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