Only byte-swap 16-bit PNGs on little-endian (#7792)

_png has some code that unconditionally byte-swaps 16-bit PNG data (which is, per the spec, stored in big-endian order). This isn't appropriate on a big-endian platform, though: this swap being done unconditionally breaks the handling of 16-bit PNGs on big-endian platforms (e.g. Fedora ppc64) (#7792). So, let's use some macros numpy kindly defines for us to decide whether to do this swap or not.
matplotlib · mdboom · Jan 11, 2017 · Jan 11, 2017 · Jan 11, 2017 · e96a2cac2fedb7f8fbab42c399f46d4e1216b0fe
commit e96a2cac2fedb7f8fbab42c399f46d4e1216b0fe
diff --git a/src/_png.cpp b/src/_png.cpp
@@ -532,10 +532,12 @@ static PyObject *_read_png(PyObject *filein, bool float_result)
         png_set_shift(png_ptr, sig_bit);
     }
 
+#if NPY_BYTE_ORDER == NPY_LITTLE_ENDIAN
     // Convert big endian to little
     if (bit_depth == 16) {
         png_set_swap(png_ptr);
     }
+#endif
 
     // Convert palletes to full RGB
     if (png_get_color_type(png_ptr, info_ptr) == PNG_COLOR_TYPE_PALETTE) {