arrayfire / arrayfire-haskell Goto Github PK

How to generate an array according to the rules defined by oneself?
I have a function like func:: Int->Int->Float, which represents calculating values based on index

I have installed arrayfire 3.6.4 from https://arrayfire.com/download/ to /opt/arrayfire as described on http://arrayfire.org/docs/installing.htm#Linux.

Running cabal install arrayfire yields the following error message:

Build profile: -w ghc-8.8.1 -O1
In order, the following will be built (use -v for more details):
 - arrayfire-0.1.0.0 (lib:arrayfire, exe:gen, exe:main) (requires build)
Starting     arrayfire-0.1.0.0 (all, legacy fallback)

Failed to build arrayfire-0.1.0.0. The failure occurred during the configure
step.
Build log (
/home/jaro/.cabal/logs/ghc-8.8.1/arrayfire-0.1.0.0-5e7527530ca25ab12a051fd4d0f7055fae7a4dba9f99268656d18fb873a6c8ce.log
):
[1 of 1] Compiling Main             ( /tmp/cabal-install.-22655/dist-newstyle/tmp/src-22655/arrayfire-0.1.0.0/dist/setup/setup.hs, /tmp/cabal-install.-22655/dist-newstyle/tmp/src-22655/arrayfire-0.1.0.0/dist/setup/Main.o )
Linking /tmp/cabal-install.-22655/dist-newstyle/tmp/src-22655/arrayfire-0.1.0.0/dist/setup/setup ...
Configuring arrayfire-0.1.0.0...
setup: Missing dependency on a foreign library:
* Missing (or bad) C library: af
This problem can usually be solved by installing the system package that
provides this library (you may need the "-dev" version). If the library is
already installed but in a non-standard location then you can use the flags
--extra-include-dirs= and --extra-lib-dirs= to specify where it is.If the
library file does exist, it may contain errors that are caught by the C
compiler at the preprocessing stage. In this case you can re-run configure
with the verbosity flag -v3 to see the error messages.

cabal: Failed to build arrayfire-0.1.0.0. See the build log above for details.

Your documentation contains an example:

>>> drawImage window ('constant' \@'Int' 1) ('Cell' 10 10 "test" 'ColorMapSpectrum')

I adapted it to:

main = do
  w <- createWindow 512 512 "Test"
  drawImage w (constant @Int [512,512] 1) (Cell 10 10 "test" ColorMapSpectrum)

That gives the error:

AFException {afExceptionType = TypeError, afExceptionCode = 204, afExceptionMsg = "Function does not support this data type"}

@chessai this would be a /very/ nice to have. Issue is that doctest can't resolve arrayfire third-party dep. when it runs.

It looks like your Show instance for Array is transposed:

λ> A.lower (A.matrix (2,2) [[1,2],[3,4]]) False
ArrayFire Array
[2 2 1 1]
    1.0000     2.0000 
    0.0000     4.0000 

λ> A.upper (A.matrix (2,2) [[1,2],[3,4]]) False
ArrayFire Array
[2 2 1 1]
    1.0000     0.0000 
    3.0000     4.0000 

Another example: exceptions thrown by matmul suggest that the dimensions are (rows,columns):

λ> A.matmul (A.constant [2,3] 5) (A.constant [1,2] 2) A.None A.None
*** Exception: AFException {afExceptionType = SizeError, afExceptionCode = 203, afExceptionMsg = "Invalid input size"}
λ> A.matmul (A.constant [2,3] 5) (A.constant [3,1] 2) A.None A.None
ArrayFire Array
[2 1 1 1]
   30.0000    30.0000 

but the Show instance suggests the opposite:

λ> (A.constant [2,3] 5)
ArrayFire Array
[2 3 1 1]
    5.0000     5.0000 
    5.0000     5.0000 
    5.0000     5.0000 

Since all examples are now printed in column-major (as opposed to row major) order, the documentation needs to reflect this change.

• randn, randu, etc. should use a list for dimensions
• Tests for file IO, reading / writing matrices
• Add more tests
• Add links to array fire things
• AFConvMode, etc. needs to be wrapped
• Features API needs to be fixed
• AFConvMode, double check to/from is correct.
• Constrain Util functions by AFType
• Add a Bits instance?
• Doctest on Darwin doesn't work since it forgets about being linked against arrayfire (cannot find .so errors abound)

Would be nice to have hackage version, build status, etc.

