src/amd/common/ac_llvm_build.h

   1 /*
   2  * Copyright 2016 Bas Nieuwenhuizen
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  13  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  14  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  15  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  16  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  17  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  18  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  19  *
  20  * The above copyright notice and this permission notice (including the
  21  * next paragraph) shall be included in all copies or substantial portions
  22  * of the Software.
  23  *
  24  */
  25 #ifndef AC_LLVM_BUILD_H
  26 #define AC_LLVM_BUILD_H
  27
  28 #include <stdbool.h>
  29 #include <llvm-c/Core.h>
  30 #include "compiler/nir/nir.h"
  31 #include "amd_family.h"
  32
  33 #ifdef __cplusplus
  34 extern "C" {
  35 #endif
  36
  37 enum {
  38         AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
  39         AC_ADDR_SPACE_GLOBAL = 1,
  40         AC_ADDR_SPACE_GDS = 2,
  41         AC_ADDR_SPACE_LDS = 3,
  42         AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
  43         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  44 };
  45
  46 #define AC_WAIT_LGKM    (1 << 0) /* LDS, GDS, constant, message */
  47 #define AC_WAIT_VLOAD   (1 << 1) /* VMEM load/sample instructions */
  48 #define AC_WAIT_VSTORE  (1 << 2) /* VMEM store instructions */
  49
  50 struct ac_llvm_flow;
  51 struct ac_llvm_compiler;
  52 enum ac_float_mode;
  53
  54 struct ac_llvm_flow_state {
  55         struct ac_llvm_flow *stack;
  56         unsigned depth_max;
  57         unsigned depth;
  58 };
  59
  60 struct ac_llvm_context {
  61         LLVMContextRef context;
  62         LLVMModuleRef module;
  63         LLVMBuilderRef builder;
  64
  65         LLVMTypeRef voidt;
  66         LLVMTypeRef i1;
  67         LLVMTypeRef i8;
  68         LLVMTypeRef i16;
  69         LLVMTypeRef i32;
  70         LLVMTypeRef i64;
  71         LLVMTypeRef intptr;
  72         LLVMTypeRef f16;
  73         LLVMTypeRef f32;
  74         LLVMTypeRef f64;
  75         LLVMTypeRef v2i16;
  76         LLVMTypeRef v2i32;
  77         LLVMTypeRef v3i32;
  78         LLVMTypeRef v4i32;
  79         LLVMTypeRef v2f32;
  80         LLVMTypeRef v3f32;
  81         LLVMTypeRef v4f32;
  82         LLVMTypeRef v8i32;
  83         LLVMTypeRef iN_wavemask;
  84         LLVMTypeRef iN_ballotmask;
  85
  86         LLVMValueRef i8_0;
  87         LLVMValueRef i8_1;
  88         LLVMValueRef i16_0;
  89         LLVMValueRef i16_1;
  90         LLVMValueRef i32_0;
  91         LLVMValueRef i32_1;
  92         LLVMValueRef i64_0;
  93         LLVMValueRef i64_1;
  94         LLVMValueRef f16_0;
  95         LLVMValueRef f16_1;
  96         LLVMValueRef f32_0;
  97         LLVMValueRef f32_1;
  98         LLVMValueRef f64_0;
  99         LLVMValueRef f64_1;
 100         LLVMValueRef i1true;
 101         LLVMValueRef i1false;
 102
 103         /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
 104          * are two ac_llvm_contexts. Declare a pointer here, so that the control
 105          * flow stack is shared by both ac_llvm_contexts.
