The ragged SPW bonanza

cubical/data_handler/ms_data_handler.py

@@ -1,3 +1,4 @@
+# -*- coding: utf-8 -*-
 # CubiCal: a radio interferometric calibration suite
 # (c) 2017 Rhodes University & Jonathan S. Kenyon
 # http://github.com/ratt-ru/CubiCal
@@ -317,7 +318,8 @@ def __init__(self, ms_name, data_column, output_column=None, output_model_column
 
             assert set(spwtabcols) <= set(
                 _spwtab.colnames()), "Measurement set conformance error - keyword table SPECTRAL_WINDOW incomplete. Perhaps disable --out-casa-gaintables or check your MS!"
-            self._spwtabcols = {t: _spwtab.getcol(t) for t in spwtabcols}
+            nrows = _spwtab.nrows()
+            self._spwtabcols = {t: [_spwtab.getcol(t, row, 1) for row in range(nrows)] for t in spwtabcols}
 
             # read observation details
             obstabcols = ["TIME_RANGE", "LOG", "SCHEDULE", "FLAG_ROW",
@@ -470,18 +472,19 @@ def __init__(self, ms_name, data_column, output_column=None, output_model_column
                             " ".join([str(ch) for ch in freqchunks + [nchan]])), file=log(0))
 
         # now accumulate list of all frequencies, and also see if selected DDIDs have a uniform rebinning and chunking map
-        all_freqs = set(self.chanfreqs[self._ddids[0]])
         self.do_freq_rebin = any([m is not None for m in list(self.rebin_chan_maps.values())])
         self._ddids_unequal = False
         ddid0_map = self.rebin_chan_maps[self._ddids[0]]
         for ddid in self._ddids[1:]:
+            if len(self.chanfreqs[0]) != len(self.chanfreqs[ddid]):
+                self._ddids_unequal = True
+                break
             map1 = self.rebin_chan_maps[ddid]
             if ddid0_map is None and map1 is None:
                 continue
             if (ddid0_map is None and map1 is not None) or (ddid0_map is not None and map1 is None) or \
                     len(ddid0_map) != len(map1) or (ddid0_map!=map1).any():
                 self._ddids_unequal = True
-            all_freqs.update(self.chanfreqs[ddid])
 
         if self._ddids_unequal:
             print("Selected DDIDs have differing channel structure. Processing may be less efficient.", file=log(0,"red"))
@@ -811,6 +814,40 @@ def fetch(self, colname, first_row=0, nrows=-1, subset=None):
         (subset or self.data).getcolnp(str(colname), prealloc, first_row, nrows)
         return prealloc
 
+    @staticmethod
+    def _get_row_chunk(array):
+        """
+        Establishes max row chunk that can be used. Workaround for https://github.com/casacore/python-casacore/issues/130
+        array: array to be written, of shape (nrows, nfreq, ncorr)
+        """
+        _maxchunk = 2**29  # max number of elements to write, see https://github.com/casacore/python-casacore/issues/130#issuecomment-748150854
+        nrows, nfreq, ncorr = array.shape
+        maxrows = max(1, _maxchunk // (nfreq*ncorr))
+        #if maxrows < nrows:
+        log(1).print(f"  table I/O request of {nrows} rows: max chunk size is {maxrows} rows")
+        return maxrows, nrows
+
+    @staticmethod
+    def _getcolnp_wrapper(table, column, array, startrow):
+        "Calls table.getcolnp() in chunks of rows. Workaround for https://github.com/casacore/python-casacore/issues/130"
+        maxrows, nrows = MSDataHandler._get_row_chunk(array)
+        for row0 in range(0, nrows, maxrows):
+            table.getcolnp(column, array[row0:row0+maxrows], startrow+row0, maxrows)
+
+    @staticmethod
+    def _getcolslicenp_wrapper(table, column, array, begin, end, incr, startrow):
+        "Calls table.getcolnp() in chunks of rows. Workaround for https://github.com/casacore/python-casacore/issues/130"
+        maxrows, nrows = MSDataHandler._get_row_chunk(array)
+        for row0 in range(0, nrows, maxrows):
+            table.getcolslicenp(column, array[row0:row0+maxrows], begin, end, incr, startrow+row0, maxrows)
+
+    @staticmethod
+    def _putcol_wrapper(table, column, array, startrow):
+        "Calls table.putcol() in chunks of rows. Workaround for https://github.com/casacore/python-casacore/issues/130"
+        maxrows, nrows = MSDataHandler._get_row_chunk(array)
+        for row0 in range(0, nrows, maxrows):
+            table.putcol(column, array[row0:row0+maxrows], startrow+row0, maxrows)
+
     def fetchslice(self, column, startrow=0, nrows=-1, subset=None):
         """
         Convenience function similar to fetch(), but assumes a column of NFREQxNCORR shape,
@@ -839,16 +876,16 @@ def fetchslice(self, column, startrow=0, nrows=-1, subset=None):
             shape = tuple([nrows] + [s for s in cell.shape]) if hasattr(cell, "shape") else nrows
 
