1 //===-- SIDefines.h - SI Helper Macros ----------------------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 /// \file 8 //===----------------------------------------------------------------------===// 9 10 #ifndef LLVM_LIB_TARGET_AMDGPU_SIDEFINES_H 11 #define LLVM_LIB_TARGET_AMDGPU_SIDEFINES_H 12 13 #include "llvm/MC/MCInstrDesc.h" 14 15 namespace llvm { 16 17 // This needs to be kept in sync with the field bits in SIRegisterClass. 18 enum SIRCFlags : uint8_t { 19 // For vector registers. 20 HasVGPR = 1 << 0, 21 HasAGPR = 1 << 1, 22 HasSGPR = 1 << 2 23 }; // enum SIRCFlags 24 25 namespace SIInstrFlags { 26 // This needs to be kept in sync with the field bits in InstSI. 27 enum : uint64_t { 28 // Low bits - basic encoding information. 29 SALU = 1 << 0, 30 VALU = 1 << 1, 31 32 // SALU instruction formats. 33 SOP1 = 1 << 2, 34 SOP2 = 1 << 3, 35 SOPC = 1 << 4, 36 SOPK = 1 << 5, 37 SOPP = 1 << 6, 38 39 // VALU instruction formats. 40 VOP1 = 1 << 7, 41 VOP2 = 1 << 8, 42 VOPC = 1 << 9, 43 44 // TODO: Should this be spilt into VOP3 a and b? 45 VOP3 = 1 << 10, 46 VOP3P = 1 << 12, 47 48 VINTRP = 1 << 13, 49 SDWA = 1 << 14, 50 DPP = 1 << 15, 51 TRANS = 1 << 16, 52 53 // Memory instruction formats. 54 MUBUF = 1 << 17, 55 MTBUF = 1 << 18, 56 SMRD = 1 << 19, 57 MIMG = 1 << 20, 58 EXP = 1 << 21, 59 FLAT = 1 << 22, 60 DS = 1 << 23, 61 62 // Pseudo instruction formats. 63 VGPRSpill = 1 << 24, 64 SGPRSpill = 1 << 25, 65 66 // LDSDIR instruction format. 67 LDSDIR = 1 << 26, 68 69 // VINTERP instruction format. 70 VINTERP = 1 << 27, 71 72 // High bits - other information. 73 VM_CNT = UINT64_C(1) << 32, 74 EXP_CNT = UINT64_C(1) << 33, 75 LGKM_CNT = UINT64_C(1) << 34, 76 77 WQM = UINT64_C(1) << 35, 78 DisableWQM = UINT64_C(1) << 36, 79 Gather4 = UINT64_C(1) << 37, 80 SOPK_ZEXT = UINT64_C(1) << 38, 81 SCALAR_STORE = UINT64_C(1) << 39, 82 FIXED_SIZE = UINT64_C(1) << 40, 83 VOPAsmPrefer32Bit = UINT64_C(1) << 41, 84 VOP3_OPSEL = UINT64_C(1) << 42, 85 maybeAtomic = UINT64_C(1) << 43, 86 renamedInGFX9 = UINT64_C(1) << 44, 87 88 // Is a clamp on FP type. 89 FPClamp = UINT64_C(1) << 45, 90 91 // Is an integer clamp 92 IntClamp = UINT64_C(1) << 46, 93 94 // Clamps lo component of register. 95 ClampLo = UINT64_C(1) << 47, 96 97 // Clamps hi component of register. 98 // ClampLo and ClampHi set for packed clamp. 99 ClampHi = UINT64_C(1) << 48, 100 101 // Is a packed VOP3P instruction. 102 IsPacked = UINT64_C(1) << 49, 103 104 // Is a D16 buffer instruction. 105 D16Buf = UINT64_C(1) << 50, 106 107 // FLAT instruction accesses FLAT_GLBL segment. 108 FlatGlobal = UINT64_C(1) << 51, 109 110 // Uses floating point double precision rounding mode 111 FPDPRounding = UINT64_C(1) << 52, 112 113 // Instruction is FP atomic. 114 FPAtomic = UINT64_C(1) << 53, 115 116 // Is a MFMA instruction. 117 IsMAI = UINT64_C(1) << 54, 118 119 // Is a DOT instruction. 120 IsDOT = UINT64_C(1) << 55, 121 122 // FLAT instruction accesses FLAT_SCRATCH segment. 123 FlatScratch = UINT64_C(1) << 56, 124 125 // Atomic without return. 126 IsAtomicNoRet = UINT64_C(1) << 57, 127 128 // Atomic with return. 129 IsAtomicRet = UINT64_C(1) << 58 130 }; 131 132 // v_cmp_class_* etc. use a 10-bit mask for what operation is checked. 