mocks: warn about small theta and large mu in float32 precision

Corrfunc/mocks/DDsmu_mocks.py

@@ -8,6 +8,7 @@
 
 from __future__ import (division, print_function, absolute_import,
                         unicode_literals)
+import warnings
 
 __author__ = ('Manodeep Sinha', 'Nick Hand')
 __all__ = ('DDsmu_mocks', )
@@ -296,6 +297,12 @@ def DDsmu_mocks(autocorr, cosmology, nthreads, mu_max, nmu_bins, binfile,
 
     integer_isa = translate_isa_string_to_enum(isa)
     sbinfile, delete_after_use = return_file_with_rbins(binfile)
+
+    warn_large_mu(mu_max,
+                  # RA and DEC are checked to be the same dtype
+                  dtype=RA1.dtype,
+                  )
+
     with sys_pipes():
         extn_results = DDsmu_extn(autocorr, cosmology, nthreads,
                                   mu_max, nmu_bins, sbinfile,
@@ -344,6 +351,26 @@ def DDsmu_mocks(autocorr, cosmology, nthreads, mu_max, nmu_bins, binfile,
     else:
         return results, api_time
 
+
+def warn_large_mu(mu_max, dtype):
+    '''
+    Small theta values (large mu) underfloat float32. Warn the user.
+    Context: https://github.com/manodeep/Corrfunc/pull/297
+    '''
+    if dtype.itemsize > 4:
+        return
+
+    if mu_max >= 0.9800666:  # cos(0.2)
+        warnings.warn("""
+Be aware that small angular pair separations (mu near 1) will suffer from loss
+of floating-point precision, as the input data is in float32 precision or
+lower. In float32, the loss of precision is 1% in mu at separations of 0.2
+degrees, and larger at smaller separations.
+For more information, see: https://github.com/manodeep/Corrfunc/pull/297
+"""
+                      )
+
+
 if __name__ == '__main__':
     import doctest
     doctest.testmod()

Corrfunc/mocks/DDtheta_mocks.py

@@ -10,6 +10,7 @@
 
 from __future__ import (division, print_function, absolute_import,
                         unicode_literals)
+import warnings
 
 __author__ = ('Manodeep Sinha', 'Kris Akira Stern')
 __all__ = ('DDtheta_mocks', 'find_fastest_DDtheta_mocks_bin_refs')
@@ -313,6 +314,11 @@ def DDtheta_mocks(autocorr, nthreads, binfile,
 
     integer_isa = translate_isa_string_to_enum(isa)
     rbinfile, delete_after_use = return_file_with_rbins(binfile)
+
+    warn_small_theta(rbinfile,
+                     RA1.dtype,  # RA and DEC are checked to be the same dtype
+                     )
+
     with sys_pipes():
         extn_results = DDtheta_mocks_extn(autocorr, nthreads, rbinfile,
                                           RA1, DEC1,
@@ -644,6 +650,40 @@ def find_fastest_DDtheta_mocks_bin_refs(autocorr, nthreads, binfile,
     return ret
 
 
+def warn_small_theta(thetabinfile, dtype):
+    '''
+    Small theta values underfloat float32. Warn the user.
+    Context: https://github.com/manodeep/Corrfunc/pull/297
+    '''
+    if dtype.itemsize > 4:
+        return
+
+    import numpy as np
+    try:
+        bins = np.loadtxt(thetabinfile)
+    except RuntimeError:
+        warnings.warn("""
+Could not load binfile "{}". Be aware that small angular pair separations
+will suffer from loss of floating-point precision, as the input data is in
+float32 precision or lower. The loss of precision is 0.5% in theta at
+separations of 0.2 degrees, and larger at smaller separations.
+For more information, see: https://github.com/manodeep/Corrfunc/pull/297
+""".format(thetabinfile)
+                      )
+        return
+
+    if bins.min() <= 0.2:
+        warnings.warn("""
+A binning with a minimum angular separation of 0.2 degrees or less was
+requested, and the input data is in float32 precision or lower. Be aware that
+small angular pair separations will suffer from loss of floating-point
+precision. In float32, the loss of precision is 0.5% in theta at separations of
+0.2 degrees, and larger at smaller separations.
+For more information, see: https://github.com/manodeep/Corrfunc/pull/297
+"""
+                      )
+
+
 if __name__ == '__main__':
     import doctest