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 #include "ac_shader_args.h"
  34 #include "ac_shader_abi.h"
  35
  36 #ifdef __cplusplus
  37 extern "C" {
  38 #endif
  39
  40 enum {
  41         AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
  42         AC_ADDR_SPACE_GLOBAL = 1,
  43         AC_ADDR_SPACE_GDS = 2,
  44         AC_ADDR_SPACE_LDS = 3,
  45         AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
  46         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  47 };
  48
  49 #define AC_WAIT_LGKM    (1 << 0) /* LDS, GDS, constant, message */
  50 #define AC_WAIT_VLOAD   (1 << 1) /* VMEM load/sample instructions */
  51 #define AC_WAIT_VSTORE  (1 << 2) /* VMEM store instructions */
  52
  53 struct ac_llvm_flow;
  54 struct ac_llvm_compiler;
  55 enum ac_float_mode;
  56
  57 struct ac_llvm_flow_state {
  58         struct ac_llvm_flow *stack;
  59         unsigned depth_max;
  60         unsigned depth;
  61 };
  62
  63 struct ac_llvm_context {
  64         LLVMContextRef context;
  65         LLVMModuleRef module;
  66         LLVMBuilderRef builder;
  67
  68         LLVMValueRef main_function;
  69
  70         LLVMTypeRef voidt;
  71         LLVMTypeRef i1;
  72         LLVMTypeRef i8;
  73         LLVMTypeRef i16;
  74         LLVMTypeRef i32;
  75         LLVMTypeRef i64;
  76         LLVMTypeRef intptr;
  77         LLVMTypeRef f16;
  78         LLVMTypeRef f32;
  79         LLVMTypeRef f64;
  80         LLVMTypeRef v2i16;
  81         LLVMTypeRef v2i32;
  82         LLVMTypeRef v3i32;
  83         LLVMTypeRef v4i32;
  84         LLVMTypeRef v2f32;
  85         LLVMTypeRef v3f32;
  86         LLVMTypeRef v4f32;
  87         LLVMTypeRef v8i32;
  88         LLVMTypeRef iN_wavemask;
  89         LLVMTypeRef iN_ballotmask;
  90
  91         LLVMValueRef i8_0;
  92         LLVMValueRef i8_1;
  93         LLVMValueRef i16_0;
  94         LLVMValueRef i16_1;
  95         LLVMValueRef i32_0;
  96         LLVMValueRef i32_1;
  97         LLVMValueRef i64_0;
  98         LLVMValueRef i64_1;
  99         LLVMValueRef f16_0;
 100         LLVMValueRef f16_1;
 101         LLVMValueRef f32_0;
 102         LLVMValueRef f32_1;
 103         LLVMValueRef f64_0;
 104         LLVMValueRef f64_1;
 105         LLVMValueRef i1true;
 106         LLVMValueRef i1false;
 107
 108         /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
 109          * are two ac_llvm_contexts. Declare a pointer here, so that the control
 110          * flow stack is shared by both ac_llvm_contexts.
 111          */
 112         struct ac_llvm_flow_state *flow;
 113
 114         unsigned range_md_kind;
 115         unsigned invariant_load_md_kind;
 116         unsigned uniform_md_kind;
 117         unsigned fpmath_md_kind;
 118         LLVMValueRef fpmath_md_2p5_ulp;
 119         LLVMValueRef empty_md;
 120
 121         enum chip_class chip_class;
 122         enum radeon_family family;
 123
 124         unsigned wave_size;
 125         unsigned ballot_mask_bits;
 126
 127         unsigned float_mode;
 128
 129         LLVMValueRef lds;
 130 };
 131
 132 void
 133 ac_llvm_context_init(struct ac_llvm_context *ctx,
 134                      struct ac_llvm_compiler *compiler,
 135                      enum chip_class chip_class, enum radeon_family family,
 136                      enum ac_float_mode float_mode, unsigned wave_size,
 137                      unsigned ballot_mask_bits);
 138
 139 void
 140 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 141
 142 int
 143 ac_get_llvm_num_components(LLVMValueRef value);
 144
 145 int
 146 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 147
 148 LLVMValueRef
 149 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 150                      LLVMValueRef value,
 151                      int index);
 152
 153 unsigned ac_get_type_size(LLVMTypeRef type);
 154
 155 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 156 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 157 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 158 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 159 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 160
 161 LLVMValueRef
 162 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 163                    LLVMTypeRef return_type, LLVMValueRef *params,
 164                    unsigned param_count, unsigned attrib_mask);
 165
 166 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 167
 168 LLVMValueRef
 169 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 170              unsigned count_incoming, LLVMValueRef *values,
 171              LLVMBasicBlockRef *blocks);
 172
 173 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 174 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 175                                    LLVMValueRef *pvgpr);
 176
 177 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 178
 179 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 180 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 181                                  LLVMValueRef value);
 182
 183 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 184
 185 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 186
 187 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 188
 189 LLVMValueRef
 190 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 191                                unsigned value_count, unsigned component);
 192
 193 LLVMValueRef
 194 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 195                                 LLVMValueRef *values,
 196                                 unsigned value_count,
 197                                 unsigned value_stride,
 198                                 bool load,
 199                                 bool always_vector);
 200 LLVMValueRef
