## GCN ISA VOP3 instructions The VOP3 instructions requires two dword to store in program code. By default, these encoding of these instructions gives all features of the VOP3 encoding: all possible modifiers, any source operand combination. In an assembler's syntax you can force this encoding by using `VOP3` modifier in instruction. List of fields for the VOP3A/VOP3B encoding (GCN 1.0/1.1): Bits | Name | Description ------|----------|------------------------------ 0-7 | VDST | Vector destination operand 8-10 | ABS | Absolute modifiers for source operands (VOP3A) 8-14 | SDST | Scalar destination operand (VOP3B) 11 | CLAMP | CLAMP modifier (VOP3A) 17-25 | OPCODE | Operation code 26-31 | ENCODING | Encoding type. Must be 0b110100 32-40 | SRC0 | First (scalar or vector) source operand 41-49 | SRC1 | Second (scalar or vector) source operand 50-58 | SRC2 | Third (scalar or vector) source operand 59-60 | OMOD | OMOD modifier. Multiplication modifier 61-63 | NEG | Negation modifier for source operands List of fields for VOP3A/VOP3B encoding (GCN 1.2/1.4): Bits | Name | Description ------|----------|------------------------------ 0-7 | VDST | Destination vector operand 8-10 | ABS | Absolute modifiers for source operands (VOP3A) 11-14 | OP_SEL | Operand selection (VOP3A) (GCN 1.4) 8-14 | SDST | Scalar destination operand (VOP3B) 15 | CLAMP | CLAMP modifier 16-25 | OPCODE | Operation code 26-31 | ENCODING | Encoding type. Must be 0b110100 32-40 | SRC0 | First (scalar or vector) source operand 41-49 | SRC1 | Second (scalar or vector) source operand 50-58 | SRC2 | Third (scalar or vector) source operand 59-60 | OMOD | OMOD modifier. Multiplication modifier 61-63 | NEG | Negation modifier for source operands Typical syntax: INSTRUCTION VDST, SRC0, SRC1, SRC2 [MODIFIERS] Modifiers: * CLAMP - clamps destination floating point value in range 0.0-1.0 * MUL:2, MUL:4, DIV:2 - OMOD modifiers. Multiply destination floating point value by 2.0, 4.0 or 0.5 respectively. Clamping applied after OMOD modifier. * -SRC - negate floating point value from source operand. Applied after ABS modifier. * ABS(SRC), |SRC| - apply absolute value to source operand * OP_SEL:VALUE|[B0,...] - operand half selection (0 - lower 16-bits, 1 - bits) NOTE: OMOD modifier doesn't work if output denormals are allowed (5 bit of MODE register for single precision or 7 bit for double precision). NOTE: OMOD and CLAMP modifier affects only for instruction that output is floating point value or for addition/subtraction instructions. NOTE: ABS and negation is applied to source operand for any instruction. Negation and absolute value can be combined: `-ABS(V0)`. Modifiers CLAMP and OMOD (MUL:2, MUL:4 and DIV:2) can be given in random order. Operand half selection (OP_SEL) take value with bits number depends of number operands. Last bit control destination operand. Zero in bit choose lower 16-bits in dword, one choose higher 16-bits. Example: op_sel:[0,1,1] - higher 16-bits in second source and in destination. List of bits of OP_SEL field: Bit | Operand | Description ----|---------|---------------------- 11 | SRC0 | Choose part of SRC0 (first source operand) 12 | SRC1 | Choose part of SRC1 (second source operand) 13 | SRC2 | Choose part of SRC2 (third source operand) 14 | VDST | Choose part of VDST (destination) Limitations for operands: * only one SGPR can be read by instruction. Multiple occurrences of this same SGPR is allowed * only one literal constant can be used, and only when a SGPR or M0 is not used in source operands * only SRC0 can holds LDS_DIRECT Unaligned pairs of SGPRs are allowed in source operands. List of the instructions by opcode (GCN 1.0/1.1): Opcode | Mnemonic (GCN 1.0) | Mnemonic (GCN 1.1) -------------|--------------------|----------------------------- 320 (0x140) | V_MAD_LEGACY_F32 | V_MAD_LEGACY_F32 321 (0x141) | V_MAD_F32 | V_MAD_F32 322 (0x142) | V_MAD_I32_I24 | V_MAD_I32_I24 323 (0x143) | V_MAD_U32_U24 | V_MAD_U32_U24 324 (0x144) | V_CUBEID_F32 | V_CUBEID_F32 325 (0x145) | V_CUBESC_F32 | V_CUBESC_F32 326 (0x146) | V_CUBETC_F32 | V_CUBETC_F32 327 (0x147) | V_CUBEMA_F32 | V_CUBEMA_F32 328 (0x148) | V_BFE_U32 | V_BFE_U32 329 (0x149) | V_BFE_I32 | V_BFE_I32 330 (0x14a) | V_BFI_B32 | V_BFI_B32 331 (0x14b) | V_FMA_F32 | V_FMA_F32 332 (0x14c) | V_FMA_F64 | V_FMA_F64 333 (0x14d) | V_LERP_U8 | V_LERP_U8 334 (0x14e) | V_ALIGNBIT_B32 | V_ALIGNBIT_B32 335 (0x14f) | V_ALIGNBYTE_B32 | V_ALIGNBYTE_B32 336 (0x150) | V_MULLIT_F32 | V_MULLIT_F32 337 (0x151) | V_MIN3_F32 | V_MIN3_F32 338 (0x152) | V_MIN3_I32 | V_MIN3_I32 339 (0x153) | V_MIN3_U32 | V_MIN3_U32 340 (0x154) | V_MAX3_F32 | V_MAX3_F32 341 (0x155) | V_MAX3_I32 | V_MAX3_I32 342 (0x156) | V_MAX3_U32 | V_MAX3_U32 343 (0x157) | V_MED3_F32 | V_MED3_F32 344 (0x158) | V_MED3_I32 | V_MED3_I32 345 (0x159) | V_MED3_U32 | V_MED3_U32 346 (0x15a) | V_SAD_U8 | V_SAD_U8 347 (0x15b) | V_SAD_HI_U8 | V_SAD_HI_U8 348 (0x15c) | V_SAD_U16 | V_SAD_U16 349 (0x15d) | V_SAD_U32 | V_SAD_U32 350 (0x15e) | V_CVT_PK_U8_F32 | V_CVT_PK_U8_F32 351 (0x15f) | V_DIV_FIXUP_F32 | V_DIV_FIXUP_F32 352 (0x160) | V_DIV_FIXUP_F64 | V_DIV_FIXUP_F64 353 (0x161) | V_LSHL_B64 | V_LSHL_B64 354 (0x162) | V_LSHR_B64 | V_LSHR_B64 355 (0x163) | V_ASHR_I64 | V_ASHR_I64 356 (0x164) | V_ADD_F64 | V_ADD_F64 357 (0x165) | V_MUL_F64 | V_MUL_F64 358 (0x166) | V_MIN_F64 | V_MIN_F64 359 (0x167) | V_MAX_F64 | V_MAX_F64 360 (0x168) | V_LDEXP_F64 | V_LDEXP_F64 361 (0x169) | V_MUL_LO_U32 | V_MUL_LO_U32 362 (0x16a) | V_MUL_HI_U32 | V_MUL_HI_U32 363 (0x16b) | V_MUL_LO_I32 | V_MUL_LO_I32 364 (0x16c) | V_MUL_HI_I32 | V_MUL_HI_I32 365 (0x16d) | V_DIV_SCALE_F32 (VOP3B) | V_DIV_SCALE_F32 (VOP3B) 366 (0x16e) | V_DIV_SCALE_F64 (VOP3B) | V_DIV_SCALE_F64 (VOP3B) 367 (0x16f) | V_DIV_FMAS_F32 | V_DIV_FMAS_F32 368 (0x170) | V_DIV_FMAS_F64 | V_DIV_FMAS_F64 369 (0x171) | V_MSAD_U8 | V_MSAD_U8 370 (0x172) | V_QSAD_U8 | V_QSAD_PK_U16_U8 371 (0x173) | V_MQSAD_U8 | V_MQSAD_PK_U16_U8 372 (0x174) | V_TRIG_PREOP_F64 | V_TRIG_PREOP_F64 373 (0x175) | -- | V_MQSAD_U32_U8 374 (0x176) | -- | V_MAD_U64_U32 (VOP3B) 375 (0x177) | -- | V_MAD_I64_I32 (VOP3B) List of the instructions by opcode (GCN 1.2/1.4): Opcode | Mnemonic (GCN 1.2) | Mnemonic (GCN 1.4) -------------|-------------------------|------------------------- 448 (0x1c0) | V_MAD_LEGACY_F32 | V_MAD_LEGACY_F32 449 (0x1c1) | V_MAD_F32 | V_MAD_F32 450 (0x1c2) | V_MAD_I32_I24 | V_MAD_I32_I24 451 (0x1c3) | V_MAD_U32_U24 | V_MAD_U32_U24 452 (0x1c4) | V_CUBEID_F32 | V_CUBEID_F32 453 (0x1c5) | V_CUBESC_F32 | V_CUBESC_F32 454 (0x1c6) | V_CUBETC_F32 | V_CUBETC_F32 455 (0x1c7) | V_CUBEMA_F32 | V_CUBEMA_F32 456 (0x1c8) | V_BFE_U32 | V_BFE_U32 457 (0x1c9) | V_BFE_I32 | V_BFE_I32 458 (0x1ca) | V_BFI_B32 | V_BFI_B32 459 (0x1cb) | V_FMA_F32 | V_FMA_F32 460 (0x1cc) | V_FMA_F64 | V_FMA_F64 461 (0x1cd) | V_LERP_U8 | V_LERP_U8 462 (0x1ce) | V_ALIGNBIT_B32 | V_ALIGNBIT_B32 463 (0x1cf) | V_ALIGNBYTE_B32 | V_ALIGNBYTE_B32 464 (0x1d0) | V_MIN3_F32 | V_MIN3_F32 465 (0x1d1) | V_MIN3_I32 | V_MIN3_I32 466 (0x1d2) | V_MIN3_U32 | V_MIN3_U32 467 (0x1d3) | V_MAX3_F32 | V_MAX3_F32 468 (0x1d4) | V_MAX3_I32 | V_MAX3_I32 469 (0x1d5) | V_MAX3_U32 | V_MAX3_U32 470 (0x1d6) | V_MED3_F32 | V_MED3_F32 471 (0x1d7) | V_MED3_I32 | V_MED3_I32 472 (0x1d8) | V_MED3_U32 | V_MED3_U32 473 (0x1d9) | V_SAD_U8 | V_SAD_U8 474 (0x1da) | V_SAD_HI_U8 | V_SAD_HI_U8 475 (0x1db) | V_SAD_U16 | V_SAD_U16 476 (0x1dc) | V_SAD_U32 | V_SAD_U32 477 (0x1dd) | V_CVT_PK_U8_F32 | V_CVT_PK_U8_F32 478 (0x1de) | V_DIV_FIXUP_F32 | V_DIV_FIXUP_F32 479 (0x1df) | V_DIV_FIXUP_F64 | V_DIV_FIXUP_F64 480 (0x1e0) | V_DIV_SCALE_F32 (VOP3B) | V_DIV_SCALE_F32 (VOP3B) 481 (0x1e1) | V_DIV_SCALE_F64 (VOP3B) | V_DIV_SCALE_F64 (VOP3B) 482 (0x1e2) | V_DIV_FMAS_F32 | V_DIV_FMAS_F32 483 (0x1e3) | V_DIV_FMAS_F64 | V_DIV_FMAS_F64 484 (0x1e4) | V_MSAD_U8 | V_MSAD_U8 485 (0x1e5) | V_QSAD_PK_U16_U8 | V_QSAD_PK_U16_U8 486 (0x1e6) | V_MQSAD_PK_U16_U8 | V_MQSAD_PK_U16_U8 487 (0x1e7) | V_MQSAD_U32_U8 | V_MQSAD_U32_U8 488 (0x1e8) | V_MAD_U64_U32 (VOP3B) | V_MAD_U64_U32 (VOP3B) 489 (0x1e9) | V_MAD_I64_I32 (VOP3B) | V_MAD_I64_I32 (VOP3B) 490 (0x1ea) | V_MAD_F16 | V_MAD_LEGACY_F16 491 (0x1eb) | V_MAD_U16 | V_MAD_LEGACY_U16 492 (0x1ec) | V_MAD_I16 | V_MAD_LEGACY_I16 493 (0x1ed) | V_PERM_B32 | V_PERM_B32 494 (0x1ee) | V_FMA_F16 | V_FMA_LEGACY_F16 495 (0x1ef) | V_DIV_FIXUP_F16 | V_DIV_FIXUP_LEGACY_F16 496 (0x1f0) | V_CVT_PKACCUM_U8_F32 | V_CVT_PKACCUM_U8_F32 497 (0x1f1) | -- | V_MAD_U32_U16 498 (0x1f2) | -- | V_MAD_I32_I16 499 (0x1f3) | -- | V_XAD_U32 500 (0x1f4) | -- | V_MIN3_F16 501 (0x1f5) | -- | V_MIN3_I16 502 (0x1f6) | -- | V_MIN3_U16 503 (0x1f7) | -- | V_MAX3_F16 504 (0x1f8) | -- | V_MAX3_I16 505 (0x1f9) | -- | V_MAX3_U16 506 (0x1fa) | -- | V_MED3_F16 507 (0x1fb) | -- | V_MED3_I16 508 (0x1fc) | -- | V_MED3_U16 509 (0x1fd) | -- | V_LSHL_ADD_U32 510 (0x1fe) | -- | V_ADD_LSHL_U32 511 (0x1ff) | -- | V_ADD3_U32 512 (0x200) | -- | V_LSHL_OR_B32 513 (0x201) | -- | V_AND_OR_B32 514 (0x202) | -- | V_OR3_B32 515 (0x203) | -- | V_MAD_F16 516 (0x204) | -- | V_MAD_U16 517 (0x205) | -- | V_MAD_I16 518 (0x206) | -- | V_FMA_F16 519 (0x207) | -- | V_DIV_FIXUP_F16 624 (0x270) | V_INTERP_P1_F32 (VINTRP) | V_INTERP_P1_F32 (VINTRP) 625 (0x271) | V_INTERP_P2_F32 (VINTRP) | V_INTERP_P2_F32 (VINTRP) 626 (0x272) | V_INTERP_MOV_F32 (VINTRP) | V_INTERP_MOV_F32 (VINTRP) 628 (0x274) | V_INTERP_P1LL_F16 (VINTRP) | V_INTERP_P1LL_F16 (VINTRP) 629 (0x275) | V_INTERP_P1LV_F16 (VINTRP) | V_INTERP_P1LV_F16 (VINTRP) 630 (0x276) | V_INTERP_P2_F16 (VINTRP)| V_INTERP_P2_F16_LEGACY (VINTRP) 631 (0x277) | -- | V_INTERP_P2_F16 (VINTRP) 640 (0x280) | V_ADD_F64 | V_ADD_F64 641 (0x281) | V_MUL_F64 | V_MUL_F64 642 (0x282) | V_MIN_F64 | V_MIN_F64 643 (0x283) | V_MAX_F64 | V_MAX_F64 644 (0x284) | V_LDEXP_F64 | V_LDEXP_F64 645 (0x285) | V_MUL_LO_U32 | V_MUL_LO_U32 646 (0x286) | V_MUL_HI_U32 | V_MUL_HI_U32 647 (0x287) | V_MUL_HI_I32 | V_MUL_HI_I32 648 (0x288) | V_LDEXP_F32 | V_LDEXP_F32 649 (0x289) | V_READLANE_B32 | V_READLANE_B32 650 (0x28a) | V_WRITELANE_B32 | V_WRITELANE_B32 651 (0x28b) | V_BCNT_U32_B32 | V_BCNT_U32_B32 652 (0x28c) | V_MBCNT_LO_U32_B32 | V_MBCNT_LO_U32_B32 653 (0x28d) | V_MBCNT_HI_U32_B32 | V_MBCNT_HI_U32_B32 654 (0x28e) | V_MAC_LEGACY_F32 | V_MAC_LEGACY_F32 655 (0x28f) | V_LSHLREV_B64 | V_LSHLREV_B64 656 (0x290) | V_LSHRREV_B64 | V_LSHRREV_B64 657 (0x291) | V_ASHRREV_I64 | V_ASHRREV_I64 658 (0x292) | V_TRIG_PREOP_F64 | V_TRIG_PREOP_F64 659 (0x293) | V_BFM_B32 | V_BFM_B32 660 (0x294) | V_CVT_PKNORM_I16_F32 | V_CVT_PKNORM_I16_F32 661 (0x295) | V_CVT_PKNORM_U16_F32 | V_CVT_PKNORM_U16_F32 662 (0x296) | V_CVT_PKRTZ_F16_F32 | V_CVT_PKRTZ_F16_F32 663 (0x297) | V_CVT_PK_U16_U32 | V_CVT_PK_U16_U32 664 (0x298) | V_CVT_PK_I16_I32 | V_CVT_PK_I16_I32 665 (0x299) | V_CVT_PKNORM_I16_F16 | V_CVT_PKNORM_I16_F16 666 (0x29a) | V_CVT_PKNORM_U16_F16 | V_CVT_PKNORM_U16_F16 667 (0x29b) | V_READLANE_REGRD_B32 | V_READLANE_REGRD_B32 668 (0x29c) | -- | V_ADD_I32 669 (0x29d) | -- | V_SUB_I32 670 (0x29e) | -- | V_ADD_I16 671 (0x29f) | -- | V_SUB_I16 672 (0x2a0) | -- | V_PACK_B32_F16 ### Instruction set Alphabetically sorted instruction list: #### V_ADD_F64 Opcode: 356 (0x164) for GCN 1.0/1.1; 640 (0x280) for GCN 1.2/1.4 Syntax: V_ADD_F64 VDST(2), SRC0(2), SRC1(2) Description: Add two double FP value from SRC0 and SRC1 and store result to VDST. Operation: ``` VDST = ASDOUBLE(SRC0) + ASDOUBLE(SRC1) ``` #### V_ADD_I16 Opcode: 670 (0x29e) for GCN 1.4 Syntax: V_ADD_I16 VDST, SRC0, SRC1 Description: Add 16-bit signed value from SRC0 to 16-bit signed value from SRC1 and store result to VDST. If CLAMP modifier supplied, then result is saturated to 16-bit signed value. Operation: ``` UINT16 result = (SRC0&0xffff) + (SRC1&0xffff) if (CLAMP) { INT32 temp = SEXT32((INT16)SRC0&0xffff) + SEXT32((INT16)SRC1&0xffff) if (temp > ((1<<16)-1)) result = 0x7fff if temp < (-1<<16) result = 0x8000 } VDST = (VDST & 0xffff0000) | result ``` #### V_ADD_I32 Opcode: 668 (0x29c) for GCN 1.4 Syntax: V_ADD_I32 VDST, SRC0, SRC1 Description: Add signed value from SRC0 to signed value from SRC1 and store result to VDST. If CLAMP modifier supplied, then result is saturated to 32-bit signed value. Operation: ``` VDST = SRC0 + SRC1 if (CLAMP) { INT64 temp = SEXT64(SRC0) + SEXT64(SRC1) if (temp > ((1LL<<31)-1)) VDST = 0x7fffffff if temp < (-1LL<<31) VDST = 0x80000000 } ``` #### V_ADD3_U32 Opcode: 511 (0x1ff) for GCN 1.4 Syntax: V_ADD3_U32 VDST, SRC0, SRC1, SRC2 Description: Make sum from SRC0, SRC1, and SRC2 and store final result to VDST. Operation: ``` VDST = SRC0 + SRC1 + SRC2 ``` #### V_ADD_LSHL_U32 Opcode: 510 (0x1fe) for GCN 1.4 Syntax: V_ADD_LSHL_U32 VDST, SRC0, SRC1, SRC2 Description: Add SRC0 and SRC1 and shift left by (SRC2&31) bits and store result to VDST. Operation: ``` VDST = (SRC0 + SRC1) << (SRC2&31) ``` #### V_ALIGNBIT_B32 Opcode: 334 (0x14e) for GCN 1.0/1.1; 462 (0x1ce) for GCN 1.2/1.4 Syntax: V_ALIGNBIT_B32 VDST, SRC0, SRC1, SRC2 Description: Align bit. Shift right bits in 64-bit stored in SRC1 (low part) and SRC0 (high part) by SRC2&31 bits, and store low 32-bit of the result in VDST. Operation: ``` VDST = (((UINT64)SRC0)<<32) | SRC1) >> (SRC2&31) ``` #### V_ALIGNBYTE_B32 Opcode: 335 (0x14f) for GCN 1.0/1.1; 463 (0x1cf) for GCN 1.2/1.4 Syntax: V_ALIGNBYTE_B32 VDST, SRC0, SRC1, SRC2 Description: Align bit. Shift right bits in 64-bit stored in SRC1 (low part) and SRC0 (high part) by (SRC2&3)*8 bits, and store low 32-bit of the result in VDST. Operation: ``` VDST = (((UINT64)SRC0)<<32) | SRC1) >> ((SRC2&3)*8) ``` #### V_AND_OR_B32 Opcode: 513 (0x201) for GCN 1.4 Syntax: V_AND_OR_B32 VDST, SRC0, SRC1, SRC2 Description: Make btwise AND with SRC0 and SRC1, make bitwise OR with result and SRC2 and store result to VDST. Operation: ``` VDST = (SRC0 & SRC1) | SRC2 ``` #### V_ASHR_I64 Opcode: 355 (0x163) for GCN 1.0/1.1 Syntax: V_ASHR_I32 VDST(2), SRC0(2), SRC1 Description: Arithmetic shift right SRC0 by (SRC1&63) bits and store result into VDST. Operation: ``` VDST = (INT64)SRC0 >> (SRC1&63) ``` #### V_ASHRREV_I64 Opcode: 657 (0x291) for GCN 1.2/1.4 Syntax: V_ASHRREV_I32 VDST(2), SRC0, SRC1(2) Description: Arithmetic shift right SRC1 by (SRC0&63) bits and store result into VDST. Operation: ``` VDST = (INT64)SRC0 >> (SRC0&63) ``` #### V_BCNT_U32_B32 Opcode: 651 (0x28b) for GCN 1.2/1.4 Syntax: V_BCNT_U32_B32 VDST, SRC0, SRC1 Description: Count bits in SRC0, adds SRC1, and store result to VDST. Operation: ``` VDST = SRC1 + BITCOUNT(SRC0) ``` #### V_BFE_I32 Opcode: 