You copied the Doc URL to your clipboard.

BFMMLA

BFloat16 floating-point matrix multiply-accumulate.

This BFloat16 floating-point (BF16) matrix multiply-accumulate instruction multiplies the 2×4 matrix of BF16 values held in each 128-bit segment of the first source vector by the 4×2 BF16 matrix in the corresponding segment of the second source vector. The resulting 2×2 single-precision (FP32) matrix product is then destructively added to the FP32 matrix accumulator held in the corresponding segment of the addend and destination vector. This is equivalent to performing a 4-way dot product per destination element.

This instruction is unpredicated and vector length agnostic.

All floating-point calculations performed by this instruction are performed with the following behaviors, irrespective of the value in FPCR:

* Uses the non-IEEE 754 Round-to-Odd mode, which forces bit 0 of an inexact result to 1, and rounds an overflow to an appropriately signed Infinity.

* The cumulative FPSR exception bits (IDC, IXC, UFC, OFC, DZC and IOC) are not modified.

* Trapped floating-point exceptions are disabled, as if the FPCR trap enable bits (IDE, IXE, UFE, OFE, DZE and IOE) are all zero.

* Denormalized inputs and results are flushed to zero, as if FPCR.FZ == 1.

* Only the Default NaN is generated, as if FPCR.DN == 1.

ID_AA64ZFR0_EL1.BF16 indicates whether this instruction is implemented.

313029282726252423222120191817161514131211109876543210
01100100011Zm111001ZnZda

SVE

BFMMLA <Zda>.S, <Zn>.H, <Zm>.H

if !HaveSVE() || !HaveBF16Ext() then UNDEFINED;
integer n = UInt(Zn);
integer m = UInt(Zm);
integer da = UInt(Zda);

Assembler Symbols

<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.

Operation

CheckSVEEnabled();
integer segments =  VL DIV 128;
bits(VL) operand1 = Z[n];
bits(VL) operand2 = Z[m];
bits(VL) operand3 = Z[da];
bits(VL) result;
bits(128) op1, op2;
bits(128) res, addend;

for s = 0 to segments-1
    op1    = Elem[operand1, s, 128];
    op2    = Elem[operand2, s, 128];
    addend = Elem[operand3, s, 128];
    res    = BFMatMulAdd(addend, op1, op2);
    Elem[result, s, 128] = res;

Z[da] = result;

Operational information

This instruction might be immediately preceded in program order by a MOVPRFX instruction that conforms to all of the following requirements, otherwise the behavior of either or both instructions is unpredictable:

  • The MOVPRFX instruction 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.
The MOVPRFX instructions that can be used with this instruction are as follows:
  • An unpredicated MOVPRFX instruction.