Arm® A64 Instruction Set Architecture: Armv8, for Armv8-A architecture profile

FEXPA

Floating-point exponential accelerator.

The FEXPA instruction accelerates the polynomial series calculation of the exp(x) function.

The double-precision variant copies the low 52 bits of an entry from a hard-wired table of 64-bit coefficients, indexed by the low 6 bits of each element of the source vector, and prepends to that the next 11 bits of the source element (src<16:6>), setting the sign bit to zero.

The single-precision variant copies the low 23 bits of an entry from hard-wired table of 32-bit coefficients, indexed by the low 6 bits of each element of the source vector, and prepends to that the next 8 bits of the source element (src<13:6>), setting the sign bit to zero.

The half-precision variant copies the low 10 bits of an entry from hard-wired table of 16-bit coefficients, indexed by the low 5 bits of each element of the source vector, and prepends to that the next 5 bits of the source element (src<9:5>), setting the sign bit to zero.

A coefficient table entry with index m holds the floating-point value 2^(m/64), or for the half-precision variant 2^(m/32). This instruction is unpredicated.

if !HaveSVE() then UNDEFINED;
if size == '00' then UNDEFINED;
integer esize = 8 << UInt(size);
integer n = UInt(Zn);
integer d = UInt(Zd);

Assembler Symbols

CheckSVEEnabled();
integer elements = VL DIV esize;
bits(VL) operand  = Z[n];
bits(VL) result;

for e = 0 to elements-1
    bits(esize) element = Elem[operand, e, esize];
    Elem[result, e, esize] = FPExpA(element);

Z[d] = result;