Re: [Haskell-cafe] Newbie Q: Overloading and type classes
On Sat, Jun 7, 2008 at 5:07 PM, Dmitri O.Kondratiev [EMAIL PROTECTED] wrote:
{--
And what does the word newbie imply to you when answering my question?
In what case using 'fundeps' and 'associated types' will make sence for this
example?
--}
Well, functional dependencies (fundeps) and associated types are
advanced extensions to the language (not Haskell98), and I
personally would not consider it a good idea to approach them until
you have a good handle on typeclasses. Actually, associated types are
not that bad, save for the peculiar restrictions on what is possible
to define (because of decidability issues). Fundeps are a bit
stranger, IMO..
However, since you asked, I'll here try to expound a situation in
which you might need associated types.
Let's say you have a CellHash type that very efficiently stores
key-value pairs, but only if the key is a string. So you have:
data CellHash v = ...
In the simple case, there is no way to make this an instance of Store,
since Store requires a constructor with two parameters. And we can't
really hack around it, because an instance of Store must support *any*
key type by definition.
Instead, we can use associated types, which associate *types* to
instances in a class. These associated types will tell us about the
acceptable keys and values of a particular instance:
class Store s where
type Key s :: *
type Value s :: *
put :: (Key s, Value s) - s - s
get :: Key s - s - Value s
The :: * annotates the kind of the result; for 'normal' value types
it is *, for constructors with one parameter it is * - *, etc.
It is straightforward to implement Store for your Cell and CellList.
instance Store (Cell k v) where -- note that this is not instance
Store Cell anymore
type Key (Cell k v) = k
type Value (Cell k v) = v
-- as before
But now we can make an instance for CellHash too:
instance Store (CellHash v) where
type Key (CellHash v) = String
type Value (CellHash v) = v
-- straightforward
There's a situation where you would want to use associated types. But
they are quite new, and can be hard to work with because of all the
restrictions on their creation. For example, IIRC something like this
would be illegal in the above instance:
type Key (CellHash v) = HashKey String StringHashPolicy
Because the right hand side is larger in terms of number of symbols
than the left. This is untested, and I'm unsure of my knowledge here,
so take the former with a grain of salt.
Fundeps could be used to achieve the same thing, but are not without
their caveats.
Luke
___
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
[Haskell-cafe] Re: Mersenne Build Problem
Bertram Felgenhauer bertram.felgenhauer at googlemail.com writes: The missing symbols are inlined functions. ghc 6.9 doesn't include the header files anymore when compiling via C. (The solution is to create C wrappers around those functions. I guess I'll whip up a patch.) Bertram, Thanks. That's done the trick. I tried 0.1.1 and it works with 6.8 and 0.1.2 works with 6.9. It certainly is a lot faster than system.random. Dominic. ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Design your modules for qualified import
On Sat, Jun 7, 2008 at 3:03 PM, Dan Doel [EMAIL PROTECTED] wrote: On Friday 06 June 2008, Andrew Coppin wrote: It's really quite frustrating that it is 100% impossible to write a single function that will process lists, arrays, sets, maps, byte strings, etc. You have to write several different versions. OK, so some functions really don't make sense for a set because it's unordered, and some functions don't make sense for a map, and so forth. But for example, if I write some complicated algorithm that uses a list to store data and I change that to a set instead, I now have to wade through the function changing every operation from a list-op into a set-op. It's really very annoying! It's not 100% impossible, depending on what exactly you're doing. For instance... Set a, ByteStrings, Seq a, Map k a and [a] are Monoids Set, Array i, Map k, Tree, Seq and [] are Foldable functors Array i, Map k, Tree, Seq, Tree and [] are Traversable functors Those can get you lots of operations (folds, maps, unions...) although there may be gaps that could be filled (Foldable is sufficient to define null, for instance, but it isn't in Data.Foldable). These do help in the case you want to traverse data structures that share a set of properties (e.g. are Functors). However, it doesn't address the problem of sharing an interface among Set, IntSet, ByteSet, etc. In other words there's a need for a Set type class. Hopefully type families will help us write such an interface and keep it efficient. Cheers, Johan ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
Den Sunday 08 June 2008 00.45.01 skrev Xiao-Yong Jin: Any delightful idea to convert my mind from a C shaped one to a Haskell shaped one? You can never go wrong with a good old fashioned hand written tail recursion when you're in doubt, they are pretty much the closest thing to for-loops there is in haskell and should be easy to grok for Imperative programmers and usually produce really fast code. sum vect = let d = dim vect in sum' (d - 1) 0 where sum' 0 s = s + (vect @ 0) sum' index s = sum' (index - 1) (s + (vect @ index)) I don't know the hmatrix api very well, but if there is a function for computing the inner product between two vectors you could always do something like the following meta code: sum v = innerProduct v 1,1,1,1,1,1 Best, Xiao-Yong ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Data.Map traversal.