329 (0x149) for GCN 1.0/1.1; 457 (0x1c9) for GCN 1.2/1.4 Syntax: V_BFE_I32 VDST, SRC0, SRC1, SRC2 Description: Extracts bits in SRC0 from range (SRC1&31) with length (SRC2&31) and extend sign from last bit of extracted value, and store result to VDST. Operation: ``` UINT8 shift = SRC1 & 31 UINT8 length = SRC2 & 31 if (length==0) VDST = 0 if (shift+length < 32) VDST = (INT32)(SRC0 << (32 - shift - length)) >> (32 - length) else VDST = (INT32)SRC0 >> shift ``` #### V_BFE_U32 Opcode: 328 (0x148) for GCN 1.0/1.1; 456 (0x1c8) for GCN 1.2/1.4 Syntax: V_BFE_U32 VDST, SRC0, SRC1, SRC2 Description: Extracts bits in SRC0 from range SRC1&31 with length SRC2&31, and store result to VDST. Operation: ``` UINT8 shift = SRC1 & 31 UINT8 length = SRC2 & 31 if (length==0) VDST = 0 if (shift+length < 32) VDST = SRC0 << (32 - shift - length) >> (32 - length) else VDST = SRC0 >> shift ``` #### V_BFI_B32 Opcode: 330 (0x14a) for GCN 1.0/1.1; 458 (0x1ca) for GCN 1.2/1.4 Syntax: V_BFI_B32 VDST, SRC0, SRC1, SRC2 Description: Replace bits in SRC2 by bits from SRC1 marked by bits in SRC0, and store result to VDST. Operation: ``` VDST = (SRC0 & SRC1) | (~SRC0 & SRC2) ``` #### V_BFM_B32 Opcode: 659 (0x293) for GCN 1.2/1.4 Syntax: V_BFM_B32 VDST, SRC0, SRC1 Description: Make 32-bit bitmask from (SRC1 & 31) bit that have length (SRC0 & 31) and store it to VDST. Operation: ``` VDST = ((1U << (SRC0&31))-1) << (SRC1&31) ``` #### V_CUBEID_F32 Opcode: 324 (0x144) for GCN 1.0/1.1; 452 (0x1c4) for GCN 1.2/1.4 Syntax: V_CUBEID_F32 VDST, SRC0, SRC1, SRC2 Description: Cubemap face identification. Determine face by comparing three single FP values: SRC0 (X), SRC1 (Y), SRC2(Z). Choose highest absolute value and check whether is negative or positive. Store floating point value of face ID: (DIM*2.0)+(V[DIM]>=0?1:0), where DIM is number of choosen dimension (X - 0, Y - 1, Z - 2); V - vector = [SRC0, SRC1, SRC2]. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) FLOAT OUT if (ABS(SF2) >= ABS(SF1) && ABS(SF2) >= ABS(SF0)) OUT = (SF2 >= 0.0) ? 4 : 5 else if (ABS(SF1) >= ABS(SF0) OUT = (SF1 >= 0.0) ? 2 : 3 else OUT = (SF0 >= 0.0) ? 0 : 1 VDST = OUT ``` #### V_CUBEMA_F32 Opcode: 327 (0x147) for GCN 1.0/1.1; 455 (0x1c7) for GCN 1.2/1.4 Syntax: V_CUBEMA_F32 VDST, SRC0, SRC1, SRC2 Description: Cubemap Major Axis. Choose highest absolute value from all three FP values (SRC0, SRC1, SRC2) and multiply choosen FP value by two. Result is stored in VDST. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ABS(SF2) >= ABS(SF1) && ABS(SF2) >= ABS(SF0)) OUT = 2*SF2 else if (ABS(SF1) >= ABS(SF0) OUT = 2*SF1 else OUT = 2*SF0 VDST = OUT ``` #### V_CUBESC_F32 Opcode: 325 (0x145) for GCN 1.0/1.1; 453 (0x1c5) for GCN 1.2/1.4 Syntax: V_CUBESC_F32 VDST, SRC0, SRC1, SRC2 Description: Cubemap S coordination. Algorithm below. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ABS(SF2) >= ABS(SF1) && ABS(SF2) >= ABS(SF0)) OUT = SIGN((SF2) * SF0 else if (ABS(SF1) >= ABS(SF0) OUT = SF0 else OUT = -SIGN((SF0) * SF2 VDST = OUT ``` #### V_CUBETC_F32 Opcode: 326 (0x146) for GCN 1.0/1.1; 454 (0x1c6) for GCN 1.2/1.4 Syntax: V_CUBETC_F32 VDST, SRC0, SRC1, SRC2 Description: Cubemap T coordination. Algorithm below. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ABS(SF2) >= ABS(SF1) && ABS(SF2) >= ABS(SF0)) OUT = -SF1 else if (ABS(SF1) >= ABS(SF0) OUT = SIGN(SF1) * SF2 else OUT = -SF1 VDST = OUT ``` #### V_CVT_PK_I16_I32 Opcode: 664 (0x298) for GCN 1.2/1.4 Syntax: V_CVT_PK_I16_I32 VDST, SRC0, SRC1 Description: Convert signed value from SRC0 and SRC1 to signed 16-bit values with clamping, and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` INT16 D0 = MAX(MIN((INT32)SRC0, 0x7fff), -0x8000) INT16 D1 = MAX(MIN((INT32)SRC1, 0x7fff), -0x8000) VDST = D0 | (((UINT32)D1) << 16) ``` #### V_CVT_PK_U16_U32 Opcode: 663 (0x297) for GCN 1.2/1.4 Syntax: V_CVT_PK_U16_U32 VDST, SRC0, SRC1 Description: Convert unsigned value from SRC0 and SRC1 to unsigned 16-bit values with clamping, and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` UINT16 D0 = MIN(SRC0, 0xffff) UINT16 D1 = MIN(SRC1, 0xffff) VDST = D0 | (((UINT32)D1) << 16) ``` #### V_CVT_PK_U8_F32 Opcode: 350 (0x15e) for GCN 1.0/1.1; 477 (0x1dd) for GCN 1.2/1.4 Syntax: V_CVT_PK_U8_F32 VDST, SRC0, SRC1, SRC2 Description: Convert floating point value from SRC0 to unsigned byte value with rounding mode from MODE register, and store this byte to (SRC1&3)'th byte with other bytes of SRC2 of VDST. Operation: ``` UINT8 shift = ((SRC1&3) * 8) UINT32 mask = 0xff << shift FLOAT f = RNDINT(ASFLOAT(SRC0)) UINT8 VAL8 = 0 if (ISNAN(f)) VAL8 = (UINT8)MAX(MIN(f, 255.0), 0.0) VDST = (SRC2&~mask) | (((UINT32)VAL8) << shift) ``` #### V_CVT_PKACCUM_U8_F32 Opcode: 496 (0x1f0) for GCN 1.2/1.4 Syntax: V_CVT_PKACCUM_U8_F32 VDST, SRC0, SRC1 Description: Convert floating point value from SRC0 to unsigned byte value with rounding mode from MODE register, and store this byte to (SRC1&3)'th byte of VDST. Operation: ``` UINT8 shift = ((SRC1&3) * 8) UINT32 mask = 0xff << shift FLOAT f = RNDINT(ASFLOAT(SRC0)) UINT8 VAL8 = 0 if (ISNAN(f)) VAL8 = (UINT8)MAX(MIN(f, 255.0), 0.0) VDST = (VDST&~mask) | (((UINT32)VAL8) << shift) ``` #### V_CVT_PKNORM_I16_F16 Opcode: 665 (0x299) for GCN 1.4 Syntax: V_CVT_PKNORM_I16_F16 VDST, SRC0, SRC1 Description: Convert normalized half FP value from SRC0 and SRC1 to signed 16-bit integers with rounding to nearest to even (??), and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` INT16 roundNorm(HALF S) { FLOAT f = RNDNEINT(S*32767) if (ISNAN(f)) return 0 return (INT16)MAX(MIN(f, 32767.0), -32767.0) } VDST = roundNorm(ASHALF(SRC0)) | ((UINT32)roundNorm(ASHALF(SRC1)) << 16) ``` #### V_CVT_PKNORM_I16_F32 Opcode: 660 (0x294) for GCN 1.2/1.4 Syntax: V_CVT_PKNORM_I16_F32 VDST, SRC0, SRC1 Description: Convert normalized FP value from SRC0 and SRC1 to signed 16-bit integers with rounding to nearest to even (??), and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` INT16 roundNorm(FLOAT S) { FLOAT f = RNDNEINT(S*32767) if (ISNAN(f)) return 0 return (INT16)MAX(MIN(f, 32767.0), -32767.0) } VDST = roundNorm(ASFLOAT(SRC0)) | ((UINT32)roundNorm(ASFLOAT(SRC1)) << 16) ``` #### V_CVT_PKNORM_U16_F16 Opcode: 666 (0x29a) for GCN 1.4 Syntax: V_CVT_PKNORM_U16_F16 VDST, SRC0, SRC1 Description: Convert normalized half FP value from SRC0 and SRC1 to unsigned 16-bit integers with rounding to nearest to even (??), and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` UINT16 roundNorm(HALF S) { HALF f = RNDNEINT(S*65535.0) if (ISNAN(f)) return 0 return (INT16)MAX(MIN(f, 65535.0), 0.0) } VDST = roundNorm(ASHALF(SRC0)) | ((UINT32)roundNorm(ASHALF(SRC1)) << 16) ``` #### V_CVT_PKNORM_U16_F32 Opcode: 661 (0x295) for GCN 1.2/1.4 Syntax: V_CVT_PKNORM_U16_F32 VDST, SRC0, SRC1 Description: Convert normalized FP value from SRC0 and SRC1 to unsigned 16-bit integers with rounding to nearest to even (??), and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` UINT16 roundNorm(FLOAT S) { FLOAT f = RNDNEINT(S*65535.0) if (ISNAN(f)) return 0 return (INT16)MAX(MIN(f, 65535.0), 0.0) } VDST = roundNorm(ASFLOAT(SRC0)) | ((UINT32)roundNorm(ASFLOAT(SRC1)) << 16) ``` #### V_CVT_PKRTZ_F16_F32 Opcode: 662 (0x296) for GCN 1.2/1.4 Syntax: V_CVT_PKRTZ_F16_F32 VDST, SRC0, SRC1 Description: Convert normalized FP value from SRC0 and SRC1 to half floating points with rounding to zero, and store first value to low 16-bit and second to high 16-bit of the VDST. Operation: ``` UINT16 D0 = ASINT16(CVT_HALF_RTZ(ASFLOAT(SRC0))) UINT16 D1 = ASINT16(CVT_HALF_RTZ(ASFLOAT(SRC1))) VDST = D0 | (((UINT32)D1) << 16) ``` #### V_DIV_FIXUP_F16 Opcode: 495 (0x1ef) for GCN 1.2; 519 (0x207) for GCN 1.4 Syntax: V_DIV_FIXUP_F16 VDST, SRC0, SRC1, SRC2 Description: Handle all exceptions requires for half floating point division. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. Correct result stored to VDST. Operation: ``` HALF SF0 = ASHALF(SRC0) HALF SF1 = ASHALF(SRC1) HALF SF2 = ASHALF(SRC2) if (ISNAN(SF1) && !ISNAN(SF2)) VDST = QUIETNAN(SF1) else if (ISNAN(SF2)) VDST = QUIETNAN(SF2) else if (SF1 == 0.0 && SF2 == 0.0) VDST = NAN_H else if (ABS(SF1)==INF && ABS(SF2)==INF) VDST = -NAN_H else if (SF1 == 0.0) VDST = INF_H*SIGN(SF1)*SIGN(SF2) else if (ABS(SF1) == INF) VDST = SIGN(SF1)*SIGN(SF2) >=0 ? 0.0 : -0.0 else if (ISNAN(SF0)) VDST = SIGN(SF1)*SIGN(SF2)*INF_H else VDST = SF0 ``` #### V_DIV_FIXUP_F32 Opcode: 351 (0x15f) for GCN 1.0/1.1; 478 (0x1de) for GCN 1.2/1.4 Syntax: V_DIV_FIXUP_F32 VDST, SRC0, SRC1, SRC2 Description: Handle all exceptions requires for single floating point division. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. Correct result stored to VDST. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ISNAN(SF1) && !ISNAN(SF2)) VDST = QUIETNAN(SF1) else if (ISNAN(SF2)) VDST = QUIETNAN(SF2) else if (SF1 == 0.0 && SF2 == 0.0) VDST = NAN else if (ABS(SF1)==INF && ABS(SF2)==INF) VDST = -NAN else if (SF1 == 0.0) VDST = INF*SIGN(SF1)*SIGN(SF2) else if (ABS(SF1) == INF) VDST = SIGN(SF1)*SIGN(SF2) >=0 ? 0.0 : -0.0 else if (ISNAN(SF0)) VDST = SIGN(SF1)*SIGN(SF2)*INF else VDST = SF0 ``` #### V_DIV_FIXUP_F64 Opcode: 352 (0x160) for GCN 1.0/1.1; 479 (0x1df) for GCN 1.2/1.4 Syntax: V_DIV_FIXUP_F64 VDST(2), SRC0(2), SRC1(2), SRC2(2) Description: Handle all exceptions requires for double floating point division. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. Correct result stored to VDST. Operation: ``` DOUBLE SF0 = ASDOUBLE(SRC0) DOUBLE SF1 = ASDOUBLE(SRC1) DOUBLE SF2 = ASDOUBLE(SRC2) if (ISNAN(SF1) && !ISNAN(SF2)) VDST = QUIETNAN(SF1) else if (ISNAN(SF2)) VDST = QUIETNAN(SF2) else if (SF1 == 0.0 && SF2 == 0.0) VDST = NAN else if (ABS(SF1)==INF && ABS(SF2)==INF) VDST = -NAN else if (SF1 == 0.0) VDST = INF*SIGN(SF1)*SIGN(SF2) else if (ABS(SF1) == INF) VDST = SIGN(SF1)*SIGN(SF2) >=0 ? 0.0 : -0.0 else if (ISNAN(SF0)) VDST = SIGN(SF1)*SIGN(SF2)*INF else VDST = SF0 ``` #### V_DIV_FIXUP_LEGACY_F16 Opcode: 495 (0x1ef) for GCN 1.4 Syntax: V_DIV_FIXUP_LEGACY_F16 VDST, SRC0, SRC1, SRC2 Description: Handle all exceptions requires for half floating point division. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. Correct result stored to VDST. Operation: ``` HALF SF0 = ASHALF(SRC0) HALF SF1 = ASHALF(SRC1) HALF SF2 = ASHALF(SRC2) if (ISNAN(SF1) && !ISNAN(SF2)) VDST = QUIETNAN(SF1) else if (ISNAN(SF2)) VDST = QUIETNAN(SF2) else if (SF1 == 0.0 && SF2 == 0.0) VDST = NAN_H else if (ABS(SF1)==INF && ABS(SF2)==INF) VDST = -NAN_H else if (SF1 == 0.0) VDST = INF_H*SIGN(SF1)*SIGN(SF2) else if (ABS(SF1) == INF) VDST = SIGN(SF1)*SIGN(SF2) >=0 ? 0.0 : -0.0 else if (ISNAN(SF0)) VDST = SIGN(SF1)*SIGN(SF2)*INF_H else VDST = SF0 ``` #### V_DIV_FMAS_F32 Opcode: 367 (0x16f) for GCN 1.0/1.1; 482 (0x1e2) for GCN 1.2/1.4 Syntax: V_DIV_FMAS_F32 VDST, SRC0, SRC1, SRC2 Description: Special case divide FMA with scale and flags. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. All input values are floating point values. Instruction does fussed multiply and addition, multiply result by POW(2.0, -64) if absolute value of the SRC2 less than 2.0, otherwise multiply by POW(2.0, 64); and store result to VDST. Operation: ``` // SRC0*SRC1+SRC2 VDST = FMA(ASFLOAT(SRC0), ASFLOAT(SRC1), ASFLOAT(SRC2)) if (ABS(ASFLOAT(SRC2)) >= 2.0) VDST = ASFLOAT(VDST)*POW(2.0,64) else VDST = ASFLOAT(VDST)*POW(2.0,-64) ``` #### V_DIV_FMAS_F64 Opcode: 368 (0x170) for GCN 1.0/1.1; 483 (0x1e3) for GCN 1.2/1.4 Syntax: V_DIV_FMAS_F64 VDST(2), SRC0(2), SRC1(2), SRC2(2) Description: Special case divide FMA with scale and flags. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. All input values are double floating point values. Instruction does fussed multiply and addition, multiply result by POW(2.0, -128) if absolute value of the SRC2 less than 2.0, otherwise multiply by POW(2.0, 128); and store result to VDST. Operation: ``` // SRC0*SRC1+SRC2 VDST = FMA(ASDOUBLE(SRC0), ASDOUBLE(SRC1), ASDOUBLE(SRC2)) if (ABS(ASDOUBLE(SRC2)) >= 2.0) VDST = ASDOUBLE(VDST)*POW(2.0,128) else VDST = ASDOUBLE(VDST)*POW(2.0,-128) ``` #### V_DIV_SCALE_F32 Opcode (VOP3B): 365 (0x16d) for GCN 1.0/1.1; 480 (0x1e0) for GCN 1.2/1.4 Syntax: V_DIV_SCALE_F32 VDST, SDST(2), SRC0, SRC1, SRC2 Description: Special case divide preop and flags. SRC0 is quotient, SRC1 is denominator, SRC2 is nominator. All input values are floating point values. SRC0 must be equal SRC1 or SRC2 (register can be different, only values must be equal). If absolute value of the input different than SRC0 is greater or equal than T2=POW(2.0, 96+EXP0) (EXP0 is exponent part (base 2) of SRC0) or this value is NaN, then instruction multiply SRC0 by POW(2.0, 64), and store that value to VDST, and set flag in bit for current lane in SDST. If SRC0 is NaN or infinity then store SRC0 to VDST and set flag. Otherwise store SRC0 to VDST and clear flag. Bits for inactive threads in SDST are always zeroed. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) FLOAT S12 = (SRC0!=SRC1) ? SF1 : SF2 SDST = 0 if (ISNAN(SF0) || ABS(SF0) == INF) { VDST = SRC0 SDST = (SDST & ~MASK) | MASK } else if (ABS(S12) >= ABS(POW(2.0, 96+FREXP_EXP(SF0)-1) || ISNAN(S12)) { VDST = SF0 * POW2(2.0, 64) UINT64 MASK = (1ULL<= ABS(POW(2.0, 768+FREXP_EXP(SD0)-1) || ISNAN(S12)) { VDST = SD0 * POW2(2.0, 128) SDST = (SDST & ~MASK) | MASK } else { VDST = SRC0 SDST = (SDST & ~MASK) } ``` #### V_FMA_F16 Opcode: 494 (0x1ee) for GCN 1.2; 518 (0x206) for GCN 1.4 Syntax: V_FMA_F16 VDST, SRC0, SRC1, SRC2 Description: Fused multiply addition on half floating point values from SRC0, SRC1 and SRC2. Result stored in VDST. Operation: ``` // SRC0*SRC1+SRC2 VDST = FMA(ASHALF(SRC0), ASHALF(SRC1), ASHALF(SRC2)) ``` #### V_FMA_F32 Opcode: 331 (0x14b) for GCN 1.0/1.1; 459 (0x1cb) for GCN 1.2/1.4 Syntax: V_FMA_F32 VDST, SRC0, SRC1, SRC2 Description: Fused multiply addition on single floating point values from SRC0, SRC1 and SRC2. Result stored in VDST. Operation: ``` // SRC0*SRC1+SRC2 VDST = FMA(ASFLOAT(SRC0), ASFLOAT(SRC1), ASFLOAT(SRC2)) ``` #### V_FMA_F64 Opcode: 332 (0x14c) for GCN 1.0/1.1; 460 (0x1cc) for GCN 1.2/1.4 Syntax: V_FMA_F64 VDST(2), SRC0(2), SRC1(2), SRC2(2) Description: Fused multiply addition on double floating point values from SRC0, SRC1 and SRC2. Result stored in VDST. Operation: ``` // SRC0*SRC1+SRC2 VDST = FMA(ASDOUBLE(SRC0), ASDOUBLE(SRC1), ASDOUBLE(SRC2)) ``` #### V_FMA_LEGACY_F16 Opcode: 494 (0x1ee) for GCN 1.4 Syntax: V_FMA_LEGACY_F16 VDST, SRC0, SRC1, SRC2 Description: Fused multiply addition on half floating point values from SRC0, SRC1 and SRC2. Result stored in VDST. Operation: ``` // SRC0*SRC1+SRC2 VDST = FMA(ASHALF(SRC0), ASHALF(SRC1), ASHALF(SRC2)) ``` #### V_INTERP_MOV_F32 Opcode: 626 (0x272) for GCN 1.2/1.4 Syntax: V_INTERP_MOV_F32 VDST, PARAMTYPE, ATTR.ATTRCHAN Description: Move parameter value into VDST. The PARAMTYPE is P0, P10 or P20. Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) if (PARAMTYPE==P0) VDST[LANEID] = ASFLOAT(LDS[S + ATTRCHAN*2]) else if (PARAMTYPE==P10) VDST[LANEID] = ASFLOAT(LDS[S + ATTRCHAN*2 + 1]) else if (PARAMTYPE==P20) VDST[LANEID] = ASFLOAT(LDS[S + ATTRCHAN + 8]) ``` #### V_INTERP_P1_F32 Opcode: 624 (0x270) for GCN 1.2/1.4 Syntax: V_INTERP_P1_F32 VDST, VSRC, ATTR.ATTRCHAN Description: Instruction does the first step of the interpolation (P0 + P10*I). The I coordinate given in VSRC register. Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. NOTE: VDST and VSRC registers must not be same. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) FLOAT P0[LANEID] = ASFLOAT(LDS[S + ATTRCHAN*2]) FLOAT P10[LANEID] = ASFLOAT(LDS[S + ATTRCHAN*2 + 1]) VDST[LANEID] = P0[LANEID] + ASFLOAT(VSRC[LANEID]) * P10[LANEID] ``` #### V_INTERP_P1LL_F16 Opcode: 628 (0x274) for GCN 1.2/1.4 Syntax: V_INTERP_P1LL_F16 VDST, VSRC, ATTR.ATTRCHAN [HIGH] Description: Instruction does the first step of the interpolation (P0 + P10*I). The I coordinate given in VSRC register. P0 and P10 factors are half floating point values stored in lower or higher (if 'HIGH' given) part of 32-bit dword. Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. NOTE: VDST and VSRC registers must not be same. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) HALF P0[LANEID], P10[LANEID] if (HIGH) { P0[LANEID] = ASHALF(LDS[S + ATTRCHAN*2] >> 16) P10[LANEID] = ASHALF(LDS[S + ATTRCHAN*2 + 1] >> 16) } else { P0[LANEID] = ASHALF(LDS[S + ATTRCHAN*2] & 0xffff) P10[LANEID] = ASHALF(LDS[S + ATTRCHAN*2 + 1] & 0xffff) } VDST[LANEID] = P0[LANEID] + ASFLOAT(VSRC[LANEID]) * P10[LANEID] ``` #### V_INTERP_P1LV_F16 Opcode: 629 (0x275) for GCN 1.2/1.4 Syntax: V_INTERP_P1LL_F16 VDST, VSRC, ATTR.ATTRCHAN, VSRC1 [HIGH] Description: Instruction does the first step of the interpolation (P0 + P10*I). The I coordinate given in VSRC register. P10 and P0 factors are half floating point values stored in lower or higher (if 'HIGH' given) part of 32-bit dword. P0 is stored in VSRC1. Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. NOTE: VDST and VSRC registers must not be same. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) HALF P0[LANEID], P10[LANEID] if (HIGH) { P0[LANEID] = ASHALF(VSRC1[LANEID] >> 16) P10[LANEID] = ASHALF(LDS[S + ATTRCHAN*2 + 1] >> 16) } else { P0[LANEID] = ASHALF(VSRC1[LANEID] & 0xffff) P10[LANEID] = ASHALF(LDS[S + ATTRCHAN*2 + 1] & 0xffff) } VDST[LANEID] = P0 + ASFLOAT(VSRC[LANEID]) * P10[LANEID] ``` #### V_INTERP_P2_F32 Opcode: 625 (0x271) for GCN 1.2/1.4 Syntax: V_INTERP_P1_F32 VDST, VSRC, ATTR.ATTRCHAN Description: Instruction does the second step of the interpolation (P20*J + D). The J coordinate given in VSRC register. Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. NOTE: VDST and VSRC registers must not be same. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) FLOAT P20[LANEID] = ASFLOAT(LDS[S + ATTRCHAN + 8]) VDST[LANEID] = ASFLOAT(VDST[LANEID]) + ASFLOAT(VSRC[LANEID]) * P20[LANEID] ``` #### V_INTERP_P2_F16 (GCN 1.2) Opcode: 630 (0x276) for GCN 1.2/1.4 Syntax: V_INTERP_P1_F16 VDST, VSRC, ATTR.ATTRCHAN, VSRC1 [HIGH] Syntax (GCN 1.4): V_INTERP_P1_F16_LEGACY VDST, VSRC, ATTR.ATTRCHAN [HIGH], VSRC1 Description: Instruction does the second step of the interpolation (P20*J + D). The J coordinate given in VSRC register. P2 factor is half floating point value stored in lower or higher (if HIGH given) part of dword. Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. NOTE: VDST and VSRC registers must not be same. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) HALF P20[LANEID] if (HIGH) HALF P20[LANEID] = ASFLOAT(LDS[S + ATTRCHAN + 8] >> 16) else HALF P20[LANEID] = ASFLOAT(LDS[S + ATTRCHAN + 8] & 0xffff) VDST[LANEID] = ASFLOAT(VSRC1[LANEID]) + ASFLOAT(VSRC[LANEID]) * P20[LANEID] ``` #### V_INTERP_P2_F16 (GCN 1.4) Opcode: 631 (0x277) for GCN 1.4 Syntax: V_INTERP_P1_F16 VDST, VSRC, ATTR.ATTRCHAN, VSRC1 [HIGH] Description: Instruction does the second step of the interpolation (P20*J + D). The J coordinate given in VSRC register. P2 factor is half floating point value stored in lower or higher (if HIGH given) part of dword. The 3-bit in OPSEL choose 16-bit part of destination (other part is preserved). Refer to [VINTRP instructions](GcnInstrsVintrp). NOTE: The indices in LDS is dword indices. NOTE: VDST and VSRC registers must not be same. Operation: ``` UINT S = 12*(ATTR*NUMPRIM + PRIMID(LANEID>>2)) HALF P20[LANEID] if (HIGH) HALF P20[LANEID] = ASFLOAT(LDS[S + ATTRCHAN + 8] >> 16) else HALF P20[LANEID] = ASFLOAT(LDS[S + ATTRCHAN + 8] & 0xffff) VDST[LANEID] = ASFLOAT(VSRC1[LANEID]) + ASFLOAT(VSRC[LANEID]) * P20[LANEID] ``` #### V_LDEXP_F32 Opcode: 648 (0x288) for GCN 1.2/1.4 Syntax: V_LDEXP_F32 VDST, SRC0, SRC1 Description: Do ldexp operation on SRC0 and SRC1 (multiply SRC0 by 2**(SRC1)). SRC1 is signed integer, SRC0 is floating point value. Operation: ``` VDST = ASFLOAT(SRC0) * POW(2.0, (INT32)SRC1) ``` #### V_LDEXP_F64 Opcode: 360 (0x168) for GCN 1.0/1.1; 644 (0x284) for GCN 1.2/1.4 Syntax: V_LDEXP_F64 VDST(2), SRC0(2), SRC1 Description: Do ldexp operation on SRC0 and SRC1 (multiply SRC0 by 2**(SRC1)). SRC1 is signed integer, SRC0 is double floating point value. Operation: ``` VDST = ASDOUBLE(SRC0) * POW(2.0, (INT32)SRC1) ``` #### V_LERP_U8 Opcode: 333 (0x14d) for GCN 1.0/1.1; 461 (0x1cd) for GCN 1.2/1.4 Syntax: V_LERP_U8 VDST, SRC0, SRC1, SRC2 Description: For each byte of dword, calculate average from SRC0 byte and SRC1 byte with rounding mode defined in first of the byte SRC2. If rounding bit is set then result for that byte is rounded, otherwise truncated. All bytes will be stored in VDST. Operation: ``` for (UINT8 i = 0; i < 4; i++) { UINT8 S0 = (SRC0 >> (i*8)) & 0xff UINT8 S1 = (SRC1 >> (i*8)) & 0xff UINT8 S2 = (SRC2 >> (i*8)) & 1 VDST = (VDST & ~(255U<<(i*8))) | (((S0+S1+S2) >> 1) << (i*8)) } ``` #### V_LSHL_ADD_U32 Opcode: 509 (0x1fd) for GCN 1.4 Syntax: V_LSHL_ADD_U32 VDST, SRC0, SRC1, SRC2 Description: Shift left SRC0 by (SRC1&31) bits and add to SRC2 and store result to VDST. Operation: ``` VDST = (SRC0 << (SRC1&31)) + SRC2 ``` #### V_LSHL_B64 Opcode: 353 (0x161) for GCN 1.0/1.1 Syntax: V_LSHL_B64 VDST(2), SRC0(2), SRC1 Description: Shift left SRC0 by (SRC1&63) bits and store result into VDST. Operation: ``` VDST = SRC0 << (SRC1&63) ``` #### V_LSHL_OR_B32 Opcode: 512 (0x200) for GCN 1.4 Syntax: V_LSHL_OR_B32 VDST, SRC0, SRC1, SRC2 Description: Shift left SRC0 by (SRC1&31) bits and make bitwise OR with SRC2 and store result to VDST. Operation: ``` VDST = (SRC0 << (SRC1&31)) | SRC2 ``` #### V_LSHLREV_B64 Opcode: 655 (0x28f) for GCN 1.2/1.4 Syntax: V_LSHLREV_B64 VDST(2), SRC0, SRC1(2) Description: Shift left SRC1 by (SRC0&63) bits and store result into VDST. Operation: ``` VDST = SRC1 << (SRC0&63) ``` #### V_LSHR_B64 Opcode: 354 (0x162) for GCN 1.0/1.1 Syntax: V_LSHR_B64 VDST(2), SRC0(2), SRC1 Description: Shift right SRC0 by (SRC1&63) bits and store result into VDST. Operation: ``` VDST = SRC0 >> (SRC1&63) ``` #### V_LSHRREV_B64 Opcode: 656 (0x290) for GCN 1.2/1.4 Syntax: V_LSHRREV_B64 VDST(2), SRC0, SRC1(2) Description: Shift right SRC1 by (SRC0&63) bits and store result into VDST. Operation: ``` VDST = SRC1 >> (SRC0&63) ``` #### V_MAC_LEGACY_F32 Opcode: 654 (0x28e) for GCN 1.2/1.4 Syntax: V_MAC_LEGACY_F32 VDST, SRC0, SRC1 Description: Multiply FP value from SRC0 by FP value from SRC1 and add result to VDST. If one of value is 0.0 then always do not change VDST (do not apply IEEE rules for 0.0*x). Operation: ``` if (ASFLOAT(SRC0)!