Optimize array operations to reduce copies and use vectorization

Co-authored-by: alexlib <747110+alexlib@users.noreply.github.com>

Author: Copilot

openpiv/filters.py

@@ -174,11 +174,12 @@ def replace_outliers(
     # regardless the grid_mask (which is a user-provided masked region)
 
 
-    if not isinstance(u, np.ma.MaskedArray):
+    # Only create masked array if needed
+    if isinstance(u, np.ma.MaskedArray):
+        grid_mask = u.mask.copy()
+    else:
         u = np.ma.masked_array(u, mask=np.ma.nomask)
-        
-    # store grid_mask for reinforcement
-    grid_mask = u.mask.copy()
+        grid_mask = np.ma.nomask
 
     u[flags] = np.nan
     v[flags] = np.nan

openpiv/pyprocess.py

@@ -335,9 +335,11 @@ def find_all_first_peaks(corr):
     ind = corr.reshape(corr.shape[0], -1).argmax(-1)
     peaks = np.array(np.unravel_index(ind, corr.shape[-2:]))
     peaks = np.vstack((peaks[0], peaks[1])).T
-    index_list = [(i, v[0], v[1]) for i, v in enumerate(peaks)]
+    # Vectorized index list creation instead of list comprehension
+    n = peaks.shape[0]
+    index_list = np.column_stack((np.arange(n), peaks))
     peaks_max = np.nanmax(corr, axis = (-2, -1))
-    return np.array(index_list), np.array(peaks_max)
+    return index_list, peaks_max
 
 
 def find_all_second_peaks(corr, width = 2):
@@ -766,7 +768,12 @@ def normalize_intensity(window):
         intensity normalized to -1 +1 and clipped if some pixels are
         extra low/high
     """
-    window = window.astype(np.float32)
+    # Convert to float32 only if needed, otherwise work in-place
+    if window.dtype != np.float32:
+        window = window.astype(np.float32)
+    else:
+        window = window.copy()  # Still need a copy to avoid modifying input
+    
     window -= window.mean(axis=(-2, -1),
                           keepdims=True, dtype=np.float32)
     tmp = window.std(axis=(-2, -1), keepdims=True)

openpiv/validation.py

@@ -110,10 +110,13 @@ def global_std(
     # def reject_outliers(data, m=2):
     #     return data[abs(data - np.mean(data)) < m * np.std(data)]
 
-    # create nan filled arrays where masks
-    # if u,v, are non-masked, ma.copy() adds false masks
-    tmpu = np.ma.copy(u).filled(np.nan)
-    tmpv = np.ma.copy(v).filled(np.nan)
+    # Avoid unnecessary copy operations - work with masked arrays directly
+    if np.ma.is_masked(u):
+        tmpu = np.where(u.mask, np.nan, u.data)
+        tmpv = np.where(v.mask, np.nan, v.data)
+    else:
+        tmpu = u
+        tmpv = v
 
     ind = np.logical_or(np.abs(tmpu - np.nanmean(tmpu)) > std_threshold * np.nanstd(tmpu),
                         np.abs(tmpv - np.nanmean(tmpv)) > std_threshold * np.nanstd(tmpv))