Hi Serguey, On Sat, Jun 07, 2008 at 05:32:46PM +0400, Serguey Zefirov wrote: So I propose to include those operations into next version of Data.Map. If anyone could point me in the right direction I could do any necessary modifications myself (just because I need it). Please see http://www.haskell.org/haskellwiki/Library_submissions for how to propose changes to the libraries. Thanks Ian ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Issues of the mailto link on the list archive page
On Sat, Jun 07, 2008 at 10:21:13AM -0400, Xiao-Yong Jin wrote: I had private conversation with Andrzej Jaworski about the fact that his reply to Alberto Ruiz's post is off thread. What he did was clicking on the mailto link beside the author's name on the list archive web page [1]. [1] http://www.haskell.org/pipermail/haskell-cafe/2008-June/044023.html I am not quite sure about what this link is supposed to do. But, if it is meant for people to click and reply to the thread, it is not doing this correctly, as the Subject and In-Reply-To fields in the link are obtained from these fields in the head of current email. It would be useful if the link shows Subject with an added Re: if there is not already one, and fill in the in-Reply-To field with the value from the Message-Id field of current email head. An optional Reference field is informative for most of the email client, too. If anyone knows how we can make mailman do this better, please let us know. Thanks Ian ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
[Haskell-cafe] Re: Quick question for a slow program
Lanny Ripple wrote: The second prime generator on this page http://www.haskell.org/haskellwiki/Prime_numbers is quick and easy. I keep it nearby for all those sudden attacks of needing to solve yet another projecteuler problem. The second prime sieve did not create an implicit heap as advertised. I've fixed that and also cleaned up the page a bit, moving this sieve to the section Implicit Heap. Regards, apfelmus ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
Tomas Andersson [EMAIL PROTECTED] writes: You can never go wrong with a good old fashioned hand written tail recursion when you're in doubt, they are pretty much the closest thing to for-loops there is in haskell and should be easy to grok for Imperative programmers and usually produce really fast code. sum vect = let d = dim vect in sum' (d - 1) 0 where sum' 0 s = s + (vect @ 0) sum' index s = sum' (index - 1) (s + (vect @ index)) Do I need the strict version of sum'? Put something like sum' a b | a `seq` b `seq` False = undefined Or ghc will optimize it automatically? I always don't know when such optimization is useful. I don't know the hmatrix api very well, but if there is a function for computing the inner product between two vectors you could always do something like the following meta code: sum v = innerProduct v 1,1,1,1,1,1 I just doubt the efficiency here. If v is a very large vector, I guess the allocation time of that intermediate vector [1,1..] is not small. But it is just my guess. Xiao-Yong -- c/*__o/* \ * (__ */\ ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Re: Patrick Perry's BLAS package
2008/6/6 Patrick Perry [EMAIL PROTECTED]: Apart from some warnings, the library compiles fine in my system. But there is a minor issue about the library it links against when `./Setup test'. I need to use `-lcblas' instead of `-lblas' to get it to link to correct libraries. I don't know other people's system. But in my system, Gentoo Linux, I use blas library provided by atlas, and libblas.so is a fortran library and libcblas.so is for C. I don't know of a good way to get around this. It seems like every system has a different convention for the location and name of the cblas libraries. So, everyone has to edit the extra-libraries and the extra-lib-dirs field in the blas.cabal file. Does anyone know of a better way of doing this? My preference is to use an autoconf script to solve that problem. (build-type: Configure in the cabal file.) For example, the editline and readline C libraries require termcap, which may be linked using one of -lcurses, -ltermcap, etc. The Haskell packages have a configure script which checks which of curses/termcap is available and outputs an editline.buildinfo file with the extra-libraries field filled in. That file is then used automatically by cabal in the build and install phases. See for example: http://code.haskell.org/editline/editline.cabal http://code.haskell.org/editline/configure.ac http://code.haskell.org/editline/editline.buildinfo.in There's also more information is in the Cabal manual: http://haskell.org/cabal/release/latest/doc/users-guide/x30.html#system-dependent Best, -Judah ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
xj2106:
Tomas Andersson [EMAIL PROTECTED] writes:
You can never go wrong with a good old fashioned hand written tail
recursion
when you're in doubt, they are pretty much the closest thing to for-loops
there is in haskell and should be easy to grok for Imperative programmers
and
usually produce really fast code.
sum vect = let d = dim vect
in sum' (d - 1) 0
where sum' 0 s = s + (vect @ 0)
sum' index s = sum' (index - 1) (s + (vect @ index))
Do I need the strict version of sum'? Put something like
sum' a b | a `seq` b `seq` False = undefined
Or ghc will optimize it automatically? I always don't know
when such optimization is useful.
If you give a type annotation, so GHC can tell its an atomic type, like
Int or Double, 's' will be inferred as strict. But if in doubt, use bang
patterns:
{-# LANGUAGE BangPatterns #-}
sum vect = sum' (d-1) 0
where
d = dim vect
go :: Int - Double - Double -- for example
go 0 s = s + (vect @ 0)
go !i !s = go (i-1) (s + (vect @ i))
See this recent post on understanding how these kind of loops are compiled,
http://cgi.cse.unsw.edu.au/~dons/blog/2008/05/16
The core for your loop should look very good, assuming @ is efficient.