133 // The result is true if any of these tests are true. 134 enum ClassFlags : unsigned { 135 S_NAN = 1 << 0, // Signaling NaN 136 Q_NAN = 1 << 1, // Quiet NaN 137 N_INFINITY = 1 << 2, // Negative infinity 138 N_NORMAL = 1 << 3, // Negative normal 139 N_SUBNORMAL = 1 << 4, // Negative subnormal 140 N_ZERO = 1 << 5, // Negative zero 141 P_ZERO = 1 << 6, // Positive zero 142 P_SUBNORMAL = 1 << 7, // Positive subnormal 143 P_NORMAL = 1 << 8, // Positive normal 144 P_INFINITY = 1 << 9 // Positive infinity 145 }; 146 } 147 148 namespace AMDGPU { 149 enum OperandType : unsigned { 150 /// Operands with register or 32-bit immediate 151 OPERAND_REG_IMM_INT32 = MCOI::OPERAND_FIRST_TARGET, 152 OPERAND_REG_IMM_INT64, 153 OPERAND_REG_IMM_INT16, 154 OPERAND_REG_IMM_FP32, 155 OPERAND_REG_IMM_FP64, 156 OPERAND_REG_IMM_FP16, 157 OPERAND_REG_IMM_FP16_DEFERRED, 158 OPERAND_REG_IMM_FP32_DEFERRED, 159 OPERAND_REG_IMM_V2FP16, 160 OPERAND_REG_IMM_V2INT16, 161 OPERAND_REG_IMM_V2INT32, 162 OPERAND_REG_IMM_V2FP32, 163 164 /// Operands with register or inline constant 165 OPERAND_REG_INLINE_C_INT16, 166 OPERAND_REG_INLINE_C_INT32, 167 OPERAND_REG_INLINE_C_INT64, 168 OPERAND_REG_INLINE_C_FP16, 169 OPERAND_REG_INLINE_C_FP32, 170 OPERAND_REG_INLINE_C_FP64, 171 OPERAND_REG_INLINE_C_V2INT16, 172 OPERAND_REG_INLINE_C_V2FP16, 173 OPERAND_REG_INLINE_C_V2INT32, 174 OPERAND_REG_INLINE_C_V2FP32, 175 176 /// Operand with 32-bit immediate that uses the constant bus. 177 OPERAND_KIMM32, 178 OPERAND_KIMM16, 179 180 /// Operands with an AccVGPR register or inline constant 181 OPERAND_REG_INLINE_AC_INT16, 182 OPERAND_REG_INLINE_AC_INT32, 183 OPERAND_REG_INLINE_AC_FP16, 184 OPERAND_REG_INLINE_AC_FP32, 185 OPERAND_REG_INLINE_AC_FP64, 186 OPERAND_REG_INLINE_AC_V2INT16, 187 OPERAND_REG_INLINE_AC_V2FP16, 188 OPERAND_REG_INLINE_AC_V2INT32, 189 OPERAND_REG_INLINE_AC_V2FP32, 190 191 OPERAND_REG_IMM_FIRST = OPERAND_REG_IMM_INT32, 192 OPERAND_REG_IMM_LAST = OPERAND_REG_IMM_V2FP32, 193 194 OPERAND_REG_INLINE_C_FIRST = OPERAND_REG_INLINE_C_INT16, 195 OPERAND_REG_INLINE_C_LAST = OPERAND_REG_INLINE_AC_V2FP32, 196 197 OPERAND_REG_INLINE_AC_FIRST = OPERAND_REG_INLINE_AC_INT16, 198 OPERAND_REG_INLINE_AC_LAST = OPERAND_REG_INLINE_AC_V2FP32, 199 200 OPERAND_SRC_FIRST = OPERAND_REG_IMM_INT32, 201 OPERAND_SRC_LAST = OPERAND_REG_INLINE_C_LAST, 202 203 // Operand for source modifiers for VOP instructions 204 OPERAND_INPUT_MODS, 205 206 // Operand for SDWA instructions 207 OPERAND_SDWA_VOPC_DST 208 209 }; 210 } 211 212 // Input operand modifiers bit-masks 213 // NEG and SEXT share same bit-mask because they can't be set simultaneously. 214 namespace SISrcMods { 215 enum : unsigned { 216 NEG = 1 << 0, // Floating-point negate modifier 217 ABS = 1 << 1, // Floating-point absolute modifier 218 SEXT = 1 << 0, // Integer sign-extend modifier 219 NEG_HI = ABS, // Floating-point negate high packed component modifier. 220 OP_SEL_0 = 1 << 2, 221 OP_SEL_1 = 1 << 3, 222 DST_OP_SEL = 1 << 3 // VOP3 dst op_sel (share mask with OP_SEL_1) 223 }; 224 } 225 226 namespace SIOutMods { 227 enum : unsigned { 228 NONE = 0, 229 MUL2 = 1, 230 MUL4 = 2, 231 DIV2 = 3 232 }; 233 } 234 235 namespace AMDGPU { 236 namespace VGPRIndexMode { 237 238 enum Id : unsigned { // id of symbolic names 239 ID_SRC0 = 0, 240 ID_SRC1, 241 ID_SRC2, 242 ID_DST, 243 244 ID_MIN = ID_SRC0, 245 ID_MAX = ID_DST 246 }; 247 248 enum EncBits : unsigned { 249 OFF = 0, 250 SRC0_ENABLE = 1 << ID_SRC0, 251 SRC1_ENABLE = 1 << ID_SRC1, 252 SRC2_ENABLE = 1 << ID_SRC2, 253 DST_ENABLE = 1 << ID_DST, 254 ENABLE_MASK = SRC0_ENABLE | SRC1_ENABLE | SRC2_ENABLE | DST_ENABLE, 255 UNDEF = 0xFFFF 256 }; 257 258 } // namespace VGPRIndexMode 259 } // namespace AMDGPU 260 261 namespace AMDGPUAsmVariants { 262 enum : unsigned { 263 DEFAULT = 0, 264 VOP3 = 1, 265 