 106          */
 107         struct ac_llvm_flow_state *flow;
 108
 109         unsigned range_md_kind;
 110         unsigned invariant_load_md_kind;
 111         unsigned uniform_md_kind;
 112         unsigned fpmath_md_kind;
 113         LLVMValueRef fpmath_md_2p5_ulp;
 114         LLVMValueRef empty_md;
 115
 116         enum chip_class chip_class;
 117         enum radeon_family family;
 118
 119         unsigned wave_size;
 120         unsigned ballot_mask_bits;
 121
 122         LLVMValueRef lds;
 123 };
 124
 125 void
 126 ac_llvm_context_init(struct ac_llvm_context *ctx,
 127                      struct ac_llvm_compiler *compiler,
 128                      enum chip_class chip_class, enum radeon_family family,
 129                      enum ac_float_mode float_mode, unsigned wave_size,
 130                      unsigned ballot_mask_bits);
 131
 132 void
 133 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 134
 135 int
 136 ac_get_llvm_num_components(LLVMValueRef value);
 137
 138 int
 139 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 140
 141 LLVMValueRef
 142 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 143                      LLVMValueRef value,
 144                      int index);
 145
 146 unsigned ac_get_type_size(LLVMTypeRef type);
 147
 148 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 149 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 150 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 151 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 152 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 153
 154 LLVMValueRef
 155 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 156                    LLVMTypeRef return_type, LLVMValueRef *params,
 157                    unsigned param_count, unsigned attrib_mask);
 158
 159 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 160
 161 LLVMValueRef
 162 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 163              unsigned count_incoming, LLVMValueRef *values,
 164              LLVMBasicBlockRef *blocks);
 165
 166 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 167 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 168                                    LLVMValueRef *pvgpr);
 169
 170 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 171
 172 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 173 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 174                                  LLVMValueRef value);
 175
 176 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 177
 178 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 179
 180 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 181
 182 LLVMValueRef
 183 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 184                                unsigned value_count, unsigned component);
 185
 186 LLVMValueRef
 187 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 188                                 LLVMValueRef *values,
 189                                 unsigned value_count,
 190                                 unsigned value_stride,
 191                                 bool load,
 192                                 bool always_vector);
 193 LLVMValueRef
 194 ac_build_gather_values(struct ac_llvm_context *ctx,
 195                        LLVMValueRef *values,
 196                        unsigned value_count);
 197
 198 LLVMValueRef
 199 ac_extract_components(struct ac_llvm_context *ctx,
 200                       LLVMValueRef value,
 201                       unsigned start,
 202                       unsigned channels);
 203
 204 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 205                                      LLVMValueRef value,
 206                                      unsigned num_channels);
 207 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 208
 209 LLVMValueRef
 210 ac_build_fdiv(struct ac_llvm_context *ctx,
 211               LLVMValueRef num,
 212               LLVMValueRef den);
 213
 214 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 215                                 LLVMValueRef num,
 216                                 LLVMValueRef multiplier,
 217                                 LLVMValueRef pre_shift,
 218                                 LLVMValueRef post_shift,
 219                                 LLVMValueRef increment);
 220 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 221                                     LLVMValueRef num,
 222                                     LLVMValueRef multiplier,
 223                                     LLVMValueRef pre_shift,
 224                                     LLVMValueRef post_shift,
 225                                     LLVMValueRef increment);
 226 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 227                                               LLVMValueRef num,
 228                                               LLVMValueRef multiplier,
 229                                               LLVMValueRef post_shift);
 230
 231 void
 232 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 233                        bool is_deriv, bool is_array, bool is_lod,
 234                        LLVMValueRef *coords_arg,
 235                        LLVMValueRef *derivs_arg);
 236
 237
 238 LLVMValueRef
 239 ac_build_fs_interp(struct ac_llvm_context *ctx,
 240                    LLVMValueRef llvm_chan,
 241                    LLVMValueRef attr_number,
 242                    LLVMValueRef params,
 243                    LLVMValueRef i,
 244                    LLVMValueRef j);
 245
 246 LLVMValueRef
 247 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 248                        LLVMValueRef llvm_chan,
 249                        LLVMValueRef attr_number,
 250                        LLVMValueRef params,
 251                        LLVMValueRef i,
 252                        LLVMValueRef j);
 253
 254 LLVMValueRef
 255 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 256                        LLVMValueRef parameter,
 257                        LLVMValueRef llvm_chan,
 258                        LLVMValueRef attr_number,
 259                        LLVMValueRef params);
 260
 261 LLVMValueRef
 262 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 263                  LLVMValueRef base_ptr,
 264                  LLVMValueRef index);
 265
 266 LLVMValueRef
 267 ac_build_gep0(struct ac_llvm_context *ctx,
 268               LLVMValueRef base_ptr,
 269               LLVMValueRef index);
 270 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 271                                   LLVMValueRef index);