=0.0 && ASFLOAT(SRC1)!=0.0) VDST = ASFLOAT(SRC0) * ASFLOAT(SRC1) + ASFLOAT(VDST) ``` #### V_MAD_F16 Opcode: 490 (0x1ea) for GCN 1.2; 515 (0x203) for GCN 1.4 Syntax: V_MAD_F16 VDST, SRC0, SRC1, SRC2 Description: Multiply half FP value from SRC0 by half FP value from SRC1 and add SRC2, and store result to VDST. It applies OMOD modifier to result and it flush denormals. Operation: ``` VDST = ASHALF(SRC0) * ASHALF(SRC1) + ASHALF(SRC2) ``` #### V_MAD_F32 Opcode: 321 (0x141) for GCN 1.0/1.1; 449 (0x1c1) for GCN 1.2/1.4 Syntax: V_MAD_F32 VDST, SRC0, SRC1, SRC2 Description: Multiply FP value from SRC0 by FP value from SRC1 and add SRC2, and store result to VDST. It applies OMOD modifier to result and it flush denormals. Operation: ``` VDST = ASFLOAT(SRC0) * ASFLOAT(SRC1) + ASFLOAT(SRC2) ``` #### V_MAD_I16 Opcode: 492 (0x1ec) for GCN 1.2; 517 (0x205) for GCN 1.4 Syntax: V_MAD_I16 VDST, SRC0, SRC1, SRC2 Description: Multiply 16-bit signed value from SRC0 by 16-bit signed value from SRC1 and add 16-bit signed value from SRC2, and store 16-bit signed result to VDST. If CLAMP modifier supplied, then result is saturated to 16-bit signed value. Operation: ``` UINT32 temp = (SEXT32((INT16)SRC0)*(INT16)SRC1 + (INT16)SRC2) VDST = CLAMP ? MIN(MAX(temp), -32768), 32767) : temp&0xffff ``` #### V_MAD_I32_I16 Opcode: 498 (0x1f2) for GCN 1.4 Syntax: V_MAD_I32_I16 VDST, SRC0, SRC1, SRC2 Description: Multiply 16-bit signed value from SRC0 by 16-bit signed value from SRC1 and add 32-bit value from SRC2, and store 32-bit result to VDST. Operation: ``` VDST = (UINT32)(SEXT32((INT16)SRC0)*(INT16)SRC1) + SRC2 ``` #### V_MAD_I32_I24 Opcode: 322 (0x142) for GCN 1.0/1.1; 450 (0x1c2) for GCN 1.2/1.4 Syntax: V_MAD_I32_I24 VDST, SRC0, SRC1, SRC2 Description: Multiply 24-bit signed integer value from SRC0 by 24-bit signed value from SRC1, add SRC2 to this product, and and store result to VDST. Operation: ``` INT32 V0 = (INT32)((SRC0&0x7fffff) | (SSRC0&0x800000 ? 0xff800000 : 0)) INT32 V1 = (INT32)((SRC1&0x7fffff) | (SSRC1&0x800000 ? 0xff800000 : 0)) VDST = V0 * V1 + SRC2 ``` #### V_MAD_I64_I32 Opcode (VOP3B): 375 (0x177) for GCN 1.1; 489 (0x1e9) for GCN 1.2/1.4 Syntax: V_MAD_I64_I32 VDST(2), SDST(2), SRC0, SRC1, SRC2(2) Description: Multiply 32-bit signed integer value from SRC0 by 32-bit signed value from SRC1 and add 64-bit unsigned value to this result, and store final result into VDST and store some value of bits to SDST (unknown behavior). Operation: ``` INT64 PROD = (INT64)SRC0*(INT32)SRC1 VDST = SRC2 + PROD SDST = 0 UINT64 mask = (1ULL< SF0 && SF2 > SF1) VDST = SF2 else VDST = MAX(SF1, SF0) ``` #### V_MAX3_F32 Opcode: 340 (0x154) for GCN 1.0/1.1; 467 (0x1d3) for GCN 1.2/1.4 Syntax: V_MAX3_F32 VDST, SRC0, SRC1, SRC2 Description: Choose largest value from FP values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ISNAN(SF0)) VDST = MAX(SF1, SF2) else if (ISNAN(SF1)) VDST = MAX(SF0, SF2) else if (ISNAN(SF2)) VDST = MAX(SF0, SF1) else if (SF2 > SF0 && SF2 > SF1) VDST = SF2 else VDST = MAX(SF1, SF0) ``` #### V_MAX3_I16 Opcode: 504 (0x1f8) for GCN 1.4 Syntax: V_MAX3_I16 VDST, SRC0, SRC1, SRC2 Description: Choose largest value from signed 16-bit integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((INT16)SRC2 > (INT16)SRC0 && (INT16)SRC2 > (INT16)SRC1) VDST = (UINT16)SRC2 else VDST = (UINT16)MAX((INT16)SRC1, (INT16)SRC0) ``` #### V_MAX3_I32 Opcode: 341 (0x155) for GCN 1.0/1.1; 468 (0x1d4) for GCN 1.2/1.4 Syntax: V_MAX3_I32 VDST, SRC0, SRC1, SRC2 Description: Choose largest value from signed integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((INT32)SRC2 > (INT32)SRC0 && (INT32)SRC2 > (INT32)SRC1) VDST = SRC2 else VDST = MAX((INT32)SRC1, (INT32)SRC0) ``` #### V_MAX3_U16 Opcode: 505 (0x1f9) for GCN 1.4 Syntax: V_MAX3_U16 VDST, SRC0, SRC1, SRC2 Description: Choose largest value from unsigned 16-bit integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((UINT16)SRC2 > (UINT16)SRC0 && (UINT16)SRC2 > (UINT16)SRC1) VDST = (UINT16)SRC2 else VDST = MAX((UINT16)SRC1, (UINT16)SRC0) ``` #### V_MAX3_U32 Opcode: 342 (0x156) for GCN 1.0/1.1; 469 (0x1d5) for GCN 1.2/1.4 Syntax: V_MAX3_U32 VDST, SRC0, SRC1, SRC2 Description: Choose largest value from unsigned integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if (SRC2 > SRC0 && SRC2 > SRC1) VDST = SRC2 else VDST = MAX(SRC1, SRC0) ``` #### V_MBCNT_HI_U32_B32 Opcode: 653 (0x28d) for GCN 1.2/1.4 Syntax: V_MBCNT_HI_U32_B32 VDST, SRC0, SRC1 Description: Make mask for all lanes ending at current lane, get from that mask higher 32-bits, use it to mask SSRC0, count bits in that value, and store result to VDST. Operation: ``` UINT32 MASK = ((1ULL << (LANEID-32)) - 1ULL) & SRC0 VDST = SRC1 + BITCOUNT(MASK) ``` #### V_MBCNT_LO_U32_B32 Opcode: 652 (0x28c) for GCN 1.2/1.4 Syntax: V_MBCNT_LO_U32_B32 VDST, SRC0, SRC1 Description: Make mask for all lanes ending at current lane, get from that mask lower 32-bits, use it to mask SSRC0, count bits in that value, and store result to VDST. Operation: ``` UINT32 MASK = ((1ULL << LANEID) - 1ULL) & SRC0 VDST = SRC1 + BITCOUNT(MASK) ``` #### V_MED3_F16 Opcode: 506 (0x1fa) for GCN 1.4 Syntax: V_MED3_F16 VDST, SRC0, SRC1, SRC2 Description: Choose medium value from half FP values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` HALF SF0 = ASHALF(SRC0) HALF SF1 = ASHALF(SRC1) HALF SF2 = ASHALF(SRC2) if (ISNAN(SF0)) VDST = MIN(SF1, SF2) else if (ISNAN(SF1)) VDST = MIN(SF0, SF2) else if (ISNAN(SF2)) VDST = MIN(SF0, SF1) else if ((SF2 > SF1 && SF2 < SF0) || (SF2 < SF1 && SF2 > SF0)) VDST = SF2 else if ((SF1 > SF2 && SF1 < SF0) || (SF1 < SF2 && SF1 > SF0)) VDST = SF1 else VDST = SF0 ``` #### V_MED3_F32 Opcode: 343 (0x157) for GCN 1.0/1.1; 470 (0x1d6) for GCN 1.2/1.4 Syntax: V_MED3_F32 VDST, SRC0, SRC1, SRC2 Description: Choose medium value from FP values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ISNAN(SF0)) VDST = MIN(SF1, SF2) else if (ISNAN(SF1)) VDST = MIN(SF0, SF2) else if (ISNAN(SF2)) VDST = MIN(SF0, SF1) else if ((SF2 > SF1 && SF2 < SF0) || (SF2 < SF1 && SF2 > SF0)) VDST = SF2 else if ((SF1 > SF2 && SF1 < SF0) || (SF1 < SF2 && SF1 > SF0)) VDST = SF1 else VDST = SF0 ``` #### V_MED3_I16 Opcode: 507 (0x1fb) for GCN 1.4 Syntax: V_MED3_I16 VDST, SRC0, SRC1, SRC2 Description: Choose medium value from signed 16-bit integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` INT16 S0 = (INT16)SRC0 INT16 S1 = (INT32)SRC1 INT16 S2 = (INT32)SRC2 if ((S2 > S1 && S2 < S0) || (S2 < S1 && S2 > S0)) VDST = (UINT16)S2 else if ((S1 > S2 && S1 < S0) || (S1 < S2 && S1 > S0)) VDST = (UINT16)S1 else VDST = (UINT16)S0 ``` #### V_MED3_I32 Opcode: 344 (0x158) for GCN 1.0/1.1; 471 (0x1d7) for GCN 1.2/1.4 Syntax: V_MED3_I32 VDST, SRC0, SRC1, SRC2 Description: Choose medium value from signed integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` INT32 S0 = (INT32)SRC0 INT32 S1 = (INT32)SRC1 INT32 S2 = (INT32)SRC2 if ((S2 > S1 && S2 < S0) || (S2 < S1 && S2 > S0)) VDST = S2 else if ((S1 > S2 && S1 < S0) || (S1 < S2 && S1 > S0)) VDST = S1 else VDST = S0 ``` #### V_MED3_U16 Opcode: 508 (0x1fc) for GCN 1.4 Syntax: V_MED3_U16 VDST, SRC0, SRC1, SRC2 Description: Choose medium value from unsigned 16-bit integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` UINT16 S0 = (UINT16)SRC0 UINT16 S1 = (UINT16)SRC1 UINT16 S2 = (UINT16)SRC2 if ((S2 > S1 && S2 < S0) || (S2 < S1 && S2 > S0)) VDST = S2 else if ((S1 > S2 && S1 < S0) || (S1 < S2 && S1 > S0)) VDST = S1 else VDST = S0 ``` #### V_MED3_U32 Opcode: 345 (0x159) for GCN 1.0/1.1; 472 (0x1d8) for GCN 1.2/1.4 Syntax: V_MED3_U32 VDST, SRC0, SRC1, SRC2 Description: Choose medium value from unsigned integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((SRC2 > SRC1 && SRC2 < SRC0) || (SRC2 < SRC1 && SRC2 > SRC0)) VDST = SRC2 else if ((SRC1 > SRC2 && SRC1 < SRC0) || (SRC1 < SRC2 && SRC1 > SRC0)) VDST = SRC1 else VDST = SRC0 ``` #### V_MIN_F64 Opcode: 358 (0x166) for GCN 1.0/1.1; 642 (0x282) for GCN 1.2/1.4 Syntax: V_MIN_F64 VDST(2), SRC0(2), SRC1(2) Description: Choose smallest double FP value from SRC0 and SRC1, and store result to VDST. Operation: ``` VDST = MIN((ASDOUBLE(SRC0), ASDOUBLE(SRC1)) ``` #### V_MIN3_F16 Opcode: 500 (0x1f4) for GCN 1.4 Syntax: V_MIN3_F16 VDST, SRC0, SRC1, SRC2 Description: Choose smallest value from half FP values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` HALF SF0 = ASHALF(SRC0) HALF SF1 = ASHALF(SRC1) HALF SF2 = ASHALF(SRC2) if (ISNAN(SF0)) VDST = MIN(SF1, SF2) else if (ISNAN(SF1)) VDST = MIN(SF0, SF2) else if (ISNAN(SF2)) VDST = MIN(SF0, SF1) else if (SF2 < SF0 && SF2 < SF1) VDST = SF2 else VDST = MIN(SF1, SF0) ``` #### V_MIN3_F32 Opcode: 337 (0x151) for GCN 1.0/1.1; 464 (0x1d0) for GCN 1.2/1.4 Syntax: V_MIN3_F32 VDST, SRC0, SRC1, SRC2 Description: Choose smallest value from FP values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` FLOAT SF0 = ASFLOAT(SRC0) FLOAT SF1 = ASFLOAT(SRC1) FLOAT SF2 = ASFLOAT(SRC2) if (ISNAN(SF0)) VDST = MIN(SF1, SF2) else if (ISNAN(SF1)) VDST = MIN(SF0, SF2) else if (ISNAN(SF2)) VDST = MIN(SF0, SF1) else if (SF2 < SF0 && SF2 < SF1) VDST = SF2 else VDST = MIN(SF1, SF0) ``` #### V_MIN3_I16 Opcode: 501 (0x1f5) for GCN 1.4 Syntax: V_MIN3_I16 VDST, SRC0, SRC1, SRC2 Description: Choose smallest value from signed 16-bit integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((INT16)SRC2 < (INT16)SRC0 && (INT16)SRC2 < (INT16)SRC1) VDST = (UINT16)SRC2 else VDST = (UINT16)MIN((INT16)SRC1, (INT16)SRC0) ``` #### V_MIN3_I32 Opcode: 338 (0x152) for GCN 1.0/1.1; 465 (0x1d1) for GCN 1.2/1.4 Syntax: V_MIN3_I32 VDST, SRC0, SRC1, SRC2 Description: Choose smallest value from signed integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((INT32)SRC2 < (INT32)SRC0 && (INT32)SRC2 < (INT32)SRC1) VDST = SRC2 else VDST = MIN((INT32)SRC1, (INT32)SRC0) ``` #### V_MIN3_U16 Opcode: 502 (0x1f6) for GCN 1.4 Syntax: V_MIN3_U16 VDST, SRC0, SRC1, SRC2 Description: Choose smallest value from unsigned 16-bit integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if ((UINT16)SRC2 < (UINT16)SRC0 && (UINT16)SRC2 < (UINT16)SRC1) VDST = (UINT16)SRC2 else VDST = MIN(S(UINT16)RC1, (UINT16)SRC0) ``` #### V_MIN3_U32 Opcode: 339 (0x153) for GCN 1.0/1.1; 466 (0x1d2) for GCN 1.2/1.4 Syntax: V_MIN3_U32 VDST, SRC0, SRC1, SRC2 Description: Choose smallest value from unsigned integer values SRC0, SRC1, SRC2, and store it to VDST. Operation: ``` if (SRC2 < SRC0 && SRC2 < SRC1) VDST = SRC2 else VDST = MIN(SRC1, SRC0) ``` #### V_MQSAD_U32_U8 Opcode: 373 (0x175) for GCN 1.1; 487 (0x1e7) for GCN 1.2/1.4 Syntax: V_MQSAD_U32_U8 VDST(4), SRC0(2), SRC1, SRC2(4) Description: Compute four masked sum of absolute differences with accumulation. Any that operation get first argument from four bytes begins from N and ends to N+3 (where N is number of operation), second argument is SRC1, and third argument is N'th 32-bit dword from SRC2. Operation: ``` void MSADU8(UINT32 S0, UINT32 S1, UINT32 S2) { UINT64 OUT = S2; for (UINT8 i = 0; i < 4; i++) if ((S1 >> (i*8)) & 0xff) != 0) OUT += ABS(((S0 >> (i*8)) & 0xff) - ((S1 >> (i*8)) & 0xff)) return (UINT32)MIN(OUT,0xffffffff); } VDST = (MSADU8((UINT32)SRC0, SRC1, SRC2) VDST |= (MSADU8((UINT32)(SRC0>>8), SRC1, SRC2>>32)<<32 VDST |= (MSADU8((UINT32)(SRC0>>16), SRC1, SRC2>>64)<<64 VDST |= (MSADU8((UINT32)(SRC0>>24), SRC1, SRC2>>96)<<96 ``` #### V_MQSAD_U8, V_MQSAD_PK_U16_U8 Opcode: 371 (0x173) for GCN 1.0/1.1; 486 (0x1e6) for GCN 1.2/1.4 Syntax (GCN 1.0): V_MQSAD_U8 VDST(2), SRC0(2), SRC1, SRC2(2) Syntax (GCN 1.1/1.2): V_MQSAD_PK_U16_U8 VDST(2), SRC0(2), SRC1, SRC2(2) Description: Compute four masked sum of absolute differences with accumulation. Any that operation get first argument from four bytes begins from N and ends to N+3 (where N is number of operation), second argument is SRC1, and third argument is N'th 16-bit dword from SRC2. Operation: ``` void MSADU8(UINT32 S0, UINT32 S1, UINT32 S2) { UINT32 OUT = S2; for (UINT8 i = 0; i < 4; i++) if ((S1 >> (i*8)) & 0xff) != 0) OUT += ABS(((S0 >> (i*8)) & 0xff) - ((S1 >> (i*8)) & 0xff)) return OUT; } VDST = (MSADU8((UINT32)SRC0, SRC1, SRC2 & 0xffff) VDST |= (MSADU8((UINT32)(SRC0>>8), SRC1, (SRC2>>16) & 0xffff)<<16 VDST |= (MSADU8((UINT32)(SRC0>>16), SRC1, (SRC2>>32) & 0xffff)<<32 VDST |= (MSADU8((UINT32)(SRC0>>24), SRC1, (SRC2>>48) & 0xffff)<<48 ``` #### V_MSAD_U8 Opcode: 369 (0x171) for GCN 1.0/1.1; 484 (0x1e4) for GCN 1.2/1.4 Syntax: V_MSAD_U8 VDST, SRC0, SRC1, SRC2 Description: Calculate sum of absolute differences in SRC0 and SRC1 for bytes that have non-zero value in SRC1; add SRC2 to result, and store result to VDST. Operation: ``` VDST = SRC2 for (UINT8 i = 0; i < 4; i++) if ((SRC1 >> (i*8)) & 0xff) != 0) VDST += ABS(((SRC0 >> (i*8)) & 0xff) - ((SRC1 >> (i*8)) & 0xff)) ``` #### V_MUL_F64 Opcode: 357 (0x165) for GCN 1.0/1.1; 641 (0x281) for GCN 1.2/1.4 Syntax: V_MUL_F64 VDST(2), SRC0(2), SRC1(2) Description: Multiply two double FP values from SRC0 and SRC1 and store result to VDST. Operation: ``` VDST = ASDOUBLE(SRC0) * ASDOUBLE(SRC1) ``` #### V_MUL_HI_I32 Opcode: 364 (0x16c) for GCN 1.0/1.1; 647 (0x287) for GCN 1.2/1.4 Syntax: V_MUL_HI_I32 VDST, SRC0, SRC1 Description: Multiply 32-bit signed value SRC0 and SRC1, and store higher part of the result to VDST. Operation: ``` VDST = ((INT64)SRC0 * (INT32)SRC1) >> 32 ``` #### V_MUL_HI_U32 Opcode: 362 (0x16a) for GCN 1.0/1.1; 646 (0x286) for GCN 1.2/1.4 Syntax: V_MUL_HI_U32 VDST, SRC0, SRC1 Description: Multiply 32-bit unsigned value SRC0 and SRC1, and store higher part of the result to VDST. Operation: ``` VDST = ((UINT64)SRC0 * SRC1) >> 32 ``` #### V_MUL_LO_I32 Opcode: 363 (0x16b) for GCN 1.0/1.1 Syntax: V_MUL_LO_I32 VDST, SRC0, SRC1 Description: Multiply 32-bit signed value SRC0 and SRC1, and store lower part of the result to VDST. Operation: ``` VDST = (INT32)SRC0 * (INT32)SRC1 ``` #### V_MUL_LO_U32 Opcode: 361 (0x169) for GCN 1.0/1.1; 645 (0x285) for GCN 1.2/1.4 Syntax: V_MUL_LO_U32 VDST, SRC0, SRC1 Description: Multiply 32-bit unsigned value SRC0 and SRC1, and store lower part of the result to VDST. Operation: ``` VDST = SRC0 * SRC1 ``` #### V_MULLIT_F32 Opcode: 336 (0x150) for GCN 1.0/1.1 Syntax: V_MULLIT_F32 VDST, SRC0, SRC1, SRC2 Description: Multiply FP value SRC0 and FP value SRC1, and store it to VDST if FP value in SRC2 is greater than zero, otherwise, store -MAX_FLOAT to VDST. If one of value is 0.0 and previous condition is satisfied then always store 0.0 to VDST (do not apply IEEE rules for 0.0*x). Operation: ``` VDST = -MAX_FLOAT if (ASFLOAT(SRC2) > 0.0 && !ISNAN(ASFLOAT(SRC2))) { VDST = 0.0 if (ASFLOAT(SRC0)!