src/amd/llvm/ac_llvm_build.h

   1 /*
   2  * Copyright 2016 Bas Nieuwenhuizen
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  13  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  14  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  15  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  16  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  17  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  18  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  19  *
  20  * The above copyright notice and this permission notice (including the
  21  * next paragraph) shall be included in all copies or substantial portions
  22  * of the Software.
  23  *
  24  */
  25 #ifndef AC_LLVM_BUILD_H
  26 #define AC_LLVM_BUILD_H
  27
  28 #include <stdbool.h>
  29 #include <llvm-c/Core.h>
  30 #include "compiler/nir/nir.h"
  31 #include "amd_family.h"
  32 #include "ac_shader_util.h"
  33 #include "ac_shader_args.h"
  34 #include "ac_shader_abi.h"
  35
  36 #ifdef __cplusplus
  37 extern "C" {
  38 #endif
  39
  40 enum {
  41         AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
  42         AC_ADDR_SPACE_GLOBAL = 1,
  43         AC_ADDR_SPACE_GDS = 2,
  44         AC_ADDR_SPACE_LDS = 3,
  45         AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
  46         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  47 };
  48
  49 #define AC_WAIT_LGKM    (1 << 0) /* LDS, GDS, constant, message */
  50 #define AC_WAIT_VLOAD   (1 << 1) /* VMEM load/sample instructions */
  51 #define AC_WAIT_VSTORE  (1 << 2) /* VMEM store instructions */
  52
  53 struct ac_llvm_flow;
  54 struct ac_llvm_compiler;
  55 enum ac_float_mode;
  56
  57 struct ac_llvm_flow_state {
  58         struct ac_llvm_flow *stack;
  59         unsigned depth_max;
  60         unsigned depth;
  61 };
  62
  63 struct ac_llvm_context {
  64         LLVMContextRef context;
  65         LLVMModuleRef module;
  66         LLVMBuilderRef builder;
  67
  68         LLVMValueRef main_function;
  69
  70         LLVMTypeRef voidt;
  71         LLVMTypeRef i1;
  72         LLVMTypeRef i8;
  73         LLVMTypeRef i16;
  74         LLVMTypeRef i32;
  75         LLVMTypeRef i64;
  76         LLVMTypeRef intptr;
  77         LLVMTypeRef f16;
  78         LLVMTypeRef f32;
  79         LLVMTypeRef f64;
  80         LLVMTypeRef v2i16;
  81         LLVMTypeRef v2i32;
  82         LLVMTypeRef v3i32;
  83         LLVMTypeRef v4i32;
  84         LLVMTypeRef v2f32;
  85         LLVMTypeRef v3f32;
  86         LLVMTypeRef v4f32;
  87         LLVMTypeRef v8i32;
  88         LLVMTypeRef iN_wavemask;
  89         LLVMTypeRef iN_ballotmask;
  90
  91         LLVMValueRef i8_0;
  92         LLVMValueRef i8_1;
  93         LLVMValueRef i16_0;
  94         LLVMValueRef i16_1;
  95         LLVMValueRef i32_0;
  96         LLVMValueRef i32_1;
  97         LLVMValueRef i64_0;
  98         LLVMValueRef i64_1;
  99         LLVMValueRef f16_0;
 100         LLVMValueRef f16_1;
 101         LLVMValueRef f32_0;
 102         LLVMValueRef f32_1;
 103         LLVMValueRef f64_0;
 104         LLVMValueRef f64_1;
 105         LLVMValueRef i1true;
 106         LLVMValueRef i1false;
 107
 108         /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
 109          * are two ac_llvm_contexts. Declare a pointer here, so that the control
 110          * flow stack is shared by both ac_llvm_contexts.
