Floating-point matrix multiply-accumulate

The floating-point matrix multiply-accumulate instruction supports single-precision and double-precision data types in a 2×2 matrix contained in segments of 128 or 256 bits, respectively. It multiplies the 2×2 matrix in each segment of the first source vector by the 2×2 matrix in the corresponding segment of the second source vector. The resulting 2×2 matrix product is then destructively added to the matrix accumulator held in the corresponding segment of the addend and destination vector. This is equivalent to performing a 2-way dot product per destination element. This instruction is unpredicated. The single-precision variant is vector length agnostic. The double-precision variant requires that the Effective SVE vector length is at least 256 bits.

ID_AA64ZFR0_EL1.F32MM indicates whether the single-precision variant is implemented.

ID_AA64ZFR0_EL1.F64MM indicates whether the double-precision variant is implemented.

This instruction is illegal when executed in Streaming SVE mode, unless FEAT_SME_FA64 is implemented and enabled.

It has encodings from 2 classes: 32-bit element and 64-bit element

(FEAT_F32MM)

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 |

0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 1 | Zm | 1 | 1 | 1 | 0 | 0 | 1 | Zn | Zda |

if !IsFeatureImplemented(FEAT_SVE) || !IsFeatureImplemented(FEAT_F32MM) then UNDEFINED; constant integer esize = 32; constant integer n = UInt(Zn); constant integer m = UInt(Zm); constant integer da = UInt(Zda);

(FEAT_F64MM)

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 |

0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 1 | Zm | 1 | 1 | 1 | 0 | 0 | 1 | Zn | Zda |

if !IsFeatureImplemented(FEAT_SVE) || !IsFeatureImplemented(FEAT_F64MM) then UNDEFINED; constant integer esize = 64; constant integer n = UInt(Zn); constant integer m = UInt(Zm); constant integer da = UInt(Zda);

<Zda> |
Is the name of the third source and destination scalable vector register, encoded in the "Zda" field. |

<Zn> |
Is the name of the first source scalable vector register, encoded in the "Zn" field. |

<Zm> |
Is the name of the second source scalable vector register, encoded in the "Zm" field. |

CheckNonStreamingSVEEnabled(); constant integer VL = CurrentVL; constant integer PL = VL DIV 8; if VL < esize * 4 then UNDEFINED; constant integer segments = VL DIV (4 * esize); constant bits(VL) operand1 = Z[n, VL]; constant bits(VL) operand2 = Z[m, VL]; constant bits(VL) operand3 = Z[da, VL]; bits(VL) result = Zeros(VL); bits(4*esize) op1, op2; bits(4*esize) res, addend; for s = 0 to segments-1 op1 = Elem[operand1, s, 4*esize]; op2 = Elem[operand2, s, 4*esize]; addend = Elem[operand3, s, 4*esize]; res = FPMatMulAdd(addend, op1, op2, esize, FPCR); Elem[result, s, 4*esize] = res; Z[da, VL] = result;

This instruction might be immediately preceded in program order by a MOVPRFX instruction. The MOVPRFX must conform to all of the following requirements, otherwise the behavior of the MOVPRFX and this instruction is constrained unpredictable:

- The MOVPRFX must be unpredicated.
- The MOVPRFX must specify the same destination register as this instruction.
- The destination register must not refer to architectural register state referenced by any other source operand register of this instruction.

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.