=0.0 && ASFLOAT(SRC1)!=0.0) VDST = ASFLOAT(SRC0) * ASFLOAT(SRC1) } ``` #### V_OR3_B32 Opcode: 514 (0x202) for GCN 1.4 Syntax: V_OR3_B32 VDST, SRC0, SRC1, SRC2 Description: Make bitwise OR with SRC0, SRC1 and SRC2 and store result to VDST. Operation: ``` VDST = SRC0 | SRC1 | SRC2 ``` #### V_PACK_B32_F16 Opcode: 672 (0x2a0) for GCN 1.4 Syntax: V_PACK_B32_F16 VDST, SRC0, SRC1 Description: Get lower 16-bits from SRC0 and put to lower 16-bits in VDST, get lower 16-bits from SRC1 and put to higher 16-bits in VDST. Operation: ``` VDST = (SRC0&0xffff) | (SRC1<<16) ``` #### V_PERM_B32 Opcode: 493 (0x1ed) for GCN 1.2/1.4 Syntax: V_PERM_B32 VDST, SRC0, SRC1, SRC2 Description: Permute bytes. Choose for every byte in dword, specified value. Bytes in SRC2 dword selects value for result dword. Value 0-7 choose byte of this index of quadword (64-bit value) built from SRC0 (higher bits) and SRC1 (lower bits). Value from 8-11 choose 0xff*BIT, where BIT is last bit from 2*N+1 from 64-bit value (SRC0,SRC1). Value 12 choose zero. Value equal or greater than 13 choose 0xff. Operation: ``` VDST = 0 UINT64 qword = (((UINT64)SRC0)<<32) | SRC1 for (int i = 0; i < 4; i++) { BYTE choice = (SRC2 >> (8*i)) & 0xff BYTE result if (choice >= 13) result = 0xff else if (choice == 12) result = 0 else if (choice >= 8) result = 0xff * qword>>((choice-8)*16 + 15) else result = (qword >> (choice*8)) & 0xff VDST |= (result << (i*8)) } ``` #### V_QSAD_U8, V_QSAD_PK_U16_U8 Opcode: 370 (0x172) for GCN 1.0/1.1; 485 (0x1e5) for GCN 1.2/1.4 Syntax (GCN 1.0): V_QSAD_U8 VDST(2), SRC0(2), SRC1, SRC2(2) Syntax (GCN 1.1/1.2): V_QSAD_PK_U16_U8 VDST(2), SRC0(2), SRC1, SRC2(2) Description: Compute four sum of absolute differences with accumulation. Any that operation get first argument from four bytes begins from N and ends to N+3 (where N is number of operation), second argument is SRC1, and third argument is N'th 16-bit dword from SRC2. Operation: ``` void SADU8(UINT32 S0, UINT32 S1, UINT32 S2) { UINT32 OUT = S2; for (UINT8 i = 0; i < 4; i++) OUT += ABS(((S0 >> (i*8)) & 0xff) - ((S1 >> (i*8)) & 0xff)) return OUT; } VDST = (SADU8((UINT32)SRC0, SRC1, SRC2 & 0xffff) VDST |= (SADU8((UINT32)(SRC0>>8), SRC1, (SRC2>>16) & 0xffff)<<16 VDST |= (SADU8((UINT32)(SRC0>>16), SRC1, (SRC2>>32) & 0xffff)<<32 VDST |= (SADU8((UINT32)(SRC0>>24), SRC1, (SRC2>>48) & 0xffff)<<48 ``` #### V_READLANE_B32 Opcode: 649 (0x289) for GCN 1.2/1.4 Syntax: V_READLANE_B32 SDST, VSRC0, SSRC1 Description: Copy one VSRC0 lane value to one SDST. Lane (thread id) choosen from SSRC1&63. SSRC1 can be SGPR or M0. Ignores EXEC mask. Operation: ``` SDST = VSRC0[SSRC1 & 63] ``` #### V_SAD_HI_U8 Opcode: 347 (0x15b) for GCN 1.0/1.1; 474 (0x1da) for GCN 1.2/1.4 Syntax: V_SAD_HI_U8 VDST, SRC0, SRC1, SRC2 Description: Calculate sum of absolute differences for all four bytes in SRC0 and SRC1, shift result to high 16-bits, add SRC2 to result, and store result to VDST. Operation: ``` VDST = SRC2 for (UINT8 i = 0; i < 4; i++) VDST += (ABS(((SRC0 >> (i*8)) & 0xff) - ((SRC1 >> (i*8)) & 0xff)))<<16 ``` #### V_SAD_U16 Opcode: 348 (0x15c) for GCN 1.0/1.1; 475 (0x1db) for GCN 1.2/1.4 Syntax: V_SAD_U16 VDST, SRC0, SRC1, SRC2 Description: Calculate sum of absolute differences for two 16-bit words in SRC0 and SRC1, add SRC2 to result, and store result to VDST. Operation: ``` VDST = SRC2 VDST += ABS((SRC0 & 0xffff) - (SRC1 & 0xffff)) VDST += ABS((SRC0 >> 16) - (SRC1 >> 16)) ``` #### V_SAD_U32 Opcode: 349 (0x15d) for GCN 1.0/1.1; 476 (0x1dc) for GCN 1.2/1.4 Syntax: V_SAD_U32 VDST, SRC0, SRC1, SRC2 Description: Calculate sum of absolute difference for SRC0 and SRC1, add SRC2 to result, and store result to VDST. Operation: ``` VDST = SRC2 + ABS(SRC0 - SRC1) ``` #### V_SAD_U8 Opcode: 346 (0x15a) for GCN 1.0/1.1; 473 (0x1d9) for GCN 1.2/1.4 Syntax: V_SAD_U8 VDST, SRC0, SRC1, SRC2 Description: Calculate sum of absolute differences for all four bytes in SRC0 and SRC1, add SRC2 to result, and store result to VDST. Operation: ``` VDST = SRC2 for (UINT8 i = 0; i < 4; i++) VDST += ABS(((SRC0 >> (i*8)) & 0xff) - ((SRC1 >> (i*8)) & 0xff)) ``` #### V_SUB_I16 Opcode: 671 (0x29f) for GCN 1.4 Syntax: V_SUB_I16 VDST, SRC0, SRC1 Description: Subtract 16-bit signed value from SRC1 from 16-bit signed value from SRC0 and store result to VDST. If CLAMP modifier supplied, then result is saturated to 16-bit signed value. Operation: ``` UINT16 result = (SRC0&0xffff) - (SRC1&0xffff) if (CLAMP) { INT32 temp = SEXT32((INT16)SRC0&0xffff) - SEXT32((INT16)SRC1&0xffff) if (temp > ((1<<16)-1)) result = 0x7fff if temp < (-1<<16) result = 0x8000 } VDST = (VDST & 0xffff0000) | result ``` #### V_SUB_I32 Opcode: 669 (0x29d) for GCN 1.4 Syntax: V_SUB_I32 VDST, SRC0, SRC1 Description: Subtract signed value from SRC1 from signed value from SRC0 and store result to VDST. If CLAMP modifier supplied, then result is saturated to 32-bit signed value. Operation: ``` VDST = SRC0 - SRC1 if (CLAMP) { INT64 temp = SEXT64(SRC0) - SEXT64(SRC1) if (temp > ((1LL<<31)-1)) VDST = 0x7fffffff if temp < (-1LL<<31) VDST = 0x80000000 } ``` #### V_TRIG_PREOP_F64 Opcode: 372 (0x174) for GCN 1.0/1.1; 658 (0x292) for GCN 1.2/1.4 Syntax: V_TRIG_PREOP_F64 VDST(2), SRC0(2), SRC1 Description: D.d = Look Up 2/PI (S0.d) with segment select S1.u[4:0]. Save choosen 53 bits of 2/PI in double floating point value in VDST. Second argument is initial segment. First argument is shift of the value (in power form). Bit are numbered from MSB to LSB, begins from value 1.0. Choosen bits begins from: 53\*SEGMENT + (FREXP_EXP(SRC0)-1)-(53\*SEGMENT if SRC0>=POW(2.0, 53\*SEGMENT), otherwise 53\*SEGMENT. Operation: ``` ASDOUBLE SD0 = ASDOUBLE(SRC0) BIT = (SRC1&31) * 53 if (SD0 >= POW(2.0, 53) { if (ABS(SD0) != INF) && !ISNAN(SD0)) BIT += (FREXP_EXP(SD0)-1) - BIT else BIT += 1024 - BIT } VDST = (DOUBLE)(TWOPERPI[BIT:BIT+52]) * POW(2.0, -BIT-53) ``` #### V_WRITELANE_B32 Opcode: 650 (0x28a) for GCN 1.2/1.4 Syntax: V_WRITELANE_B32 VDST, VSRC0, SSRC1 Description: Copy SGPR to one lane of VDST. Lane choosen (thread id) from SSRC1&63. SSRC1 can be SGPR or M0. Ignores EXEC mask. Operation: ``` VDST[SSRC1 & 63] = SSRC0 ``` #### V_XAD_U32 Opcode: 499 (0x1f3) for GCN 1.4 Syntax: V_XAD_U32 VDST, SRC0, SRC1, SRC2 Description: Make XOR bitwise operation on SRC0 and SRC1, add SRC2 and store result to VDST. Instruction added to speed up SHA256 sum. Operation: ``` VDST = (SRC0 ^ SRC1) + SRC2 ```