             prealloc = np.empty(shape, dtype=dtype)
-            subset.getcolnp(str(column), prealloc, startrow, nrows)
+            self._getcolnp_wrapper(subset, str(column), prealloc, startrow)
             return prealloc
         # ugly hack because getcell returns a different dtype to getcol
-        cell = (subset or self.data).getcol(str(column), startrow, nrow=1)[0, ...]
+        cell = subset.getcol(str(column), startrow, nrow=1)[0, ...]
         dtype = getattr(cell, "dtype", type(cell))
 
         shape = tuple([len(list(range(l, r + 1, i))) #inclusive in cc
                        for l, r, i in zip(self._ms_blc, self._ms_trc, self._ms_incr)])
         prealloc = np.empty(shape, dtype=dtype)
-        subset.getcolslicenp(str(column), prealloc, self._ms_blc, self._ms_trc, self._ms_incr, startrow, nrows)
+        self._getcolslicenp_wrapper(subset, str(column), prealloc, self._ms_blc, self._ms_trc, self._ms_incr, startrow)
         return prealloc
 
     def fetchslicenp(self, column, data, startrow=0, nrows=-1, subset=None):
@@ -869,8 +906,8 @@ def fetchslicenp(self, column, data, startrow=0, nrows=-1, subset=None):
         """
         subset = subset or self.data
         if self._ms_blc == None:
-            return subset.getcolnp(column, data, startrow, nrows)
-        return subset.getcolslicenp(column, data, self._ms_blc, self._ms_trc, self._ms_incr, startrow, nrows)
+            return self._getcolnp_wrapper(subset, column, data, startrow)
+        return self._getcolslicenp_wrapper(subset, column, data, self._ms_blc, self._ms_trc, self._ms_incr, startrow)
 
     def putslice(self, column, value, startrow=0, nrows=-1, subset=None):
         """
@@ -895,7 +932,7 @@ def putslice(self, column, value, startrow=0, nrows=-1, subset=None):
         # if no slicing, just use putcol to put the whole thing. This always works,
         # unless the MS is screwed up
         if self._ms_blc == None:
-            return subset.putcol(str(column), value, startrow, nrows)
+            return self._putcol_wrapper(subset, str(column), value, startrow)
         if nrows<0:
             nrows = subset.nrows()
 
@@ -919,19 +956,21 @@ def putslice(self, column, value, startrow=0, nrows=-1, subset=None):
             value[:] = np.bitwise_or.reduce(value, axis=2)[:,:,np.newaxis]
 
         if self._channel_slice == slice(None) and self._corr_slice == slice(None):
-            return subset.putcol(column, value, startrow, nrows)
+            return self._putcol_wrapper(subset, value, startrow)
         else:
             # for bitflags, we want to preserve flags we haven't touched -- read the column
             if column == "BITFLAG" or column == "FLAG":
-                value0 = subset.getcol(column)
+                value0 = np.empty_like(value)
+                self._getcolnp_wrapper(subset, column, value0, startrow)
+                # cheekily propagate per-corr flags to all corrs
                 value0[:] = np.bitwise_or.reduce(value0, axis=2)[:,:,np.newaxis]
             # otherwise, init empty column
             else:
                 ddid = subset.getcol("DATA_DESC_ID", 0, 1)[0]
                 shape = (nrows, self._nchan0_orig[ddid], self.nmscorrs)
                 value0 = np.zeros(shape, value.dtype)
             value0[:, self._channel_slice, self._corr_slice] = value
-            return subset.putcol(str(column), value0, startrow, nrows)
+            return self._putcol_wrapper(subset, str(column), value0, startrow, nrows)
 
     def define_chunk(self, chunk_time, rebin_time, fdim=1, chunk_by=None, chunk_by_jump=0, chunks_per_tile=4, max_chunks_per_tile=0):
         """