 201 ac_build_gather_values(struct ac_llvm_context *ctx,
 202                        LLVMValueRef *values,
 203                        unsigned value_count);
 204
 205 LLVMValueRef
 206 ac_extract_components(struct ac_llvm_context *ctx,
 207                       LLVMValueRef value,
 208                       unsigned start,
 209                       unsigned channels);
 210
 211 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 212                                      LLVMValueRef value,
 213                                      unsigned num_channels);
 214 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 215
 216 LLVMValueRef
 217 ac_build_fdiv(struct ac_llvm_context *ctx,
 218               LLVMValueRef num,
 219               LLVMValueRef den);
 220
 221 LLVMValueRef ac_build_fast_udiv(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_nuw(struct ac_llvm_context *ctx,
 228                                     LLVMValueRef num,
 229                                     LLVMValueRef multiplier,
 230                                     LLVMValueRef pre_shift,
 231                                     LLVMValueRef post_shift,
 232                                     LLVMValueRef increment);
 233 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 234                                               LLVMValueRef num,
 235                                               LLVMValueRef multiplier,
 236                                               LLVMValueRef post_shift);
 237
 238 void
 239 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 240                        bool is_deriv, bool is_array, bool is_lod,
 241                        LLVMValueRef *coords_arg,
 242                        LLVMValueRef *derivs_arg);
 243
 244
 245 LLVMValueRef
 246 ac_build_fs_interp(struct ac_llvm_context *ctx,
 247                    LLVMValueRef llvm_chan,
 248                    LLVMValueRef attr_number,
 249                    LLVMValueRef params,
 250                    LLVMValueRef i,
 251                    LLVMValueRef j);
 252
 253 LLVMValueRef
 254 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 255                        LLVMValueRef llvm_chan,
 256                        LLVMValueRef attr_number,
 257                        LLVMValueRef params,
 258                        LLVMValueRef i,
 259                        LLVMValueRef j);
 260
 261 LLVMValueRef
 262 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 263                        LLVMValueRef parameter,
 264                        LLVMValueRef llvm_chan,
 265                        LLVMValueRef attr_number,
 266                        LLVMValueRef params);
 267
 268 LLVMValueRef
 269 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 270                  LLVMValueRef base_ptr,
 271                  LLVMValueRef index);
 272
 273 LLVMValueRef
 274 ac_build_gep0(struct ac_llvm_context *ctx,
 275               LLVMValueRef base_ptr,
 276               LLVMValueRef index);
 277 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 278                                   LLVMValueRef index);
 279
 280 void
 281 ac_build_indexed_store(struct ac_llvm_context *ctx,
 282                        LLVMValueRef base_ptr, LLVMValueRef index,
 283                        LLVMValueRef value);
 284
 285 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 286                            LLVMValueRef index);
 287 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 288                                      LLVMValueRef base_ptr, LLVMValueRef index);
 289 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 290                                    LLVMValueRef base_ptr, LLVMValueRef index);
 291 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 292                                    LLVMValueRef base_ptr, LLVMValueRef index);
 293
 294 void
 295 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 296                             LLVMValueRef rsrc,
 297                             LLVMValueRef vdata,
 298                             unsigned num_channels,
 299                             LLVMValueRef voffset,
 300                             LLVMValueRef soffset,
 301                             unsigned inst_offset,
 302                             unsigned cache_policy,
 303                             bool swizzle_enable_hint);
 304
 305 void
 306 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 307                              LLVMValueRef rsrc,
 308                              LLVMValueRef data,
 309                              LLVMValueRef vindex,
 310                              LLVMValueRef voffset,
 311                              unsigned num_channels,
 312                              unsigned cache_policy);
 313
 314 LLVMValueRef
 315 ac_build_buffer_load(struct ac_llvm_context *ctx,
 316                      LLVMValueRef rsrc,
 317                      int num_channels,
 318                      LLVMValueRef vindex,
 319                      LLVMValueRef voffset,
 320                      LLVMValueRef soffset,
 321                      unsigned inst_offset,
 322                      unsigned cache_policy,
 323                      bool can_speculate,
 324                      bool allow_smem);
 325
 326 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 327                                          LLVMValueRef rsrc,
 328                                          LLVMValueRef vindex,
 329                                          LLVMValueRef voffset,
 330                                          unsigned num_channels,
 331                                          unsigned cache_policy,
 332                                          bool can_speculate);
 333
 334 LLVMValueRef
 335 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 336                             LLVMValueRef rsrc,
