Multi-vector BFloat16 floating-point dot-product

The instruction computes the dot product of a pair of BF16 values held in the corresponding 32-bit elements of the two or four first and second source vectors. 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.

It has encodings from 2 classes: Two ZA single-vectors and Four ZA single-vectors

(FEAT_SME2)

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 | 1 | 0 | 1 | Zm | 0 | 0 | Rv | 1 | 0 | 0 | Zn | 0 | 1 | 0 | off3 |

if !IsFeatureImplemented(FEAT_SME2) then UNDEFINED; constant integer v = UInt('010':Rv); constant integer n = UInt(Zn:'0'); constant integer m = UInt(Zm:'0'); constant integer offset = UInt(off3); constant integer nreg = 2;

(FEAT_SME2)

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 | 1 | 0 | 1 | Zm | 0 | 1 | 0 | Rv | 1 | 0 | 0 | Zn | 0 | 0 | 1 | 0 | off3 |

if !IsFeatureImplemented(FEAT_SME2) then UNDEFINED; constant integer v = UInt('010':Rv); constant integer n = UInt(Zn:'00'); constant integer m = UInt(Zm:'00'); constant integer offset = UInt(off3); constant integer nreg = 4;

<Wv> |
Is the 32-bit name of the vector select register W8-W11, encoded in the "Rv" field. |

<offs> |
Is the vector select offset, in the range 0 to 7, encoded in the "off3" field. |

<Zn4> |
Is the name of the fourth scalable vector register of the first source multi-vector group, encoded as "Zn" times 4 plus 3. |

<Zn2> |
Is the name of the second scalable vector register of the first source multi-vector group, encoded as "Zn" times 2 plus 1. |

<Zm4> |
Is the name of the fourth scalable vector register of the second source multi-vector group, encoded as "Zm" times 4 plus 3. |

<Zm2> |
Is the name of the second scalable vector register of the second source multi-vector group, encoded as "Zm" times 2 plus 1. |

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, VL]; constant bits(VL) operand2 = Z[m+r, 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;

Internal version only: aarchmrs v2024-03_relA, pseudocode v2024-03_rel, sve v2024-03_rel ; Build timestamp: 2024-03-26T09:45

Copyright © 2010-2024 Arm Limited or its affiliates. All rights reserved. This document is Non-Confidential.