Windows CI (pandas-dev#23182)

Author: TomAugspurger

ci/azure-windows-36.yaml

@@ -5,12 +5,14 @@ channels:
 dependencies:
   - blosc
   - bottleneck
+  - boost-cpp<1.67
   - fastparquet
   - feather-format
   - matplotlib
   - numexpr
   - numpy=1.14*
   - openpyxl=2.5.5
+  - parquet-cpp
   - pyarrow
   - pytables
   - python-dateutil

pandas/_libs/src/headers/cmath

@@ -1,16 +1,36 @@
 #ifndef _PANDAS_MATH_H_
 #define _PANDAS_MATH_H_
 
+// MSVC 2017 has a bug where `x == x` can be true for NaNs.
+// MSC_VER from https://stackoverflow.com/a/70630/1889400
+// Place upper bound on this check once a fixed MSVC is released.
+#if defined(_MSC_VER) && (_MSC_VER < 1800)
+#include <cmath>
 // In older versions of Visual Studio there wasn't a std::signbit defined
 // This defines it using _copysign
-#if defined(_MSC_VER) && (_MSC_VER < 1800)
+namespace std {
+   __inline int isnan(double x) { return _isnan(x); }
+   __inline int signbit(double num) { return _copysign(1.0, num) < 0; }
+   __inline int notnan(double x) { return !isnan(x); }
+}
+#elif defined(_MSC_VER) && (_MSC_VER >= 1900)
+#include <cmath>
+namespace std {
+  __inline int isnan(double x) { return _isnan(x); }
+  __inline int notnan(double x) { return !isnan(x); }
+}
+#elif defined(_MSC_VER)
 #include <cmath>
 namespace std {
   __inline int isnan(double x) { return _isnan(x); }
-  __inline int signbit(double num) { return _copysign(1.0, num) < 0; }
+  __inline int notnan(double x) { return x == x; }
 }
 #else
 #include <cmath>
-#endif
 
+namespace std {
+  __inline int notnan(double x) { return x == x; }
+}
+
+#endif
 #endif

pandas/_libs/window.pyx

@@ -15,6 +15,7 @@ cnp.import_array()
 
 cdef extern from "src/headers/cmath" namespace "std":
     bint isnan(double) nogil
+    bint notnan(double) nogil
     int signbit(double) nogil
     double sqrt(double x) nogil
 
@@ -381,21 +382,21 @@ def roll_count(ndarray[double_t] input, int64_t win, int64_t minp,
                 count_x = 0.0
                 for j in range(s, e):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         count_x += 1.0
 
             else:
 
                 # calculate deletes
                 for j in range(start[i - 1], s):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         count_x -= 1.0
 
                 # calculate adds
                 for j in range(end[i - 1], e):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         count_x += 1.0
 
             if count_x >= minp:
@@ -424,15 +425,15 @@ cdef inline void add_sum(double val, int64_t *nobs, double *sum_x) nogil:
     """ add a value from the sum calc """
 
     # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] + 1
         sum_x[0] = sum_x[0] + val
 
 
 cdef inline void remove_sum(double val, int64_t *nobs, double *sum_x) nogil:
     """ remove a value from the sum calc """
 
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] - 1
         sum_x[0] = sum_x[0] - val
 
@@ -538,7 +539,7 @@ cdef inline void add_mean(double val, Py_ssize_t *nobs, double *sum_x,
     """ add a value from the mean calc """
 
     # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] + 1
         sum_x[0] = sum_x[0] + val
         if signbit(val):
@@ -549,7 +550,7 @@ cdef inline void remove_mean(double val, Py_ssize_t *nobs, double *sum_x,
                              Py_ssize_t *neg_ct) nogil:
     """ remove a value from the mean calc """
 
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] - 1
         sum_x[0] = sum_x[0] - val
         if signbit(val):
@@ -671,8 +672,7 @@ cdef inline void remove_var(double val, double *nobs, double *mean_x,
     """ remove a value from the var calc """
     cdef double delta
 
-    # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] - 1
         if nobs[0]:
             # a part of Welford's method for the online variance-calculation
@@ -760,7 +760,7 @@ def roll_var(ndarray[double_t] input, int64_t win, int64_t minp,
                 val = input[i]
                 prev = input[i - win]
 
-                if val == val:
+                if notnan(val):
                     if prev == prev:
 
                         # Adding one observation and removing another one
@@ -822,7 +822,7 @@ cdef inline void add_skew(double val, int64_t *nobs, double *x, double *xx,
     """ add a value from the skew calc """
 
     # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] + 1
 
         # seriously don't ask me why this is faster
@@ -836,7 +836,7 @@ cdef inline void remove_skew(double val, int64_t *nobs, double *x, double *xx,
     """ remove a value from the skew calc """
 
     # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] - 1
 
         # seriously don't ask me why this is faster
@@ -959,7 +959,7 @@ cdef inline void add_kurt(double val, int64_t *nobs, double *x, double *xx,
     """ add a value from the kurotic calc """
 
     # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] + 1
 
         # seriously don't ask me why this is faster
@@ -974,7 +974,7 @@ cdef inline void remove_kurt(double val, int64_t *nobs, double *x, double *xx,
     """ remove a value from the kurotic calc """
 
     # Not NaN
-    if val == val:
+    if notnan(val):
         nobs[0] = nobs[0] - 1
 
         # seriously don't ask me why this is faster
@@ -1089,7 +1089,7 @@ def roll_median_c(ndarray[float64_t] input, int64_t win, int64_t minp,
 
                 # setup
                 val = input[i]
-                if val == val:
+                if notnan(val):
                     nobs += 1
                     err = skiplist_insert(sl, val) != 1
                     if err:
@@ -1100,14 +1100,14 @@ def roll_median_c(ndarray[float64_t] input, int64_t win, int64_t minp,
                 # calculate deletes
                 for j in range(start[i - 1], s):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         skiplist_remove(sl, val)
                         nobs -= 1
 
                 # calculate adds
                 for j in range(end[i - 1], e):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         nobs += 1
                         err = skiplist_insert(sl, val) != 1
                         if err:
@@ -1472,7 +1472,7 @@ def roll_quantile(ndarray[float64_t, cast=True] input, int64_t win,
 
                 # setup
                 val = input[i]
-                if val == val:
+                if notnan(val):
                     nobs += 1
                     skiplist_insert(skiplist, val)
 
@@ -1481,14 +1481,14 @@ def roll_quantile(ndarray[float64_t, cast=True] input, int64_t win,
                 # calculate deletes
                 for j in range(start[i - 1], s):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         skiplist_remove(skiplist, val)
                         nobs -= 1
 
                 # calculate adds
                 for j in range(end[i - 1], e):
                     val = input[j]
-                    if val == val:
+                    if notnan(val):
                         nobs += 1
                         skiplist_insert(skiplist, val)