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 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 194                                      LLVMValueRef value,
 195                                      unsigned num_channels);
 196 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 197
 198 LLVMValueRef
 199 ac_build_fdiv(struct ac_llvm_context *ctx,
 200               LLVMValueRef num,
 201               LLVMValueRef den);
 202
 203 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 204                                 LLVMValueRef num,
 205                                 LLVMValueRef multiplier,
 206                                 LLVMValueRef pre_shift,
 207                                 LLVMValueRef post_shift,
 208                                 LLVMValueRef increment);
 209 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 210                                     LLVMValueRef num,
 211                                     LLVMValueRef multiplier,
 212                                     LLVMValueRef pre_shift,
 213                                     LLVMValueRef post_shift,
 214                                     LLVMValueRef increment);
 215 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 216                                               LLVMValueRef num,
 217                                               LLVMValueRef multiplier,
 218                                               LLVMValueRef post_shift);
 219
 220 void
 221 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 222                        bool is_deriv, bool is_array, bool is_lod,
 223                        LLVMValueRef *coords_arg,
 224                        LLVMValueRef *derivs_arg);
 225
 226
 227 LLVMValueRef
 228 ac_build_fs_interp(struct ac_llvm_context *ctx,
 229                    LLVMValueRef llvm_chan,
 230                    LLVMValueRef attr_number,
 231                    LLVMValueRef params,
 232                    LLVMValueRef i,
 233                    LLVMValueRef j);
 234
 235 LLVMValueRef
 236 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 237                        LLVMValueRef llvm_chan,
 238                        LLVMValueRef attr_number,
 239                        LLVMValueRef params,
 240                        LLVMValueRef i,
 241                        LLVMValueRef j);
 242
 243 LLVMValueRef
 244 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 245                        LLVMValueRef parameter,
 246                        LLVMValueRef llvm_chan,
 247                        LLVMValueRef attr_number,
 248                        LLVMValueRef params);
 249
 250 LLVMValueRef
 251 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 252                  LLVMValueRef base_ptr,
 253                  LLVMValueRef index);
 254
 255 LLVMValueRef
 256 ac_build_gep0(struct ac_llvm_context *ctx,
 257               LLVMValueRef base_ptr,
 258               LLVMValueRef index);
 259 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 260                                   LLVMValueRef index);
 261
 262 void
 263 ac_build_indexed_store(struct ac_llvm_context *ctx,
 264                        LLVMValueRef base_ptr, LLVMValueRef index,
 265                        LLVMValueRef value);
 266
 267 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 268                            LLVMValueRef index);
 269 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 270                                      LLVMValueRef base_ptr, LLVMValueRef index);
 271 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 272                                    LLVMValueRef base_ptr, LLVMValueRef index);
 273 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 274                                    LLVMValueRef base_ptr, LLVMValueRef index);
 275
 276 void
 277 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 278                             LLVMValueRef rsrc,
 279                             LLVMValueRef vdata,
 280                             unsigned num_channels,
 281                             LLVMValueRef voffset,
 282                             LLVMValueRef soffset,
 283                             unsigned inst_offset,
 284                             unsigned cache_policy,
 285                             bool swizzle_enable_hint);
 286
 287 void
 288 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 289                              LLVMValueRef rsrc,
 290                              LLVMValueRef data,
 291                              LLVMValueRef vindex,
 292                              LLVMValueRef voffset,
 293                              unsigned num_channels,
 294                              unsigned cache_policy);
 295
 296 LLVMValueRef
 297 ac_build_buffer_load(struct ac_llvm_context *ctx,
 298                      LLVMValueRef rsrc,
 299                      int num_channels,
 300                      LLVMValueRef vindex,
 301                      LLVMValueRef voffset,
 302                      LLVMValueRef soffset,
 303                      unsigned inst_offset,
 304                      unsigned cache_policy,
 305                      bool can_speculate,
 306                      bool allow_smem);
 307
 308 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 309                                          LLVMValueRef rsrc,
 310                                          LLVMValueRef vindex,
 311                                          LLVMValueRef voffset,
 312                                          unsigned num_channels,
