Turn on the apodized FFT interpolation

The original Goddard code effectively had this turned off,
so in order to replicate those results this was turned off here
too even though was part of the Goddard code.
This was reported to Goddard as a potential bug, and was confirmed to
be a bug in the reference implementation, so it should be turned
on by default.

CHANGES.rst

@@ -90,6 +90,8 @@ refpix
 
 - Update cal_step, add suffix and add to the exposure pipeline [#890]
 
+- Enable apodized FFT interpolation by default. [#1017]
+
 resample
 --------
 

romancal/refpix/refpix.py

@@ -18,7 +18,7 @@ def run_steps(
     remove_offset: bool = True,
     remove_trends: bool = True,
     cosine_interpolate: bool = True,
-    fft_interpolate: bool = False,
+    fft_interpolate: bool = True,
 ) -> RampModel:
     """
     Organize the steps to run the reference pixel correction.

romancal/refpix/refpix_step.py

@@ -24,7 +24,7 @@ class RefPixStep(RomanStep):
     # linear trends
     cosine_interpolate = boolean(default=True) # Turn on or off the cosine
     # interpolation of the reference pixels
-    fft_interpolate = boolean(default=False) # Turn on or off the FFT interpolation
+    fft_interpolate = boolean(default=True) # Turn on or off the FFT interpolation
     # of the reference pixel padded values.
     """
 

romancal/refpix/tests/reference_utils.py

@@ -255,7 +255,7 @@ def interp_zeros_channel_fun(
 
 def cos_interp_reference(dataUniformTime, numFrames):
     numRows = dataUniformTime.shape[2]
-    return interp_zeros_channel_fun(
+    interp_zeros_channel_fun(
         REFERENCE_CHAN,
         getTrigInterpolationFunction(dataUniformTime),
         dataUniformTime,
@@ -506,8 +506,10 @@ def apply_correction(data, alpha, gamma, zeta):
     # Perform cosine weighted interpolation
     cos_interp_reference(dataUniformTime, dataUniformTime.shape[1])
 
-    # Perform fft interpolation (does nothing right now)
-    fft_interp_amp33(dataUniformTime, dataUniformTime.shape[1])
+    # Perform fft interpolation
+    dataUniformTime[REFERENCE_CHAN, :, :, :] = fft_interp_amp33(
+        dataUniformTime, dataUniformTime.shape[1]
+    )
 
     # Perform the correction
     data0 = apply_correction_to_data(

romancal/refpix/tests/test_data.py

@@ -535,25 +535,6 @@ def test_fft_interpolate(self, channels):
     def test_fft_interpolate_regression(self, channels):
         """
         Run fft interpolation regression test
-
-        NOTE:
-            The reference code assumes the data will be changed in-place for all
-            its major operations.However, the reference code violates this assumption
-            for the Amp33 FFT interpolation step. It does make an in-place change to
-            a sub-array, `dataReferenceChan_FramesFlat`. However
-            `dataReferenceChan_FramesFlat` loses its view on the original data array
-            because it unnecessarily recasts the dtype, which is a non-view compatible
-            change.
-
-            romancal does not make this mistake because it does not recast the dtype.
-
-            The reference utils are modified to output the data array which is modified
-            as a functional return value. This is so that the reference code's
-            computation can be tests against the romancal code's computation.
-
-            After this step is computed, the reference code an romancal code will
-            diverge in values, because the changes made by this will propagate in
-            romancal but the do not in the reference code.
         """
         non_view_data = channels.data.copy()
 

romancal/refpix/tests/test_refpix.py

@@ -12,7 +12,7 @@ def test_run_steps_regression(datamodel, ref_pix_ref):
         regression, ref_pix_ref.alpha, ref_pix_ref.gamma, ref_pix_ref.zeta
     )
 
-    result = run_steps(datamodel, ref_pix_ref, True, True, True, False)
+    result = run_steps(datamodel, ref_pix_ref, True, True, True, True)
 
     assert (result.data.value == regression_out).all()
     # regression_out does not return amp33 data