SDWA = 2, 266 SDWA9 = 3, 267 DPP = 4 268 }; 269 } 270 271 namespace AMDGPU { 272 namespace EncValues { // Encoding values of enum9/8/7 operands 273 274 enum : unsigned { 275 SGPR_MIN = 0, 276 SGPR_MAX_SI = 101, 277 SGPR_MAX_GFX10 = 105, 278 TTMP_VI_MIN = 112, 279 TTMP_VI_MAX = 123, 280 TTMP_GFX9PLUS_MIN = 108, 281 TTMP_GFX9PLUS_MAX = 123, 282 INLINE_INTEGER_C_MIN = 128, 283 INLINE_INTEGER_C_POSITIVE_MAX = 192, // 64 284 INLINE_INTEGER_C_MAX = 208, 285 INLINE_FLOATING_C_MIN = 240, 286 INLINE_FLOATING_C_MAX = 248, 287 LITERAL_CONST = 255, 288 VGPR_MIN = 256, 289 VGPR_MAX = 511, 290 IS_VGPR = 256 // Indicates VGPR or AGPR 291 }; 292 293 } // namespace EncValues 294 } // namespace AMDGPU 295 296 namespace AMDGPU { 297 namespace CPol { 298 299 enum CPol { 300 GLC = 1, 301 SLC = 2, 302 DLC = 4, 303 SCC = 16, 304 SC0 = GLC, 305 SC1 = SCC, 306 NT = SLC, 307 ALL = GLC | SLC | DLC | SCC 308 }; 309 310 } // namespace CPol 311 312 namespace SendMsg { // Encoding of SIMM16 used in s_sendmsg* insns. 313 314 enum Id { // Message ID, width(4) [3:0]. 315 ID_INTERRUPT = 1, 316 317 ID_GS_PreGFX11 = 2, // replaced in GFX11 318 ID_GS_DONE_PreGFX11 = 3, // replaced in GFX11 319 320 ID_HS_TESSFACTOR_GFX11Plus = 2, // reused in GFX11 321 ID_DEALLOC_VGPRS_GFX11Plus = 3, // reused in GFX11 322 323 ID_SAVEWAVE = 4, // added in GFX8, removed in GFX11 324 ID_STALL_WAVE_GEN = 5, // added in GFX9 325 ID_HALT_WAVES = 6, // added in GFX9 326 ID_ORDERED_PS_DONE = 7, // added in GFX9 327 ID_EARLY_PRIM_DEALLOC = 8, // added in GFX9, removed in GFX10 328 ID_GS_ALLOC_REQ = 9, // added in GFX9 329 ID_GET_DOORBELL = 10, // added in GFX9, removed in GFX11 330 ID_GET_DDID = 11, // added in GFX10, removed in GFX11 331 ID_SYSMSG = 15, 332 333 ID_RTN_GET_DOORBELL = 128, 334 ID_RTN_GET_DDID = 129, 335 ID_RTN_GET_TMA = 130, 336 ID_RTN_GET_REALTIME = 131, 337 ID_RTN_SAVE_WAVE = 132, 338 ID_RTN_GET_TBA = 133, 339 340 ID_MASK_PreGFX11_ = 0xF, 341 ID_MASK_GFX11Plus_ = 0xFF 342 }; 343 344 enum Op { // Both GS and SYS operation IDs. 345 OP_UNKNOWN_ = -1, 346 OP_SHIFT_ = 4, 347 OP_NONE_ = 0, 348 // Bits used for operation encoding 349 OP_WIDTH_ = 3, 350 OP_MASK_ = (((1 << OP_WIDTH_) - 1) << OP_SHIFT_), 351 // GS operations are encoded in bits 5:4 352 OP_GS_NOP = 0, 353 OP_GS_CUT = 1, 354 OP_GS_EMIT = 2, 355 OP_GS_EMIT_CUT = 3, 356 OP_GS_LAST_, 357 OP_GS_FIRST_ = OP_GS_NOP, 358 // SYS operations are encoded in bits 6:4 359 OP_SYS_ECC_ERR_INTERRUPT = 1, 360 OP_SYS_REG_RD = 2, 361 OP_SYS_HOST_TRAP_ACK = 3, 362 OP_SYS_TTRACE_PC = 4, 363 OP_SYS_LAST_, 364 OP_SYS_FIRST_ = OP_SYS_ECC_ERR_INTERRUPT, 365 }; 366 367 enum StreamId : unsigned { // Stream ID, (2) [9:8]. 368 STREAM_ID_NONE_ = 0, 369 STREAM_ID_DEFAULT_ = 0, 370 STREAM_ID_LAST_ = 4, 371 STREAM_ID_FIRST_ = STREAM_ID_DEFAULT_, 372 STREAM_ID_SHIFT_ = 8, 373 STREAM_ID_WIDTH_= 2, 374 STREAM_ID_MASK_ = (((1 << STREAM_ID_WIDTH_) - 1) << STREAM_ID_SHIFT_) 375 }; 376 377 } // namespace SendMsg 378 379 namespace Hwreg { // Encoding of SIMM16 used in s_setreg/getreg* insns. 