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/TargetMachine.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 /* Combine these with & instead of |. */
  47 #define NOOP_WAITCNT    0xcf7f
  48 #define LGKM_CNT        0xc07f
  49 #define EXP_CNT         0xcf0f
  50 #define VM_CNT          0x0f70 /* On GFX9, vmcnt has 6 bits in [0:3] and [14:15] */
  51
  52 struct ac_llvm_flow;
  53
  54 struct ac_llvm_context {
  55         LLVMContextRef context;
  56         LLVMModuleRef module;
  57         LLVMBuilderRef builder;
  58
  59         LLVMTypeRef voidt;
  60         LLVMTypeRef i1;
  61         LLVMTypeRef i8;
  62         LLVMTypeRef i16;
  63         LLVMTypeRef i32;
  64         LLVMTypeRef i64;
  65         LLVMTypeRef intptr;
  66         LLVMTypeRef f16;
  67         LLVMTypeRef f32;
  68         LLVMTypeRef f64;
  69         LLVMTypeRef v2i16;
  70         LLVMTypeRef v2i32;
  71         LLVMTypeRef v3i32;
  72         LLVMTypeRef v4i32;
  73         LLVMTypeRef v2f32;
  74         LLVMTypeRef v4f32;
  75         LLVMTypeRef v8i32;
  76
  77         LLVMValueRef i16_0;
  78         LLVMValueRef i16_1;
  79         LLVMValueRef i32_0;
  80         LLVMValueRef i32_1;
  81         LLVMValueRef i64_0;
  82         LLVMValueRef i64_1;
  83         LLVMValueRef f32_0;
  84         LLVMValueRef f32_1;
  85         LLVMValueRef f64_0;
  86         LLVMValueRef f64_1;
  87         LLVMValueRef i1true;
  88         LLVMValueRef i1false;
  89
  90         struct ac_llvm_flow *flow;
  91         unsigned flow_depth;
  92         unsigned flow_depth_max;
  93
  94         unsigned range_md_kind;
  95         unsigned invariant_load_md_kind;
  96         unsigned uniform_md_kind;
  97         unsigned fpmath_md_kind;
  98         LLVMValueRef fpmath_md_2p5_ulp;
  99         LLVMValueRef empty_md;
 100
 101         enum chip_class chip_class;
 102         enum radeon_family family;
 103
 104         LLVMValueRef lds;
 105 };
 106
 107 void
 108 ac_llvm_context_init(struct ac_llvm_context *ctx,
 109                      enum chip_class chip_class, enum radeon_family family);
 110
 111 void
 112 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 113
 114 int
 115 ac_get_llvm_num_components(LLVMValueRef value);
 116
 117 int
 118 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 119
 120 LLVMValueRef
 121 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 122                      LLVMValueRef value,
 123                      int index);
 124
 125 unsigned ac_get_type_size(LLVMTypeRef type);
 126
 127 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 128 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 129 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 130 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 131 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 132
 133 LLVMValueRef
 134 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 135                    LLVMTypeRef return_type, LLVMValueRef *params,
 136                    unsigned param_count, unsigned attrib_mask);
 137
 138 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 139
 140 LLVMValueRef
 141 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 142              unsigned count_incoming, LLVMValueRef *values,
 143              LLVMBasicBlockRef *blocks);
 144
 145 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 146 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 147                                    LLVMValueRef *pvgpr);
 148
 149 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 150
 151 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 152
 153 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 154
 155 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 156
 157 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 158
 159 LLVMValueRef
 160 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 161                                unsigned value_count, unsigned component);
 162
 163 LLVMValueRef
 164 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 165                                 LLVMValueRef *values,
 166                                 unsigned value_count,
 167                                 unsigned value_stride,
 168                                 bool load,
 169                                 bool always_vector);
 170 LLVMValueRef
 171 ac_build_gather_values(struct ac_llvm_context *ctx,
 172                        LLVMValueRef *values,
 173                        unsigned value_count);
 174 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 175                                      LLVMValueRef value,
 176                                      unsigned num_channels);
 177 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 178
 179 LLVMValueRef
 180 ac_build_fdiv(struct ac_llvm_context *ctx,
 181               LLVMValueRef num,
 182               LLVMValueRef den);
 183
 184 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 185                                 LLVMValueRef num,
 186                                 LLVMValueRef multiplier,
 187                                 LLVMValueRef pre_shift,
 188                                 LLVMValueRef post_shift,
 189                                 LLVMValueRef increment);
 190 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 191                                     LLVMValueRef num,
 192                                     LLVMValueRef multiplier,
 193                                     LLVMValueRef pre_shift,
 194                                     LLVMValueRef post_shift,
 195                                     LLVMValueRef increment);
