PERF: faster unstacking

closes #15503

Author: jreback

asv_bench/benchmarks/reshape.py

@@ -59,6 +59,27 @@ def time_reshape_unstack_simple(self):
         self.df.unstack(1)
 
 
+class reshape_unstack_large_single_dtype(object):
+    goal_time = 0.2
+
+    def setup(self):
+        m = 100
+        n = 1000
+
+        levels = np.arange(m)
+        index = pd.MultiIndex.from_product([levels]*2)
+        columns = np.arange(n)
+        values = np.arange(m*m*n).reshape(m*m, n)
+        self.df = pd.DataFrame(values, index, columns)
+        self.df2 = self.df.iloc[:-1]
+
+    def time_unstack_full_product(self):
+        self.df.unstack()
+
+    def time_unstack_with_mask(self):
+        self.df2.unstack()
+
+
 class unstack_sparse_keyspace(object):
     goal_time = 0.2
 

doc/source/whatsnew/v0.20.0.txt

@@ -512,7 +512,7 @@ Performance Improvements
 - Improved performance of ``groupby().cummin()`` and ``groupby().cummax()`` (:issue:`15048`, :issue:`15109`)
 - Improved performance and reduced memory when indexing with a ``MultiIndex`` (:issue:`15245`)
 - When reading buffer object in ``read_sas()`` method without specified format, filepath string is inferred rather than buffer object. (:issue:`14947`)
-
+- Improved performance when using ``.unstack()`` (:issue:`15503`)
 
 
 .. _whatsnew_0200.bug_fixes:

pandas/core/reshape.py

@@ -7,7 +7,9 @@
 
 import numpy as np
 
-from pandas.types.common import _ensure_platform_int, is_list_like
+from pandas.types.common import (_ensure_platform_int,
+                                 is_list_like, is_bool_dtype,
+                                 needs_i8_conversion)
 from pandas.types.cast import _maybe_promote
 from pandas.types.missing import notnull
 import pandas.types.concat as _concat
@@ -25,6 +27,7 @@
 
 import pandas.core.algorithms as algos
 import pandas.algos as _algos
+import pandas._reshape as _reshape
 
 from pandas.core.index import MultiIndex, _get_na_value
 
@@ -182,9 +185,21 @@ def get_new_values(self):
         stride = values.shape[1]
         result_width = width * stride
         result_shape = (length, result_width)
+        mask = self.mask
+        mask_all = mask.all()
+
+        # we can simply reshape if we don't have a mask
+        if mask_all and len(values):
+            new_values = (self.sorted_values
+                              .reshape(length, width, stride)
+                              .swapaxes(1, 2)
+                              .reshape(result_shape)
+                          )
+            new_mask = np.ones(result_shape, dtype=bool)
+            return new_values, new_mask
 
         # if our mask is all True, then we can use our existing dtype
-        if self.mask.all():
+        if mask_all:
             dtype = values.dtype
             new_values = np.empty(result_shape, dtype=dtype)
         else:
@@ -194,13 +209,36 @@ def get_new_values(self):
 
         new_mask = np.zeros(result_shape, dtype=bool)
 
-        # is there a simpler / faster way of doing this?
-        for i in range(values.shape[1]):
-            chunk = new_values[:, i * width:(i + 1) * width]
-            mask_chunk = new_mask[:, i * width:(i + 1) * width]
-
-            chunk.flat[self.mask] = self.sorted_values[:, i]
-            mask_chunk.flat[self.mask] = True
+        name = np.dtype(dtype).name
+        sorted_values = self.sorted_values
+
+        # we need to convert to a basic dtype
+        # and possibly coerce an input to our output dtype
+        # e.g. ints -> floats
+        if needs_i8_conversion(values):
+            sorted_values = sorted_values.view('i8')
+            new_values = new_values.view('i8')
+            name = 'int64'
+        elif is_bool_dtype(values):
+            sorted_values = sorted_values.astype('object')
+            new_values = new_values.astype('object')
+            name = 'object'
+        else:
+            sorted_values = sorted_values.astype(name, copy=False)
+
+        # fill in our values & mask
+        f = getattr(_reshape, "reshape_{}".format(name))
+        f(sorted_values,
+          mask.view('u1'),
+          stride,
+          length,
+          width,
+          new_values,
+          new_mask.view('u1'))
+
+        # reconstruct dtype if needed
+        if needs_i8_conversion(values):
+            new_values = new_values.view(values.dtype)
 
