Getting to know Repa in Haskell

Posted on December 6, 2013

Repa, or REgular Parallel Arrays, is a fast unboxed multi-dimensional array implementation in Haskell that can help you achieve speeds near what you would expect from C libraries.

It also has combinators that make it easier to run parallel computation, though I was not able to use them effectively in my initial use.

My reason to learn it

For my Thanksgiving break, I was working on a significant project for my introduction to probability for machine learning course.

I realized that many of my computations throughout this project will be re-using many of the computations from before, so why not break this down into a pipeline with matrices?

It did turn out to be much faster than lazily computed list-monad implementations, which was necessary to process a more data for better results.

Equations

As part of the project, I rendered equations like $$P(c|\underline{x}_i) = \frac{\lambda_c \Pi_{j=1}^V \beta_{c,j}^{x_{i,j}}} {\sum_{k=1}^{K}\lambda_k \Pi_{j=1}^V \beta_{k,j}^{x_{i,j}}}$$ As you can see, the $$\beta_{c,j}^{x_{i,j}}$$ can be considered a result of a function that takes the $$\beta$$ matrix, $$x$$ matrix, and indices $$c$$, $$i$$, and $$j$$. Since this particular part happens both in the top and bottom, it would be senseless to recompute it every time. But, also considering the $$\Pi_{j=1}^V$$ portion, we can actually do a matrix operation here and collapse the results down to just two dimensions, $$c$$, and $$i$$.

Most operations were put into the log domain due to the precision issue. I did try to only use ratios as an experiment, but one EM-Step took 55 minutes.

I’ve heard of accelerate, but it seems tailored mostly to CUDA, which I do not have support for on my laptop. So, I opted for repa, hoping for a short learning curve.

The basics

For the rest of this document, refer to the following type synonyms:

type DoubleMatrix = Array U (Z :. Int :. Int) Double
type IntMatrix = Array U (Z :. Int :. Int) Int
type DynMatrix = Array D (Z :. Int :. Int) Double
type DoubleVector = Array U (Z :. Int) Double
type BetaMap = DoubleMatrix
type Lambda = DoubleVector

So, given the top half of the fun equation, namely $$\Pi_{j=1}^V \beta_{c,j}^{x_{i,j}}$$ (without the lambda), we see that a loop is involved, and $$c$$ and $$i$$ should be given. However, instead of making it a one-off function, let’s just return a matrix where for any coordinate c,i, we have stored the result for this.

However, since this will be computed in the log domain, the respective equation will be $$\sum_{j=1}^V x_{i,j} \cdot log \beta_{c,j}$$

In Haskell, using the list monad, we can create a 2D repa array, which acts like a matrix:

-- | Calculate the matrix, where entry c,j is as if
-- the answer to the single function call using c,j
logBetaXi2  :: BetaMap -- ^ The beta matrix
-> DoubleMatrix -- ^ Our X_i,j count vector,
-- but as doubles.
-> DoubleMatrix -- ^ Response matrix
logBetaXi2 b x = fromListUnboxed (Z :. cs :. is) $do c <- [0..cs-1] -- Start with the first dimension (cs) i <- [0..is-1] -- Then the next dimension (is) let bs = do -- We loop over js times to get each piece -- which we later sum. j <- [0..js-1] -- do some lookups let xij = Repa.index x (Z :. i :. j) bcj = Repa.index b (Z :. c :. j) -- return the entry for this part of (c,i,j) return$ xij * log bcj
-- Now fold the result with a sum to make the singular
-- entry for (c,i)
-- and return it.
return $sum bs where (Z :. cs :. js) = Repa.extent b -- Use the extents to get the current dimensions (Z :. is :. _) = Repa.extent x Now that we have that kind of naive representation out of the way, we can focus on making an equivalent version just using repa operations! logBetaXi :: BetaMap -> DoubleMatrix -> DoubleMatrix logBetaXi bs xs = mmultS lbs xst where lbs = Repa.computeS$ Repa.map log bs
xst = Repa.computeS \$ Repa.transpose xs

W-wait, that was it!? Also, where did the mmultS matrix multiply come from? Well, there’s a repa-algortihms library which has this available. Though I really dislike the restriction that the only things that can go in are already-computed matrices. It also has no warnings if the extents don’t match up for proper multiplication. To this end, I stole the implementation out of the source and added my own check to stop me from making stupid logical mistakes.