 196 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 197                                               LLVMValueRef num,
 198                                               LLVMValueRef multiplier,
 199                                               LLVMValueRef post_shift);
 200
 201 void
 202 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 203                        bool is_deriv, bool is_array, bool is_lod,
 204                        LLVMValueRef *coords_arg,
 205                        LLVMValueRef *derivs_arg);
 206
 207
 208 LLVMValueRef
 209 ac_build_fs_interp(struct ac_llvm_context *ctx,
 210                    LLVMValueRef llvm_chan,
 211                    LLVMValueRef attr_number,
 212                    LLVMValueRef params,
 213                    LLVMValueRef i,
 214                    LLVMValueRef j);
 215
 216 LLVMValueRef
 217 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 218                        LLVMValueRef llvm_chan,
 219                        LLVMValueRef attr_number,
 220                        LLVMValueRef params,
 221                        LLVMValueRef i,
 222                        LLVMValueRef j);
 223
 224 LLVMValueRef
 225 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 226                        LLVMValueRef parameter,
 227                        LLVMValueRef llvm_chan,
 228                        LLVMValueRef attr_number,
 229                        LLVMValueRef params);
 230
 231 LLVMValueRef
 232 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 233                  LLVMValueRef base_ptr,
 234                  LLVMValueRef index);
 235
 236 LLVMValueRef
 237 ac_build_gep0(struct ac_llvm_context *ctx,
 238               LLVMValueRef base_ptr,
 239               LLVMValueRef index);
 240 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 241                                   LLVMValueRef index);
 242
 243 void
 244 ac_build_indexed_store(struct ac_llvm_context *ctx,
 245                        LLVMValueRef base_ptr, LLVMValueRef index,
 246                        LLVMValueRef value);
 247
 248 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 249                            LLVMValueRef index);
 250 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 251                                      LLVMValueRef base_ptr, LLVMValueRef index);
 252 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 253                                    LLVMValueRef base_ptr, LLVMValueRef index);
 254 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 255                                    LLVMValueRef base_ptr, LLVMValueRef index);
 256
 257 void
 258 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 259                             LLVMValueRef rsrc,
 260                             LLVMValueRef vdata,
 261                             unsigned num_channels,
 262                             LLVMValueRef voffset,
 263                             LLVMValueRef soffset,
 264                             unsigned inst_offset,
 265                             bool glc,
 266                             bool slc,
 267                             bool writeonly_memory,
 268                             bool swizzle_enable_hint);
 269
 270 void
 271 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 272                              LLVMValueRef rsrc,
 273                              LLVMValueRef data,
 274                              LLVMValueRef vindex,
 275                              LLVMValueRef voffset,
 276                              unsigned num_channels,
 277                              bool glc,
 278                              bool writeonly_memory);
 279
 280 LLVMValueRef
 281 ac_build_buffer_load(struct ac_llvm_context *ctx,
 282                      LLVMValueRef rsrc,
 283                      int num_channels,
 284                      LLVMValueRef vindex,
 285                      LLVMValueRef voffset,
 286                      LLVMValueRef soffset,
 287                      unsigned inst_offset,
 288                      unsigned glc,
 289                      unsigned slc,
 290                      bool can_speculate,
 291                      bool allow_smem);
 292
 293 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 294                                          LLVMValueRef rsrc,
 295                                          LLVMValueRef vindex,
 296                                          LLVMValueRef voffset,
 297                                          unsigned num_channels,
 298                                          bool glc,
 299                                          bool can_speculate);
 300
 301 /* load_format that handles the stride & element count better if idxen is
 302  * disabled by LLVM. */
 303 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 304                                                   LLVMValueRef rsrc,
 305                                                   LLVMValueRef vindex,
 306                                                   LLVMValueRef voffset,
 307                                                   unsigned num_channels,
 308                                                   bool glc,
 309                                                   bool can_speculate);
 310
 311 LLVMValueRef
 312 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 313                             LLVMValueRef rsrc,
 314                             LLVMValueRef voffset,
 315                             LLVMValueRef soffset,
 316                             LLVMValueRef immoffset,
 317                             bool glc);
 318
 319 LLVMValueRef
 320 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 321                              LLVMValueRef rsrc,
