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