380 381 enum Id { // HwRegCode, (6) [5:0] 382 ID_MODE = 1, 383 ID_STATUS = 2, 384 ID_TRAPSTS = 3, 385 ID_HW_ID = 4, 386 ID_GPR_ALLOC = 5, 387 ID_LDS_ALLOC = 6, 388 ID_IB_STS = 7, 389 ID_MEM_BASES = 15, 390 ID_TBA_LO = 16, 391 ID_TBA_HI = 17, 392 ID_TMA_LO = 18, 393 ID_TMA_HI = 19, 394 ID_XCC_ID = 20, 395 ID_SQ_PERF_SNAPSHOT_DATA = 21, 396 ID_SQ_PERF_SNAPSHOT_DATA1 = 22, 397 ID_SQ_PERF_SNAPSHOT_PC_LO = 23, 398 ID_SQ_PERF_SNAPSHOT_PC_HI = 24, 399 ID_FLAT_SCR_LO = 20, 400 ID_FLAT_SCR_HI = 21, 401 ID_XNACK_MASK = 22, 402 ID_HW_ID1 = 23, 403 ID_HW_ID2 = 24, 404 ID_POPS_PACKER = 25, 405 ID_SHADER_CYCLES = 29, 406 407 ID_SHIFT_ = 0, 408 ID_WIDTH_ = 6, 409 ID_MASK_ = (((1 << ID_WIDTH_) - 1) << ID_SHIFT_) 410 }; 411 412 enum Offset : unsigned { // Offset, (5) [10:6] 413 OFFSET_DEFAULT_ = 0, 414 OFFSET_SHIFT_ = 6, 415 OFFSET_WIDTH_ = 5, 416 OFFSET_MASK_ = (((1 << OFFSET_WIDTH_) - 1) << OFFSET_SHIFT_), 417 418 OFFSET_MEM_VIOL = 8, 419 420 OFFSET_SRC_SHARED_BASE = 16, 421 OFFSET_SRC_PRIVATE_BASE = 0 422 }; 423 424 enum WidthMinusOne : unsigned { // WidthMinusOne, (5) [15:11] 425 WIDTH_M1_DEFAULT_ = 31, 426 WIDTH_M1_SHIFT_ = 11, 427 WIDTH_M1_WIDTH_ = 5, 428 WIDTH_M1_MASK_ = (((1 << WIDTH_M1_WIDTH_) - 1) << WIDTH_M1_SHIFT_), 429 430 WIDTH_M1_SRC_SHARED_BASE = 15, 431 WIDTH_M1_SRC_PRIVATE_BASE = 15 432 }; 433 434 // Some values from WidthMinusOne mapped into Width domain. 435 enum Width : unsigned { 436 WIDTH_DEFAULT_ = WIDTH_M1_DEFAULT_ + 1, 437 }; 438 439 enum ModeRegisterMasks : uint32_t { 440 FP_ROUND_MASK = 0xf << 0, // Bits 0..3 441 FP_DENORM_MASK = 0xf << 4, // Bits 4..7 442 DX10_CLAMP_MASK = 1 << 8, 443 IEEE_MODE_MASK = 1 << 9, 444 LOD_CLAMP_MASK = 1 << 10, 445 DEBUG_MASK = 1 << 11, 446 447 // EXCP_EN fields. 448 EXCP_EN_INVALID_MASK = 1 << 12, 449 EXCP_EN_INPUT_DENORMAL_MASK = 1 << 13, 450 EXCP_EN_FLOAT_DIV0_MASK = 1 << 14, 451 EXCP_EN_OVERFLOW_MASK = 1 << 15, 452 EXCP_EN_UNDERFLOW_MASK = 1 << 16, 453 EXCP_EN_INEXACT_MASK = 1 << 17, 454 EXCP_EN_INT_DIV0_MASK = 1 << 18, 455 456 GPR_IDX_EN_MASK = 1 << 27, 457 VSKIP_MASK = 1 << 28, 458 CSP_MASK = 0x7u << 29 // Bits 29..31 459 }; 460 461 } // namespace Hwreg 462 463 namespace MTBUFFormat { 464 465 enum DataFormat : int64_t { 466 DFMT_INVALID = 0, 467 DFMT_8, 468 DFMT_16, 469 DFMT_8_8, 470 DFMT_32, 471 DFMT_16_16, 472 DFMT_10_11_11, 473 DFMT_11_11_10, 474 DFMT_10_10_10_2, 475 DFMT_2_10_10_10, 476 DFMT_8_8_8_8, 477 DFMT_32_32, 478 DFMT_16_16_16_16, 479 DFMT_32_32_32, 480 DFMT_32_32_32_32, 481 DFMT_RESERVED_15, 482 483 DFMT_MIN = DFMT_INVALID, 484 DFMT_MAX = DFMT_RESERVED_15, 485 486 DFMT_UNDEF = -1, 487 DFMT_DEFAULT = DFMT_8, 488 489 DFMT_SHIFT = 0, 490 DFMT_MASK = 0xF 491 }; 492 493 enum NumFormat : int64_t { 494 NFMT_UNORM = 0, 495 NFMT_SNORM, 496 NFMT_USCALED, 497 NFMT_SSCALED, 498 NFMT_UINT, 499 NFMT_SINT, 500 NFMT_RESERVED_6, // VI and GFX9 501 NFMT_SNORM_OGL = NFMT_RESERVED_6, // SI and CI only 502 NFMT_FLOAT, 503 504 NFMT_MIN = NFMT_UNORM, 505 NFMT_MAX = NFMT_FLOAT, 506 507 NFMT_UNDEF = -1, 508 NFMT_DEFAULT = NFMT_UNORM, 509 510 NFMT_SHIFT = 4, 511 NFMT_MASK = 7 512 }; 513 514 enum MergedFormat : int64_t { 515 DFMT_NFMT_UNDEF = -1, 516 DFMT_NFMT_DEFAULT = ((DFMT_DEFAULT & DFMT_MASK) << DFMT_SHIFT) | 517 ((NFMT_DEFAULT & NFMT_MASK) << NFMT_SHIFT), 518 519 520 DFMT_NFMT_MASK = (DFMT_MASK << DFMT_SHIFT) | (NFMT_MASK << NFMT_SHIFT), 521 522 DFMT_NFMT_MAX = DFMT_NFMT_MASK 523 }; 524 525 enum UnifiedFormatCommon : int64_t { 526 UFMT_MAX = 127, 527 UFMT_UNDEF = -1, 528 UFMT_DEFAULT = 1 529 }; 530 531 } // namespace MTBUFFormat 532 533 namespace UfmtGFX10 { 534 enum UnifiedFormat : int64_t { 535 UFMT_INVALID = 0, 536 537 UFMT_8_UNORM, 538 UFMT_8_SNORM, 539 UFMT_8_USCALED, 540 UFMT_8_SSCALED, 541 UFMT_8_UINT, 542 UFMT_8_SINT, 543 544 UFMT_16_UNORM, 545 UFMT_16_SNORM, 546 UFMT_16_USCALED, 547 UFMT_16_SSCALED, 548 UFMT_16_UINT, 549 UFMT_16_SINT, 550 UFMT_16_FLOAT, 551 552 UFMT_8_8_UNORM, 553 UFMT_8_8_SNORM, 554 UFMT_8_8_USCALED, 555 UFMT_8_8_SSCALED, 556 UFMT_8_8_UINT, 557 UFMT_8_8_SINT, 558 559 