 272
 273 void
 274 ac_build_indexed_store(struct ac_llvm_context *ctx,
 275                        LLVMValueRef base_ptr, LLVMValueRef index,
 276                        LLVMValueRef value);
 277
 278 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 279                            LLVMValueRef index);
 280 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 281                                      LLVMValueRef base_ptr, LLVMValueRef index);
 282 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 283                                    LLVMValueRef base_ptr, LLVMValueRef index);
 284 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 285                                    LLVMValueRef base_ptr, LLVMValueRef index);
 286
 287 void
 288 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 289                             LLVMValueRef rsrc,
 290                             LLVMValueRef vdata,
 291                             unsigned num_channels,
 292                             LLVMValueRef voffset,
 293                             LLVMValueRef soffset,
 294                             unsigned inst_offset,
 295                             unsigned cache_policy,
 296                             bool swizzle_enable_hint);
 297
 298 void
 299 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 300                              LLVMValueRef rsrc,
 301                              LLVMValueRef data,
 302                              LLVMValueRef vindex,
 303                              LLVMValueRef voffset,
 304                              unsigned num_channels,
 305                              unsigned cache_policy);
 306
 307 LLVMValueRef
 308 ac_build_buffer_load(struct ac_llvm_context *ctx,
 309                      LLVMValueRef rsrc,
 310                      int num_channels,
 311                      LLVMValueRef vindex,
 312                      LLVMValueRef voffset,
 313                      LLVMValueRef soffset,
 314                      unsigned inst_offset,
 315                      unsigned cache_policy,
 316                      bool can_speculate,
 317                      bool allow_smem);
 318
 319 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 320                                          LLVMValueRef rsrc,
 321                                          LLVMValueRef vindex,
 322                                          LLVMValueRef voffset,
 323                                          unsigned num_channels,
 324                                          unsigned cache_policy,
 325                                          bool can_speculate);
 326
 327 /* load_format that handles the stride & element count better if idxen is
 328  * disabled by LLVM. */
 329 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 330                                                   LLVMValueRef rsrc,
 331                                                   LLVMValueRef vindex,
 332                                                   LLVMValueRef voffset,
 333                                                   unsigned num_channels,
 334                                                   unsigned cache_policy,
 335                                                   bool can_speculate);
 336
 337 LLVMValueRef
 338 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 339                             LLVMValueRef rsrc,
 340                             LLVMValueRef voffset,
 341                             LLVMValueRef soffset,
 342                             LLVMValueRef immoffset,
 343                             unsigned cache_policy);
 344
 345 LLVMValueRef
 346 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 347                            LLVMValueRef rsrc,
 348                            LLVMValueRef voffset,
 349                            LLVMValueRef soffset,
 350                            LLVMValueRef immoffset,
 351                            unsigned cache_policy);
 352
 353 LLVMValueRef
 354 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 355                              LLVMValueRef rsrc,
 356                              LLVMValueRef vindex,
 357                              LLVMValueRef voffset,
 358                              LLVMValueRef soffset,
 359                              LLVMValueRef immoffset,
 360                              unsigned num_channels,
 361                              unsigned dfmt,
 362                              unsigned nfmt,
 363                              unsigned cache_policy,
 364                              bool can_speculate);
 365
 366 LLVMValueRef
 367 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 368                           LLVMValueRef rsrc,
 369                           LLVMValueRef voffset,
 370                           LLVMValueRef soffset,
 371                           LLVMValueRef immoffset,
 372                           unsigned num_channels,
 373                           unsigned dfmt,
 374                           unsigned nfmt,
 375                           unsigned cache_policy,
 376                           bool can_speculate);
 377
 378 /* For ac_build_fetch_format.
 379  *
 380  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 381  */
 382 enum {
 383         AC_FETCH_FORMAT_FLOAT = 0,
 384         AC_FETCH_FORMAT_FIXED,
 385         AC_FETCH_FORMAT_UNORM,
 386         AC_FETCH_FORMAT_SNORM,
 387         AC_FETCH_FORMAT_USCALED,
 388         AC_FETCH_FORMAT_SSCALED,
 389         AC_FETCH_FORMAT_UINT,
 390         AC_FETCH_FORMAT_SINT,
 391 };
 392
 393 LLVMValueRef
 394 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 395                                unsigned log_size,
 396                                unsigned num_channels,
 397                                unsigned format,
 398                                bool reverse,
 399                                bool known_aligned,
 400                                LLVMValueRef rsrc,
 401                                LLVMValueRef vindex,
 402                                LLVMValueRef voffset,
 403                                LLVMValueRef soffset,
 404                                unsigned cache_policy,
 405                                bool can_speculate);
 406
 407 void
 408 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 409                              LLVMValueRef rsrc,
 410                              LLVMValueRef vdata,
 411                              LLVMValueRef voffset,
 412                              LLVMValueRef soffset,
 413                              unsigned cache_policy);