         return new_values, new_mask
 

pandas/src/reshape.pyx

@@ -0,0 +1,35 @@
+# cython: profile=False
+
+from numpy cimport *
+cimport numpy as np
+import numpy as np
+
+cimport cython
+
+import_array()
+
+cimport util
+
+from numpy cimport NPY_INT8 as NPY_int8
+from numpy cimport NPY_INT16 as NPY_int16
+from numpy cimport NPY_INT32 as NPY_int32
+from numpy cimport NPY_INT64 as NPY_int64
+from numpy cimport NPY_FLOAT16 as NPY_float16
+from numpy cimport NPY_FLOAT32 as NPY_float32
+from numpy cimport NPY_FLOAT64 as NPY_float64
+
+from numpy cimport (int8_t, int16_t, int32_t, int64_t, uint8_t, uint16_t,
+                    uint32_t, uint64_t, float16_t, float32_t, float64_t)
+
+int8 = np.dtype(np.int8)
+int16 = np.dtype(np.int16)
+int32 = np.dtype(np.int32)
+int64 = np.dtype(np.int64)
+float16 = np.dtype(np.float16)
+float32 = np.dtype(np.float32)
+float64 = np.dtype(np.float64)
+
+cdef double NaN = <double> np.NaN
+cdef double nan = NaN
+
+include "reshape_helper.pxi"

pandas/src/reshape_helper.pxi.in

@@ -0,0 +1,65 @@
+"""
+Template for each `dtype` helper function for take
+
+WARNING: DO NOT edit .pxi FILE directly, .pxi is generated from .pxi.in
+"""
+
+# ----------------------------------------------------------------------
+# reshape
+# ----------------------------------------------------------------------
+
+{{py:
+
+# name, c_type
+dtypes = [('uint8', 'uint8_t'),
+          ('uint16', 'uint16_t'),
+          ('uint32', 'uint32_t'),
+          ('uint64', 'uint64_t'),
+          ('int8', 'int8_t'),
+          ('int16', 'int16_t'),
+          ('int32', 'int32_t'),
+          ('int64', 'int64_t'),
+          ('float32', 'float32_t'),
+          ('float64', 'float64_t'),
+          ('object', 'object')]
+}}
+
+{{for dtype, c_type in dtypes}}
+
+
+@cython.wraparound(False)
+@cython.boundscheck(False)
+def reshape_{{dtype}}(ndarray[{{c_type}}, ndim=2] sorted_values,
+                      ndarray[uint8_t, ndim=1] mask,
+                      Py_ssize_t stride,
+                      Py_ssize_t length,
+                      Py_ssize_t width,
+                      ndarray[{{c_type}}, ndim=2] new_values,
+                      ndarray[uint8_t, ndim=2] new_mask):
+
+    cdef:
+        Py_ssize_t i, j, w, nulls, s, offset
+
+    {{if dtype == 'object'}}
+    if True:
+    {{else}}
+    with nogil:
+    {{endif}}
+
+        for i in range(stride):
+
+            nulls = 0
+            for j in range(length):
+
+                for w in range(width):
+
+                    offset = j * width + w
+
+                    if mask[offset]:
+                        s = i * width + w
+                        new_values[j, s] = sorted_values[offset - nulls, i]
+                        new_mask[j, s] = 1
+                    else:
+                        nulls += 1
+
+{{endfor}}

pandas/tests/frame/test_reshape.py

@@ -121,19 +121,22 @@ def test_pivot_index_none(self):
         assert_frame_equal(result, expected)
 
     def test_stack_unstack(self):
-        stacked = self.frame.stack()
+        f = self.frame.copy()
+        f[:] = np.arange(np.prod(f.shape)).reshape(f.shape)
+
+        stacked = f.stack()
         stacked_df = DataFrame({'foo': stacked, 'bar': stacked})
 
         unstacked = stacked.unstack()
         unstacked_df = stacked_df.unstack()
 
-        assert_frame_equal(unstacked, self.frame)
-        assert_frame_equal(unstacked_df['bar'], self.frame)
+        assert_frame_equal(unstacked, f)
+        assert_frame_equal(unstacked_df['bar'], f)
 
         unstacked_cols = stacked.unstack(0)
         unstacked_cols_df = stacked_df.unstack(0)
-        assert_frame_equal(unstacked_cols.T, self.frame)
-        assert_frame_equal(unstacked_cols_df['bar'].T, self.frame)
+        assert_frame_equal(unstacked_cols.T, f)
+        assert_frame_equal(unstacked_cols_df['bar'].T, f)
 
     def test_unstack_fill(self):
 

setup.py

@@ -113,6 +113,7 @@ def is_platform_mac():
 _pxi_dep_template = {
     'algos': ['algos_common_helper.pxi.in', 'algos_groupby_helper.pxi.in',
               'algos_take_helper.pxi.in', 'algos_rank_helper.pxi.in'],
+    '_reshape': ['reshape_helper.pxi.in'],
     '_join': ['join_helper.pxi.in', 'joins_func_helper.pxi.in'],
     'hashtable': ['hashtable_class_helper.pxi.in',
                   'hashtable_func_helper.pxi.in'],
@@ -496,6 +497,8 @@ def pxd(name):
     algos={'pyxfile': 'algos',
            'pxdfiles': ['src/util', 'hashtable'],
            'depends': _pxi_dep['algos']},
+    _reshape={'pyxfile': 'src/reshape',
+              'depends': _pxi_dep['_reshape']},
     _join={'pyxfile': 'src/join',
            'pxdfiles': ['src/util', 'hashtable'],
            'depends': _pxi_dep['_join']},