 111          */
 112         struct ac_llvm_flow_state *flow;
 113
 114         unsigned range_md_kind;
 115         unsigned invariant_load_md_kind;
 116         unsigned uniform_md_kind;
 117         unsigned fpmath_md_kind;
 118         LLVMValueRef fpmath_md_2p5_ulp;
 119         LLVMValueRef empty_md;
 120
 121         enum chip_class chip_class;
 122         enum radeon_family family;
 123
 124         unsigned wave_size;
 125         unsigned ballot_mask_bits;
 126
 127         unsigned float_mode;
 128
 129         LLVMValueRef lds;
 130 };
 131
 132 void
 133 ac_llvm_context_init(struct ac_llvm_context *ctx,
 134                      struct ac_llvm_compiler *compiler,
 135                      enum chip_class chip_class, enum radeon_family family,
 136                      enum ac_float_mode float_mode, unsigned wave_size,
 137                      unsigned ballot_mask_bits);
 138
 139 void
 140 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 141
 142 int
 143 ac_get_llvm_num_components(LLVMValueRef value);
 144
 145 int
 146 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 147
 148 LLVMValueRef
 149 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 150                      LLVMValueRef value,
 151                      int index);
 152
 153 unsigned ac_get_type_size(LLVMTypeRef type);
 154
 155 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 156 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 157 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 158 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 159 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 160
 161 LLVMValueRef
 162 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 163                    LLVMTypeRef return_type, LLVMValueRef *params,
 164                    unsigned param_count, unsigned attrib_mask);
 165
 166 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 167
 168 LLVMValueRef
 169 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 170              unsigned count_incoming, LLVMValueRef *values,
 171              LLVMBasicBlockRef *blocks);
 172
 173 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 174 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 175                                    LLVMValueRef *pvgpr);
 176
 177 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 178
 179 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 180 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 181                                  LLVMValueRef value);
 182
 183 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 184
 185 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 186
 187 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 188
 189 LLVMValueRef
 190 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 191                                unsigned value_count, unsigned component);
 192
 193 LLVMValueRef
 194 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 195                                 LLVMValueRef *values,
 196                                 unsigned value_count,
 197                                 unsigned value_stride,
 198                                 bool load,
 199                                 bool always_vector);
 200 LLVMValueRef
 201 ac_build_gather_values(struct ac_llvm_context *ctx,
 202                        LLVMValueRef *values,
 203                        unsigned value_count);
 204
 205 LLVMValueRef
 206 ac_extract_components(struct ac_llvm_context *ctx,
 207                       LLVMValueRef value,
 208                       unsigned start,
 209                       unsigned channels);
 210
 211 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 212                                      LLVMValueRef value,
 213                                      unsigned num_channels);
 214 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 215
 216 LLVMValueRef
 217 ac_build_fdiv(struct ac_llvm_context *ctx,
 218               LLVMValueRef num,
 219               LLVMValueRef den);
 220
 221 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 222                                 LLVMValueRef num,
 223                                 LLVMValueRef multiplier,
 224                                 LLVMValueRef pre_shift,
 225                                 LLVMValueRef post_shift,
 226                                 LLVMValueRef increment);
 227 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 228                                     LLVMValueRef num,
 229                                     LLVMValueRef multiplier,
 230                                     LLVMValueRef pre_shift,
 231                                     LLVMValueRef post_shift,
 232                                     LLVMValueRef increment);
 233 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 234                                               LLVMValueRef num,
 235                                               LLVMValueRef multiplier,
 236                                               LLVMValueRef post_shift);
 237
 238 void
 239 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 240                        bool is_deriv, bool is_array, bool is_lod,
 241                        LLVMValueRef *coords_arg,
 242                        LLVMValueRef *derivs_arg);
 243
 244
 245 LLVMValueRef
 246 ac_build_fs_interp(struct ac_llvm_context *ctx,
 247                    LLVMValueRef llvm_chan,
 248                    LLVMValueRef attr_number,
 249                    LLVMValueRef params,
 250                    LLVMValueRef i,
 251                    LLVMValueRef j);
 252
 253 LLVMValueRef
 254 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 255                        LLVMValueRef llvm_chan,
 256                        LLVMValueRef attr_number,
 257                        LLVMValueRef params,
 258                        LLVMValueRef i,
 259                        LLVMValueRef j);
 260
 261 LLVMValueRef
 262 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 263                        LLVMValueRef parameter,
 264                        LLVMValueRef llvm_chan,
 265                        LLVMValueRef attr_number,
 266                        LLVMValueRef params);
 267
 268 LLVMValueRef
 269 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 270                  LLVMValueRef base_ptr,
 271                  LLVMValueRef index);
 272
 273 LLVMValueRef
 274 ac_build_gep0(struct ac_llvm_context *ctx,
 275               LLVMValueRef base_ptr,
 276               LLVMValueRef index);
 277 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 278                                   LLVMValueRef index);
 279
 280 void
 281 ac_build_indexed_store(struct ac_llvm_context *ctx,
 282                        LLVMValueRef base_ptr, LLVMValueRef index,
 283                        LLVMValueRef value);
 284
 285 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 286                            LLVMValueRef index);
 287 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 288                                      LLVMValueRef base_ptr, LLVMValueRef index);
 289 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 290                                    LLVMValueRef base_ptr, LLVMValueRef index);
 291 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 292                                    LLVMValueRef base_ptr, LLVMValueRef index);
 293
 294 void
 295 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 296                             LLVMValueRef rsrc,
 297                             LLVMValueRef vdata,
 298                             unsigned num_channels,
