* Pop off */
 321         ctx->break_block = saved_break;
 322         ctx->continue_block = saved_continue;
 323         ++ctx->loop_count;
 324 }
 325
 326 static bi_block *
 327 emit_cf_list(bi_context *ctx, struct exec_list *list)
 328 {
 329         bi_block *start_block = NULL;
 330
 331         foreach_list_typed(nir_cf_node, node, node, list) {
 332                 switch (node->type) {
 333                 case nir_cf_node_block: {
 334                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 335
 336                         if (!start_block)
 337                                 start_block = block;
 338
 339                         break;
 340                 }
 341
 342                 case nir_cf_node_if:
 343                         emit_if(ctx, nir_cf_node_as_if(node));
 344                         break;
 345
 346                 case nir_cf_node_loop:
 347                         emit_loop(ctx, nir_cf_node_as_loop(node));
 348                         break;
 349
 350                 default:
 351                         unreachable("Unknown control flow");
 352                 }
 353         }
 354
 355         return start_block;
 356 }
 357
 358 static int
 359 glsl_type_size(const struct glsl_type *type, bool bindless)
 360 {
 361         return glsl_count_attribute_slots(type, false);
 362 }
 363
 364 static void
 365 bi_optimize_nir(nir_shader *nir)
 366 {
 367         bool progress;
 368         unsigned lower_flrp = 16 | 32 | 64;
 369
 370         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 371         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 372
 373         nir_lower_tex_options lower_tex_options = {
 374                 .lower_txs_lod = true,
 375                 .lower_txp = ~0,
 376                 .lower_tex_without_implicit_lod = true,
 377                 .lower_txd = true,
 378         };
 379
 380         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 381
 382         do {
 383                 progress = false;
 384
 385                 NIR_PASS(progress, nir, nir_lower_var_copies);
 386                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 387
 388                 NIR_PASS(progress, nir, nir_copy_prop);
 389                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 390                 NIR_PASS(progress, nir, nir_opt_dce);
 391                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 392                 NIR_PASS(progress, nir, nir_opt_cse);
 393                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 394                 NIR_PASS(progress, nir, nir_opt_algebraic);
 395                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 396
 397                 if (lower_flrp != 0) {
 398                         bool lower_flrp_progress = false;
 399                         NIR_PASS(lower_flrp_progress,
 400                                  nir,
 401                                  nir_lower_flrp,
 402                                  lower_flrp,
 403                                  false /* always_precise */,
 404                                  nir->options->lower_ffma);
 405                         if (lower_flrp_progress) {
 406                                 NIR_PASS(progress, nir,
 407                                          nir_opt_constant_folding);
 408                                 progress = true;
 409                         }
 410
 411                         /* Nothing should rematerialize any flrps, so we only
 412                          * need to do this lowering once.
 413                          */
 414                         lower_flrp = 0;
 415                 }
 416
 417                 NIR_PASS(progress, nir, nir_opt_undef);
 418                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 419                          nir_var_shader_in |
 420                          nir_var_shader_out |
 421                          nir_var_function_temp);
 422         } while (progress);
 423
 424         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 425
 426         /* Take us out of SSA */
 427         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 428         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 429 }
 430
 431 void
 432 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 433 {
 434         bi_context *ctx = rzalloc(NULL, bi_context);
 435         ctx->nir = nir;
 436         ctx->stage = nir->info.stage;
 437         ctx->quirks = bifrost_get_quirks(product_id);
 438         list_inithead(&ctx->blocks);
 439
 440         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 441          * (so we don't accidentally duplicate the epilogue since mesa/st has
 442          * messed with our I/O quite a bit already) */
 443
 444         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 445
 446         if (ctx->stage == MESA_SHADER_VERTEX) {
 447                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 448                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 449         }
 450
 451         NIR_PASS_V(nir, nir_split_var_copies);
 452         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 453         NIR_PASS_V(nir, nir_lower_var_copies);
 454         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 455         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 456         NIR_PASS_V(nir, nir_lower_ssbo);
 457
 458         /* We have to lower ALU to scalar ourselves since viewport
 459          * transformations produce vector ops */
 460         NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL, NULL);
 461
 462         bi_optimize_nir(nir);
 463         nir_print_shader(nir, stdout);
 464
 465         nir_foreach_function(func, nir) {
 466                 if (!func->impl)
 467                         continue;
 468
 469                 emit_cf_list(ctx, &func->impl->body);
 470                 break; /* TODO: Multi-function shaders */
 471         }
 472
 473         bi_print_shader(ctx, stdout);
 474
 475         ralloc_free(ctx);
 476 }