 337                             LLVMValueRef voffset,
 338                             LLVMValueRef soffset,
 339                             LLVMValueRef immoffset,
 340                             unsigned cache_policy);
 341
 342 LLVMValueRef
 343 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 344                            LLVMValueRef rsrc,
 345                            LLVMValueRef voffset,
 346                            LLVMValueRef soffset,
 347                            LLVMValueRef immoffset,
 348                            unsigned cache_policy);
 349
 350 LLVMValueRef
 351 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 352                              LLVMValueRef rsrc,
 353                              LLVMValueRef vindex,
 354                              LLVMValueRef voffset,
 355                              LLVMValueRef soffset,
 356                              LLVMValueRef immoffset,
 357                              unsigned num_channels,
 358                              unsigned dfmt,
 359                              unsigned nfmt,
 360                              unsigned cache_policy,
 361                              bool can_speculate);
 362
 363 LLVMValueRef
 364 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 365                           LLVMValueRef rsrc,
 366                           LLVMValueRef voffset,
 367                           LLVMValueRef soffset,
 368                           LLVMValueRef immoffset,
 369                           unsigned num_channels,
 370                           unsigned dfmt,
 371                           unsigned nfmt,
 372                           unsigned cache_policy,
 373                           bool can_speculate);
 374
 375 /* For ac_build_fetch_format.
 376  *
 377  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 378  */
 379 enum {
 380         AC_FETCH_FORMAT_FLOAT = 0,
 381         AC_FETCH_FORMAT_FIXED,
 382         AC_FETCH_FORMAT_UNORM,
 383         AC_FETCH_FORMAT_SNORM,
 384         AC_FETCH_FORMAT_USCALED,
 385         AC_FETCH_FORMAT_SSCALED,
 386         AC_FETCH_FORMAT_UINT,
 387         AC_FETCH_FORMAT_SINT,
 388 };
 389
 390 LLVMValueRef
 391 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 392                                unsigned log_size,
 393                                unsigned num_channels,
 394                                unsigned format,
 395                                bool reverse,
 396                                bool known_aligned,
 397                                LLVMValueRef rsrc,
 398                                LLVMValueRef vindex,
 399                                LLVMValueRef voffset,
 400                                LLVMValueRef soffset,
 401                                unsigned cache_policy,
 402                                bool can_speculate);
 403
 404 void
 405 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 406                              LLVMValueRef rsrc,
 407                              LLVMValueRef vdata,
 408                              LLVMValueRef voffset,
 409                              LLVMValueRef soffset,
 410                              unsigned cache_policy);
 411
 412 void
 413 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 414                             LLVMValueRef rsrc,
 415                             LLVMValueRef vdata,
 416                             LLVMValueRef voffset,
 417                             LLVMValueRef soffset,
 418                             unsigned cache_policy);
 419
 420 void
 421 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 422                               LLVMValueRef rsrc,
 423                               LLVMValueRef vdata,
 424                               LLVMValueRef vindex,
 425                               LLVMValueRef voffset,
 426                               LLVMValueRef soffset,
 427                               LLVMValueRef immoffset,
 428                               unsigned num_channels,
 429                               unsigned dfmt,
 430                               unsigned nfmt,
 431                               unsigned cache_policy);
 432
 433 void
 434 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 435                            LLVMValueRef rsrc,
 436                            LLVMValueRef vdata,
 437                            LLVMValueRef voffset,
 438                            LLVMValueRef soffset,
 439                            LLVMValueRef immoffset,
 440                            unsigned num_channels,
 441                            unsigned dfmt,
 442                            unsigned nfmt,
 443                            unsigned cache_policy);
 444
 445 LLVMValueRef
 446 ac_get_thread_id(struct ac_llvm_context *ctx);
 447
 448 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 449 #define AC_TID_MASK_TOP      0xfffffffd
 450 #define AC_TID_MASK_LEFT     0xfffffffe
 451
 452 LLVMValueRef
 453 ac_build_ddxy(struct ac_llvm_context *ctx,
 454               uint32_t mask,
 455               int idx,
 456               LLVMValueRef val);
 457
 458 #define AC_SENDMSG_GS 2
 459 #define AC_SENDMSG_GS_DONE 3
 460 #define AC_SENDMSG_GS_ALLOC_REQ 9
 461
 462 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 463 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 464 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 465 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 466
 467 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 468                       uint32_t msg,
 469                       LLVMValueRef wave_id);
 470
 471 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 472                            LLVMValueRef arg,