 299                             LLVMValueRef voffset,
 300                             LLVMValueRef soffset,
 301                             unsigned inst_offset,
 302                             unsigned cache_policy);
 303
 304 void
 305 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 306                              LLVMValueRef rsrc,
 307                              LLVMValueRef data,
 308                              LLVMValueRef vindex,
 309                              LLVMValueRef voffset,
 310                              unsigned num_channels,
 311                              unsigned cache_policy);
 312
 313 LLVMValueRef
 314 ac_build_buffer_load(struct ac_llvm_context *ctx,
 315                      LLVMValueRef rsrc,
 316                      int num_channels,
 317                      LLVMValueRef vindex,
 318                      LLVMValueRef voffset,
 319                      LLVMValueRef soffset,
 320                      unsigned inst_offset,
 321                      unsigned cache_policy,
 322                      bool can_speculate,
 323                      bool allow_smem);
 324
 325 LLVMValueRef ac_build_buffer_load_format(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 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 531 enum ac_image_cache_policy {
 532         ac_glc = 1 << 0, /* per-CU cache control */
 533         ac_slc = 1 << 1, /* global L2 cache control */
 534         ac_dlc = 1 << 2, /* per-shader-array cache control */
 535         ac_swizzled = 1 << 3, /* the access is swizzled, disabling load/store merging */
 536 };
 537
 538 struct ac_image_args {
 539         enum ac_image_opcode opcode : 4;
 540         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 541         enum ac_image_dim dim : 3;
 542         unsigned dmask : 4;
 543         unsigned cache_policy : 3;
 544         bool unorm : 1;
 545         bool level_zero : 1;
 546         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 547
 548         LLVMValueRef resource;
 549         LLVMValueRef sampler;
 550         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 551         LLVMValueRef offset;
 552         LLVMValueRef bias;
 553         LLVMValueRef compare;
 554         LLVMValueRef derivs[6];
 555         LLVMValueRef coords[4];
 556         LLVMValueRef lod; // also used by ac_image_get_resinfo
 557 };
 558
 559 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 560                                    struct ac_image_args *a);
 561 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
 562                                              LLVMValueRef rsrc);
 563 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 564                                     LLVMValueRef args[2]);
 565 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 566                                      LLVMValueRef args[2]);
 567 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 568                                      LLVMValueRef args[2]);
 569 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 570                                  LLVMValueRef args[2], unsigned bits, bool hi);
 571 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 572                                  LLVMValueRef args[2], unsigned bits, bool hi);
 573 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 574 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 575 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 576                           LLVMValueRef offset, LLVMValueRef width,
 577                           bool is_signed);
 578 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 579                            LLVMValueRef s1, LLVMValueRef s2);
 580 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 581                            LLVMValueRef s1, LLVMValueRef s2);
 582
 583 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 584
 585 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 586                            unsigned bitsize);
 587
 588 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 589                             LLVMValueRef src1, LLVMValueRef src2,
 590                             unsigned bitsize);
 591
 592 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 593                             unsigned bitsize);
 594
 595 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 596                             unsigned bitsize);
 597
 598 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 599
 600 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 601                                        LLVMValueRef src0);
 602
 603 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 604                             LLVMValueRef main_fn,
 605                             uint8_t *vs_output_param_offset,
 606                             uint32_t num_outputs,
 607                             uint8_t *num_param_exports);
 608 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 609
 610 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 611 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 612                          LLVMValueRef dw_addr);
 613 void ac_lds_store(struct ac_llvm_context *ctx,
 614                   LLVMValueRef dw_addr, LLVMValueRef value);
 615
 616 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 617                          LLVMTypeRef dst_type,
 618                          LLVMValueRef src0);
 619
 620 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 621 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 622
 623 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 624 void ac_build_break(struct ac_llvm_context *ctx);
 625 void ac_build_continue(struct ac_llvm_context *ctx);
 626 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 627 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 628 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 629 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 630 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 631                  int lable_id);
 632 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 633                   int lable_id);
 634
 635 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 636                              const char *name);
 637 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 638                                    const char *name);
 639
 640 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 641                          LLVMTypeRef type);
 642
 643 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 644                             unsigned count);
 645
 646 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 647                              unsigned rshift, unsigned bitwidth);
 648
 649 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 650                               LLVMValueRef *addr, bool is_array_tex);
 651
 652 LLVMValueRef
 653 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 654
 655 LLVMValueRef
 656 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 657
 658 LLVMValueRef
 659 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 660
 661 LLVMValueRef
 662 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 663
 664 LLVMValueRef
 665 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 666
 667 LLVMValueRef
 668 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 669
 670 LLVMValueRef
 671 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 672
 673 /**
 674  * Common arguments for a scan/reduce operation that accumulates per-wave
 675  * values across an entire workgroup, while respecting the order of waves.