UFMT_32_UINT, 560 UFMT_32_SINT, 561 UFMT_32_FLOAT, 562 563 UFMT_16_16_UNORM, 564 UFMT_16_16_SNORM, 565 UFMT_16_16_USCALED, 566 UFMT_16_16_SSCALED, 567 UFMT_16_16_UINT, 568 UFMT_16_16_SINT, 569 UFMT_16_16_FLOAT, 570 571 UFMT_10_11_11_UNORM, 572 UFMT_10_11_11_SNORM, 573 UFMT_10_11_11_USCALED, 574 UFMT_10_11_11_SSCALED, 575 UFMT_10_11_11_UINT, 576 UFMT_10_11_11_SINT, 577 UFMT_10_11_11_FLOAT, 578 579 UFMT_11_11_10_UNORM, 580 UFMT_11_11_10_SNORM, 581 UFMT_11_11_10_USCALED, 582 UFMT_11_11_10_SSCALED, 583 UFMT_11_11_10_UINT, 584 UFMT_11_11_10_SINT, 585 UFMT_11_11_10_FLOAT, 586 587 UFMT_10_10_10_2_UNORM, 588 UFMT_10_10_10_2_SNORM, 589 UFMT_10_10_10_2_USCALED, 590 UFMT_10_10_10_2_SSCALED, 591 UFMT_10_10_10_2_UINT, 592 UFMT_10_10_10_2_SINT, 593 594 UFMT_2_10_10_10_UNORM, 595 UFMT_2_10_10_10_SNORM, 596 UFMT_2_10_10_10_USCALED, 597 UFMT_2_10_10_10_SSCALED, 598 UFMT_2_10_10_10_UINT, 599 UFMT_2_10_10_10_SINT, 600 601 UFMT_8_8_8_8_UNORM, 602 UFMT_8_8_8_8_SNORM, 603 UFMT_8_8_8_8_USCALED, 604 UFMT_8_8_8_8_SSCALED, 605 UFMT_8_8_8_8_UINT, 606 UFMT_8_8_8_8_SINT, 607 608 UFMT_32_32_UINT, 609 UFMT_32_32_SINT, 610 UFMT_32_32_FLOAT, 611 612 UFMT_16_16_16_16_UNORM, 613 UFMT_16_16_16_16_SNORM, 614 UFMT_16_16_16_16_USCALED, 615 UFMT_16_16_16_16_SSCALED, 616 UFMT_16_16_16_16_UINT, 617 UFMT_16_16_16_16_SINT, 618 UFMT_16_16_16_16_FLOAT, 619 620 UFMT_32_32_32_UINT, 621 UFMT_32_32_32_SINT, 622 UFMT_32_32_32_FLOAT, 623 UFMT_32_32_32_32_UINT, 624 UFMT_32_32_32_32_SINT, 625 UFMT_32_32_32_32_FLOAT, 626 627 UFMT_FIRST = UFMT_INVALID, 628 UFMT_LAST = UFMT_32_32_32_32_FLOAT, 629 }; 630 631 } // namespace UfmtGFX10 632 633 namespace UfmtGFX11 { 634 enum UnifiedFormat : int64_t { 635 UFMT_INVALID = 0, 636 637 UFMT_8_UNORM, 638 UFMT_8_SNORM, 639 UFMT_8_USCALED, 640 UFMT_8_SSCALED, 641 UFMT_8_UINT, 642 UFMT_8_SINT, 643 644 UFMT_16_UNORM, 645 UFMT_16_SNORM, 646 UFMT_16_USCALED, 647 UFMT_16_SSCALED, 648 UFMT_16_UINT, 649 UFMT_16_SINT, 650 UFMT_16_FLOAT, 651 652 UFMT_8_8_UNORM, 653 UFMT_8_8_SNORM, 654 UFMT_8_8_USCALED, 655 UFMT_8_8_SSCALED, 656 UFMT_8_8_UINT, 657 UFMT_8_8_SINT, 658 659 UFMT_32_UINT, 660 UFMT_32_SINT, 661 UFMT_32_FLOAT, 662 663 UFMT_16_16_UNORM, 664 UFMT_16_16_SNORM, 665 UFMT_16_16_USCALED, 666 UFMT_16_16_SSCALED, 667 UFMT_16_16_UINT, 668 UFMT_16_16_SINT, 669 UFMT_16_16_FLOAT, 670 671 UFMT_10_11_11_FLOAT, 672 673 UFMT_11_11_10_FLOAT, 674 675 UFMT_10_10_10_2_UNORM, 676 UFMT_10_10_10_2_SNORM, 677 UFMT_10_10_10_2_UINT, 678 UFMT_10_10_10_2_SINT, 679 680 UFMT_2_10_10_10_UNORM, 681 UFMT_2_10_10_10_SNORM, 682 UFMT_2_10_10_10_USCALED, 683 UFMT_2_10_10_10_SSCALED, 684 UFMT_2_10_10_10_UINT, 685 UFMT_2_10_10_10_SINT, 686 687 UFMT_8_8_8_8_UNORM, 688 UFMT_8_8_8_8_SNORM, 689 UFMT_8_8_8_8_USCALED, 690 UFMT_8_8_8_8_SSCALED, 691 UFMT_8_8_8_8_UINT, 692 UFMT_8_8_8_8_SINT, 693 694 UFMT_32_32_UINT, 695 UFMT_32_32_SINT, 696 UFMT_32_32_FLOAT, 697 698 UFMT_16_16_16_16_UNORM, 699 UFMT_16_16_16_16_SNORM, 700 UFMT_16_16_16_16_USCALED, 701 UFMT_16_16_16_16_SSCALED, 702 UFMT_16_16_16_16_UINT, 703 UFMT_16_16_16_16_SINT, 704 UFMT_16_16_16_16_FLOAT, 705 706 UFMT_32_32_32_UINT, 707 UFMT_32_32_32_SINT, 708 UFMT_32_32_32_FLOAT, 709 UFMT_32_32_32_32_UINT, 710 UFMT_32_32_32_32_SINT, 711 UFMT_32_32_32_32_FLOAT, 712 713 UFMT_FIRST = UFMT_INVALID, 714 UFMT_LAST = UFMT_32_32_32_32_FLOAT, 715 }; 716 717 } // namespace UfmtGFX11 718 719 namespace Swizzle { // Encoding of swizzle macro used in ds_swizzle_b32. 