src/freedreno/ir3/ir3_nir.c

   1 /*
   2  * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
   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:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27
  28 #include "util/debug.h"
  29 #include "util/u_math.h"
  30
  31 #include "ir3_nir.h"
  32 #include "ir3_compiler.h"
  33 #include "ir3_shader.h"
  34
  35 static void ir3_setup_const_state(struct ir3_shader *shader, nir_shader *nir);
  36
  37 static const nir_shader_compiler_options options = {
  38                 .lower_fpow = true,
  39                 .lower_scmp = true,
  40                 .lower_flrp16 = true,
  41                 .lower_flrp32 = true,
  42                 .lower_flrp64 = true,
  43                 .lower_ffract = true,
  44                 .lower_fmod = true,
  45                 .lower_fdiv = true,
  46                 .lower_isign = true,
  47                 .lower_ldexp = true,
  48                 .lower_uadd_carry = true,
  49                 .lower_usub_borrow = true,
  50                 .lower_mul_high = true,
  51                 .lower_mul_2x32_64 = true,
  52                 .fuse_ffma = true,
  53                 .vertex_id_zero_based = true,
  54                 .lower_extract_byte = true,
  55                 .lower_extract_word = true,
  56                 .lower_all_io_to_elements = true,
  57                 .lower_helper_invocation = true,
  58                 .lower_bitfield_insert_to_shifts = true,
  59                 .lower_bitfield_extract_to_shifts = true,
  60                 .lower_pack_half_2x16 = true,
  61                 .lower_pack_snorm_4x8 = true,
  62                 .lower_pack_snorm_2x16 = true,
  63                 .lower_pack_unorm_4x8 = true,
  64                 .lower_pack_unorm_2x16 = true,
  65                 .lower_unpack_half_2x16 = true,
  66                 .lower_unpack_snorm_4x8 = true,
  67                 .lower_unpack_snorm_2x16 = true,
  68                 .lower_unpack_unorm_4x8 = true,
  69                 .lower_unpack_unorm_2x16 = true,
  70                 .lower_pack_split = true,
  71                 .use_interpolated_input_intrinsics = true,
  72                 .lower_rotate = true,
  73                 .lower_to_scalar = true,
  74                 .has_imul24 = true,
  75 };
  76
  77 /* we don't want to lower vertex_id to _zero_based on newer gpus: */
  78 static const nir_shader_compiler_options options_a6xx = {
  79                 .lower_fpow = true,
  80                 .lower_scmp = true,
  81                 .lower_flrp16 = true,
  82                 .lower_flrp32 = true,
  83                 .lower_flrp64 = true,
  84                 .lower_ffract = true,
  85                 .lower_fmod = true,
  86                 .lower_fdiv = true,
  87                 .lower_isign = true,
  88                 .lower_ldexp = true,
  89                 .lower_uadd_carry = true,
  90                 .lower_usub_borrow = true,
  91                 .lower_mul_high = true,
  92                 .lower_mul_2x32_64 = true,
  93                 .fuse_ffma = true,
  94                 .vertex_id_zero_based = false,
  95                 .lower_extract_byte = true,
  96                 .lower_extract_word = true,
  97                 .lower_all_io_to_elements = true,
  98                 .lower_helper_invocation = true,
  99                 .lower_bitfield_insert_to_shifts = true,
 100                 .lower_bitfield_extract_to_shifts = true,
 101                 .lower_pack_half_2x16 = true,
 102                 .lower_pack_snorm_4x8 = true,
 103                 .lower_pack_snorm_2x16 = true,
 104                 .lower_pack_unorm_4x8 = true,
 105                 .lower_pack_unorm_2x16 = true,
 106                 .lower_unpack_half_2x16 = true,
 107                 .lower_unpack_snorm_4x8 = true,
 108                 .lower_unpack_snorm_2x16 = true,
 109                 .lower_unpack_unorm_4x8 = true,
 110                 .lower_unpack_unorm_2x16 = true,
 111                 .lower_pack_split = true,