 313                                          unsigned cache_policy,
 314                                          bool can_speculate);
 315
 316 /* load_format that handles the stride & element count better if idxen is
 317  * disabled by LLVM. */
 318 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 319                                                   LLVMValueRef rsrc,
 320                                                   LLVMValueRef vindex,
 321                                                   LLVMValueRef voffset,
 322                                                   unsigned num_channels,
 323                                                   unsigned cache_policy,
 324                                                   bool can_speculate);
 325
 326 LLVMValueRef
 327 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 328                             LLVMValueRef rsrc,
 329                             LLVMValueRef voffset,
 330                             LLVMValueRef soffset,
 331                             LLVMValueRef immoffset,
 332                             unsigned cache_policy);
 333
 334 LLVMValueRef
 335 ac_build_tbuffer_load_byte(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_struct_tbuffer_load(struct ac_llvm_context *ctx,
 344                              LLVMValueRef rsrc,
 345                              LLVMValueRef vindex,
 346                              LLVMValueRef voffset,
 347                              LLVMValueRef soffset,
 348                              LLVMValueRef immoffset,
 349                              unsigned num_channels,
 350                              unsigned dfmt,
 351                              unsigned nfmt,
 352                              unsigned cache_policy,
 353                              bool can_speculate);
 354
 355 LLVMValueRef
 356 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 357                           LLVMValueRef rsrc,
 358                           LLVMValueRef voffset,
 359                           LLVMValueRef soffset,
 360                           LLVMValueRef immoffset,
 361                           unsigned num_channels,
 362                           unsigned dfmt,
 363                           unsigned nfmt,
 364                           unsigned cache_policy,
 365                           bool can_speculate);
 366
 367 /* For ac_build_fetch_format.
 368  *
 369  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 370  */
 371 enum {
 372         AC_FETCH_FORMAT_FLOAT = 0,
 373         AC_FETCH_FORMAT_FIXED,
 374         AC_FETCH_FORMAT_UNORM,
 375         AC_FETCH_FORMAT_SNORM,
 376         AC_FETCH_FORMAT_USCALED,
 377         AC_FETCH_FORMAT_SSCALED,
 378         AC_FETCH_FORMAT_UINT,
 379         AC_FETCH_FORMAT_SINT,
 380 };
 381
 382 LLVMValueRef
 383 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 384                                unsigned log_size,
 385                                unsigned num_channels,
 386                                unsigned format,
 387                                bool reverse,
 388                                bool known_aligned,
 389                                LLVMValueRef rsrc,
 390                                LLVMValueRef vindex,
 391                                LLVMValueRef voffset,
 392                                LLVMValueRef soffset,
 393                                unsigned cache_policy,
 394                                bool can_speculate);
 395
 396 void
 397 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 398                              LLVMValueRef rsrc,
 399                              LLVMValueRef vdata,
 400                              LLVMValueRef voffset,
 401                              LLVMValueRef soffset,
 402                              unsigned cache_policy);
 403
 404 void
 405 ac_build_tbuffer_store_byte(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_struct_tbuffer_store(struct ac_llvm_context *ctx,
 414                               LLVMValueRef rsrc,
 415                               LLVMValueRef vdata,
 416                               LLVMValueRef vindex,
 417                               LLVMValueRef voffset,
 418                               LLVMValueRef soffset,
 419                               LLVMValueRef immoffset,
 420                               unsigned num_channels,
 421                               unsigned dfmt,
 422                               unsigned nfmt,
 423                               unsigned cache_policy);
 424
 425 void
 426 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 427                            LLVMValueRef rsrc,
 428                            LLVMValueRef vdata,
 429                            LLVMValueRef voffset,
 430                            LLVMValueRef soffset,
 431                            LLVMValueRef immoffset,
 432                            unsigned num_channels,
 433                            unsigned dfmt,
 434                            unsigned nfmt,
 435                            unsigned cache_policy);
 436
 437 LLVMValueRef
 438 ac_get_thread_id(struct ac_llvm_context *ctx);
 439
 440 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 441 #define AC_TID_MASK_TOP      0xfffffffd
 442 #define AC_TID_MASK_LEFT     0xfffffffe
 443
 444 LLVMValueRef
 445 ac_build_ddxy(struct ac_llvm_context *ctx,
 446               uint32_t mask,
 447               int idx,
 448               LLVMValueRef val);
 449
 450 #define AC_SENDMSG_GS 2
 451 #define AC_SENDMSG_GS_DONE 3
 452 #define AC_SENDMSG_GS_ALLOC_REQ 9