 676  */
 677 struct ac_wg_scan {
 678         bool enable_reduce;
 679         bool enable_exclusive;
 680         bool enable_inclusive;
 681         nir_op op;
 682         LLVMValueRef src; /* clobbered! */
 683         LLVMValueRef result_reduce;
 684         LLVMValueRef result_exclusive;
 685         LLVMValueRef result_inclusive;
 686         LLVMValueRef extra;
 687         LLVMValueRef waveidx;
 688         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 689
 690         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 691         LLVMValueRef scratch;
 692         unsigned maxwaves;
 693 };
 694
 695 void
 696 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 697 void
 698 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 699 void
 700 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 701
 702 void
 703 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 704 void
 705 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 706 void
 707 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 708
 709 LLVMValueRef
 710 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 711                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 712
 713 LLVMValueRef
 714 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 715
 716 LLVMValueRef
 717 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 718                    unsigned bitsize);
 719
 720 LLVMValueRef
 721 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 722                     unsigned bitsize);
 723
 724 LLVMValueRef
 725 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
 726                       unsigned bitsize);
 727
 728 LLVMValueRef
 729 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 730
 731 LLVMValueRef
 732 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 733
 734 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 735                            LLVMValueRef *args, unsigned num_args);
 736
 737 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 738                                  LLVMValueRef ptr, LLVMValueRef val,
 739                                  const char *sync_scope);
 740
 741 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 742                                       LLVMValueRef cmp, LLVMValueRef val,
 743                                       const char *sync_scope);
 744
 745 void
 746 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
 747                 LLVMValueRef stencil, LLVMValueRef samplemask,
 748                 struct ac_export_args *args);
 749
 750 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
 751                                    LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt);
 752
 753 struct ac_ngg_prim {
 754         unsigned num_vertices;
 755         LLVMValueRef isnull;
 756         LLVMValueRef index[3];
 757         LLVMValueRef edgeflag[3];
 758         LLVMValueRef passthrough;
 759 };
 760
 761 LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
 762                                  const struct ac_ngg_prim *prim);
 763 void ac_build_export_prim(struct ac_llvm_context *ctx,
 764                           const struct ac_ngg_prim *prim);
 765
 766 static inline LLVMValueRef
 767 ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
 768 {
 769         assert(arg.used);
 770         return LLVMGetParam(ctx->main_function, arg.arg_index);
 771 }
 772
 773 enum ac_llvm_calling_convention {
 774         AC_LLVM_AMDGPU_VS = 87,
 775         AC_LLVM_AMDGPU_GS = 88,
 776         AC_LLVM_AMDGPU_PS = 89,
 777         AC_LLVM_AMDGPU_CS = 90,
 778         AC_LLVM_AMDGPU_HS = 93,
 779 };
 780
 781 LLVMValueRef ac_build_main(const struct ac_shader_args *args,
 782                            struct ac_llvm_context *ctx,
 783                            enum ac_llvm_calling_convention convention,
 784                            const char *name, LLVMTypeRef ret_type,
 785                            LLVMModuleRef module);
 786
 787 #ifdef __cplusplus
 788 }
 789 #endif
 790
 791 #endif