add an interface for dask.fft

.github/workflows/build-with-clang.yml

@@ -73,5 +73,5 @@ jobs:
       - name: Run mkl_fft tests
         run: |
           source ${{ env.ONEAPI_ROOT }}/setvars.sh
-          pip install scipy mkl-service pytest
+          pip install pytest mkl-service scipy dask
           pytest -s -v --pyargs mkl_fft

CHANGELOG.md

@@ -10,6 +10,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 * Added Hermitian FFT functions to SciPy interface `mkl_fft.interfaces.scipy_fft`: `hfft`, `ihfft`, `hfftn`, `ihfftn`, `hfft2`, and `ihfft2` [gh-161](https://github.com/IntelPython/mkl_fft/pull/161)
 * Added support for `out` kwarg to all FFT functions in `mkl_fft` and `mkl_fft.interfaces.numpy_fft` [gh-157](https://github.com/IntelPython/mkl_fft/pull/157)
 * Added `fftfreq`, `fftshift`, `ifftshift`, and `rfftfreq` to both NumPy and SciPy interfaces [gh-179](https://github.com/IntelPython/mkl_fft/pull/179)
+* Added a new interface for FFT module of Dask accessible through `mkl_fft.interfaces.dask_fft` [gh-184](https://github.com/IntelPython/mkl_fft/pull/184)
 
 ### Changed
 * NumPy interface `mkl_fft.interfaces.numpy_fft` is aligned with numpy-2.x.x [gh-139](https://github.com/IntelPython/mkl_fft/pull/139), [gh-157](https://github.com/IntelPython/mkl_fft/pull/157)

README.md

@@ -47,7 +47,7 @@ More details can be found in [SciPy 2017 conference proceedings](https://github.
 
 ---
 
-The `mkl_fft` package offers interfaces that act as drop-in replacements for equivalent functions in NumPy and SciPy. Learn more about these interfaces [here](https://github.com/IntelPython/mkl_fft/blob/master/mkl_fft/interfaces/README.md).
+The `mkl_fft` package offers interfaces that act as drop-in replacements for equivalent functions in NumPy, SciPy, and Dask. Learn more about these interfaces [here](https://github.com/IntelPython/mkl_fft/blob/master/mkl_fft/interfaces/README.md).
 
 While using these interfaces is the easiest way to leverage `mk_fft`, one can also use `mkl_fft` directly with the following FFT functions:
 

conda-recipe-cf/meta.yaml

@@ -27,6 +27,7 @@ requirements:
       - mkl-service
       - numpy
       - scipy >=1.10.1
+      - dask
 
 test:
     commands:
@@ -38,6 +39,7 @@ test:
       - mkl_fft.interfaces
       - mkl_fft.interfaces.numpy_fft
       - mkl_fft.interfaces.scipy_fft
+      - mkl_fft.interfaces.dask_fft
 
 about:
     home: http://github.com/IntelPython/mkl_fft

conda-recipe/meta.yaml

@@ -27,6 +27,7 @@ requirements:
       - mkl-service
       - {{ pin_compatible('numpy') }}
       - scipy >=1.10.1
+      - dask
 
 test:
     commands:
@@ -38,6 +39,7 @@ test:
       - mkl_fft.interfaces
       - mkl_fft.interfaces.numpy_fft
       - mkl_fft.interfaces.scipy_fft
+      - mkl_fft.interfaces.dask_fft
 
 about:
     home: http://github.com/IntelPython/mkl_fft

mkl_fft/interfaces/README.md

@@ -1,5 +1,5 @@
 # Interfaces
-The `mkl_fft` package provides interfaces that serve as drop-in replacements for equivalent functions in NumPy and SciPy.
+The `mkl_fft` package provides interfaces that serve as drop-in replacements for equivalent functions in NumPy, SciPy, and Dask.
 
