BFCVT

Single-precision convert to BFloat16 (predicated)

This instruction converts to BFloat16 from single-precision in each active floating-point element of the source vector, and places the results in the corresponding elements of the destination vector. Inactive elements in the destination vector register remain unmodified or are set to zero, depending on whether merging or zeroing predication is selected.

Since the result type is smaller than the input type, the results are zero-extended to fill each destination element.

ID_AA64ZFR0_EL1.BF16 indicates whether this instruction is implemented.

It has encodings from 2 classes: Merging and Zeroing

Merging
((FEAT_SVE || FEAT_SME) && FEAT_BF16)

313029282726252423222120191817161514131211109876543210
0110010110001010101PgZnZd
opcopc2

Encoding

BFCVT <Zd>.H, <Pg>/M, <Zn>.S

Decode for this encoding

if ((!IsFeatureImplemented(FEAT_SVE) && !IsFeatureImplemented(FEAT_SME)) || !IsFeatureImplemented(FEAT_BF16)) then EndOfDecode(Decode_UNDEF); end; let g : integer = UInt(Pg); let n : integer = UInt(Zn); let d : integer = UInt(Zd); let merging : boolean = TRUE;

Zeroing
(FEAT_SVE2p2 || FEAT_SME2p2)

313029282726252423222120191817161514131211109876543210
0110010010011010110PgZnZd
opcopc2

Encoding

BFCVT <Zd>.H, <Pg>/Z, <Zn>.S

Decode for this encoding

if !IsFeatureImplemented(FEAT_SVE2p2) && !IsFeatureImplemented(FEAT_SME2p2) then EndOfDecode(Decode_UNDEF); end; let g : integer = UInt(Pg); let n : integer = UInt(Zn); let d : integer = UInt(Zd); let merging : boolean = FALSE;

Assembler Symbols

<Zd>

Is the name of the destination scalable vector register, encoded in the "Zd" field.

<Pg>

Is the name of the governing scalable predicate register P0-P7, encoded in the "Pg" field.

<Zn>

Is the name of the source scalable vector register, encoded in the "Zn" field.

Operation

CheckSVEEnabled(); let VL : integer{} = CurrentVL(); let PL : integer{} = VL DIV 8; let elements : integer = VL DIV 32; let mask : bits(PL) = P{}(g); let operand : bits(VL) = if AnyActiveElement{PL}(mask, 32) then Z{VL}(n) else Zeros{VL}; var result : bits(VL) = if merging then Z{VL}(d) else Zeros{VL}; for e = 0 to elements-1 do if ActivePredicateElement{PL}(mask, e, 32) then let element : bits(32) = operand[e*:32]; result[(2*e)*:16] = FPConvertBF(element, FPCR()); result[(2*e+1)*:16] = Zeros{16}; end; end; Z{VL}(d) = result;

Operational information

For the "Merging" variant:

The merging variant of 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 the merging variant of this instruction is CONSTRAINED UNPREDICTABLE:


2026-03_rel 2026-03-26 20:48:11

Copyright © 2010-2026 Arm Limited or its affiliates. All rights reserved. This document is Non-Confidential.