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
  43 static void
  44 emit_jump(bi_context *ctx, nir_jump_instr *instr)
  45 {
  46         bi_instruction *branch = bi_emit_branch(ctx);
  47
  48         switch (instr->type) {
  49         case nir_jump_break:
  50                 branch->branch.target = ctx->break_block;
  51                 break;
  52         case nir_jump_continue:
  53                 branch->branch.target = ctx->continue_block;
  54                 break;
  55         default:
  56                 unreachable("Unhandled jump type");
  57         }
  58
  59         bi_block_add_successor(ctx->current_block, branch->branch.target);
  60 }
  61
  62 static void
  63 bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
  64 {
  65         bi_instruction ins = {
  66                 .type = BI_LOAD_VAR,
  67                 .load_vary = {
  68                         .load = {
  69                                 .location = nir_intrinsic_base(instr),
  70                                 .channels = instr->num_components,
  71                         },
  72                         .interp_mode = BIFROST_INTERP_DEFAULT, /* TODO */
  73                         .reuse = false, /* TODO */
  74                         .flat = instr->intrinsic != nir_intrinsic_load_interpolated_input
  75                 },
  76                 .dest = bir_dest_index(&instr->dest),
  77                 .dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
  78         };
  79
  80         nir_src *offset = nir_get_io_offset_src(instr);
  81
  82         if (nir_src_is_const(*offset))
  83                 ins.load_vary.load.location += nir_src_as_uint(*offset);
  84         else
  85                 ins.src[0] = bir_src_index(offset);
  86
  87         bi_emit(ctx, ins);
  88 }
  89
  90 static void
  91 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
  92 {
  93
  94         switch (instr->intrinsic) {
  95         case nir_intrinsic_load_barycentric_pixel:
  96                 /* stub */
  97                 break;
  98         case nir_intrinsic_load_interpolated_input:
  99                 bi_emit_ld_vary(ctx, instr);
 100                 break;
 101         default:
 102                 /* todo */
 103                 break;
 104         }
 105 }
 106
 107 static void
 108 emit_instr(bi_context *ctx, struct nir_instr *instr)
 109 {
 110         switch (instr->type) {
 111 #if 0
 112         case nir_instr_type_load_const:
 113                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 114                 break;
 115 #endif
 116
 117         case nir_instr_type_intrinsic:
 118                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 119                 break;
 120
 121 #if 0
 122         case nir_instr_type_alu:
 123                 emit_alu(ctx, nir_instr_as_alu(instr));
 124                 break;
 125
 126         case nir_instr_type_tex:
 127                 emit_tex(ctx, nir_instr_as_tex(instr));
 128                 break;
 129 #endif
 130
 131         case nir_instr_type_jump:
 132                 emit_jump(ctx, nir_instr_as_jump(instr));
 133                 break;
 134
 135         case nir_instr_type_ssa_undef:
 136                 /* Spurious */
 137                 break;
 138
 139         default:
 140                 //unreachable("Unhandled instruction type");
 141                 break;
 142         }
 143 }
 144
 145
 146
 147 static bi_block *
 148 create_empty_block(bi_context *ctx)
 149 {
 150         bi_block *blk = rzalloc(ctx, bi_block);
 151
 152         blk->predecessors = _mesa_set_create(blk,
 153                         _mesa_hash_pointer,
 154                         _mesa_key_pointer_equal);
 155
 156         blk->name = ctx->block_name_count++;
 157
 158         return blk;
 159 }
 160
 161 static void
 162 bi_block_add_successor(bi_block *block, bi_block *successor)
 163 {
 164         assert(block);
 165         assert(successor);
 166
 167         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 168                 if (block->successors[i]) {
 169                        if (block->successors[i] == successor)
 170                                return;
 171                        else
 172                                continue;
 173                 }
 174
 175                 block->successors[i] = successor;
 176                 _mesa_set_add(successor->predecessors, block);
 177                 return;
 178         }
 179
 180         unreachable("Too many successors");
 181 }
 182
 183 static bi_block *
 184 emit_block(bi_context *ctx, nir_block *block)
 185 {
 186         if (ctx->after_block) {
 187                 ctx->current_block = ctx->after_block;
 188                 ctx->after_block = NULL;
 189         } else {
 190                 ctx->current_block = create_empty_block(ctx);
 191         }
 192
 193         list_addtail(&ctx->current_block->link, &ctx->blocks);
 194         list_inithead(&ctx->current_block->instructions);
 195
 196         nir_foreach_instr(instr, block) {
 197                 emit_instr(ctx, instr);
 198                 ++ctx->instruction_count;
 199         }
 200
 201         return ctx->current_block;
 202 }
 203
 204 /* Emits an unconditional branch to the end of the current block, returning a
 205  * pointer so the user can fill in details */
 206
 207 static bi_instruction *
 208 bi_emit_branch(bi_context *ctx)
 209 {
 210         bi_instruction branch = {
 211                 .type = BI_BRANCH,
 212                 .branch = {
 213                         .cond = BI_COND_ALWAYS
 214                 }
 215         };
 216
 217         return bi_emit(ctx, branch);
 218 }
 219
 220 /* Sets a condition for a branch by examing the NIR condition. If we're
 221  * familiar with the condition, we unwrap it to fold it into the branch
 222  * instruction. Otherwise, we consume the condition directly. We
 223  * generally use 1-bit booleans which allows us to use small types for
 224  * the conditions.
