Fused multiple-add instruction about nim HOT 6 OPEN

A quick google search gives:
https://gist.github.com/Yaffle/fb47de4c18b63147699e0b621f1031f7
and
https://gist.github.com/munrocket/adfbca362bf710f53621e483cb208f0e
They are more accurate albeit probably much slower than a pure axpy (a*b+c) computation.
One test that illustrates the difference:

assert( 0.1 * 10 - 1 == 0)
assert( fma(0.1, 10, 1) == 5.55112e-17) # Example from en.cppreference.com

I would expect a different rounding behaviour from a fma function than from an axpy computation.

P.S: The axpy term comes from the BLAS (basic linear algebra subprogram) specification.

from nim.

most recent high-perf x86 CPUs should do a uop-fuse on axpy, the FMA3 instructions really exist for increased accuracy. not really certain about ARM-land.

from nim.

For JS, would something like this be enough?

function fma(a, b, c) {
  return (a * b) + c
}

And an explanation on the documentation for JS targets.

from nim.

Then make it inside a when not defined(js): block and mention in docs is not available in JS.

from nim.

Feel free to open a pull request; see also nim-lang/RFCs#92

from nim.

@ringabout Sorry, I do not have time to implement this.

from nim.