 112                 .use_interpolated_input_intrinsics = true,
 113                 .lower_rotate = true,
 114                 .vectorize_io = true,
 115                 .lower_to_scalar = true,
 116                 .has_imul24 = true,
 117                 .max_unroll_iterations = 32,
 118 };
 119
 120 const nir_shader_compiler_options *
 121 ir3_get_compiler_options(struct ir3_compiler *compiler)
 122 {
 123         if (compiler->gpu_id >= 600)
 124                 return &options_a6xx;
 125         return &options;
 126 }
 127
 128 /* for given shader key, are any steps handled in nir? */
 129 bool
 130 ir3_key_lowers_nir(const struct ir3_shader_key *key)
 131 {
 132         return key->fsaturate_s | key->fsaturate_t | key->fsaturate_r |
 133                         key->vsaturate_s | key->vsaturate_t | key->vsaturate_r |
 134                         key->ucp_enables | key->color_two_side |
 135                         key->fclamp_color | key->vclamp_color |
 136                         key->tessellation | key->has_gs;
 137 }
 138
 139 #define OPT(nir, pass, ...) ({                             \
 140    bool this_progress = false;                             \
 141    NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
 142    this_progress;                                          \
 143 })
 144
 145 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
 146
 147 static void
 148 ir3_optimize_loop(nir_shader *s)
 149 {
 150         bool progress;
 151         unsigned lower_flrp =
 152                 (s->options->lower_flrp16 ? 16 : 0) |
 153                 (s->options->lower_flrp32 ? 32 : 0) |
 154                 (s->options->lower_flrp64 ? 64 : 0);
 155
 156         do {
 157                 progress = false;
 158
 159                 OPT_V(s, nir_lower_vars_to_ssa);
 160                 progress |= OPT(s, nir_opt_copy_prop_vars);
 161                 progress |= OPT(s, nir_opt_dead_write_vars);
 162                 progress |= OPT(s, nir_lower_alu_to_scalar, NULL, NULL);
 163                 progress |= OPT(s, nir_lower_phis_to_scalar);
 164
 165                 progress |= OPT(s, nir_copy_prop);
 166                 progress |= OPT(s, nir_opt_dce);
 167                 progress |= OPT(s, nir_opt_cse);
 168                 static int gcm = -1;
 169                 if (gcm == -1)
 170                         gcm = env_var_as_unsigned("GCM", 0);
 171                 if (gcm == 1)
 172                         progress |= OPT(s, nir_opt_gcm, true);
 173                 else if (gcm == 2)
 174                         progress |= OPT(s, nir_opt_gcm, false);
 175                 progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
 176                 progress |= OPT(s, nir_opt_intrinsics);
 177                 progress |= OPT(s, nir_opt_algebraic);
 178                 progress |= OPT(s, nir_lower_alu);
 179                 progress |= OPT(s, nir_lower_pack);
 180                 progress |= OPT(s, nir_opt_constant_folding);
 181
 182                 if (lower_flrp != 0) {
 183                         if (OPT(s, nir_lower_flrp,
 184                                         lower_flrp,
 185                                         false /* always_precise */,
 186                                         s->options->lower_ffma)) {
 187                                 OPT(s, nir_opt_constant_folding);
 188                                 progress = true;
 189                         }
 190
 191                         /* Nothing should rematerialize any flrps, so we only
 192                          * need to do this lowering once.
 193                          */
 194                         lower_flrp = 0;
 195                 }
 196
 197                 progress |= OPT(s, nir_opt_dead_cf);
 198                 if (OPT(s, nir_opt_trivial_continues)) {
 199                         progress |= true;
 200                         /* If nir_opt_trivial_continues makes progress, then we need to clean
 201                          * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
 202                          * to make progress.
