Trevor L. McDonell

Hey, thanks for opening that ticket! Yeah, accelerate takes _ages_ to compile. I have been meaning to look into it for a while but never found the time; hopefully the...

Hey @tdammers, thanks for poking on this, I've been eagerly following the ghc ticket... Honestly that module probably doesn't need all the inlining. I think I just read in the...

In the standard build the test files do get built, unfortunately. The test programs are exported so they can be used for the different backends (`accelerate-llvm-native` and `accelerate-llvm-ptx`) but are...

Hi Jan, sorry for the slow reply... To try and narrow down the problem a bit first: It looks like for each step of your solver you need to iterate...

> OK so I made a little example. Here it is: > http://www.fileswap.com/dl/EW0i7HOfty/ > > Great, thanks! > > So it basically does 2 loops -- one inside, and one...

I pushed some patches that attempt to reduce the number of allocations. For your example it doesn't change things too much, but once you scale up the number of iterations...

I see what you mean about cranking up the iterations. As a workaround, if you change the definition of `loop1` to ``` loop1 steps = run1 $ foldl1 (>->) (P.replicate...

One problem might be that because of the way kernel caching works at the moment, it still needs to generate the code for the kernel before it can check whether...

We should also get an `accelerate-blas` package up, so that the individual `mult`, `dot`, etc., operations are efficient.

If you turn on debugging mode, you can get it to spit out the generated code with `-ddump-cc`. You could compile those yourself and turn it into a foreign function...