This was mentioned in #22, but I'll make an Issue for it here. The data provided to the smart constructors matrix, cube, and tensor is validated (in terms of row and column sizes) in row-major order, but the constructors all flatten the input lists and use mkArray, which enters the data in the column-major order. This leads to, first, properly arranged rows and columns being invalid:

λ> A.matrix (3,2) [[1,2,3],[4,5,6]]
*** Exception: AFException {afExceptionType = SizeError, afExceptionCode = 203, afExceptionMsg = "Invalid elements provided. Expected 6 elements received 4"}

and, second, if the input lists are transposed so they validate OK, the data in the Array doesn't even follow the rows and columns provided:

λ> A.matrix (3,2) [[1,2],[3,4],[5,6]]
ArrayFire Array
[3 2 1 1]
    1.0000     4.0000 
    2.0000     5.0000 
    3.0000     6.0000 

Running cabal test fails because the linker can't find these functions:

In Image.hsc:

• af_iterative_deconv
• af_iverse_deconv

In LAPACK.hsc:

• af_pinverse

In Signal.hsc:

• af_approx1_uniform
• af_approx2_uniform

Should these functions be removed?

Currently the installation instructions for arrayfire-haskell target users who install binaries into /lib / lib64 on OSX/ Linux respectively. For "power users" (those who build from source), there should be instructions on how to install.

cabal configure -f-disable-default-paths --extra-lib-dirs=/path/to/arrayfire/lib --extra-include-dirs=/path/to/arrayfire/include

Thanks for your excellent work!
I'm very interesting in this library. But I want to know how to interacting with other library. For example , When I read an image with JuicyPixels, how to transfer this data to arrayfire.

I often use accelerate library, and it has many auxiliary libraries.

Thanks again.

Error: setup.exe: Missing dependency on a foreign library:

• Missing (or bad) C library: af
  This problem can usually be solved by installing the system package that
  provides this library (you may need the "-dev" version). If the library is
  already installed but in a non-standard location then you can use the flags
  --extra-include-dirs= and --extra-lib-dirs= to specify where it is.If the
  library file does exist, it may contain errors that are caught by the C
  compiler at the preprocessing stage. In this case you can re-run configure
  with the verbosity flag -v3 to see the error messages.

After cabal install.

Is this an issue with path configuration in Windows? Can I work around this with mingw? The problem with mingw, though, is that my CUDA drivers are on Windows.

Theoretically, there's no reason this can't be supported.

instance Bits (ArrayFire CBool) where
  (.&.) = A.and
  (.|.) = A.or
  xor = A.bitXor
  shiftL x n = A.shiftL x (scalar n)
  shiftR x n = A.shiftR x (scalar n)

Still need to look into implementing the following:

(rotate | rotateL, rotateR), bitSize, bitSizeMaybe, isSigned, testBit, bit, popCount

Would be very nice to have comparisons of ArrayFire vs. libraries like hmatrix, accelerate, etc. This might even warrant its own package due to the difficulty in procuring all the dependencies.

pending on NixOS/nixpkgs#72142

Currently, imag and real have type AFType a => Array a -> Array a. But imag and real convert complex numbers to real numbers, so the type should be (AFType a, AFType (Complex a)) => Array (Complex a) -> Array a.

I am not sure this is entirely helpful, feel free to close. I would like to use arrayfire for a project I am starting. It is included in my nix setup via https://github.com/chiroptical/homing-pigeon/blob/main/default.nix.

However, I am getting errors from these tests after setting my LD_LIBRARY_PATH appropriately. Here is the output from the test,

  test/ArrayFire/ArithSpec.hs:31:7: 
  1) ArrayFire.Arith, Arith tests, Should take cubed root
       expected: ArrayFire Array
                 [1 1 1 1]
                     3.0000 
                 
        but got: ArrayFire Array
                 [1 1 1 1]
                     3.0000 

I am wondering if it is comparing the pointers and not doing an element-wise comparison? AFAICT, I tried some C++ code and arrayA == arrayB seems to generate an element-wise comparison but I am unsure how equals is handled in Haskell-land.

The functions which take a dimension to reduce along should return a full matrix, with only that dimension flattened, rather than a single scalar. These are (at least) sum, sumNaN, product, productNaN, min, max, allTrue, anyTrue, and count. The correct array is being computed by the FFI, but the getScalar function is being applied and just extracts the first element, rather than returning the entire array. The current behaviour of (for instance) sum is

λ> A.mkArray @Int [2,3] [1..]
ArrayFire Array
[2 3 1 1]
         1          3          5 
         2          4          6 
λ> A.sum it 0
3
λ> A.sum it 1
9

as opposed to the expected output,

ArrayFire Array
[1 3 1 1]
         3          7         11 

and

ArrayFire Array
[2 1 1 1]
         9 
        12 

Polymorphic across any kind of an array. Combine all constant functions into one, use castPtr

For example lt, gt, le, ge, eq, neq should have type AFType a => Array a -> Array a -> Array CBool.

And functions like and, or should have type Array CBool -> Array CBool -> Array CBool.

And not should have type Array CBool -> Array CBool.

By the way why is Bool not an instance of AFType? Would it be possible to write an instance that uses CBool under the hood?

Please, sleep over the releases. No need to spam Hackage.

(edit: only three are in rapid succession, the first one is already two days ago...)