So, let’s see, what magic allowed me to do this? First of all, we were doing a summation, and a multiplication over an intermediary dimension $$j$$. This is a sure sign that what you’re doing can be a matrix multiply!

But we can’t just multiply $$x_{i,j}$$ as is. You cannot multiply a $$[i,j]$$ matrix to a $$[c,j]$$ matrix. However you can multiply a $$[c,j]$$ matrix to a $$[j,i]$$. How do we go from a $$[i,j]$$ to a $$[j,i]$$? It is really just a matter of a matrix transpose. This results in a $$[c,i]$$ matrix, which is exactly what we want!

Let’s refer to the result of this operation as $$BX$$.

The not as intuitive basics

Now back to that $$\lambda_c$$ from earlier. We did not simply add it to the equation because it was only a single dimensional vector, and not a matrix. How then, do we add it to $$BX$$?

In the naive list-based generation version, it is trivial to add this to the looping. But as a matrix operation, how to we ensure that for each $$c$$, the correct corresponding $$\lambda_c$$ is added to each entry on $$c,i$$?

It turns out that this is really not difficult at all. We take our vector $$\lambda$$, which is of size $$c$$ (in the sense of $$BD$$ being $$c,i$$ in it’s size), and create a new matrix which has $$\lambda$$ replicated $$i$$ times. We now have a matrix of size $$c,i$$ that we can add to $$BD$$.

For the entire equation that computes $$P(c|\underline{x}_i)$$, we have the following:

logP_C_Given_Xi :: Lambda -> BetaMap -> DoubleMatrix -> DynMatrix
logP_C_Given_Xi l b x = s2
where
-- Get our dimensions, the last parameter
-- being the js, but we don't care.
(Z :. cs :. _) = Repa.extent b
(Z :. is :. _) = Repa.extent x
-- Operation of adding the extended lambda matrix
-- mentioned to BX.
s1 = repaZipAssure (+) lcs bxs
-- Because we are in the log domain, we do a
-- subtraction of the denominator.
s2 = repaZipAssure (-) s1 lsbxse
-- perform the log on lambda
ll = Repa.map log l
-- Now extend it to be i wide.
lcs = Repa.extend (Any :. is) ll
-- Now get our BX matrix
bxs = logBetaXi b x
-- The bottom iterates over the c's and
-- not the i's. So we need to transpose before
-- we sum into a vector (and not a matrix).
-- Folds take place on the "innermost" dimension.
s1t = Repa.transpose s1
-- 'logadd' is a special function that
-- assists with adding in the log domain
-- don't worry about it's implementation
-- only the fact that e^-infinity is 0.
lsbxs = Repa.foldS logadd (-1/0) s1t
-- Now extend it back so we can properly
-- subtract it.
lsbxse = Repa.extend (Any :. cs :. All) lsbxs

Note that I am using some yet-to-be defined repaZipAssure function. This is merely a zip that requires the dimensions to be exactly equal, unlike the available one which goes with the minimum. Otherwise, it is literally just Repa.zipWith f a b, where f is a function, and a and b are equivalently sized arrays.

Conclusion

Previous attempts used maps and sets in attempt to solve this problem, and folding over maps and sets, while building a map or set using an accumulator quickly became

• Less easy to understand
• Painful error messages when typing became an issue
• Performance was terrible

Repa was magnitudes faster than a naive implementation in my class, where we used our language of choice.

I used Haskell because it seemed (and was) easier to express problems like these equations, and test them than what you might expect in something like C++ or Java. Python has it much easier thanks to numPy and the various containers available.

Repa is wonderful for when you know lazy evaluation will only get in the way. In the case shown above, each and every entry in the $$\beta$$ matrix was used to construct the values in the resultant matrix for the vector of probability distributions (which is can be thought of as a matrix).

In case you plan to do numerical processing, and your data can be represented as vectors of vectors–or a matrix–then repa is a good starting place for you.