-- Don
___
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
Xiao-Yong Jin wrote: Tomas Andersson [EMAIL PROTECTED] writes: You can never go wrong with a good old fashioned hand written tail recursion when you're in doubt, they are pretty much the closest thing to for-loops there is in haskell and should be easy to grok for Imperative programmers and usually produce really fast code. sum vect = let d = dim vect in sum' (d - 1) 0 where sum' 0 s = s + (vect @ 0) sum' index s = sum' (index - 1) (s + (vect @ index)) Do I need the strict version of sum'? Put something like sum' a b | a `seq` b `seq` False = undefined Or ghc will optimize it automatically? I always don't know when such optimization is useful. I don't know the hmatrix api very well, but if there is a function for computing the inner product between two vectors you could always do something like the following meta code: sum v = innerProduct v 1,1,1,1,1,1 I just doubt the efficiency here. If v is a very large vector, I guess the allocation time of that intermediate vector [1,1..] is not small. But it is just my guess. My experience is that Haskell allocation time is very fast and usually negligible in most non trivial matrix computations. A good thing about sum v = constant 1 (dim v) . v is that a constant vector is efficiently created internally (not from an intermediate Haskell list), and the inner product will be computed by the possibly optimized blas version available in your machine. You can also write simple definitions like the next one for the average of the rows of a matrix as a simple vector-matrix product: mean m = w m where w = constant k r r = rows m k = 1 / fromIntegral r I prefer high level index free matrix operations to C-like code. Definitions are clearer, have less bugs, and are also more efficient if you use optimized numerical libs. In any case, many algorithms can be solved by the recursive approach described by Tomas. Alberto Xiao-Yong ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Re: Quick question for a slow program
At least when I teased apart why the first one worked it looked heap-like. Each step of the foldr pulled off the smallest nonprime and merged the next two lists guaranteeing that the next smallest nonprime would be at the head of the next step. Can't argue with results though. The version you have up is about twice as fast as the old one (if a bit harder to read). -ljr apfelmus wrote: Lanny Ripple wrote: The second prime generator on this page http://www.haskell.org/haskellwiki/Prime_numbers is quick and easy. I keep it nearby for all those sudden attacks of needing to solve yet another projecteuler problem. The second prime sieve did not create an implicit heap as advertised. I've fixed that and also cleaned up the page a bit, moving this sieve to the section Implicit Heap. Regards, apfelmus ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
[Haskell-cafe] Printing a random list
I'm just starting out with Haskell, and I could use some help. I'm trying to create a random list and print it out, which seems simple enough, but has been giving me problems. Here's what I have: module Main where import IO import Random randomList :: Random a = a - a- [IO a] randomList lbound ubound = randomRIO(lbound, ubound) : randomList lbound ubound main = do myRandomList - sequence(randomList(0::Int 255)) putStrLn(show(take(10 myRandomList))) - So, I have tried to make a randomList action which defines an infinite random list, bounded by lbound and ubound. It seems that to print this, I need to convert between randomList, which is of type [IO a] to something like IO [a], which is what sequence should do for me. Then I just want to print out the first 10 elements. I'm currently getting the error Only unit numeric type pattern is valid, pointing to 0::Int 255 in the code. I'm not sure what this means. I'm sure I'm looking at this the wrong way, since I'm new to Haskell and haven't quite wrapped my head around it yet. Maybe you can fix the problem by showing me a more Haskell approach to creating a random list and printing it... =) Thanks! - bryan catanzaro ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Printing a random list
catanzar: I'm just starting out with Haskell, and I could use some help. I'm trying to create a random list and print it out, which seems simple enough, but has been giving me problems. Here's what I have: module Main where import IO import Random randomList :: Random a = a - a- [IO a] randomList lbound ubound = randomRIO(lbound, ubound) : randomList lbound ubound main = do myRandomList - sequence(randomList(0::Int 255)) putStrLn(show(take(10 myRandomList))) - So, I have tried to make a randomList action which defines an infinite random list, bounded by lbound and ubound. It seems that to print this, I need to convert between randomList, which is of type [IO a] to something like IO [a], which is what sequence should do for me. Then I just want to print out the first 10 elements. I'm currently getting the error Only unit numeric type pattern is valid, pointing to 0::Int 255 in the code. I'm not sure what this means. Missing parenthesis around the (0 :: Int) type annotation. I'm sure I'm looking at this the wrong way, since I'm new to Haskell and haven't quite wrapped my head around it yet. Maybe you can fix the problem by showing me a more Haskell approach to creating a random list and printing it... =) For lists, best to use the randomRs function, import System.Random main = do g - newStdGen print (take 10 (randomRs (0,255) g :: [Int])) Running it: $ runhaskell A.hs [11,90,187,119,240,57,241,52,143,86] Cheers, Don ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
