Multi-vector BFloat16 dot-product by vector
This instruction computes the dot product of a pair of BF16 values held in the corresponding 32-bit elements of the two or four first source vectors and the second source vector. The single-precision dot product results are destructively added to the corresponding single-precision elements of the ZA single-vector groups.
The single-vector group within each half of or each quarter of the ZA array is selected by the sum of the vector select register and offset, modulo half or quarter the number of ZA array vectors.
The vector group symbol, VGx2 or VGx4, indicates that the ZA operand consists of two or four ZA single-vector groups respectively. The vector group symbol is preferred for disassembly, but optional in assembler source code.
This instruction follows SME2 ZA-targeting BFloat16 numerical behaviors.
This instruction is unpredicated.
Variants: FEAT_SME2 (ARMv9.3)
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
1 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | ||||||||||||||
Zm | Rv | Zn | opc | off3 |
---|
BFDOT ZA.S[<Wv>, <offs>{, VGx2}], { <Zn1>.H-<Zn2>.H }, <Zm>.H
if !IsFeatureImplemented(FEAT_SME2) then EndOfDecode(Decode_UNDEF); constant integer v = UInt('010':Rv); constant integer n = UInt(Zn); constant integer m = UInt('0':Zm); constant integer offset = UInt(off3); constant integer nreg = 2;
Variants: FEAT_SME2 (ARMv9.3)
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
1 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 0 | 0 | 1 | 0 | ||||||||||||||
Zm | Rv | Zn | opc | off3 |
---|
BFDOT ZA.S[<Wv>, <offs>{, VGx4}], { <Zn1>.H-<Zn4>.H }, <Zm>.H
if !IsFeatureImplemented(FEAT_SME2) then EndOfDecode(Decode_UNDEF); constant integer v = UInt('010':Rv); constant integer n = UInt(Zn); constant integer m = UInt('0':Zm); constant integer offset = UInt(off3); constant integer nreg = 4;
CheckStreamingSVEAndZAEnabled(); constant integer VL = CurrentVL; constant integer elements = VL DIV 32; constant integer vectors = VL DIV 8; constant integer vstride = vectors DIV nreg; constant bits(32) vbase = X[v, 32]; integer vec = (UInt(vbase) + offset) MOD vstride; bits(VL) result; for r = 0 to nreg-1 constant bits(VL) operand1 = Z[(n+r) MOD 32, VL]; constant bits(VL) operand2 = Z[m, VL]; constant bits(VL) operand3 = ZAvector[vec, VL]; for e = 0 to elements-1 constant bits(16) elt1_a = Elem[operand1, 2 * e + 0, 16]; constant bits(16) elt1_b = Elem[operand1, 2 * e + 1, 16]; constant bits(16) elt2_a = Elem[operand2, 2 * e + 0, 16]; constant bits(16) elt2_b = Elem[operand2, 2 * e + 1, 16]; bits(32) sum = Elem[operand3, e, 32]; sum = BFDotAdd(sum, elt1_a, elt1_b, elt2_a, elt2_b, FPCR); Elem[result, e, 32] = sum; ZAvector[vec, VL] = result; vec = vec + vstride;