720 721 enum Id : unsigned { // id of symbolic names 722 ID_QUAD_PERM = 0, 723 ID_BITMASK_PERM, 724 ID_SWAP, 725 ID_REVERSE, 726 ID_BROADCAST 727 }; 728 729 enum EncBits : unsigned { 730 731 // swizzle mode encodings 732 733 QUAD_PERM_ENC = 0x8000, 734 QUAD_PERM_ENC_MASK = 0xFF00, 735 736 BITMASK_PERM_ENC = 0x0000, 737 BITMASK_PERM_ENC_MASK = 0x8000, 738 739 // QUAD_PERM encodings 740 741 LANE_MASK = 0x3, 742 LANE_MAX = LANE_MASK, 743 LANE_SHIFT = 2, 744 LANE_NUM = 4, 745 746 // BITMASK_PERM encodings 747 748 BITMASK_MASK = 0x1F, 749 BITMASK_MAX = BITMASK_MASK, 750 BITMASK_WIDTH = 5, 751 752 BITMASK_AND_SHIFT = 0, 753 BITMASK_OR_SHIFT = 5, 754 BITMASK_XOR_SHIFT = 10 755 }; 756 757 } // namespace Swizzle 758 759 namespace SDWA { 760 761 enum SdwaSel : unsigned { 762 BYTE_0 = 0, 763 BYTE_1 = 1, 764 BYTE_2 = 2, 765 BYTE_3 = 3, 766 WORD_0 = 4, 767 WORD_1 = 5, 768 DWORD = 6, 769 }; 770 771 enum DstUnused : unsigned { 772 UNUSED_PAD = 0, 773 UNUSED_SEXT = 1, 774 UNUSED_PRESERVE = 2, 775 }; 776 777 enum SDWA9EncValues : unsigned { 778 SRC_SGPR_MASK = 0x100, 779 SRC_VGPR_MASK = 0xFF, 780 VOPC_DST_VCC_MASK = 0x80, 781 VOPC_DST_SGPR_MASK = 0x7F, 782 783 SRC_VGPR_MIN = 0, 784 SRC_VGPR_MAX = 255, 785 SRC_SGPR_MIN = 256, 786 SRC_SGPR_MAX_SI = 357, 787 SRC_SGPR_MAX_GFX10 = 361, 788 SRC_TTMP_MIN = 364, 789 SRC_TTMP_MAX = 379, 790 }; 791 792 } // namespace SDWA 793 794 namespace DPP { 795 796 // clang-format off 797 enum DppCtrl : unsigned { 798 QUAD_PERM_FIRST = 0, 799 QUAD_PERM_ID = 0xE4, // identity permutation 800 QUAD_PERM_LAST = 0xFF, 801 DPP_UNUSED1 = 0x100, 802 ROW_SHL0 = 0x100, 803 ROW_SHL_FIRST = 0x101, 804 ROW_SHL_LAST = 0x10F, 805 DPP_UNUSED2 = 0x110, 806 ROW_SHR0 = 0x110, 807 ROW_SHR_FIRST = 0x111, 808 ROW_SHR_LAST = 0x11F, 809 DPP_UNUSED3 = 0x120, 810 ROW_ROR0 = 0x120, 811 ROW_ROR_FIRST = 0x121, 812 ROW_ROR_LAST = 0x12F, 813 WAVE_SHL1 = 0x130, 814 DPP_UNUSED4_FIRST = 0x131, 815 DPP_UNUSED4_LAST = 0x133, 816 WAVE_ROL1 = 0x134, 817 DPP_UNUSED5_FIRST = 0x135, 818 DPP_UNUSED5_LAST = 0x137, 819 WAVE_SHR1 = 0x138, 820 DPP_UNUSED6_FIRST = 0x139, 821 DPP_UNUSED6_LAST = 0x13B, 822 WAVE_ROR1 = 0x13C, 823 DPP_UNUSED7_FIRST = 0x13D, 824 DPP_UNUSED7_LAST = 0x13F, 825 ROW_MIRROR = 0x140, 826 ROW_HALF_MIRROR = 0x141, 827 BCAST15 = 0x142, 828 BCAST31 = 0x143, 829 DPP_UNUSED8_FIRST = 0x144, 830 DPP_UNUSED8_LAST = 0x14F, 831 ROW_NEWBCAST_FIRST= 0x150, 832 ROW_NEWBCAST_LAST = 0x15F, 833 ROW_SHARE0 = 0x150, 834 ROW_SHARE_FIRST = 0x150, 835 ROW_SHARE_LAST = 0x15F, 836 ROW_XMASK0 = 0x160, 837 ROW_XMASK_FIRST = 0x160, 838 ROW_XMASK_LAST = 0x16F, 839 DPP_LAST = ROW_XMASK_LAST 840 }; 841 // clang-format on 842 843 enum DppFiMode { 844 DPP_FI_0 = 0, 845 DPP_FI_1 = 1, 846 DPP8_FI_0 = 0xE9, 847 DPP8_FI_1 = 0xEA, 848 }; 849 850 } // namespace DPP 851 852 namespace Exp { 853 854 enum Target : unsigned { 855 ET_MRT0 = 0, 856 ET_MRT7 = 7, 857 ET_MRTZ = 8, 858 ET_NULL = 9, // Pre-GFX11 859 ET_POS0 = 12, 860 ET_POS3 = 15, 861 ET_POS4 = 16, // GFX10+ 862 ET_POS_LAST = ET_POS4, // Highest pos used on any subtarget 863 ET_PRIM = 20, // GFX10+ 864 ET_DUAL_SRC_BLEND0 = 21, // GFX11+ 865 ET_DUAL_SRC_BLEND1 = 22, // GFX11+ 866 ET_PARAM0 = 32, // Pre-GFX11 867 ET_PARAM31 = 63, // Pre-GFX11 868 869 ET_NULL_MAX_IDX = 0, 870 ET_MRTZ_MAX_IDX = 0, 871 ET_PRIM_MAX_IDX = 0, 872 ET_MRT_MAX_IDX = 7, 873 ET_POS_MAX_IDX = 4, 874 ET_DUAL_SRC_BLEND_MAX_IDX = 1, 875 ET_PARAM_MAX_IDX = 31, 876 877 ET_INVALID = 255, 878 }; 879 880 } // namespace Exp 881 882 namespace VOP3PEncoding { 883 884 enum OpSel : uint64_t { 885 OP_SEL_HI_0 = UINT64_C(1) << 59, 886 OP_SEL_HI_1 = UINT64_C(1) << 60, 887 OP_SEL_HI_2 = UINT64_C(1) << 14, 888 }; 889 890 } // namespace VOP3PEncoding 891 892 namespace ImplicitArg { 893 // Implicit kernel argument offset for code object version 5. 