Alberto Ruiz [EMAIL PROTECTED] writes: My experience is that Haskell allocation time is very fast and usually negligible in most non trivial matrix computations. A good thing about sum v = constant 1 (dim v) . v is that a constant vector is efficiently created internally (not from an intermediate Haskell list), and the inner product will be computed by the possibly optimized blas version available in your machine. You can also write simple definitions like the next one for the average of the rows of a matrix as a simple vector-matrix product: mean m = w m where w = constant k r r = rows m k = 1 / fromIntegral r I prefer high level index free matrix operations to C-like code. Definitions are clearer, have less bugs, and are also more efficient if you use optimized numerical libs. In any case, many algorithms can be solved by the recursive approach described by Tomas. After reading don's blog, I tried to make a test with both methods. The code is very simple, as following module Main where import System import Numeric.LinearAlgebra vsum1 :: Vector Double - Double vsum1 v = constant 1 (dim v) . v vsum2 :: Vector Double - Double vsum2 v = go (d - 1) 0 where d = dim v go :: Int - Double - Double go 0 s = s + (v @ 0) go !j !s = go (j - 1) (s + (v @ j)) mean :: Vector Double - Double mean v = vsum v / fromIntegral (dim v) where vsum = vsum1 main :: IO () main = do fn:nrow:ncol:_ - getArgs print . mean . flatten = fromFile fn (read nrow, read ncol) Compile it with -O2 -optc-O2, and run with a data set of length 500. The results are with vsum1: 80,077,984 bytes allocated in the heap 2,208 bytes copied during GC (scavenged) 64 bytes copied during GC (not scavenged) 40,894,464 bytes maximum residency (2 sample(s)) %GC time 0.0% (0.0% elapsed) Alloc rate35,235,448 bytes per MUT second ./vsum1 huge 500 1 2.25s user 0.09s system 99% cpu 2.348 total This is reasonable, exactly two copies of vector with size of 40MB. with vsum2: 560,743,120 bytes allocated in the heap 19,160 bytes copied during GC (scavenged) 15,920 bytes copied during GC (not scavenged) 40,919,040 bytes maximum residency (2 sample(s)) %GC time 0.3% (0.3% elapsed) Alloc rate222,110,261 bytes per MUT second ./mean2 huge 500 1 2.53s user 0.06s system 99% cpu 2.598 total This is strange. A lot of extra usage of heap? Probably because '@' is not efficient? So it looks like the inner-product approach wins with a fairly margin. -- c/*__o/* \ * (__ */\ ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Printing a random list
Thanks for the response, it does compile after I juggled some parentheses around. And also I appreciate the pointer to the better way of making a random list. So that problem is solved. However, when I ran my random list generator, the interpreter had a stack overflow. Here's my code again: --- module Main where import IO import Random randomList :: Random a = a - a- [IO a] randomList lbound ubound = randomRIO(lbound, ubound) : randomList lbound ubound main = do myRandomList - sequence(randomList (0::Int) 255) putStrLn(show(take 10 myRandomList)) --- It seems that this code somehow tries to evaluate every element of the infinite list defined by randomList. Can you tell me why it is not lazily evaluating this list? I can get around this by changing main to do this instead: --- main = do myRandomList - sequence(take 10 (randomList (0::Int) 255)) putStrLn(show(myRandomList)) --- But I don't understand why sequence(randomList (0::Int) 255) actually tries to evaluate the entire infinite list, instead of just lazily defining a list with the proper types, that I evaluate later when I take elements from it. Thanks for your help! - bryan On Jun 8, 2008, at 4:33 PM, Don Stewart wrote: catanzar: I'm just starting out with Haskell, and I could use some help. I'm trying to create a random list and print it out, which seems simple enough, but has been giving me problems. Here's what I have: module Main where import IO import Random randomList :: Random a = a - a- [IO a] randomList lbound ubound = randomRIO(lbound, ubound) : randomList lbound ubound main = do myRandomList - sequence(randomList(0::Int 255)) putStrLn(show(take(10 myRandomList))) - So, I have tried to make a randomList action which defines an infinite random list, bounded by lbound and ubound. It seems that to print this, I need to convert between randomList, which is of type [IO a] to something like IO [a], which is what sequence should do for me. Then I just want to print out the first 10 elements. I'm currently getting the error Only unit numeric type pattern is valid, pointing to 0::Int 255 in the code. I'm not sure what this means. Missing parenthesis around the (0 :: Int) type annotation. I'm sure I'm looking at this the wrong way, since I'm new to Haskell and haven't quite wrapped my head around it yet. Maybe you can fix the problem by showing me a more Haskell approach to creating a random list and printing it... =) For lists, best to use the randomRs function, import System.Random main = do g - newStdGen print (take 10 (randomRs (0,255) g :: [Int])) Running it: $ runhaskell A.hs [11,90,187,119,240,57,241,52,143,86] Cheers, Don ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Re: Patrick Perry's BLAS package
Judah Jacobson wrote: My preference is to use an autoconf script to solve that problem. (build-type: Configure in the cabal file.) That approach would not work well for BLAS. The various BLAS libraries have profoundly different performance characteristics, and you wouldn't want to get the wrong one for your system, if you had both installed. b ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
Below are some notes on this for Simon PJ and Alberto.
In general, GHC is doing very well here, with only one small wibble preventing
the
recursive version running as fast as the C version.
xj2106:
Alberto Ruiz [EMAIL PROTECTED] writes:
My experience is that Haskell allocation time is very fast and usually
negligible in most non trivial matrix computations.
A good thing about
sum v = constant 1 (dim v) . v
is that a constant vector is efficiently created internally (not from
an intermediate Haskell list), and the inner product will be computed
by the possibly optimized blas version available in your machine.
You can also write simple definitions like the next one for the
average of the rows of a matrix as a simple vector-matrix product:
mean m = w m
where w = constant k r
r = rows m
k = 1 / fromIntegral r
I prefer high level index free matrix operations to C-like
code. Definitions are clearer, have less bugs, and are also more
efficient if you use optimized numerical libs.
In any case, many algorithms can be solved by the recursive approach
described by Tomas.
