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 LLVMValueRef
 328 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 329                             LLVMValueRef rsrc,
 330                             LLVMValueRef voffset,
 331                             LLVMValueRef soffset,
 332                             LLVMValueRef immoffset,
 333                             unsigned cache_policy);
 334
 335 LLVMValueRef
 336 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 337                            LLVMValueRef rsrc,
 338                            LLVMValueRef voffset,
 339                            LLVMValueRef soffset,
 340                            LLVMValueRef immoffset,
 341                            unsigned cache_policy);
 342
 343 LLVMValueRef
 344 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 345                              LLVMValueRef rsrc,
 346                              LLVMValueRef vindex,
 347                              LLVMValueRef voffset,
 348                              LLVMValueRef soffset,
 349                              LLVMValueRef immoffset,
 350                              unsigned num_channels,
 351                              unsigned dfmt,
 352                              unsigned nfmt,
 353                              unsigned cache_policy,
 354                              bool can_speculate);
 355
 356 LLVMValueRef
 357 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 358                           LLVMValueRef rsrc,
 359                           LLVMValueRef voffset,
 360                           LLVMValueRef soffset,
 361                           LLVMValueRef immoffset,
 362                           unsigned num_channels,
 363                           unsigned dfmt,
 364                           unsigned nfmt,
 365                           unsigned cache_policy,
 366                           bool can_speculate);
 367
 368 /* For ac_build_fetch_format.
 369  *
 370  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 371  */
 372 enum {
 373         AC_FETCH_FORMAT_FLOAT = 0,
 374         AC_FETCH_FORMAT_FIXED,
 375         AC_FETCH_FORMAT_UNORM,
 376         AC_FETCH_FORMAT_SNORM,
 377         AC_FETCH_FORMAT_USCALED,
 378         AC_FETCH_FORMAT_SSCALED,
 379         AC_FETCH_FORMAT_UINT,
 380         AC_FETCH_FORMAT_SINT,
 381 };
 382
 383 LLVMValueRef
 384 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 385                                unsigned log_size,
 386                                unsigned num_channels,
 387                                unsigned format,
 388                                bool reverse,
 389                                bool known_aligned,
 390                                LLVMValueRef rsrc,
 391                                LLVMValueRef vindex,
 392                                LLVMValueRef voffset,
 393                                LLVMValueRef soffset,
 394                                unsigned cache_policy,
 395                                bool can_speculate);
 396
 397 void
 398 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 399                              LLVMValueRef rsrc,
 400                              LLVMValueRef vdata,
 401                              LLVMValueRef voffset,
 402                              LLVMValueRef soffset,
 403                              unsigned cache_policy);
 404
 405 void
 406 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 407                             LLVMValueRef rsrc,
 408                             LLVMValueRef vdata,
 409                             LLVMValueRef voffset,
 410                             LLVMValueRef soffset,
 411                             unsigned cache_policy);
 412
 413 void
 414 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 415                               LLVMValueRef rsrc,
 416                               LLVMValueRef vdata,
 417                               LLVMValueRef vindex,
 418                               LLVMValueRef voffset,
 419                               LLVMValueRef soffset,
 420                               LLVMValueRef immoffset,
 421                               unsigned num_channels,
 422                               unsigned dfmt,
 423                               unsigned nfmt,
 424                               unsigned cache_policy);
 425
 426 void
 427 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 428                            LLVMValueRef rsrc,
 429                            LLVMValueRef vdata,
 430                            LLVMValueRef voffset,
 431                            LLVMValueRef soffset,
 432                            LLVMValueRef immoffset,
 433                            unsigned num_channels,
 434                            unsigned dfmt,
 435                            unsigned nfmt,
 436                            unsigned cache_policy);
 437
 438 LLVMValueRef
 439 ac_get_thread_id(struct ac_llvm_context *ctx);
 440
 441 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 442 #define AC_TID_MASK_TOP      0xfffffffd
 443 #define AC_TID_MASK_LEFT     0xfffffffe
 444
 445 LLVMValueRef
 446 ac_build_ddxy(struct ac_llvm_context *ctx,
 447               uint32_t mask,
 448               int idx,
 449               LLVMValueRef val);
 450
 451 #define AC_SENDMSG_GS 2
 452 #define AC_SENDMSG_GS_DONE 3
 453 #define AC_SENDMSG_GS_ALLOC_REQ 9
 454
