One thing I forgot to make clear in the announcement is that the language is still highly experimental, and some obvious things, such as complex numbers, are currently missing. So this first release should probably not be used for real applications. However, while I don't know how autotuning works, I don't see why you shouldn't be able to code it in Feldspar a few releases from now. I don't know if it will be "easy" though :) / Emil Warren Henning skrev:

I see that section 4.1 of the user guide - http://feldspar.sourceforge.net/documents/language/FeldsparLanguage.html#hto... - includes an example involving autocorrelation.

Does this mean I could use Feldspare to easily build my own Autotune program? I love T-Pain and Autotune the News!

Warren

On Tue, Nov 3, 2009 at 7:39 PM, Emil Axelsson wrote:

...

I'm happy to announce the first release of Feldspar, which is an embedded domain-specific language with associated code generator mainly targeting DSP algorithms. The language is developed in cooperation by Ericsson, Chalmers University and Eötvös Loránd University.

Feldspar stands for *F*unctional *E*mbedded *L*anguage for *DSP* and *PAR*allelism.

The language front-end is available on Hackage:

http://hackage.haskell.org/package/feldspar-language

The back-end C code generator will be uploaded and announced shortly. For more information, see:

http://feldspar.sourceforge.net/

/ Emil

_______________________________________________ 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

On Tue, 3 Nov 2009, Warren Henning wrote:

I see that section 4.1 of the user guide - http://feldspar.sourceforge.net/documents/language/FeldsparLanguage.html#hto... - includes an example involving autocorrelation.

Does this mean I could use Feldspare to easily build my own Autotune program? I love T-Pain and Autotune the News!

There are several packages on hackage for performing signal processing: like dsp and fftw, that can assist doing autocorrelation.

Henning Thielemann skrev:

On Wed, 4 Nov 2009, Emil Axelsson wrote:

...

I'm happy to announce the first release of Feldspar, which is an embedded domain-specific language with associated code generator mainly targeting DSP algorithms. The language is developed in cooperation by Ericsson, Chalmers University and Eötvös Loránd University.

Feldspar stands for *F*unctional *E*mbedded *L*anguage for *DSP* and *PAR*allelism.

The language front-end is available on Hackage:

http://hackage.haskell.org/package/feldspar-language

I'm trying to get realtime signal processing with Haskell for long. I make progress, but slowly. Has Ericsson ever thought about using Haskell itself for signal processing? (But I think they already have Erlang?)

No, using Haskell directly is not an option (at least with current compiler technology). Their performance requirements are very high, and the signal processors have quite limited memory, so putting a Haskell RTS on them wouldn't work. Yes, Erlang is used in some applications, but that's more for "control programs", not for numerical computations. I hope you will succeed in making real-time signal processing in Haskell work! / Emil

Henning Thielemann skrev:

On Fri, 6 Nov 2009, Emil Axelsson wrote:

...

...

I'm trying to get realtime signal processing with Haskell for long. I make progress, but slowly. Has Ericsson ever thought about using Haskell itself for signal processing? (But I think they already have Erlang?) No, using Haskell directly is not an option (at least with current compiler technology). Their performance requirements are very high, and the signal

Henning Thielemann skrev: processors have quite limited memory, so putting a Haskell RTS on them wouldn't work.

Yes, Erlang is used in some applications, but that's more for "control programs", not for numerical computations.

I hope you will succeed in making real-time signal processing in Haskell work!

I'm currently testing JHC to that end. It produces relatively small C programs without a precompiled run-time system.

Cool! I'd be very interested to see how that works out. / Emil

On Fri, Nov 6, 2009 at 6:28 AM, Emil Axelsson wrote:

...

I'm trying to get realtime signal processing with Haskell for long. I make progress, but slowly. Has Ericsson ever thought about using Haskell itself for signal processing? (But I think they already have Erlang?)

No, using Haskell directly is not an option (at least with current compiler technology). Their performance requirements are very high, and the signal processors have quite limited memory, so putting a Haskell RTS on them wouldn't work.