 473                            LLVMTypeRef dst_type);
 474
 475 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 476                           LLVMValueRef arg,
 477                           LLVMTypeRef dst_type);
 478 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 479                            LLVMValueRef b);
 480 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 481                            LLVMValueRef b);
 482 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 483                            LLVMValueRef b);
 484 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 485                            LLVMValueRef b);
 486 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 487 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 488 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 489
 490 struct ac_export_args {
 491         LLVMValueRef out[4];
 492         unsigned target;
 493         unsigned enabled_channels;
 494         bool compr;
 495         bool done;
 496         bool valid_mask;
 497 };
 498
 499 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 500
 501 void ac_build_export_null(struct ac_llvm_context *ctx);
 502
 503 enum ac_image_opcode {
 504         ac_image_sample,
 505         ac_image_gather4,
 506         ac_image_load,
 507         ac_image_load_mip,
 508         ac_image_store,
 509         ac_image_store_mip,
 510         ac_image_get_lod,
 511         ac_image_get_resinfo,
 512         ac_image_atomic,
 513         ac_image_atomic_cmpswap,
 514 };
 515
 516 enum ac_atomic_op {
 517         ac_atomic_swap,
 518         ac_atomic_add,
 519         ac_atomic_sub,
 520         ac_atomic_smin,
 521         ac_atomic_umin,
 522         ac_atomic_smax,
 523         ac_atomic_umax,
 524         ac_atomic_and,
 525         ac_atomic_or,
 526         ac_atomic_xor,
 527         ac_atomic_inc_wrap,
 528         ac_atomic_dec_wrap,
 529 };
 530
 531 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 532 enum ac_image_cache_policy {
 533         ac_glc = 1 << 0, /* per-CU cache control */
 534         ac_slc = 1 << 1, /* global L2 cache control */
 535         ac_dlc = 1 << 2, /* per-shader-array cache control */
 536 };
 537
 538 struct ac_image_args {
 539         enum ac_image_opcode opcode : 4;
 540         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 541         enum ac_image_dim dim : 3;
 542         unsigned dmask : 4;
 543         unsigned cache_policy : 3;
 544         bool unorm : 1;
 545         bool level_zero : 1;
 546         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 547
 548         LLVMValueRef resource;
 549         LLVMValueRef sampler;
 550         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 551         LLVMValueRef offset;
 552         LLVMValueRef bias;
 553         LLVMValueRef compare;
 554         LLVMValueRef derivs[6];
 555         LLVMValueRef coords[4];
 556         LLVMValueRef lod; // also used by ac_image_get_resinfo
 557 };
 558
 559 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 560                                    struct ac_image_args *a);
 561 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
 562                                              LLVMValueRef rsrc);
 563 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 564                                     LLVMValueRef args[2]);
 565 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 566                                      LLVMValueRef args[2]);
 567 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 568                                      LLVMValueRef args[2]);
 569 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 570                                  LLVMValueRef args[2], unsigned bits, bool hi);
 571 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 572                                  LLVMValueRef args[2], unsigned bits, bool hi);
 573 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 574 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 575 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 576                           LLVMValueRef offset, LLVMValueRef width,
 577                           bool is_signed);
 578 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 579                            LLVMValueRef s1, LLVMValueRef s2);
 580 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 581                            LLVMValueRef s1, LLVMValueRef s2);
 582
 583 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 584
 585 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 586                            unsigned bitsize);
 587
 588 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 589                             LLVMValueRef src1, LLVMValueRef src2,
 590                             unsigned bitsize);
 591
 592 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 593                             unsigned bitsize);
 594
 595 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 596                             unsigned bitsize);
 597
 598 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 599
 600 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 601                                        LLVMValueRef src0);
 602
 603 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 604                             LLVMValueRef main_fn,
 605                             uint8_t *vs_output_param_offset,
 606                             uint32_t num_outputs,
 607                             uint8_t *num_param_exports);
 608 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 609
 610 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 611 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 612                          LLVMValueRef dw_addr);
 613 void ac_lds_store(struct ac_llvm_context *ctx,