After reading don's blog, I tried to make a test with both
methods. The code is very simple, as following
module Main where
import System
import Numeric.LinearAlgebra
vsum1 :: Vector Double - Double
vsum1 v = constant 1 (dim v) . v
vsum2 :: Vector Double - Double
vsum2 v = go (d - 1) 0
where
d = dim v
go :: Int - Double - Double
go 0 s = s + (v @ 0)
go !j !s = go (j - 1) (s + (v @ j))
mean :: Vector Double - Double
mean v = vsum v / fromIntegral (dim v)
where vsum = vsum1
main :: IO ()
main = do
fn:nrow:ncol:_ - getArgs
print . mean . flatten = fromFile fn (read nrow, read ncol)
Compile it with -O2 -optc-O2, and run with a data set of
length 500. The results are
with vsum1:
80,077,984 bytes allocated in the heap
2,208 bytes copied during GC (scavenged)
64 bytes copied during GC (not scavenged)
40,894,464 bytes maximum residency (2 sample(s))
%GC time 0.0% (0.0% elapsed)
Alloc rate35,235,448 bytes per MUT second
./vsum1 huge 500 1 2.25s user 0.09s system 99% cpu 2.348 total
This is reasonable, exactly two copies of vector with size
of 40MB.
with vsum2:
560,743,120 bytes allocated in the heap
19,160 bytes copied during GC (scavenged)
15,920 bytes copied during GC (not scavenged)
40,919,040 bytes maximum residency (2 sample(s))
%GC time 0.3% (0.3% elapsed)
Alloc rate222,110,261 bytes per MUT second
./mean2 huge 500 1 2.53s user 0.06s system 99% cpu 2.598 total
This is strange. A lot of extra usage of heap? Probably
because '@' is not efficient? So it looks like the
inner-product approach wins with a fairly margin.
Yes, I'd suspect that @ isn't fully unlifted, so you get some heap
allocation on each index, each time around the loop? We'd have to look
at how @ was defined to spot why.
I began with:
$ cabal install hmatrix
which failed, due to missing linking against gfortran and glscblas
Added those to the cabal file, and hmatrix was installed.
Looking at the core we get from vsum2, we see:
vsum2 compiles quite well
$wgo_s1Nn :: Int# - Double# - Double#
$wgo_s1Nn =
\ (ww3_s1Mc :: Int#) (ww4_s1Mg :: Double#) -
case ww3_s1Mc of ds_X17R {
__DEFAULT -
case Data.Packed.Internal.Vector.$wat
@ Double Foreign.Storable.$f9 ww_s1Ms ww1_s1Mu (I# ds_X17R)
of { D# y_a1KQ -
$wgo_s1Nn (-# ds_X17R 1) (+## ww4_s1Mg y_a1KQ)
};
0 - +## ww4_s1Mg ww2_s1M7
};
} in $wgo_s1Nn (-# ww_s1Ms 1) 0.0
But note the return value from $wat is boxed, only to be immediately
unboxed. That looks to be the source of the heap allocations.
Let's see if we can help that.
Vector is defined as:
data Vector t = V { dim :: Int -- ^ number of elements
, fptr :: ForeignPtr t -- ^ foreign pointer to the
memory block
}
While a much better definition would be:
data Vector t = V { dim :: {-# UNPACK #-} !Int -- ^ number of
elements
, fptr :: {-# UNPACK #-} !(ForeignPtr t)-- ^ foreign
pointer to the memory block
}
And we can add some inlining to at' and at.
That might be enough for GHC to see through to the D# boxing.
Now they're fully unfolded and specialised into our source program,
True -
case unsafeDupablePerformIO
@ Double
((\ (eta1_a1KA :: State# RealWorld) -
case noDuplicate# eta1_a1KA of s'_a1KB {
__DEFAULT -
case readDoubleOffAddr# @ RealWorld ww1_s1NB
ds_X184 s'_a1KB
Re: [Haskell-cafe] Re: Patrick Perry's BLAS package
Bryan O'Sullivan [EMAIL PROTECTED] writes: Judah Jacobson wrote: My preference is to use an autoconf script to solve that problem. (build-type: Configure in the cabal file.) That approach would not work well for BLAS. The various BLAS libraries have profoundly different performance characteristics, and you wouldn't want to get the wrong one for your system, if you had both installed. Is the blas library linked statically? It looks to me that the binary is dynamically linked. linux-vdso.so.1 = (0x7fff003fe000) libgsl.so.0 = /usr/lib/libgsl.so.0 (0x2af4aa8cc000) libblas.so.0 = /usr/lib/libblas.so.0 (0x2af4aac91000) liblapack.so.0 = /usr/lib/liblapack.so.0 (0x2af4aaeb) libutil.so.1 = /lib/libutil.so.1 (0x2af4ab612000) libdl.so.2 = /lib/libdl.so.2 (0x2af4ab815000) libm.so.6 = /lib/libm.so.6 (0x2af4aba19000) libgmp.so.3 = /usr/lib/libgmp.so.3 (0x2af4abc9b000) librt.so.1 = /lib/librt.so.1 (0x2af4abedb000) libc.so.6 = /lib/libc.so.6 (0x2af4ac0e4000) libgslcblas.so.0 = /usr/lib/libgslcblas.so.0 (0x2af4ac424000) libgfortran.so.1 = /usr/lib/gcc/x86_64-pc-linux-gnu/4.1.2/libgfortran.so.1 (0x2af4ac654000) libatlas.so.0 = /usr/lib/libatlas.so.0 (0x2af4ac8eb000) libpthread.so.0 = /lib/libpthread.so.0 (0x2af4ad123000) libcblas.so.0 = /usr/lib/libcblas.so.0 (0x2af4ad33e000) libgcc_s.so.1 = /lib/libgcc_s.so.1 (0x2af4ad55d000) /lib64/ld-linux-x86-64.so.2 (0x2af4aa6b) And presumably the various BLAS implementations share identical interface. You can still change the library at run-time with LD_LIBRARY_PATH. Or am I missing something? Xiao-Yong -- c/*__o/* \ * (__ */\ ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Printing a random list
