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