 203                          */
 204                         OPT(s, nir_copy_prop);
 205                         OPT(s, nir_opt_dce);
 206                 }
 207                 progress |= OPT(s, nir_opt_if, false);
 208                 progress |= OPT(s, nir_opt_remove_phis);
 209                 progress |= OPT(s, nir_opt_undef);
 210         } while (progress);
 211 }
 212
 213 static bool
 214 should_split_wrmask(const nir_instr *instr, const void *data)
 215 {
 216         nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 217
 218         switch (intr->intrinsic) {
 219         case nir_intrinsic_store_ssbo:
 220         case nir_intrinsic_store_shared:
 221         case nir_intrinsic_store_global:
 222                 return true;
 223         default:
 224                 return false;
 225         }
 226 }
 227
 228 void
 229 ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 230                 const struct ir3_shader_key *key)
 231 {
 232         struct nir_lower_tex_options tex_options = {
 233                         .lower_rect = 0,
 234                         .lower_tg4_offsets = true,
 235         };
 236
 237         if (key && (key->has_gs || key->tessellation)) {
 238                 switch (shader->type) {
 239                 case MESA_SHADER_VERTEX:
 240                         NIR_PASS_V(s, ir3_nir_lower_to_explicit_output, shader, key->tessellation);
 241                         break;
 242                 case MESA_SHADER_TESS_CTRL:
 243                         NIR_PASS_V(s, ir3_nir_lower_tess_ctrl, shader, key->tessellation);
 244                         NIR_PASS_V(s, ir3_nir_lower_to_explicit_input);
 245                         break;
 246                 case MESA_SHADER_TESS_EVAL:
 247                         NIR_PASS_V(s, ir3_nir_lower_tess_eval, key->tessellation);
 248                         if (key->has_gs)
 249                                 NIR_PASS_V(s, ir3_nir_lower_to_explicit_output, shader, key->tessellation);
 250                         break;
 251                 case MESA_SHADER_GEOMETRY:
 252                         NIR_PASS_V(s, ir3_nir_lower_to_explicit_input);
 253                         break;
 254                 default:
 255                         break;
 256                 }
 257         }
 258
 259         if (key) {
 260                 switch (shader->type) {
 261                 case MESA_SHADER_FRAGMENT:
 262                         tex_options.saturate_s = key->fsaturate_s;
 263                         tex_options.saturate_t = key->fsaturate_t;
 264                         tex_options.saturate_r = key->fsaturate_r;
 265                         break;
 266                 case MESA_SHADER_VERTEX:
 267                         tex_options.saturate_s = key->vsaturate_s;
 268                         tex_options.saturate_t = key->vsaturate_t;
 269                         tex_options.saturate_r = key->vsaturate_r;
 270                         break;
 271                 default:
 272                         /* TODO */
 273                         break;
 274                 }
 275         }
 276
 277         if (shader->compiler->gpu_id >= 400) {
 278                 /* a4xx seems to have *no* sam.p */
 279                 tex_options.lower_txp = ~0;  /* lower all txp */
 280         } else {
 281                 /* a3xx just needs to avoid sam.p for 3d tex */
 282                 tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
 283         }
 284
 285         if (ir3_shader_debug & IR3_DBG_DISASM) {
 286                 debug_printf("----------------------\n");
 287                 nir_print_shader(s, stdout);
 288                 debug_printf("----------------------\n");
 289         }
 290
 291         OPT_V(s, nir_lower_regs_to_ssa);
 292         OPT_V(s, nir_lower_wrmasks, should_split_wrmask, s);
 293
 294         if (key) {
 295                 if (s->info.stage == MESA_SHADER_VERTEX) {
 296                         OPT_V(s, nir_lower_clip_vs, key->ucp_enables, false, false, NULL);
 297                         if (key->vclamp_color)
 298                                 OPT_V(s, nir_lower_clamp_color_outputs);
 299                 } else if (s->info.stage == MESA_SHADER_FRAGMENT) {
 300                         OPT_V(s, nir_lower_clip_fs, key->ucp_enables, false);
 301                         if (key->fclamp_color)
 302                                 OPT_V(s, nir_lower_clamp_color_outputs);
 303                 }
 304                 if (key->color_two_side) {
 305                         OPT_V(s, nir_lower_two_sided_color);
 306                 }
 307         } else {
 308                 /* only want to do this the first time (when key is null)
 309                  * and not again on any potential 2nd variant lowering pass:
 310                  */
 311                 OPT_V(s, ir3_nir_apply_trig_workarounds);
 312
 313                 /* This wouldn't hurt to run multiple times, but there is
 314                  * no need to:
 315                  */
 316                 if (shader->type == MESA_SHADER_FRAGMENT)
 317                         OPT_V(s, nir_lower_fb_read);
 318         }
 319
 320         OPT_V(s, nir_lower_tex, &tex_options);
 321         OPT_V(s, nir_lower_load_const_to_scalar);
 322         if (shader->compiler->gpu_id < 500)
 323                 OPT_V(s, ir3_nir_lower_tg4_to_tex);
 324
 325         ir3_optimize_loop(s);
 326
 327         /* do ubo load and idiv lowering after first opt loop to get a chance to
 328          * propagate constants for divide by immed power-of-two and constant ubo
 329          * block/offsets:
 330          *
 331          * NOTE that UBO analysis pass should only be done once, before variants
 332          */
 333         const bool ubo_progress = !key && OPT(s, ir3_nir_analyze_ubo_ranges, shader);
 334         const bool idiv_progress = OPT(s, nir_lower_idiv, nir_lower_idiv_fast);
 335         /* UBO offset lowering has to come after we've decided what will be left as load_ubo */
 336         OPT_V(s, ir3_nir_lower_io_offsets, shader->compiler->gpu_id);
 337
 338         if (ubo_progress || idiv_progress)
 339                 ir3_optimize_loop(s);
 340
 341         /* Do late algebraic optimization to turn add(a, neg(b)) back into
 342         * subs, then the mandatory cleanup after algebraic.  Note that it may
 343         * produce fnegs, and if so then we need to keep running to squash
 344         * fneg(fneg(a)).