 ---
 
@@ -125,3 +125,43 @@ with mkl_fft.set_workers(4):
     y = scipy.signal.fftconvolve(a, a)  # Note that Nthr:4
 # MKL_VERBOSE FFT(dcbo256x128,input_strides:{0,128,1},output_strides:{0,128,1},bScale:3.05176e-05,tLim:4,unaligned_output,desc:0x563aefe86180) 187.37us CNR:OFF Dyn:1 FastMM:1 TID:0  NThr:4
 ```
+
+---
+
+## Dask interface - `mkl_fft.interfaces.dask_fft`
+
+This interface is a drop-in replacement for the [`dask.fft`](https://dask.pydata.org/en/latest/array-api.html#fast-fourier-transforms) module and includes **all** the functions available there:
+
+* complex-to-complex FFTs: `fft`, `ifft`, `fft2`, `ifft2`, `fftn`, `ifftn`.
+
+* real-to-complex and complex-to-real FFTs: `rfft`, `irfft`, `rfft2`, `irfft2`, `rfftn`, `irfftn`.
+
+* Hermitian FFTs: `hfft`, `ihfft`.
+
+* Helper routines: `fft_wrap`, `fftfreq`, `rfftfreq`, `fftshift`, `ifftshift`. These routines serve as a fallback to the Dask implementation and are included for completeness.
+
+The following example shows how to use this interface for calculating a 2D FFT.
+
+```python
+import numpy, dask
+import mkl_fft.interfaces.dask_fft as dask_fft
+
+a = numpy.random.randn(128, 64) + 1j*numpy.random.randn(128, 64)
+x = dask.array.from_array(a, chunks=(64, 64))
+lazy_res = dask_fft.fft(x)
+mkl_res = lazy_res.compute()
+np_res = numpy.fft.fft(a)
+numpy.allclose(mkl_res, np_res)
+# True
+
+# There are two chunks in this example based on the size of input array (128, 64) and chunk size (64, 64)
+# to confirm that MKL FFT is called twice, turn on verbosity
+import mkl
+mkl.verbose(1)
+# True
+
+mkl_res = lazy_res.compute()  # MKL_VERBOSE FFT is shown twice below which means MKL FFT is called twice
+# MKL_VERBOSE oneMKL 2024.0 Update 2 Patch 2 Product build 20240823 for Intel(R) 64 architecture Intel(R) Advanced Vector Extensions 512 (Intel(R) AVX-512) with support for INT8, BF16, FP16 (limited) instructions, and Intel(R) Advanced Matrix Extensions (Intel(R) AMX) with INT8 and BF16, Lnx 3.80GHz intel_thread
+# MKL_VERBOSE FFT(dcfo64*64,input_strides:{0,1},output_strides:{0,1},input_distance:64,output_distance:64,bScale:0.015625,tLim:32,unaligned_input,desc:0x7fd000010e40) 432.84us CNR:OFF Dyn:1 FastMM:1 TID:0  NThr:112
+# MKL_VERBOSE FFT(dcfo64*64,input_strides:{0,1},output_strides:{0,1},input_distance:64,output_distance:64,bScale:0.015625,tLim:32,unaligned_input,desc:0x7fd480011300) 499.00us CNR:OFF Dyn:1 FastMM:1 TID:0  NThr:112
+```

mkl_fft/interfaces/__init__.py

@@ -24,3 +24,5 @@
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 from . import numpy_fft, scipy_fft
+
+from . import dask_fft  # isort: skip

mkl_fft/interfaces/dask_fft.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python
+# Copyright (c) 2025, Intel Corporation
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+#     * Redistributions of source code must retain the above copyright notice,
+#       this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above copyright
+#       notice, this list of conditions and the following disclaimer in the
+#       documentation and/or other materials provided with the distribution.
+#     * Neither the name of Intel Corporation nor the names of its contributors
+#       may be used to endorse or promote products derived from this software
+#       without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from dask.array.fft import fft_wrap, fftfreq, fftshift, ifftshift, rfftfreq
+
+from . import numpy_fft as _numpy_fft
+
+__all__ = [