 414
 415 void
 416 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 417                             LLVMValueRef rsrc,
 418                             LLVMValueRef vdata,
 419                             LLVMValueRef voffset,
 420                             LLVMValueRef soffset,
 421                             unsigned cache_policy);
 422
 423 void
 424 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 425                               LLVMValueRef rsrc,
 426                               LLVMValueRef vdata,
 427                               LLVMValueRef vindex,
 428                               LLVMValueRef voffset,
 429                               LLVMValueRef soffset,
 430                               LLVMValueRef immoffset,
 431                               unsigned num_channels,
 432                               unsigned dfmt,
 433                               unsigned nfmt,
 434                               unsigned cache_policy);
 435
 436 void
 437 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 438                            LLVMValueRef rsrc,
 439                            LLVMValueRef vdata,
 440                            LLVMValueRef voffset,
 441                            LLVMValueRef soffset,
 442                            LLVMValueRef immoffset,
 443                            unsigned num_channels,
 444                            unsigned dfmt,
 445                            unsigned nfmt,
 446                            unsigned cache_policy);
 447
 448 LLVMValueRef
 449 ac_get_thread_id(struct ac_llvm_context *ctx);
 450
 451 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 452 #define AC_TID_MASK_TOP      0xfffffffd
 453 #define AC_TID_MASK_LEFT     0xfffffffe
 454
 455 LLVMValueRef
 456 ac_build_ddxy(struct ac_llvm_context *ctx,
 457               uint32_t mask,
 458               int idx,
 459               LLVMValueRef val);
 460
 461 #define AC_SENDMSG_GS 2
 462 #define AC_SENDMSG_GS_DONE 3
 463 #define AC_SENDMSG_GS_ALLOC_REQ 9
 464
 465 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 466 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 467 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 468 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 469
 470 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 471                       uint32_t msg,
 472                       LLVMValueRef wave_id);
 473
 474 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 475                            LLVMValueRef arg,
 476                            LLVMTypeRef dst_type);
 477
 478 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 479                           LLVMValueRef arg,
 480                           LLVMTypeRef dst_type);
 481 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 482                            LLVMValueRef b);
 483 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 484                            LLVMValueRef b);
 485 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 486                            LLVMValueRef b);
 487 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 488                            LLVMValueRef b);
 489 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 490 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 491 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 492
 493 struct ac_export_args {
 494         LLVMValueRef out[4];
 495         unsigned target;
 496         unsigned enabled_channels;
 497         bool compr;
 498         bool done;
 499         bool valid_mask;
 500 };
 501
 502 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 503
 504 void ac_build_export_null(struct ac_llvm_context *ctx);
 505
 506 enum ac_image_opcode {
 507         ac_image_sample,
 508         ac_image_gather4,
 509         ac_image_load,
 510         ac_image_load_mip,
 511         ac_image_store,
 512         ac_image_store_mip,
 513         ac_image_get_lod,
 514         ac_image_get_resinfo,
 515         ac_image_atomic,
 516         ac_image_atomic_cmpswap,
 517 };
 518
 519 enum ac_atomic_op {
 520         ac_atomic_swap,
 521         ac_atomic_add,
 522         ac_atomic_sub,
 523         ac_atomic_smin,
 524         ac_atomic_umin,
 525         ac_atomic_smax,
 526         ac_atomic_umax,
 527         ac_atomic_and,
 528         ac_atomic_or,
 529         ac_atomic_xor,
 530         ac_atomic_inc_wrap,
 531         ac_atomic_dec_wrap,
 532 };
 533
 534 enum ac_image_dim {
 535         ac_image_1d,
 536         ac_image_2d,
 537         ac_image_3d,
 538         ac_image_cube, // includes cube arrays
 539         ac_image_1darray,
 540         ac_image_2darray,
 541         ac_image_2dmsaa,
 542         ac_image_2darraymsaa,
 543 };
 544
 545 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 546 enum ac_image_cache_policy {
 547         ac_glc = 1 << 0, /* per-CU cache control */
 548         ac_slc = 1 << 1, /* global L2 cache control */
 549         ac_dlc = 1 << 2, /* per-shader-array cache control */
 550 };
 551
 552 struct ac_image_args {
 553         enum ac_image_opcode opcode : 4;
 554         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 555         enum ac_image_dim dim : 3;
 556         unsigned dmask : 4;
 557         unsigned cache_policy : 3;
 558         bool unorm : 1;
 559         bool level_zero : 1;
 560         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 561
 562         LLVMValueRef resource;
 563         LLVMValueRef sampler;
 564         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 565         LLVMValueRef offset;
 566         LLVMValueRef bias;
 567         LLVMValueRef compare;
 568         LLVMValueRef derivs[6];
 569         LLVMValueRef coords[4];
 570         LLVMValueRef lod; // also used by ac_image_get_resinfo
 571 };
 572
 573 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 574                                    struct ac_image_args *a);
 575 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 576                                     LLVMValueRef args[2]);
 577 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 578                                      LLVMValueRef args[2]);
 579 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 580                                      LLVMValueRef args[2]);
 581 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 582                                  LLVMValueRef args[2], unsigned bits, bool hi);
 583 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 584                                  LLVMValueRef args[2], unsigned bits, bool hi);