 345         */
 346         bool more_late_algebraic = true;
 347         while (more_late_algebraic) {
 348                 more_late_algebraic = OPT(s, nir_opt_algebraic_late);
 349                 OPT_V(s, nir_opt_constant_folding);
 350                 OPT_V(s, nir_copy_prop);
 351                 OPT_V(s, nir_opt_dce);
 352                 OPT_V(s, nir_opt_cse);
 353         }
 354
 355         OPT_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
 356
 357         OPT_V(s, nir_opt_sink, nir_move_const_undef);
 358
 359         if (ir3_shader_debug & IR3_DBG_DISASM) {
 360                 debug_printf("----------------------\n");
 361                 nir_print_shader(s, stdout);
 362                 debug_printf("----------------------\n");
 363         }
 364
 365         nir_sweep(s);
 366
 367         /* The first time thru, when not creating variant, do the one-time
 368          * const_state layout setup.  This should be done after ubo range
 369          * analysis.
 370          */
 371         if (!key) {
 372                 ir3_setup_const_state(shader, s);
 373         }
 374 }
 375
 376 static void
 377 ir3_nir_scan_driver_consts(nir_shader *shader,
 378                 struct ir3_const_state *layout)
 379 {
 380         nir_foreach_function (function, shader) {
 381                 if (!function->impl)
 382                         continue;
 383
 384                 nir_foreach_block (block, function->impl) {
 385                         nir_foreach_instr (instr, block) {
 386                                 if (instr->type != nir_instr_type_intrinsic)
 387                                         continue;
 388
 389                                 nir_intrinsic_instr *intr =
 390                                         nir_instr_as_intrinsic(instr);
 391                                 unsigned idx;
 392
 393                                 switch (intr->intrinsic) {
 394                                 case nir_intrinsic_get_buffer_size:
 395                                         idx = nir_src_as_uint(intr->src[0]);
 396                                         if (layout->ssbo_size.mask & (1 << idx))
 397                                                 break;
 398                                         layout->ssbo_size.mask |= (1 << idx);
 399                                         layout->ssbo_size.off[idx] =
 400                                                 layout->ssbo_size.count;
 401                                         layout->ssbo_size.count += 1; /* one const per */
 402                                         break;
 403                                 case nir_intrinsic_image_atomic_add:
 404                                 case nir_intrinsic_image_atomic_imin:
 405                                 case nir_intrinsic_image_atomic_umin:
 406                                 case nir_intrinsic_image_atomic_imax:
 407                                 case nir_intrinsic_image_atomic_umax:
 408                                 case nir_intrinsic_image_atomic_and:
 409                                 case nir_intrinsic_image_atomic_or:
 410                                 case nir_intrinsic_image_atomic_xor:
 411                                 case nir_intrinsic_image_atomic_exchange:
 412                                 case nir_intrinsic_image_atomic_comp_swap:
 413                                 case nir_intrinsic_image_store:
 414                                 case nir_intrinsic_image_size:
 415                                         idx = nir_src_as_uint(intr->src[0]);
 416                                         if (layout->image_dims.mask & (1 << idx))
 417                                                 break;
 418                                         layout->image_dims.mask |= (1 << idx);
 419                                         layout->image_dims.off[idx] =
 420                                                 layout->image_dims.count;
 421                                         layout->image_dims.count += 3; /* three const per */
 422                                         break;
 423                                 case nir_intrinsic_load_base_vertex:
 424                                 case nir_intrinsic_load_first_vertex:
 425                                         layout->num_driver_params =
 426                                                 MAX2(layout->num_driver_params, IR3_DP_VTXID_BASE + 1);
 427                                         break;
 428                                 case nir_intrinsic_load_base_instance:
 429                                         layout->num_driver_params =
 430                                                 MAX2(layout->num_driver_params, IR3_DP_INSTID_BASE + 1);