+    "fft",
+    "ifft",
+    "fft2",
+    "ifft2",
+    "fftn",
+    "ifftn",
+    "rfft",
+    "irfft",
+    "rfft2",
+    "irfft2",
+    "rfftn",
+    "irfftn",
+    "hfft",
+    "ihfft",
+    "fftshift",
+    "ifftshift",
+    "fftfreq",
+    "rfftfreq",
+    "fft_wrap",
+]
+
+
+fft = fft_wrap(_numpy_fft.fft)
+ifft = fft_wrap(_numpy_fft.ifft)
+fft2 = fft_wrap(_numpy_fft.fft2)
+ifft2 = fft_wrap(_numpy_fft.ifft2)
+fftn = fft_wrap(_numpy_fft.fftn)
+ifftn = fft_wrap(_numpy_fft.ifftn)
+rfft = fft_wrap(_numpy_fft.rfft)
+irfft = fft_wrap(_numpy_fft.irfft)
+rfft2 = fft_wrap(_numpy_fft.rfft2)
+irfft2 = fft_wrap(_numpy_fft.irfft2)
+rfftn = fft_wrap(_numpy_fft.rfftn)
+irfftn = fft_wrap(_numpy_fft.irfftn)
+hfft = fft_wrap(_numpy_fft.hfft)
+ihfft = fft_wrap(_numpy_fft.ihfft)

mkl_fft/tests/test_interfaces.py

@@ -167,10 +167,16 @@ def test_axes(func):
 
 
 @pytest.mark.parametrize(
-    "interface", [mfi.scipy_fft, mfi.numpy_fft], ids=["scipy", "numpy"]
+    "interface",
+    [mfi.scipy_fft, mfi.numpy_fft, mfi.dask_fft],
+    ids=["scipy", "numpy", "dask"],
 )
 @pytest.mark.parametrize(
     "func", ["fftshift", "ifftshift", "fftfreq", "rfftfreq"]
 )
 def test_interface_helper_functions(interface, func):
     assert hasattr(interface, func)
+
+
+def test_dask_fftwrap():
+    assert hasattr(mfi.dask_fft, "fft_wrap")

mkl_fft/tests/third_party/dask/test_dask_fft.py

@@ -0,0 +1,161 @@
+# This file includes tests from dask.fft module:
+# https://github.com/dask/dask/blob/main/dask/array/tests/test_fft.py
+
+import contextlib
+from itertools import combinations_with_replacement
+
+import dask
+import dask.array as da
+import numpy as np
+import pytest
+from dask.array.numpy_compat import NUMPY_GE_200
+from dask.array.utils import assert_eq, same_keys
+
+import mkl_fft.interfaces.dask_fft as dask_fft
+
+requires_dask_2024_8_2 = pytest.mark.skipif(
+    dask.__version__ < "2024.8.2",
+    reason="norm kwarg requires Dask >= 2024.8.2",
+)
+
+all_1d_funcnames = ["fft", "ifft", "rfft", "irfft", "hfft", "ihfft"]
+
+all_nd_funcnames = [
+    "fft2",
+    "ifft2",
+    "fftn",
+    "ifftn",
+    "rfft2",
+    "irfft2",
+    "rfftn",
+    "irfftn",
+]
+
+if not da._array_expr_enabled():
+
+    nparr = np.arange(100).reshape(10, 10)
+    darr = da.from_array(nparr, chunks=(1, 10))
+    darr2 = da.from_array(nparr, chunks=(10, 1))
+    darr3 = da.from_array(nparr, chunks=(10, 10))
+
+
+@pytest.mark.parametrize("funcname", all_1d_funcnames)
+def test_cant_fft_chunked_axis(funcname):
+    da_fft = getattr(dask_fft, funcname)
+
+    bad_darr = da.from_array(nparr, chunks=(5, 5))
+    for i in range(bad_darr.ndim):
+        with pytest.raises(ValueError):
+            da_fft(bad_darr, axis=i)
+
+
+@pytest.mark.parametrize("funcname", all_1d_funcnames)
+def test_fft(funcname):
+    da_fft = getattr(dask_fft, funcname)
+    np_fft = getattr(np.fft, funcname)
+
+    # pylint: disable=possibly-used-before-assignment
+    assert_eq(da_fft(darr), np_fft(nparr))
+
+
+@pytest.mark.parametrize("funcname", all_nd_funcnames)
+def test_fft2n_shapes(funcname):
+    da_fft = getattr(dask_fft, funcname)
+    np_fft = getattr(np.fft, funcname)
+
+    # pylint: disable=possibly-used-before-assignment
+    assert_eq(da_fft(darr3), np_fft(nparr))
+    assert_eq(
+        da_fft(darr3, (8, 9), axes=(1, 0)), np_fft(nparr, (8, 9), axes=(1, 0))
+    )
+    assert_eq(
+        da_fft(darr3, (12, 11), axes=(1, 0)),
+        np_fft(nparr, (12, 11), axes=(1, 0)),
+    )
+
