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