 431                                         break;
 432                                 case nir_intrinsic_load_user_clip_plane:
 433                                         layout->num_driver_params =
 434                                                 MAX2(layout->num_driver_params, IR3_DP_UCP7_W + 1);
 435                                         break;
 436                                 case nir_intrinsic_load_num_work_groups:
 437                                         layout->num_driver_params =
 438                                                 MAX2(layout->num_driver_params, IR3_DP_NUM_WORK_GROUPS_Z + 1);
 439                                         break;
 440                                 case nir_intrinsic_load_local_group_size:
 441                                         layout->num_driver_params =
 442                                                 MAX2(layout->num_driver_params, IR3_DP_LOCAL_GROUP_SIZE_Z + 1);
 443                                         break;
 444                                 default:
 445                                         break;
 446                                 }
 447                         }
 448                 }
 449         }
 450 }
 451
 452 static void
 453 ir3_setup_const_state(struct ir3_shader *shader, nir_shader *nir)
 454 {
 455         struct ir3_compiler *compiler = shader->compiler;
 456         struct ir3_const_state *const_state = &shader->const_state;
 457
 458         memset(&const_state->offsets, ~0, sizeof(const_state->offsets));
 459
 460         ir3_nir_scan_driver_consts(nir, const_state);
 461
 462         if ((compiler->gpu_id < 500) &&
 463                         (shader->stream_output.num_outputs > 0)) {
 464                 const_state->num_driver_params =
 465                         MAX2(const_state->num_driver_params, IR3_DP_VTXCNT_MAX + 1);
 466         }
 467
 468         /* On a6xx, we use UBO descriptors and LDC instead of UBO pointers in the
 469          * constbuf.
 470          */
 471         if (compiler->gpu_id >= 600)
 472                 shader->num_ubos = nir->info.num_ubos;
 473         else
 474                 const_state->num_ubos = nir->info.num_ubos;
 475
 476         /* num_driver_params is scalar, align to vec4: */
 477         const_state->num_driver_params = align(const_state->num_driver_params, 4);
 478
 479         debug_assert((shader->ubo_state.size % 16) == 0);
 480         unsigned constoff = align(shader->ubo_state.size / 16, 8);
 481         unsigned ptrsz = ir3_pointer_size(compiler);
 482
 483         if (const_state->num_ubos > 0) {
 484                 const_state->offsets.ubo = constoff;
 485                 constoff += align(const_state->num_ubos * ptrsz, 4) / 4;
 486         }
 487
 488         if (const_state->ssbo_size.count > 0) {
 489                 unsigned cnt = const_state->ssbo_size.count;
 490                 const_state->offsets.ssbo_sizes = constoff;
 491                 constoff += align(cnt, 4) / 4;
 492         }
 493
 494         if (const_state->image_dims.count > 0) {
 495                 unsigned cnt = const_state->image_dims.count;
 496                 const_state->offsets.image_dims = constoff;
 497                 constoff += align(cnt, 4) / 4;
 498         }
 499
 500         if (const_state->num_driver_params > 0)
 501                 const_state->offsets.driver_param = constoff;
 502         constoff += const_state->num_driver_params / 4;
 503
 504         if ((shader->type == MESA_SHADER_VERTEX) &&
 505                         (compiler->gpu_id < 500) &&
 506                         shader->stream_output.num_outputs > 0) {
 507                 const_state->offsets.tfbo = constoff;
 508                 constoff += align(IR3_MAX_SO_BUFFERS * ptrsz, 4) / 4;
 509         }
 510
 511         switch (shader->type) {
 512         case MESA_SHADER_VERTEX:
 513                 const_state->offsets.primitive_param = constoff;
 514                 constoff += 1;
 515                 break;
 516         case MESA_SHADER_TESS_CTRL:
 517         case MESA_SHADER_TESS_EVAL:
 518                 constoff = align(constoff - 1, 4) + 3;
 519                 const_state->offsets.primitive_param = constoff;
 520                 const_state->offsets.primitive_map = constoff + 5;
 521                 constoff += 5 + DIV_ROUND_UP(nir->num_inputs, 4);
 522                 break;
 523         case MESA_SHADER_GEOMETRY:
 524                 const_state->offsets.primitive_param = constoff;
 525                 const_state->offsets.primitive_map = constoff + 1;
 526                 constoff += 1 + DIV_ROUND_UP(nir->num_inputs, 4);
 527                 break;
 528         default:
 529                 break;
 530         }
 531
 532         const_state->offsets.immediate = constoff;
 533 }