FFT of 2D arrays - Githubissues

Nodd commented 11 years ago

While investigating for #8, I tried to compute the FFT on a 2D array, but I get:

Traceback (most recent call last):
  File "/home/joseph/prog/libs.git/python_JO/EFUSION/test_reinka.py", line 27, in <module>
    main()
  File "/home/joseph/prog/libs.git/python_JO/EFUSION/test_reinka.py", line 18, in main
    fft = FFT(thr).prepare_for(cl_data_out, cl_data_in, -1, axes=(0,))
  File "/usr/lib/python2.7/site-packages/reikna/core/computation.py", line 214, in prepare_for
    self._operations.finalize()
  File "/usr/lib/python2.7/site-packages/reikna/core/operation.py", line 212, in finalize
    value.shape, value.dtype, dependencies=dependencies)
  File "/usr/lib/python2.7/site-packages/reikna/cluda/api.py", line 267, in temp_array
    return self.temp_alloc.array(shape, dtype, dependencies=dependencies)
  File "/usr/lib/python2.7/site-packages/reikna/cluda/tempalloc.py", line 74, in array
    self.update_buffer(new_id)
  File "/usr/lib/python2.7/site-packages/reikna/cluda/tempalloc.py", line 82, in update_buffer
    array.data = buf
AttributeError: can't set attribute

My code is:

import numpy as np
from reikna import cluda
from reikna.fft import FFT

def main():
    api = cluda.ocl_api()
    thr = api.Thread.create()

    N = 256
    M = 10000

    data_in = np.random.rand(N, N) + 1j*np.random.rand(N, N)
    cl_data_in = thr.to_device(data_in)
    cl_data_out = thr.empty_like(cl_data_in)
    fft = FFT(thr).prepare_for(cl_data_out, cl_data_in, -1, axes=(0,))

if __name__ == "__main__":
    main()

It passes if I use: data_in = np.random.rand(N) + 1j*np.random.rand(N)

As a side note, I have an (other) exception if data_in is not complex. Is it a bug or a feature ?

fjarri commented 11 years ago

FFT for a large enough problem size (256x256 in your case) needs several kernel calls, and has to allocate some temporary arrays (that's why 1D works — it needs only one kernel, no temporary arrays required). Reikna tries to pack these arrays into the same location of physical memory, if their dependencies allow it. On PyOpenCL it means reassigning the array's pointer to the physical buffer dynamically — that's what array.data = buf was for. In PyOpenCL 2013.1 this attribute became read-only, so I fixed this part in the current development version (0.3.0).

Now for the hard part. You have two choices:

1) Fix this in your local version. In reikna/cluda/tempalloc.py change

array.data = buf

to

array.base_data = buf

or 2) Switch to the development version. May not be the best idea, as its API is not completely stable, and you'll have to rewrite some of your code.

As a side note, I have an (other) exception if data_in is not complex. Is it a bug or a feature ?

Rather, the lack of feature. FFT itself only does complex-to-complex transformation; if you need only real numbers on either side, you have to attach transformations:

import numpy as np
from reikna import cluda, Transformation
from reikna.cluda import dtypes
from reikna.fft import FFT
from reikna.cluda.tempalloc import TrivialManager

tr = Transformation(
    inputs=['in_re'], outputs=['out_c'],
    derive_o_from_is=lambda in_re: dtypes.complex_for(in_re),
    snippet="${out_c.store}(COMPLEX_CTR(${out_c.ctype})(${in_re.load}, 0));")

def main():
    api = cluda.ocl_api()
    thr = api.Thread.create(temp_alloc=dict(cls=TrivialManager))

    N = 256
    M = 10000

    data_in = np.random.rand(N)
    data_in = data_in.astype(np.float32)

    cl_data_in = thr.to_device(data_in)

    cl_data_out = thr.array(data_in.shape, np.complex64)

    fft = FFT(thr)
    fft.connect(tr, 'input', ['input_re'])
    fft.prepare_for(cl_data_out, cl_data_in, -1, axes=(0,))

if __name__ == "__main__":
    main()

fjarri commented 11 years ago

Closing this one too, 0.3.0 is released.

fjarri / reikna

FFT of 2D arrays #9