Bryan Catanzaro wrote: However, when I ran my random list generator, the interpreter had a stack overflow. Here's my code again: --- module Main where import IO import Random randomList :: Random a = a - a- [IO a] randomList lbound ubound = randomRIO(lbound, ubound) : randomList lbound ubound main = do myRandomList - sequence(randomList (0::Int) 255) putStrLn(show(take 10 myRandomList)) --- It seems that this code somehow tries to evaluate every element of the infinite list defined by randomList. You are correct. Can you tell me why it is not lazily evaluating this list? Whenever you use IO, there is a baton being passed along behind the scenes. The baton is called RealWorld# and it represents the fact that interactions with global state and the outside world have to be serialized. In particular, whenever you use the global random number generator, a global state variable has to be updated. This has to be serialized, thus the baton has to be passed along from one action to the next. When you sequence a list of IO actions, you are effectively sending the baton along that list, and you don't get it back until the end of the list is reached. Your code is sending the baton into an infinite list of actions, never to be returned. I can get around this by changing main to do this instead: --- main = do myRandomList - sequence(take 10 (randomList (0::Int) 255)) putStrLn(show(myRandomList)) --- Now you are sending the baton into a list of only 10 actions. The baton comes back, and the program goes on. If you don't know in advance how many random numbers you need, and if you are satisfied with the global random number generator, then Don Stewart's solution is a better approach. Don Stewart wrote: main = do g - newStdGen print (take 10 (randomRs (0,255) g :: [Int])) If you want to be able to reproduce the same sequence of random numbers, for example for testing and debugging purposes, then you can use mkStdGen to create your own random number generator, independent of the global one. The catch is that you will have to thread the state of the random number generator through your code. Once you learn about monads (if you haven't already), you'll recognize that you can use the State monad for your random number generator. ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] sum in hmatrix and blas?
Problem solved. The Haskell loop now wins.
After reading don's blog, I tried to make a test with both
methods. The code is very simple, as following
module Main where
import System
import Numeric.LinearAlgebra
vsum1 :: Vector Double - Double
vsum1 v = constant 1 (dim v) . v
vsum2 :: Vector Double - Double
vsum2 v = go (d - 1) 0
where
d = dim v
go :: Int - Double - Double
go 0 s = s + (v @ 0)
go !j !s = go (j - 1) (s + (v @ j))
mean :: Vector Double - Double
mean v = vsum v / fromIntegral (dim v)
where vsum = vsum1
main :: IO ()
main = do
fn:nrow:ncol:_ - getArgs
print . mean . flatten = fromFile fn (read nrow, read ncol)
Compile it with -O2 -optc-O2, and run with a data set of
length 500. The results are
with vsum1:
80,077,984 bytes allocated in the heap
2,208 bytes copied during GC (scavenged)
64 bytes copied during GC (not scavenged)
40,894,464 bytes maximum residency (2 sample(s))
%GC time 0.0% (0.0% elapsed)
Alloc rate35,235,448 bytes per MUT second
./vsum1 huge 500 1 2.25s user 0.09s system 99% cpu 2.348 total
This is reasonable, exactly two copies of vector with size
of 40MB.
with vsum2:
560,743,120 bytes allocated in the heap
19,160 bytes copied during GC (scavenged)
15,920 bytes copied during GC (not scavenged)
40,919,040 bytes maximum residency (2 sample(s))
%GC time 0.3% (0.3% elapsed)
Alloc rate222,110,261 bytes per MUT second
./mean2 huge 500 1 2.53s user 0.06s system 99% cpu 2.598 total
This is strange. A lot of extra usage of heap? Probably
because '@' is not efficient? So it looks like the
inner-product approach wins with a fairly margin.
Looking at the core we get from vsum2, we see:
vsum2 compiles quite well
$wgo_s1Nn :: Int# - Double# - Double#
$wgo_s1Nn =
\ (ww3_s1Mc :: Int#) (ww4_s1Mg :: Double#) -
case ww3_s1Mc of ds_X17R {
__DEFAULT -
case Data.Packed.Internal.Vector.$wat
@ Double Foreign.Storable.$f9 ww_s1Ms ww1_s1Mu (I#
ds_X17R)
of { D# y_a1KQ -
$wgo_s1Nn (-# ds_X17R 1) (+## ww4_s1Mg y_a1KQ)
};
0 - +## ww4_s1Mg ww2_s1M7
};
} in $wgo_s1Nn (-# ww_s1Ms 1) 0.0
But note the return value from $wat is boxed, only to be immediately
unboxed. That looks to be the source of the heap allocations.
Let's see if we can help that.
Ah, of course. The problem is the unsafePerformIO used to wrap up the
'peek'. This blocks the optimisations somewhat, after we unpack the Vector
type..