 322                              LLVMValueRef vindex,
 323                              LLVMValueRef voffset,
 324                              LLVMValueRef soffset,
 325                              LLVMValueRef immoffset,
 326                              unsigned num_channels,
 327                              unsigned dfmt,
 328                              unsigned nfmt,
 329                              bool glc,
 330                              bool slc,
 331                              bool can_speculate);
 332
 333 LLVMValueRef
 334 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 335                           LLVMValueRef rsrc,
 336                           LLVMValueRef voffset,
 337                           LLVMValueRef soffset,
 338                           LLVMValueRef immoffset,
 339                           unsigned num_channels,
 340                           unsigned dfmt,
 341                           unsigned nfmt,
 342                           bool glc,
 343                           bool slc,
 344                           bool can_speculate);
 345
 346 LLVMValueRef
 347 ac_get_thread_id(struct ac_llvm_context *ctx);
 348
 349 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 350 #define AC_TID_MASK_TOP      0xfffffffd
 351 #define AC_TID_MASK_LEFT     0xfffffffe
 352
 353 LLVMValueRef
 354 ac_build_ddxy(struct ac_llvm_context *ctx,
 355               uint32_t mask,
 356               int idx,
 357               LLVMValueRef val);
 358
 359 #define AC_SENDMSG_GS 2
 360 #define AC_SENDMSG_GS_DONE 3
 361
 362 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 363 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 364 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 365 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 366
 367 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 368                       uint32_t msg,
 369                       LLVMValueRef wave_id);
 370
 371 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 372                            LLVMValueRef arg,
 373                            LLVMTypeRef dst_type);
 374
 375 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 376                           LLVMValueRef arg,
 377                           LLVMTypeRef dst_type);
 378 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 379                            LLVMValueRef b);
 380 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 381                            LLVMValueRef b);
 382 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 383                            LLVMValueRef b);
 384 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 385                            LLVMValueRef b);
 386 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 387 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 388
 389 struct ac_export_args {
 390         LLVMValueRef out[4];
 391         unsigned target;
 392         unsigned enabled_channels;
 393         bool compr;
 394         bool done;
 395         bool valid_mask;
 396 };
 397
 398 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 399
 400 void ac_build_export_null(struct ac_llvm_context *ctx);
 401
 402 enum ac_image_opcode {
 403         ac_image_sample,
 404         ac_image_gather4,
 405         ac_image_load,
 406         ac_image_load_mip,
 407         ac_image_store,
 408         ac_image_store_mip,
 409         ac_image_get_lod,
 410         ac_image_get_resinfo,
 411         ac_image_atomic,
 412         ac_image_atomic_cmpswap,
 413 };
 414
 415 enum ac_atomic_op {
 416         ac_atomic_swap,
 417         ac_atomic_add,
 418         ac_atomic_sub,
 419         ac_atomic_smin,
 420         ac_atomic_umin,
 421         ac_atomic_smax,
 422         ac_atomic_umax,
 423         ac_atomic_and,
 424         ac_atomic_or,
 425         ac_atomic_xor,
 426 };
 427
 428 enum ac_image_dim {
 429         ac_image_1d,
 430         ac_image_2d,
 431         ac_image_3d,
 432         ac_image_cube, // includes cube arrays
 433         ac_image_1darray,
 434         ac_image_2darray,
 435         ac_image_2dmsaa,
 436         ac_image_2darraymsaa,
 437 };
 438
 439 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 440 enum ac_image_cache_policy {
 441         ac_glc = 1 << 0,
 442         ac_slc = 1 << 1,
 443 };
 444
 445 struct ac_image_args {
 446         enum ac_image_opcode opcode : 4;
 447         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 448         enum ac_image_dim dim : 3;
 449         unsigned dmask : 4;
 450         unsigned cache_policy : 2;
 451         bool unorm : 1;
 452         bool level_zero : 1;
 453         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 454
 455         LLVMValueRef resource;
 456         LLVMValueRef sampler;
 457         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 458         LLVMValueRef offset;
 459         LLVMValueRef bias;
 460         LLVMValueRef compare;
 461         LLVMValueRef derivs[6];
 462         LLVMValueRef coords[4];
 463         LLVMValueRef lod; // also used by ac_image_get_resinfo
 464 };
 465
 466 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 467                                    struct ac_image_args *a);
 468 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 469                                     LLVMValueRef args[2]);
 470 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 471                                      LLVMValueRef args[2]);
 472 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 473                                      LLVMValueRef args[2]);
 474 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 475                                  LLVMValueRef args[2], unsigned bits, bool hi);