894 enum Offset_COV5 : unsigned { 895 HOSTCALL_PTR_OFFSET = 80, 896 MULTIGRID_SYNC_ARG_OFFSET = 88, 897 HEAP_PTR_OFFSET = 96, 898 PRIVATE_BASE_OFFSET = 192, 899 SHARED_BASE_OFFSET = 196, 900 QUEUE_PTR_OFFSET = 200, 901 }; 902 903 } // namespace ImplicitArg 904 } // namespace AMDGPU 905 906 #define R_00B028_SPI_SHADER_PGM_RSRC1_PS 0x00B028 907 #define S_00B028_VGPRS(x) (((x) & 0x3F) << 0) 908 #define S_00B028_SGPRS(x) (((x) & 0x0F) << 6) 909 #define S_00B028_MEM_ORDERED(x) (((x) & 0x1) << 25) 910 #define G_00B028_MEM_ORDERED(x) (((x) >> 25) & 0x1) 911 #define C_00B028_MEM_ORDERED 0xFDFFFFFF 912 913 #define R_00B02C_SPI_SHADER_PGM_RSRC2_PS 0x00B02C 914 #define S_00B02C_EXTRA_LDS_SIZE(x) (((x) & 0xFF) << 8) 915 #define R_00B128_SPI_SHADER_PGM_RSRC1_VS 0x00B128 916 #define S_00B128_MEM_ORDERED(x) (((x) & 0x1) << 27) 917 #define G_00B128_MEM_ORDERED(x) (((x) >> 27) & 0x1) 918 #define C_00B128_MEM_ORDERED 0xF7FFFFFF 919 920 #define R_00B228_SPI_SHADER_PGM_RSRC1_GS 0x00B228 921 #define S_00B228_WGP_MODE(x) (((x) & 0x1) << 27) 922 #define G_00B228_WGP_MODE(x) (((x) >> 27) & 0x1) 923 #define C_00B228_WGP_MODE 0xF7FFFFFF 924 #define S_00B228_MEM_ORDERED(x) (((x) & 0x1) << 25) 925 #define G_00B228_MEM_ORDERED(x) (((x) >> 25) & 0x1) 926 #define C_00B228_MEM_ORDERED 0xFDFFFFFF 927 928 #define R_00B328_SPI_SHADER_PGM_RSRC1_ES 0x00B328 929 #define R_00B428_SPI_SHADER_PGM_RSRC1_HS 0x00B428 930 #define S_00B428_WGP_MODE(x) (((x) & 0x1) << 26) 931 #define G_00B428_WGP_MODE(x) (((x) >> 26) & 0x1) 932 #define C_00B428_WGP_MODE 0xFBFFFFFF 933 #define S_00B428_MEM_ORDERED(x) (((x) & 0x1) << 24) 934 #define G_00B428_MEM_ORDERED(x) (((x) >> 24) & 0x1) 935 #define C_00B428_MEM_ORDERED 0xFEFFFFFF 936 937 #define R_00B528_SPI_SHADER_PGM_RSRC1_LS 0x00B528 938 939 #define R_00B84C_COMPUTE_PGM_RSRC2 0x00B84C 940 #define S_00B84C_SCRATCH_EN(x) (((x) & 0x1) << 0) 941 #define G_00B84C_SCRATCH_EN(x) (((x) >> 0) & 0x1) 942 #define C_00B84C_SCRATCH_EN 0xFFFFFFFE 943 #define S_00B84C_USER_SGPR(x) (((x) & 0x1F) << 1) 944 #define G_00B84C_USER_SGPR(x) (((x) >> 1) & 0x1F) 945 #define C_00B84C_USER_SGPR 0xFFFFFFC1 946 #define S_00B84C_TRAP_HANDLER(x) (((x) & 0x1) << 6) 947 #define G_00B84C_TRAP_HANDLER(x) (((x) >> 6) & 0x1) 948 #define C_00B84C_TRAP_HANDLER 0xFFFFFFBF 949 #define S_00B84C_TGID_X_EN(x) (((x) & 0x1) << 7) 950 #define G_00B84C_TGID_X_EN(x) (((x) >> 7) & 0x1) 951 #define C_00B84C_TGID_X_EN 0xFFFFFF7F 952 #define S_00B84C_TGID_Y_EN(x) (((x) & 0x1) << 8) 953 #define G_00B84C_TGID_Y_EN(x) (((x) >> 8) & 0x1) 954 #define C_00B84C_TGID_Y_EN 0xFFFFFEFF 955 #define S_00B84C_TGID_Z_EN(x) (((x) & 0x1) << 9) 956 #define G_00B84C_TGID_Z_EN(x) (((x) >> 9) & 0x1) 957 #define C_00B84C_TGID_Z_EN 0xFFFFFDFF 958 #define S_00B84C_TG_SIZE_EN(x) (((x) & 0x1) << 10) 959 #define G_00B84C_TG_SIZE_EN(x) (((x) >> 10) & 0x1) 960 #define C_00B84C_TG_SIZE_EN 0xFFFFFBFF 961 #define S_00B84C_TIDIG_COMP_CNT(x) (((x) & 0x03) << 11) 962 #define G_00B84C_TIDIG_COMP_CNT(x) (((x) >> 11) & 0x03) 963 #define C_00B84C_TIDIG_COMP_CNT 0xFFFFE7FF 964 /* CIK */ 965 #define S_00B84C_EXCP_EN_MSB(x) (((x) & 0x03) << 13) 966 #define G_00B84C_EXCP_EN_MSB(x) (((x) >> 13) & 0x03) 967 #define C_00B84C_EXCP_EN_MSB 0xFFFF9FFF 968 /* */ 969 #define S_00B84C_LDS_SIZE(x) (((x) & 0x1FF) << 15) 970 #define G_00B84C_LDS_SIZE(x) (((x) >> 15) & 0x1FF) 971 #define C_00B84C_LDS_SIZE 