 453
 454 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 455 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 456 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 457 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 458
 459 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 460                       uint32_t msg,
 461                       LLVMValueRef wave_id);
 462
 463 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 464                            LLVMValueRef arg,
 465                            LLVMTypeRef dst_type);
 466
 467 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 468                           LLVMValueRef arg,
 469                           LLVMTypeRef dst_type);
 470 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 471                            LLVMValueRef b);
 472 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 473                            LLVMValueRef b);
 474 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 475                            LLVMValueRef b);
 476 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 477                            LLVMValueRef b);
 478 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 479 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 480 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 481
 482 struct ac_export_args {
 483         LLVMValueRef out[4];
 484         unsigned target;
 485         unsigned enabled_channels;
 486         bool compr;
 487         bool done;
 488         bool valid_mask;
 489 };
 490
 491 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 492
 493 void ac_build_export_null(struct ac_llvm_context *ctx);
 494
 495 enum ac_image_opcode {
 496         ac_image_sample,
 497         ac_image_gather4,
 498         ac_image_load,
 499         ac_image_load_mip,
 500         ac_image_store,
 501         ac_image_store_mip,
 502         ac_image_get_lod,
 503         ac_image_get_resinfo,
 504         ac_image_atomic,
 505         ac_image_atomic_cmpswap,
 506 };
 507
 508 enum ac_atomic_op {
 509         ac_atomic_swap,
 510         ac_atomic_add,
 511         ac_atomic_sub,
 512         ac_atomic_smin,
 513         ac_atomic_umin,
 514         ac_atomic_smax,
 515         ac_atomic_umax,
 516         ac_atomic_and,
 517         ac_atomic_or,
 518         ac_atomic_xor,
 519 };
 520
 521 enum ac_image_dim {
 522         ac_image_1d,
 523         ac_image_2d,
 524         ac_image_3d,
 525         ac_image_cube, // includes cube arrays
 526         ac_image_1darray,
 527         ac_image_2darray,
 528         ac_image_2dmsaa,
 529         ac_image_2darraymsaa,
 530 };
 531
 532 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 533 enum ac_image_cache_policy {
 534         ac_glc = 1 << 0, /* per-CU cache control */
 535         ac_slc = 1 << 1, /* global L2 cache control */
 536         ac_dlc = 1 << 2, /* per-shader-array cache control */
 537 };
 538
 539 struct ac_image_args {
 540         enum ac_image_opcode opcode : 4;
 541         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 542         enum ac_image_dim dim : 3;
 543         unsigned dmask : 4;
 544         unsigned cache_policy : 3;
 545         bool unorm : 1;
 546         bool level_zero : 1;
 547         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 548
 549         LLVMValueRef resource;
 550         LLVMValueRef sampler;
 551         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 552         LLVMValueRef offset;
 553         LLVMValueRef bias;
 554         LLVMValueRef compare;
 555         LLVMValueRef derivs[6];
 556         LLVMValueRef coords[4];
 557         LLVMValueRef lod; // also used by ac_image_get_resinfo
 558 };
 559
 560 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 561                                    struct ac_image_args *a);
 562 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 563                                     LLVMValueRef args[2]);
 564 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 565                                      LLVMValueRef args[2]);
 566 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 567                                      LLVMValueRef args[2]);
 568 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 569                                  LLVMValueRef args[2], unsigned bits, bool hi);
 570 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 571                                  LLVMValueRef args[2], unsigned bits, bool hi);
 572 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 573 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 574 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 575                           LLVMValueRef offset, LLVMValueRef width,
 576                           bool is_signed);
 577 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 578                            LLVMValueRef s1, LLVMValueRef s2);
 579 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 580                            LLVMValueRef s1, LLVMValueRef s2);
 581
 582 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 583
 584 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 585                            unsigned bitsize);
 586
 587 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 588                             LLVMValueRef src1, LLVMValueRef src2,
 589                             unsigned bitsize);
 590
 591 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 592                             unsigned bitsize);
 593
