Agena

agena >> `The Power of Procedural Programming` 7.1.4 Cygnus, March 10, 2026 - `muladd`, the fused multiply-add operator, when given at least one complex number, by default no longer uses 80-bit precision. Only if you pass the optional last argument `true` will the operator switch to 80 bits. The default mode is up to three times faster than 80-bit precision. - `muladd` now also allows to easily `chain` multiple fused multiply-add operations together in just one call: > muladd(2, 3, 4!0, 5!0, 1!0): 27 This equals: > 2*3 + 4!0*5!0 + 1!0: 27 and is short for: > muladd(2, 3, muladd(4!0, 5!0, 1!0)): 27 - Likewise, `fma` supports `chaining`: > fma(2, 3, 4!0, 5!0, 6!1, 7!2, 1): 67+19*I which is short for: > fma(2, 3, fma(4!0, 5!0, fma(6!1, 7!2, 1))): 67+19*I - `squareadd` no longer uses 80-bit precision when given at least one complex number, boosting execution speed by a factor of almost three. If you need increased accuracy, you may use the `muladd` operator with the fourth argument `true`. - `calc.savgolcoeffs`, `dual.arccsc`, `dual.arccsch`, `linalg.crossprod` and `linalg.innerprod` have become sightly more accurate by avoiding round-off errors. - The source code has again been cleansed a bit. 7.1.3 Cygnus, March 10, 2026 - The fused multiply-add operator `muladd` now accepts both numbers and complex numbers as input. When given at least one complex number, it internally computes the result with 80-bit precision which `fma` does not. That is why the operator is a bit more accurate, but slower than `fma`. - On Intel platforms when given complex numbers, the result of `squareadd` will now be computed in 80-bit precision, at the expense of speed (around -40 percent). You might want to call `fma` instead which is much faster with complex numbers. - In DOS and with complex numbers, `cis`, `cube`, `foreach`, `roll`, `square`, `squareadd` and `sumup` have been tweaked a tiny bit. - The accuracy of `calc.savgolcoeffs`, `divs.add` and `divs.subtract` has been improved slightly. - Multiplication and division of a `linalg` vector and a scalar with the `*` and `/` operators now work correctly and faulty zero vectors are no longer returned. - The source code has been cleansed a bit.