 225  */
 226
 227 static void
 228 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 229 {
 230         /* TODO: Try to unwrap instead of always bailing */
 231         branch->src[0] = bir_src_index(cond);
 232         branch->src[1] = BIR_INDEX_ZERO;
 233         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 234         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 235 }
 236
 237 static void
 238 emit_if(bi_context *ctx, nir_if *nif)
 239 {
 240         bi_block *before_block = ctx->current_block;
 241
 242         /* Speculatively emit the branch, but we can't fill it in until later */
 243         bi_instruction *then_branch = bi_emit_branch(ctx);
 244         bi_set_branch_cond(then_branch, &nif->condition, true);
 245
 246         /* Emit the two subblocks. */
 247         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 248         bi_block *end_then_block = ctx->current_block;
 249
 250         /* Emit a jump from the end of the then block to the end of the else */
 251         bi_instruction *then_exit = bi_emit_branch(ctx);
 252
 253         /* Emit second block, and check if it's empty */
 254
 255         int count_in = ctx->instruction_count;
 256         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 257         bi_block *end_else_block = ctx->current_block;
 258         ctx->after_block = create_empty_block(ctx);
 259
 260         /* Now that we have the subblocks emitted, fix up the branches */
 261
 262         assert(then_block);
 263         assert(else_block);
 264
 265         if (ctx->instruction_count == count_in) {
 266                 /* The else block is empty, so don't emit an exit jump */
 267                 bi_remove_instruction(then_exit);
 268                 then_branch->branch.target = ctx->after_block;
 269         } else {
 270                 then_branch->branch.target = else_block;
 271                 then_exit->branch.target = ctx->after_block;
 272                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 273         }
 274
 275         /* Wire up the successors */
 276
 277         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 278
 279         bi_block_add_successor(before_block, then_block); /* fallthrough */
 280         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 281 }
 282
 283 static void
 284 emit_loop(bi_context *ctx, nir_loop *nloop)
 285 {
 286         /* Remember where we are */
 287         bi_block *start_block = ctx->current_block;
 288
 289         bi_block *saved_break = ctx->break_block;
 290         bi_block *saved_continue = ctx->continue_block;
 291
 292         ctx->continue_block = create_empty_block(ctx);
 293         ctx->break_block = create_empty_block(ctx);
 294         ctx->after_block = ctx->continue_block;
 295
 296         /* Emit the body itself */
 297         emit_cf_list(ctx, &nloop->body);
 298
 299         /* Branch back to loop back */
 300         bi_instruction *br_back = bi_emit_branch(ctx);
 301         br_back->branch.target = ctx->continue_block;
 302         bi_block_add_successor(start_block, ctx->continue_block);
 303         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 304
 305         ctx->after_block = ctx->break_block;
 306
 307         /* Pop off */
 308         ctx->break_block = saved_break;
 309         ctx->continue_block = saved_continue;
 310         ++ctx->loop_count;
 311 }
 312
 313 static bi_block *
 314 emit_cf_list(bi_context *ctx, struct exec_list *list)
 315 {
 316         bi_block *start_block = NULL;
 317
 318         foreach_list_typed(nir_cf_node, node, node, list) {
 319                 switch (node->type) {
 320                 case nir_cf_node_block: {
 321                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 322
 323                         if (!start_block)
 324                                 start_block = block;
 325
 326                         break;
 327                 }
 328
 329                 case nir_cf_node_if:
 330                         emit_if(ctx, nir_cf_node_as_if(node));
 331                         break;
 332
 333                 case nir_cf_node_loop:
 334                         emit_loop(ctx, nir_cf_node_as_loop(node));
 335                         break;
 336
 337                 default:
 338                         unreachable("Unknown control flow");
 339                 }
 340         }
 341
 342         return start_block;
 343 }
 344
 345 static int
 346 glsl_type_size(const struct glsl_type *type, bool bindless)
 347 {
 348         return glsl_count_attribute_slots(type, false);
 349 }
 350
 351 static void
 352 bi_optimize_nir(nir_shader *nir)
 353 {