 455 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 456 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 457 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 458 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 459
 460 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 461                       uint32_t msg,
 462                       LLVMValueRef wave_id);
 463
 464 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 465                            LLVMValueRef arg,
 466                            LLVMTypeRef dst_type);
 467
 468 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 469                           LLVMValueRef arg,
 470                           LLVMTypeRef dst_type);
 471 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 472                            LLVMValueRef b);
 473 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 474                            LLVMValueRef b);
 475 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 476                            LLVMValueRef b);
 477 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 478                            LLVMValueRef b);
 479 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 480 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 481 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 482
 483 struct ac_export_args {
 484         LLVMValueRef out[4];
 485         unsigned target;
 486         unsigned enabled_channels;
 487         bool compr;
 488         bool done;
 489         bool valid_mask;
 490 };
 491
 492 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 493
 494 void ac_build_export_null(struct ac_llvm_context *ctx);
 495
 496 enum ac_image_opcode {
 497         ac_image_sample,
 498         ac_image_gather4,
 499         ac_image_load,
 500         ac_image_load_mip,
 501         ac_image_store,
 502         ac_image_store_mip,
 503         ac_image_get_lod,
 504         ac_image_get_resinfo,
 505         ac_image_atomic,
 506         ac_image_atomic_cmpswap,
 507 };
 508
 509 enum ac_atomic_op {
 510         ac_atomic_swap,
 511         ac_atomic_add,
 512         ac_atomic_sub,
 513         ac_atomic_smin,
 514         ac_atomic_umin,
 515         ac_atomic_smax,
 516         ac_atomic_umax,
 517         ac_atomic_and,
 518         ac_atomic_or,
 519         ac_atomic_xor,
 520         ac_atomic_inc_wrap,
 521         ac_atomic_dec_wrap,
 522 };
 523
 524 enum ac_image_dim {
 525         ac_image_1d,
 526         ac_image_2d,
 527         ac_image_3d,
 528         ac_image_cube, // includes cube arrays
 529         ac_image_1darray,
 530         ac_image_2darray,
 531         ac_image_2dmsaa,
 532         ac_image_2darraymsaa,
 533 };
 534
 535 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 536 enum ac_image_cache_policy {
 537         ac_glc = 1 << 0, /* per-CU cache control */
 538         ac_slc = 1 << 1, /* global L2 cache control */
 539         ac_dlc = 1 << 2, /* per-shader-array cache control */
 540 };
 541
 542 struct ac_image_args {
 543         enum ac_image_opcode opcode : 4;
 544         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 545         enum ac_image_dim dim : 3;
 546         unsigned dmask : 4;
 547         unsigned cache_policy : 3;
 548         bool unorm : 1;
 549         bool level_zero : 1;
 550         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 551
 552         LLVMValueRef resource;
 553         LLVMValueRef sampler;
 554         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 555         LLVMValueRef offset;
 556         LLVMValueRef bias;
 557         LLVMValueRef compare;
 558         LLVMValueRef derivs[6];
 559         LLVMValueRef coords[4];
 560         LLVMValueRef lod; // also used by ac_image_get_resinfo
 561 };
 562
 563 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 564                                    struct ac_image_args *a);
 565 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 566                                     LLVMValueRef args[2]);
 567 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 568                                      LLVMValueRef args[2]);
 569 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 570                                      LLVMValueRef args[2]);
 571 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 572                                  LLVMValueRef args[2], unsigned bits, bool hi);
 573 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 574                                  LLVMValueRef args[2], unsigned bits, bool hi);
 575 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 576 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 577 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 578                           LLVMValueRef offset, LLVMValueRef width,
 579                           bool is_signed);
 580 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 581                            LLVMValueRef s1, LLVMValueRef s2);
 582 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 583                            LLVMValueRef s1, LLVMValueRef s2);
 584
 585 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 586
 587 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 588                            unsigned bitsize);
 589
 590 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 591                             LLVMValueRef src1, LLVMValueRef src2,
 592                             unsigned bitsize);
 593
 594 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 595                             unsigned bitsize);
 596
 597 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 598                             unsigned bitsize);
