src/amd/common/ac_llvm_build.h

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