 354         bool progress;
 355         unsigned lower_flrp = 16 | 32 | 64;
 356
 357         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 358         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 359
 360         nir_lower_tex_options lower_tex_options = {
 361                 .lower_txs_lod = true,
 362                 .lower_txp = ~0,
 363                 .lower_tex_without_implicit_lod = true,
 364                 .lower_txd = true,
 365         };
 366
 367         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 368
 369         do {
 370                 progress = false;
 371
 372                 NIR_PASS(progress, nir, nir_lower_var_copies);
 373                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 374
 375                 NIR_PASS(progress, nir, nir_copy_prop);
 376                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 377                 NIR_PASS(progress, nir, nir_opt_dce);
 378                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 379                 NIR_PASS(progress, nir, nir_opt_cse);
 380                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 381                 NIR_PASS(progress, nir, nir_opt_algebraic);
 382                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 383
 384                 if (lower_flrp != 0) {
 385                         bool lower_flrp_progress = false;
 386                         NIR_PASS(lower_flrp_progress,
 387                                  nir,
 388                                  nir_lower_flrp,
 389                                  lower_flrp,
 390                                  false /* always_precise */,
 391                                  nir->options->lower_ffma);
 392                         if (lower_flrp_progress) {
 393                                 NIR_PASS(progress, nir,
 394                                          nir_opt_constant_folding);
 395                                 progress = true;
 396                         }
 397
 398                         /* Nothing should rematerialize any flrps, so we only
 399                          * need to do this lowering once.
 400                          */
 401                         lower_flrp = 0;
 402                 }
 403
 404                 NIR_PASS(progress, nir, nir_opt_undef);
 405                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 406                          nir_var_shader_in |
 407                          nir_var_shader_out |
 408                          nir_var_function_temp);
 409         } while (progress);
 410
 411         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 412
 413         /* Take us out of SSA */
 414         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 415         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 416 }
 417
 418 void
 419 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 420 {
 421         bi_context *ctx = rzalloc(NULL, bi_context);
 422         ctx->nir = nir;
 423         ctx->stage = nir->info.stage;
 424         ctx->quirks = bifrost_get_quirks(product_id);
 425         list_inithead(&ctx->blocks);
 426
 427         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 428          * (so we don't accidentally duplicate the epilogue since mesa/st has
 429          * messed with our I/O quite a bit already) */
 430
 431         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 432
 433         if (ctx->stage == MESA_SHADER_VERTEX) {
 434                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 435                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 436         }
 437
 438         NIR_PASS_V(nir, nir_split_var_copies);
 439         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 440         NIR_PASS_V(nir, nir_lower_var_copies);
 441         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 442         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 443         NIR_PASS_V(nir, nir_lower_ssbo);
 444
 445         /* We have to lower ALU to scalar ourselves since viewport
 446          * transformations produce vector ops */
 447         NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL, NULL);
 448
 449         bi_optimize_nir(nir);
 450         nir_print_shader(nir, stdout);
 451
 452         nir_foreach_function(func, nir) {
 453                 if (!func->impl)
 454                         continue;
 455
 456                 emit_cf_list(ctx, &func->impl->body);
 457                 break; /* TODO: Multi-function shaders */
 458         }
 459
 460         bi_print_shader(ctx, stdout);
 461
 462         ralloc_free(ctx);
 463 }