 585 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 586 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 587 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 588                           LLVMValueRef offset, LLVMValueRef width,
 589                           bool is_signed);
 590 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 591                            LLVMValueRef s1, LLVMValueRef s2);
 592 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 593                            LLVMValueRef s1, LLVMValueRef s2);
 594
 595 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 596
 597 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 598                            unsigned bitsize);
 599
 600 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 601                             LLVMValueRef src1, LLVMValueRef src2,
 602                             unsigned bitsize);
 603
 604 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 605                             unsigned bitsize);
 606
 607 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 608                             unsigned bitsize);
 609
 610 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 611
 612 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 613                                        LLVMValueRef src0);
 614
 615 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 616                             LLVMValueRef main_fn,
 617                             uint8_t *vs_output_param_offset,
 618                             uint32_t num_outputs,
 619                             uint8_t *num_param_exports);
 620 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 621
 622 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 623 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 624                          LLVMValueRef dw_addr);
 625 void ac_lds_store(struct ac_llvm_context *ctx,
 626                   LLVMValueRef dw_addr, LLVMValueRef value);
 627
 628 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 629                          LLVMTypeRef dst_type,
 630                          LLVMValueRef src0);
 631
 632 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 633 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 634
 635 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 636 void ac_build_break(struct ac_llvm_context *ctx);
 637 void ac_build_continue(struct ac_llvm_context *ctx);
 638 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 639 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 640 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 641 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 642 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 643                  int lable_id);
 644 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 645                   int lable_id);
 646
 647 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 648                              const char *name);
 649 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 650                                    const char *name);
 651
 652 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 653                          LLVMTypeRef type);
 654
 655 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 656                             unsigned count);
 657
 658 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 659                              unsigned rshift, unsigned bitwidth);
 660
 661 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 662                               LLVMValueRef *addr, bool is_array_tex);
 663
 664 LLVMValueRef
 665 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 666
 667 LLVMValueRef
 668 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 669
 670 LLVMValueRef
 671 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 672
 673 LLVMValueRef
 674 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 675
 676 LLVMValueRef
 677 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 678
 679 LLVMValueRef
 680 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 681
 682 LLVMValueRef
 683 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 684
 685 /**
 686  * Common arguments for a scan/reduce operation that accumulates per-wave
 687  * values across an entire workgroup, while respecting the order of waves.
 688  */
 689 struct ac_wg_scan {
 690         bool enable_reduce;
 691         bool enable_exclusive;
 692         bool enable_inclusive;
 693         nir_op op;
 694         LLVMValueRef src; /* clobbered! */
 695         LLVMValueRef result_reduce;
 696         LLVMValueRef result_exclusive;
 697         LLVMValueRef result_inclusive;
 698         LLVMValueRef extra;
 699         LLVMValueRef waveidx;
 700         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 701
 702         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 703         LLVMValueRef scratch;
 704         unsigned maxwaves;
 705 };
 706
 707 void
 708 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 709 void
 710 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 711 void
 712 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 713
 714 void
 715 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 716 void
 717 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 718 void
 719 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 720
 721 LLVMValueRef
 722 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 723                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 724
 725 LLVMValueRef
 726 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 727
 728 LLVMValueRef
 729 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 730                    unsigned bitsize);
 731
 732 LLVMValueRef
 733 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 734                     unsigned bitsize);
 735
 736 LLVMValueRef
 737 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 738
 739 LLVMValueRef
 740 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 741
 742 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 743                            LLVMValueRef *args, unsigned num_args);
 744
 745 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 746                                  LLVMValueRef ptr, LLVMValueRef val,
 747                                  const char *sync_scope);
 748
 749 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 750                                       LLVMValueRef cmp, LLVMValueRef val,
 751                                       const char *sync_scope);
 752
 753 #ifdef __cplusplus
 754 }
 755 #endif
 756
 757 #endif