 594 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 595                             unsigned bitsize);
 596
 597 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 598
 599 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 600                                        LLVMValueRef src0);
 601
 602 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 603                             LLVMValueRef main_fn,
 604                             uint8_t *vs_output_param_offset,
 605                             uint32_t num_outputs,
 606                             uint8_t *num_param_exports);
 607 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 608
 609 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 610 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 611                          LLVMValueRef dw_addr);
 612 void ac_lds_store(struct ac_llvm_context *ctx,
 613                   LLVMValueRef dw_addr, LLVMValueRef value);
 614
 615 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 616                          LLVMTypeRef dst_type,
 617                          LLVMValueRef src0);
 618
 619 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 620 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 621
 622 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 623 void ac_build_break(struct ac_llvm_context *ctx);
 624 void ac_build_continue(struct ac_llvm_context *ctx);
 625 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 626 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 627 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 628 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 629 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 630                  int lable_id);
 631 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 632                   int lable_id);
 633
 634 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 635                              const char *name);
 636 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 637                                    const char *name);
 638
 639 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 640                          LLVMTypeRef type);
 641
 642 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 643                             unsigned count);
 644
 645 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 646                              unsigned rshift, unsigned bitwidth);
 647
 648 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 649                               LLVMValueRef *addr, bool is_array_tex);
 650
 651 LLVMValueRef
 652 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 653
 654 LLVMValueRef
 655 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 656
 657 LLVMValueRef
 658 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 659
 660 LLVMValueRef
 661 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 662
 663 LLVMValueRef
 664 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 665
 666 LLVMValueRef
 667 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 668
 669 LLVMValueRef
 670 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 671
 672 /**
 673  * Common arguments for a scan/reduce operation that accumulates per-wave
 674  * values across an entire workgroup, while respecting the order of waves.
 675  */
 676 struct ac_wg_scan {
 677         bool enable_reduce;
 678         bool enable_exclusive;
 679         bool enable_inclusive;
 680         nir_op op;
 681         LLVMValueRef src; /* clobbered! */
 682         LLVMValueRef result_reduce;
 683         LLVMValueRef result_exclusive;
 684         LLVMValueRef result_inclusive;
 685         LLVMValueRef extra;
 686         LLVMValueRef waveidx;
 687         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 688
 689         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 690         LLVMValueRef scratch;
 691         unsigned maxwaves;
 692 };
 693
 694 void
 695 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 696 void
 697 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 698 void
 699 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 700
 701 void
 702 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 703 void
 704 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 705 void
 706 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 707
 708 LLVMValueRef
 709 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 710                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 711
 712 LLVMValueRef
 713 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 714
 715 LLVMValueRef
 716 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 717                    unsigned bitsize);
 718
 719 LLVMValueRef
 720 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 721                     unsigned bitsize);
 722
 723 LLVMValueRef
 724 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 725
 726 LLVMValueRef
 727 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 728
 729 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 730                            LLVMValueRef *args, unsigned num_args);
 731
 732 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 733                                  LLVMValueRef ptr, LLVMValueRef val,
 734                                  const char *sync_scope);
 735
 736 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 737                                       LLVMValueRef cmp, LLVMValueRef val,
 738                                       const char *sync_scope);
 739
 740 #ifdef __cplusplus
 741 }
 742 #endif
 743
 744 #endif