+    if NUMPY_GE_200 and funcname.endswith("fftn"):
+        ctx = pytest.warns(
+            DeprecationWarning,
+            match="`axes` should not be `None` if `s` is not `None`",
+        )
+    else:
+        ctx = contextlib.nullcontext()
+    with ctx:
+        expect = np_fft(nparr, (8, 9))
+    with ctx:
+        actual = da_fft(darr3, (8, 9))
+    assert_eq(expect, actual)
+
+
+@requires_dask_2024_8_2
+@pytest.mark.parametrize("funcname", all_1d_funcnames)
+def test_fft_n_kwarg(funcname):
+    da_fft = getattr(dask_fft, funcname)
+    np_fft = getattr(np.fft, funcname)
+
+    assert_eq(da_fft(darr, 5), np_fft(nparr, 5))
+    assert_eq(da_fft(darr, 13), np_fft(nparr, 13))
+    assert_eq(
+        da_fft(darr, 13, norm="backward"), np_fft(nparr, 13, norm="backward")
+    )
+    assert_eq(da_fft(darr, 13, norm="ortho"), np_fft(nparr, 13, norm="ortho"))
+    assert_eq(
+        da_fft(darr, 13, norm="forward"), np_fft(nparr, 13, norm="forward")
+    )
+    # pylint: disable=possibly-used-before-assignment
+    assert_eq(da_fft(darr2, axis=0), np_fft(nparr, axis=0))
+    assert_eq(da_fft(darr2, 5, axis=0), np_fft(nparr, 5, axis=0))
+    assert_eq(
+        da_fft(darr2, 13, axis=0, norm="backward"),
+        np_fft(nparr, 13, axis=0, norm="backward"),
+    )
+    assert_eq(
+        da_fft(darr2, 12, axis=0, norm="ortho"),
+        np_fft(nparr, 12, axis=0, norm="ortho"),
+    )
+    assert_eq(
+        da_fft(darr2, 12, axis=0, norm="forward"),
+        np_fft(nparr, 12, axis=0, norm="forward"),
+    )
+
+
+@pytest.mark.parametrize("funcname", all_1d_funcnames)
+def test_fft_consistent_names(funcname):
+    da_fft = getattr(dask_fft, funcname)
+
+    assert same_keys(da_fft(darr, 5), da_fft(darr, 5))
+    assert same_keys(da_fft(darr2, 5, axis=0), da_fft(darr2, 5, axis=0))
+    assert not same_keys(da_fft(darr, 5), da_fft(darr, 13))
+
+
+@pytest.mark.parametrize("funcname", all_nd_funcnames)
+@pytest.mark.parametrize("dtype", ["float32", "float64"])
+def test_nd_ffts_axes(funcname, dtype):
+    np_fft = getattr(np.fft, funcname)
+    da_fft = getattr(dask_fft, funcname)
+
+    shape = (7, 8, 9)
+    chunk_size = (3, 3, 3)
+    a = np.arange(np.prod(shape), dtype=dtype).reshape(shape)
+    d = da.from_array(a, chunks=chunk_size)
+
+    for num_axes in range(1, d.ndim):
+        for axes in combinations_with_replacement(range(d.ndim), num_axes):
+            cs = list(chunk_size)
+            for i in axes:
+                cs[i] = shape[i]
+            d2 = d.rechunk(cs)
+            if len(set(axes)) < len(axes):
+                with pytest.raises(ValueError):
+                    da_fft(d2, axes=axes)
+            else:
+                r = da_fft(d2, axes=axes)
+                er = np_fft(a, axes=axes)
+                if np.lib.NumpyVersion(np.__version__) >= "2.0.0":
+                    check_dtype = True
+                    assert r.dtype == er.dtype
+                else:
+                    check_dtype = False
+                assert r.shape == er.shape
+
+                assert_eq(r, er, check_dtype=check_dtype, rtol=1e-6, atol=1e-4)

pyproject.toml

@@ -49,7 +49,7 @@ classifiers = [
   "Operating System :: POSIX",
   "Operating System :: Unix"
 ]
-dependencies = ["numpy>=1.26.4", "mkl-service", "scipy>=1.10.1"]
+dependencies = ["numpy>=1.26.4", "mkl-service", "scipy>=1.10.1", "dask"]
 description = "MKL-based FFT transforms for NumPy arrays"
 dynamic = ["version"]
 keywords = ["DFTI", "FFT", "Fourier", "MKL"]