Code Overview:
Numpy currently decouples the math function definitions in `npy_math.c.src`
from the function declarations found in `npy_math.h`. This patch allows
definitions to be included along with the inclusion of the `npy_math.h`
header.
Keeping the declarations and definitions separate is usually the right
approach, but mathematical code like this might be better off as an
exception to this common practice. Because the definitions are in the source
file instead of the header, the compiler does not have any clue what lies
underneath these math functions. This means the compiler can't make
important optimizations like inlining and vectorization. Extensions that
utilize these functions could greatly benefit from this, specifically
`loops.c.src` from the umath extension.
Implementation Details:
+ Renames `npy_math.c.src` to `npy_math_internal.h.src`
+ Generates `npy_math_internal.h` from template by adding to
`npymath_sources` list and adding `npymath` directory to include paths in
`generate_numpyconfig_h` function of `numpy/core/setup.py`
+ Numpy's core distutils defines `#NPY_INTERNAL_BUILD` macro to make sure
`npy_math_internal.h` is not included when other modules try to include
public header `npy_math.h`
- Currently do not know how to ship headers generated from template
files
+ Adds `npy_math.c`, a file that includes the `npy_math_internal.h.src`
file (but does not add NPY_INLINE static)
- This is to keep the same static npy_math library as it exists now
+ Appends `numpy/npy_math.h` with `npy_math_internal.h` under condition that
it's not being included in npy_math.c.src
- The conditional macros mean `loops.c.src` will have definitions
included, and the compiler will vectorize accordingly
+ Adds `NPY_INLINE` static to function declarations and definitions when
necessary
+ Replaces `sqrtf` with `npy_sqrtf` in `numpy/core/src/umath/umath_tests.c`
to make function portable
- `_sqrtf` was not found on certain Windows environments compiling with
Py2