Hello Steve, I'm not up to date on the new broadcast functionality, but I can say that the function norm does not heap allocate any memory, the heap allocation is coming from the x[p:p+2], which creates a new array of length 3 and copies the values from x into it. There are a few solutions for creating sub-vectors that reference the existing array rather creating a new one. To do this without allocation, I use the `unsafe_view` from the ArrayViews package. You can do `nr = norm(unsafe_view( x, p:p+2)). The downside is there is no boundschecking with `unsafe_view`s, and you have to ensure a reference to the original array x exists somewhere (otherwise it might get garbage collected). You can also use `view` from the same package, which does not have those downsides, but it does get heap allocated. You can also use `sub` (renamed `view` in 0.5), which is part of Base, and works on any AbstractArray (ArrayViews only work on Arrays), but also gets heap allocated.
Jared Crean On Sunday, September 25, 2016 at 11:19:39 AM UTC-4, vav...@uwaterloo.ca wrote: > > My code has many operations on small matrices and vectors. In Julia 0.4, > carrying these out with subvector operations causes a needless heap > allocation, so I've done them all with loops. In other words, instead of > nr = norm(x[p:p+2]) > > I write > > nr = 0.0 > for j = p : p + 2 > nr += x[j]^2 > end > nr = sqrt(nr) > > I've also used the Einsum.jl macro package, which implicitly writes the > loops. > > My question is: do the new broadcast and generator-comprehension > operations available in 0.5 or 0.6 make it possible to accomplish these > small operations using one-line solutions that don't allocate heap memory? > > Here are some other examples of operations for which I would like > efficient one-liners: > > t = dot(v[p:p+3], w[p:p+3]) > c += A[p:p+2, q:q+2] * v[q:q+2] # "gaxpy"; c is a length-3 vector > > Thanks, > Steve Vavasis > >
