FCVT (narrowing, FP16 to FP8)
Multi-vector floating-point convert from half-precision to packed 8-bit floating-point format
Convert each half-precision element of the two source vectors to 8-bit floating-point while scaling the value by 2SInt(FPMR.NSCALE[4:0]), and place the results in the half-width elements of the destination vector. The 8-bit floating-point encoding format is selected by FPMR.F8D.
This instruction is unpredicated.
Green
False
SM_1_only
1
1
0
0
0
0
0
1
0
0
1
0
0
1
0
0
1
1
1
0
0
0
0
FCVT <Zd>.B, { <Zn1>.H-<Zn2>.H }
if !IsFeatureImplemented(FEAT_SME2) || !IsFeatureImplemented(FEAT_FP8) then UNDEFINED;
constant integer n = UInt(Zn:'0');
constant integer d = UInt(Zd);
<Zd>
Is the name of the destination scalable vector register, encoded in the "Zd" field.
<Zn1>
Is the name of the first scalable vector register of the source multi-vector group, encoded as "Zn" times 2.
<Zn2>
Is the name of the second scalable vector register of the source multi-vector group, encoded as "Zn" times 2 plus 1.
CheckFPMREnabled();
CheckStreamingSVEEnabled();
constant integer VL = CurrentVL;
constant integer elements = VL DIV 16;
bits(VL) result;
constant bits(VL) operand1 = Z[n+0, VL];
constant bits(VL) operand2 = Z[n+1, VL];
for e = 0 to elements-1
constant bits(16) element1 = Elem[operand1, e, 16];
constant bits(16) element2 = Elem[operand2, e, 16];
Elem[result, 0*elements + e, 8] = FPConvertFP8(element1, FPCR, FPMR, 8);
Elem[result, 1*elements + e, 8] = FPConvertFP8(element2, FPCR, FPMR, 8);
Z[d, VL] = result;