btmit

I've started thinking about this a bit, at least in the context of overslap-save convolution. The way I think about efficient linear convolution is break it down into a number...

@galenlynch Thank you for taking a look. Your observations are almost entirely correct. - The simple example above does require more memory than a serial approach, specifically by the same...

A specific example that would benefit from this change: ``` n = 8 a = rand(n, n) a = PDMat(a * a') b = UpperTriangular(rand(n, n)) X_A_Xt(a, b) ``` should...

It might be worth extending addition as well. Below uses a full rank update to add two matrices directly using their Cholesky decompositions. Obviously slower than just adding the matrices...

That's a great workaround that will keep me going for now. Thanks!

There shouldn't be any reason you can't do an in-place (i)rfft. FFTW supports this. The trick is that the real input must be padded by either one or two values...

> To all those who expressed interest, this should be much simpler now with the rewrite in #226. No need to figure out all the metaprogramming stuff. Checking in to...

Thanks for the pointer! Do you remember what the last version that worked was? The output appears the same as far back as ~~1.6.7~~ 1.0.5.

Okay, it looks to me that 0.5 broke this by making a comprehension wrap a generator. ### 0.4.7 (2016) I believe this is the last one that would have worked...

Unfortunately this is the best I can do. ```julia macro tuplegen(expr) if expr.head != :comprehension || length(expr.args[1].args) != 2 || expr.args[1].args[2].head != :(=) || typeof(expr.args[1].args[2].args[1]) != Symbol error("Expression following tuplegen...