So if we rewrite 'safeRead' to not use unsafePerformIO, but instead:
safeRead v = inlinePerformIO . withForeignPtr (fptr v)
inlinePerformIO (IO m) = case m realWorld# of (# _, r #) - r
Along with this change to Vector:
data Vector t = V { dim :: {-# UNPACK #-} !Int
, fptr :: {-# UNPACK #-} !(ForeignPtr t)
}
and some inlining for the bounds checks.
Following bytestrings, we see the performance go from, with a 180M input
file:
Goal:
vsum1:
$ ghc-core A.hs -optc-O2 -fvia-C -fbang-patterns
$ time ./A data 4000 2500
0.727441161678
./A data 4000 2500 5.37s user 0.20s system 99% cpu 5.609 total
160,081,136 bytes allocated in the heap
155 Mb total memory in use
Before:
vsum2-old:
$wgo_s1Ns =
\ (ww3_s1Me :: Int#) (ww4_s1Mi :: Double#) -
case ww3_s1Me of ds_X17Y {
__DEFAULT -
case Data.Packed.Internal.Vector.$wat
@ Double Foreign.Storable.$f9 ww_s1Mu ww1_s1Mw (I# ds_X17Y)
of wild1_a1Hg { D# y_a1Hi -
$wgo_s1Ns (-# ds_X17Y 1) (+## ww4_s1Mi y_a1Hi)
};
0 - +## ww4_s1Mi ww2_s1M9
};
} in $wgo_s1Ns (-# ww_s1Mu 1) 0.0
./A data 4000 2500 +RTS -sstderr
0.72744115876
./A data 4000 2500 +RTS -sstderr 6.04s user 0.15s system 99% cpu 6.203
total
1,121,416,400 bytes allocated in the heap
78 Mb total memory in use
After:
True -
case readDoubleOffAddr#
@ RealWorld ww1_s1Nr ds_X180 realWorld#
of wild2_a1HS { (# s2_a1HU, x_a1HV #) -
case touch# @ ForeignPtrContents ww2_s1Ns s2_a1HU
of s_a1HG { __DEFAULT -
$wgo_s1Ok (-# ds_X180 1) (+## ww4_s1Ng x_a1HV)
}
No needless boxing.
$ time ./A data 4000 2500 +RTS -sstderr
./A data 4000 2500 +RTS -sstderr
0.72744115876
./A data 4000 2500 +RTS -sstderr 5.41s user 0.10s system 99% cpu 5.548
total
80,071,072 bytes allocated in the heap
78 Mb total memory in use
And the total runtime and allocation now is better than the fully 'C' version.
Yay.
Patch attached to the darcs
Re: [Haskell-cafe] Vancouver Haskell users meeting
Same deal but i'm in Ottawa for the summer. I'll be back around september. --ryan 2008/6/6 Asumu Takikawa [EMAIL PROTECTED]: Hi. I'd be interested in a meeting like this, but unfortunately since UBC is done for winter term I'm out of Canada for the summer. If anyone organizes a meet-up come fall I'd happily attend. Cheers, AT On 12:48 Mon 02 Jun , Jon Strait wrote: Anyone else here from Vancouver (Canada)? I thought it would be great to have a little informal get-together at a local cafe and share how we're currently using Haskell, or really anything (problems, comparisons, useful software tools, etc.) in relation to Haskell. I'm scheduling a meeting for this Thursday, June 5th. for 7PM at [1]Waazubee Cafe. (At Commercial Dr. and 1st Ave.) They have wireless internet access. I'll get a table near the back, bring my laptop, and will have a copy of Hudak's SOE book (the front cover is impossible to miss) out on the table. If anyone wants to meet, but this Thursday is not a good day for you, let me know what days are better and we'll move the meeting. If anyone is sure that they will come this Thursday, you might let me know, so I can have an idea about the resistance in changing the day, if needed. Thanks, Jon References 1. http://www.waazubee.com/content/directions.php ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe -BEGIN PGP SIGNATURE- Version: GnuPG v1.4.6 (GNU/Linux) iD8DBQFISL4aPVZMXBlgx7ERAv/OAJwP/1bfduqEa6bTBEaOV3420puRKACfU+Pa sZtx9R39ZlrrjUp8/zMlNhk= =+LbA -END PGP SIGNATURE- ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
[Haskell-cafe] FunPtr error?
Hello,
I am getting what is to me a mysterious error in a test case that I am
writing:
[EMAIL PROTECTED]:~/FTP/Haskell/unix-2.2.0.0/tests/timer$ runhaskell
Setup.lhs build
Preprocessing executables for Test-1.0...
Building Test-1.0...
[1 of 1] Compiling Main ( ./timer.hs,
dist/build/timer/timer-tmp/Main.o )
./timer.hs:11:45: Not in scope: data constructor `FunPtr'
It seems like the compiler is complaining about the lack of FunPtr in it's
symbol table but System.Posix is imported:
module Main where
import System.Posix
import Foreign
import Foreign.C
import Foreign.Ptr
main = do
let event = Sigevent{sigevFunction=(FunPtr
(notifyFunc))} error here
timerId - timerCreate Clock_Realtime Nothing
timerDelete timerId
return ()
notifyFunc :: Sigval - IO ()
notifyFunc sigval = do
putStrLn timer POP!!!
return ()
I am probably looking right at the answer and not seeing it. ??