0xFF007FFF 972 #define S_00B84C_EXCP_EN(x) (((x) & 0x7F) << 24) 973 #define G_00B84C_EXCP_EN(x) (((x) >> 24) & 0x7F) 974 #define C_00B84C_EXCP_EN 975 976 #define R_0286CC_SPI_PS_INPUT_ENA 0x0286CC 977 #define R_0286D0_SPI_PS_INPUT_ADDR 0x0286D0 978 979 #define R_00B848_COMPUTE_PGM_RSRC1 0x00B848 980 #define S_00B848_VGPRS(x) (((x) & 0x3F) << 0) 981 #define G_00B848_VGPRS(x) (((x) >> 0) & 0x3F) 982 #define C_00B848_VGPRS 0xFFFFFFC0 983 #define S_00B848_SGPRS(x) (((x) & 0x0F) << 6) 984 #define G_00B848_SGPRS(x) (((x) >> 6) & 0x0F) 985 #define C_00B848_SGPRS 0xFFFFFC3F 986 #define S_00B848_PRIORITY(x) (((x) & 0x03) << 10) 987 #define G_00B848_PRIORITY(x) (((x) >> 10) & 0x03) 988 #define C_00B848_PRIORITY 0xFFFFF3FF 989 #define S_00B848_FLOAT_MODE(x) (((x) & 0xFF) << 12) 990 #define G_00B848_FLOAT_MODE(x) (((x) >> 12) & 0xFF) 991 #define C_00B848_FLOAT_MODE 0xFFF00FFF 992 #define S_00B848_PRIV(x) (((x) & 0x1) << 20) 993 #define G_00B848_PRIV(x) (((x) >> 20) & 0x1) 994 #define C_00B848_PRIV 0xFFEFFFFF 995 #define S_00B848_DX10_CLAMP(x) (((x) & 0x1) << 21) 996 #define G_00B848_DX10_CLAMP(x) (((x) >> 21) & 0x1) 997 #define C_00B848_DX10_CLAMP 0xFFDFFFFF 998 #define S_00B848_DEBUG_MODE(x) (((x) & 0x1) << 22) 999 #define G_00B848_DEBUG_MODE(x) (((x) >> 22) & 0x1) 1000 #define C_00B848_DEBUG_MODE 0xFFBFFFFF 1001 #define S_00B848_IEEE_MODE(x) (((x) & 0x1) << 23) 1002 #define G_00B848_IEEE_MODE(x) (((x) >> 23) & 0x1) 1003 #define C_00B848_IEEE_MODE 0xFF7FFFFF 1004 #define S_00B848_WGP_MODE(x) (((x) & 0x1) << 29) 1005 #define G_00B848_WGP_MODE(x) (((x) >> 29) & 0x1) 1006 #define C_00B848_WGP_MODE 0xDFFFFFFF 1007 #define S_00B848_MEM_ORDERED(x) (((x) & 0x1) << 30) 1008 #define G_00B848_MEM_ORDERED(x) (((x) >> 30) & 0x1) 1009 #define C_00B848_MEM_ORDERED 0xBFFFFFFF 1010 #define S_00B848_FWD_PROGRESS(x) (((x) & 0x1) << 31) 1011 #define G_00B848_FWD_PROGRESS(x) (((x) >> 31) & 0x1) 1012 #define C_00B848_FWD_PROGRESS 0x7FFFFFFF 1013 1014 1015 // Helpers for setting FLOAT_MODE 1016 #define FP_ROUND_ROUND_TO_NEAREST 0 1017 #define FP_ROUND_ROUND_TO_INF 1 1018 #define FP_ROUND_ROUND_TO_NEGINF 2 1019 #define FP_ROUND_ROUND_TO_ZERO 3 1020 1021 // Bits 3:0 control rounding mode. 1:0 control single precision, 3:2 double 1022 // precision. 1023 #define FP_ROUND_MODE_SP(x) ((x) & 0x3) 1024 #define FP_ROUND_MODE_DP(x) (((x) & 0x3) << 2) 1025 1026 #define FP_DENORM_FLUSH_IN_FLUSH_OUT 0 1027 #define FP_DENORM_FLUSH_OUT 1 1028 #define FP_DENORM_FLUSH_IN 2 1029 #define FP_DENORM_FLUSH_NONE 3 1030 1031 1032 // Bits 7:4 control denormal handling. 5:4 control single precision, 6:7 double 1033 // precision. 1034 #define FP_DENORM_MODE_SP(x) (((x) & 0x3) << 4) 1035 #define FP_DENORM_MODE_DP(x) (((x) & 0x3) << 6) 1036 1037 #define R_00B860_COMPUTE_TMPRING_SIZE 0x00B860 1038 #define S_00B860_WAVESIZE(x) (((x) & 0x1FFF) << 12) 1039 1040 #define R_0286E8_SPI_TMPRING_SIZE 0x0286E8 1041 #define S_0286E8_WAVESIZE(x) (((x) & 0x1FFF) << 12) 1042 1043 #define R_028B54_VGT_SHADER_STAGES_EN 0x028B54 1044 #define S_028B54_HS_W32_EN(x) (((x) & 0x1) << 21) 1045 #define S_028B54_GS_W32_EN(x) (((x) & 0x1) << 22) 1046 #define S_028B54_VS_W32_EN(x) (((x) & 0x1) << 23) 1047 #define R_0286D8_SPI_PS_IN_CONTROL 0x0286D8 1048 #define S_0286D8_PS_W32_EN(x) (((x) & 0x1) << 15) 1049 #define R_00B800_COMPUTE_DISPATCH_INITIATOR 0x00B800 1050 #define S_00B800_CS_W32_EN(x) (((x) & 0x1) << 15) 1051 1052 #define R_SPILLED_SGPRS 0x4 1053 #define R_SPILLED_VGPRS 0x8 1054 } // End namespace llvm 1055 1056 #endif 1057