Atom may be another option. Though it is not intended for high performance DSP, we do use it for basic signal processing. Here is an IIR filter that is used is some fault detection logic on our application: -- | IIR filter implemented using direct form 2. iirFilter :: Name -> Float -> [(Float, Float)] -> E Float -> Atom (E Float) iirFilter name b0 coeffs x = do -- Create the filter taps. vs <- mapM (\ i -> float (name ++ show i) 0) [1 .. length coeffs] -- Cascade the filter taps together. mapM_ (\ (vA, vB) -> vA <== value vB) $ zip (tail vs) vs -- Calculate the input to the chain of taps. let w0 = sum ( x : [ (value v) * Const (-a) | (v, (a, _)) <- zip vs coeffs ]) bs = b0 : (snd $ unzip coeffs) ws = w0 : map value vs us = [ w * Const b | (w, b) <- zip ws bs ] head vs <== w0 -- Return the output. return $ sum us http://hackage.haskell.org/package/atom

On Mon, Nov 9, 2009 at 10:09 AM, Emil Axelsson wrote:

Nice!

One of our project members has been looking at Atom, not for numerical computations, but for real-time scheduling (which Feldspar should deal with eventually).

What kind of code (in terms of efficiency) does the above description compile to?

Here's and example: module Main (main) where import Language.Atom main :: IO () main = do compile "filter" defaults design return () design :: Atom () design = atom "filter" $ do input <- float' "input" output <- float' "output" x <- iirFilter "filter" 1 [(2,3), (4,5)] (value input) output <== x -- | IIR filter implemented using direct form 2. iirFilter :: Name -> Float -> [(Float, Float)] -> E Float -> Atom (E Float) iirFilter name b0 coeffs x = do -- Create the filter taps. vs <- mapM (\ i -> float (name ++ show i) 0) [1 .. length coeffs] -- Cascade the filter taps together. mapM_ (\ (vA, vB) -> vA <== value vB) $ zip (tail vs) vs -- Calculate the input to the chain of taps. let w0 = sum ( x : [ (value v) * Const (-a) | (v, (a, _)) <- zip vs coeffs ]) bs = b0 : (snd $ unzip coeffs) ws = w0 : map value vs us = [ w * Const b | (w, b) <- zip ws bs ] head vs <== w0 -- Return the output. return $ sum us Here's the generated C. Note the filter calculation is done entirely by function __r0: static unsigned long long __global_clock = 0; static const unsigned long __coverage_len = 1; static unsigned long __coverage[1] = {0}; static unsigned long __coverage_index = 0; static float __v1 = 0; /* filter.filter.filter2 */ static float __v0 = 0; /* filter.filter.filter1 */ /* filter.filter */ static void __r0(void) { unsigned char __0 = 1; float __1 = 0.0; float __2 = input; float __3 = __1 + __2; float __4 = __v0 /* filter.filter.filter1 */ ; float __5 = -2.0; float __6 = __4 * __5; float __7 = __3 + __6; float __8 = __v1 /* filter.filter.filter2 */ ; float __9 = -4.0; float __10 = __8 * __9; float __11 = __7 + __10; float __12 = 1.0; float __13 = __11 * __12; float __14 = __1 + __13; float __15 = 3.0; float __16 = __4 * __15; float __17 = __14 + __16; float __18 = 5.0; float __19 = __8 * __18; float __20 = __17 + __19; if (__0) { __coverage[0] = __coverage[0] | (1 << 0); } output = __20; __v0 /* filter.filter.filter1 */ = __11; __v1 /* filter.filter.filter2 */ = __4; } void filter(void) { { static unsigned char __scheduling_clock = 0; if (__scheduling_clock == 0) { __r0(); /* filter.filter */ __scheduling_clock = 0; } else { __scheduling_clock = __scheduling_clock - 1; } } __global_clock = __global_clock + 1; }