Thanks, Vasili
___
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] FunPtr error?
2008/6/8 Galchin, Vasili [EMAIL PROTECTED]: Hello, I am getting what is to me a mysterious error in a test case that I am writing: [EMAIL PROTECTED]:~/FTP/Haskell/unix-2.2.0.0/tests/timer$ runhaskell Setup.lhs build Preprocessing executables for Test-1.0... Building Test-1.0... [1 of 1] Compiling Main ( ./timer.hs, dist/build/timer/timer-tmp/Main.o ) ./timer.hs:11:45: Not in scope: data constructor `FunPtr' There is a *type* called FunPtr in scope, but not a data constructor as you are using it. That is, you could say: foo :: FunPtr (Int - IO ()) That is, use the type called FunPtr, but you may not use a *function* called FunPtr, because it doesn't exist. You need to use functions like nullFunPtr, castPtrToFunPtr, etc. to construct FunPtrs. Luke ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] FunPtr error?
ah ..,. right ,. my bad. Vasili On Sun, Jun 8, 2008 at 10:01 PM, Luke Palmer [EMAIL PROTECTED] wrote: 2008/6/8 Galchin, Vasili [EMAIL PROTECTED]: Hello, I am getting what is to me a mysterious error in a test case that I am writing: [EMAIL PROTECTED]:~/FTP/Haskell/unix-2.2.0.0/tests/timer$ runhaskell Setup.lhs build Preprocessing executables for Test-1.0... Building Test-1.0... [1 of 1] Compiling Main ( ./timer.hs, dist/build/timer/timer-tmp/Main.o ) ./timer.hs:11:45: Not in scope: data constructor `FunPtr' There is a *type* called FunPtr in scope, but not a data constructor as you are using it. That is, you could say: foo :: FunPtr (Int - IO ()) That is, use the type called FunPtr, but you may not use a *function* called FunPtr, because it doesn't exist. You need to use functions like nullFunPtr, castPtrToFunPtr, etc. to construct FunPtrs. Luke ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] Package updates on haskell.org
On Thu, Jun 5, 2008 at 6:47 PM, Don Stewart [EMAIL PROTECTED] wrote: The HWN, which I'm sadly too busy to maintain now, Does this imply that you're looking for someone to take over the HWN? I'd be willing. -Brent ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Fwd: [Haskell-cafe] FunPtr error?
So Luke .. how do I go from (Sigval - ()), i..e notifyFunc, to FunPtr using the suggested data constructors? On Sun, Jun 8, 2008 at 10:01 PM, Luke Palmer [EMAIL PROTECTED] wrote: 2008/6/8 Galchin, Vasili [EMAIL PROTECTED]: Hello, I am getting what is to me a mysterious error in a test case that I am writing: [EMAIL PROTECTED]:~/FTP/Haskell/unix-2.2.0.0/tests/timer$ runhaskell Setup.lhs build Preprocessing executables for Test-1.0... Building Test-1.0... [1 of 1] Compiling Main ( ./timer.hs, dist/build/timer/timer-tmp/Main.o ) ./timer.hs:11:45: Not in scope: data constructor `FunPtr' There is a *type* called FunPtr in scope, but not a data constructor as you are using it. That is, you could say: foo :: FunPtr (Int - IO ()) That is, use the type called FunPtr, but you may not use a *function* called FunPtr, because it doesn't exist. You need to use functions like nullFunPtr, castPtrToFunPtr, etc. to construct FunPtrs. Luke ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] FunPtr error?
Hello Vasili,
Monday, June 9, 2008, 6:17:14 AM, you wrote:
1. standard place to import FunPtr from is Foreign.Ptr, not System.Posix
2. FunPtr is exported as abstract type, without constructors. you
can't construct values of this type directly. instead you should use
wrapper generators as in the example that Clause has wrote. read it
carefully :)
Hello,
I am getting what is to me a mysterious error in a test case that I am
writing:
[EMAIL PROTECTED]:~/FTP/Haskell/unix-2.2.0.0/tests/timer$ runhaskell
Setup.lhs build
Preprocessing executables for Test-1.0...
Building Test-1.0...
[1 of 1] Compiling Main ( ./timer.hs,
dist/build/timer/timer-tmp/Main.o )
./timer.hs:11:45: Not in scope: data constructor `FunPtr'
It seems like the compiler is complaining about the lack of FunPtr
in it's symbol table but System.Posix is imported:
module Main where
import System.Posix
import Foreign
import Foreign.C
import Foreign.Ptr
main = do
let event = Sigevent{sigevFunction=(FunPtr (notifyFunc))}
error here
timerId - timerCreate Clock_Realtime Nothing
timerDelete timerId
return ()
notifyFunc :: Sigval - IO ()
notifyFunc sigval = do
putStrLn timer POP!!!
return ()
I am probably looking right at the answer and not seeing it. ??
Thanks, Vasili
--
Best regards,
Bulatmailto:[EMAIL PROTECTED]
___
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Re[2]: [Haskell-cafe] Package updates on haskell.org
Hello Brent, Monday, June 9, 2008, 7:43:58 AM, you wrote: The HWN, which I'm sadly too busy to maintain now, Does this imply that you're looking for someone to take over the HWN? I'd be willing. it will be cool! -- Best regards, Bulatmailto:[EMAIL PROTECTED] ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