 476 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 477                                  LLVMValueRef args[2], unsigned bits, bool hi);
 478 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 479 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 480 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 481                           LLVMValueRef offset, LLVMValueRef width,
 482                           bool is_signed);
 483 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 484                            LLVMValueRef s1, LLVMValueRef s2);
 485 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 486                            LLVMValueRef s1, LLVMValueRef s2);
 487
 488 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
 489
 490 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 491                            unsigned bitsize);
 492
 493 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 494                             unsigned bitsize);
 495
 496 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 497                             unsigned bitsize);
 498
 499 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 500
 501 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 502                                        LLVMValueRef src0);
 503
 504 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 505                             LLVMValueRef main_fn,
 506                             uint8_t *vs_output_param_offset,
 507                             uint32_t num_outputs,
 508                             uint8_t *num_param_exports);
 509 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 510
 511 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 512 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 513                          LLVMValueRef dw_addr);
 514 void ac_lds_store(struct ac_llvm_context *ctx,
 515                   LLVMValueRef dw_addr, LLVMValueRef value);
 516
 517 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 518                          LLVMTypeRef dst_type,
 519                          LLVMValueRef src0);
 520
 521 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 522 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 523
 524 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 525 void ac_build_break(struct ac_llvm_context *ctx);
 526 void ac_build_continue(struct ac_llvm_context *ctx);
 527 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 528 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 529 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 530 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 531 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 532                  int lable_id);
 533 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 534                   int lable_id);
 535
 536 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 537                              const char *name);
 538 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 539                                    const char *name);
 540
 541 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 542                          LLVMTypeRef type);
 543
 544 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 545                             unsigned count);
 546
 547 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 548                              unsigned rshift, unsigned bitwidth);
 549
 550 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 551                               LLVMValueRef *addr, bool is_array_tex);
 552
 553 LLVMValueRef
 554 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 555
 556 LLVMValueRef
 557 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 558
 559 LLVMValueRef
 560 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 561
 562 LLVMValueRef
 563 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 564
 565 LLVMValueRef
 566 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 567
 568 LLVMValueRef
 569 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 570
 571 LLVMValueRef
 572 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 573
 574 /**
 575  * Common arguments for a scan/reduce operation that accumulates per-wave
 576  * values across an entire workgroup, while respecting the order of waves.
 577  */
 578 struct ac_wg_scan {
 579         bool enable_reduce;
 580         bool enable_exclusive;
 581         bool enable_inclusive;
 582         nir_op op;
 583         LLVMValueRef src; /* clobbered! */
 584         LLVMValueRef result_reduce;
 585         LLVMValueRef result_exclusive;
 586         LLVMValueRef result_inclusive;
 587         LLVMValueRef extra;
 588         LLVMValueRef waveidx;
 589         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 590
 591         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 592         LLVMValueRef scratch;
 593         unsigned maxwaves;
 594 };
 595
 596 void
 597 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 598 void
 599 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 600 void
 601 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 602
 603 void
 604 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 605 void
 606 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 607 void
 608 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 609
 610 LLVMValueRef
 611 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 612                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 613
 614 LLVMValueRef
 615 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 616
 617 #ifdef __cplusplus
 618 }
 619 #endif
 620
 621 #endif