 614                   LLVMValueRef dw_addr, LLVMValueRef value);
 615
 616 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 617                          LLVMTypeRef dst_type,
 618                          LLVMValueRef src0);
 619
 620 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 621 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 622
 623 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 624 void ac_build_break(struct ac_llvm_context *ctx);
 625 void ac_build_continue(struct ac_llvm_context *ctx);
 626 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 627 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 628 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 629 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 630 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 631                  int lable_id);
 632 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 633                   int lable_id);
 634
 635 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 636                              const char *name);
 637 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 638                                    const char *name);
 639
 640 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 641                          LLVMTypeRef type);
 642
 643 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 644                             unsigned count);
 645
 646 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 647                              unsigned rshift, unsigned bitwidth);
 648
 649 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 650                               LLVMValueRef *addr, bool is_array_tex);
 651
 652 LLVMValueRef
 653 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 654
 655 LLVMValueRef
 656 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 657
 658 LLVMValueRef
 659 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 660
 661 LLVMValueRef
 662 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 663
 664 LLVMValueRef
 665 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 666
 667 LLVMValueRef
 668 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 669
 670 LLVMValueRef
 671 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 672
 673 /**
 674  * Common arguments for a scan/reduce operation that accumulates per-wave
 675  * values across an entire workgroup, while respecting the order of waves.
 676  */
 677 struct ac_wg_scan {
 678         bool enable_reduce;
 679         bool enable_exclusive;
 680         bool enable_inclusive;
 681         nir_op op;
 682         LLVMValueRef src; /* clobbered! */
 683         LLVMValueRef result_reduce;
 684         LLVMValueRef result_exclusive;
 685         LLVMValueRef result_inclusive;
 686         LLVMValueRef extra;
 687         LLVMValueRef waveidx;
 688         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 689
 690         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 691         LLVMValueRef scratch;
 692         unsigned maxwaves;
 693 };
 694
 695 void
 696 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 697 void
 698 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 699 void
 700 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 701
 702 void
 703 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 704 void
 705 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 706 void
 707 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 708
 709 LLVMValueRef
 710 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 711                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 712
 713 LLVMValueRef
 714 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 715
 716 LLVMValueRef
 717 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 718                    unsigned bitsize);
 719
 720 LLVMValueRef
 721 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 722                     unsigned bitsize);
 723
 724 LLVMValueRef
 725 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
 726                       unsigned bitsize);
 727
 728 LLVMValueRef
 729 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 730
 731 LLVMValueRef
 732 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 733
 734 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 735                            LLVMValueRef *args, unsigned num_args);
 736
 737 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 738                                  LLVMValueRef ptr, LLVMValueRef val,
 739                                  const char *sync_scope);
 740
 741 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 742                                       LLVMValueRef cmp, LLVMValueRef val,
 743                                       const char *sync_scope);
 744
 745 void
 746 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
 747                 LLVMValueRef stencil, LLVMValueRef samplemask,
 748                 struct ac_export_args *args);
 749
 750 static inline LLVMValueRef
 751 ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
 752 {
 753         assert(arg.used);
 754         return LLVMGetParam(ctx->main_function, arg.arg_index);
 755 }
 756
 757 enum ac_llvm_calling_convention {
 758         AC_LLVM_AMDGPU_VS = 87,
 759         AC_LLVM_AMDGPU_GS = 88,
 760         AC_LLVM_AMDGPU_PS = 89,
 761         AC_LLVM_AMDGPU_CS = 90,
 762         AC_LLVM_AMDGPU_HS = 93,
 763 };
 764
 765 LLVMValueRef ac_build_main(const struct ac_shader_args *args,
 766                            struct ac_llvm_context *ctx,
 767                            enum ac_llvm_calling_convention convention,
 768                            const char *name, LLVMTypeRef ret_type,
 769                            LLVMModuleRef module);
 770
 771 #ifdef __cplusplus
 772 }
 773 #endif
 774
 775 #endif