assuming n = 2^x elements in Vec，
adding elements in Vec using
vec.reduce(_ +& _)
will end up with element width = w + 2^x - 1
if adding them for every two elements like a tree， the width will be w + x
seems
vec.par.reduce(_ +& _)
works for 4 elements, but not for 8 or 16.
Any workaround?
I suggest you take a look at the experimental Interval type in chisel3. There's some limited documentation here. Intervals were designed for just this situation. You shouldn't even have to change everything, just map the elements of your Vec to Intervals and then reduce them.
chisel3.Vec has a reduceTree method that will do what you want:
class Example extends MultiIOModule {
val in = IO(Input(Vec(8, UInt(8.W))))
val out = IO(Output(UInt())) // Infers width to be 11
out := in.reduceTree(_ +& _)
}
Scastie link: https://scastie.scala-lang.org/rhm4je13T0ykEtRS7P7mfQ
Also see the Vec API docs: https://www.chisel-lang.org/api/latest/chisel3/Vec.html