 599
 600 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 601
 602 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 603                                        LLVMValueRef src0);
 604
 605 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 606                             LLVMValueRef main_fn,
 607                             uint8_t *vs_output_param_offset,
 608                             uint32_t num_outputs,
 609                             uint8_t *num_param_exports);
 610 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 611
 612 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 613 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 614                          LLVMValueRef dw_addr);
 615 void ac_lds_store(struct ac_llvm_context *ctx,
 616                   LLVMValueRef dw_addr, LLVMValueRef value);
 617
 618 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 619                          LLVMTypeRef dst_type,
 620                          LLVMValueRef src0);
 621
 622 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 623 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 624
 625 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 626 void ac_build_break(struct ac_llvm_context *ctx);
 627 void ac_build_continue(struct ac_llvm_context *ctx);
 628 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 629 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 630 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 631 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 632 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 633                  int lable_id);
 634 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 635                   int lable_id);
 636
 637 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 638                              const char *name);
 639 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 640                                    const char *name);
 641
 642 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 643                          LLVMTypeRef type);
 644
 645 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 646                             unsigned count);
 647
 648 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 649                              unsigned rshift, unsigned bitwidth);
 650
 651 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 652                               LLVMValueRef *addr, bool is_array_tex);
 653
 654 LLVMValueRef
 655 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 656
 657 LLVMValueRef
 658 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 659
 660 LLVMValueRef
 661 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 662
 663 LLVMValueRef
 664 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 665
 666 LLVMValueRef
 667 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 668
 669 LLVMValueRef
 670 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 671
 672 LLVMValueRef
 673 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 674
 675 /**
 676  * Common arguments for a scan/reduce operation that accumulates per-wave
 677  * values across an entire workgroup, while respecting the order of waves.
 678  */
 679 struct ac_wg_scan {
 680         bool enable_reduce;
 681         bool enable_exclusive;
 682         bool enable_inclusive;
 683         nir_op op;
 684         LLVMValueRef src; /* clobbered! */
 685         LLVMValueRef result_reduce;
 686         LLVMValueRef result_exclusive;
 687         LLVMValueRef result_inclusive;
 688         LLVMValueRef extra;
 689         LLVMValueRef waveidx;
 690         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 691
 692         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 693         LLVMValueRef scratch;
 694         unsigned maxwaves;
 695 };
 696
 697 void
 698 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 699 void
 700 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 701 void
 702 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 703
 704 void
 705 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 706 void
 707 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 708 void
 709 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 710
 711 LLVMValueRef
 712 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 713                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 714
 715 LLVMValueRef
 716 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 717
 718 LLVMValueRef
 719 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 720                    unsigned bitsize);
 721
 722 LLVMValueRef
 723 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 724                     unsigned bitsize);
 725
 726 LLVMValueRef
 727 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 728
 729 LLVMValueRef
 730 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 731
 732 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 733                            LLVMValueRef *args, unsigned num_args);
 734
 735 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 736                                  LLVMValueRef ptr, LLVMValueRef val,
 737                                  const char *sync_scope);
 738
 739 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 740                                       LLVMValueRef cmp, LLVMValueRef val,
 741                                       const char *sync_scope);
 742
 743 #ifdef __cplusplus
 744 }
 745 #endif
 746
 747 #endif