update ldexp types

[vtavana]

With stock NumPy 2.2.5 on Windows we have

>>> numpy.ldexp.types
# ['ei->e', 'fi->f', 'eq->e', 'fq->f', 'di->d', 'dq->d', 'gi->g', 'gq->g']`

While for Stock NumPy 2.2.5 on Linux we have

>>> numpy.ldexp.types 
# ['ei->e', 'fi->f', 'el->e', 'fl->f', 'di->d', 'dl->d', 'gi->g', 'gl->g']`

The difference is that we have 'q' on Windows instead of 'l' on Linux.

With Intel/Stock NumPy 1.26.4, on both Windows and Linux we have 'l'.

mkl_umath currently has functions for type 'fl->f' and 'dl->d' which have no equivalent in Stock NumPy 2.2.5 on Windows.
This PR resolves this issue. The changes I made is based on what is done in numpy-2.x.x.

With this branch

# Python 3.11.12 | packaged by conda-forge | [MSC v.1943 64 bit (AMD64)]
>>> import mkl_umath, numpy
>>> mkl_umath.use_in_numpy()

>>> mkl_umath.ldexp.types
# ['fi->f', 'fq->f', 'di->d', 'dq->d']
>>> numpy.ldexp.types
# ['ei->e', 'fi->f', 'eq->e', 'fq->f', 'di->d', 'dq->d', 'gi->g', 'gq->g']

[AndresGuzman-Ballen]

Since this branch is currently being used to debug numpy test failures, I wonder if it'd be worth making this change in this PR:

diff --git a/CMakeLists.txt b/CMakeLists.txt
index af4e480..c421df7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -45,6 +45,7 @@ if(WIN32)
    string(CONCAT PRECISION_FLAGS
      "/fp:fast=2 "
      "/Qimf-precision=high "
+     "/Qprec-sqrt "
      "/Qprotect-parens "
    )
    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /Ox ${WARNING_FLAGS} ${SDL_FLAGS} ${PRECISION_FLAGS}")
@@ -81,6 +82,7 @@ elseif(UNIX)
      "${SDL_FLAGS}"
    )
    string(CONCAT PRECISION_FLAGS
+     "-prec-sqrt "
      "-fprotect-parens "
      "-fimf-precision=high "
      "-fp-model fast=2 "

Using assert_almost_equal didn't address the issues I was seeing in the test failures, and it was only when I saw "sqrt" in the test failure did I realize that the failures were probably stemming from some change in the use of the -prec-sqrt flag and sure enough, it was removed in 77ef1a5

[AndresGuzman-Ballen]

Within the context of fixing the ldexp issue, I think this PR is good!

[ndgrigorian]

@AndresGuzman-Ballen
we had discussion in #69 about removing the option and the possibility of test failures.

What's interesting is that it is apparently no longer going to be supported—but if this means test failures, do we know what alternative they recommend?

Note that following their recommendation to use icx -qnextgen-diag does list -prec-sqrt as being removed.

[vtavana]

What's interesting is that it is apparently no longer going to be supported—but if this means test failures, do we know what alternative they recommend?

@ndgrigorian
I talked to Andres yesterday. They use frankenstein internal ICC compiler not icx for Intel numpy recipe. They should continue to use the patched CMakeLists.txt or use icx, if possible.

[AndresGuzman-Ballen]

@vtavana correct! @ndgrigorian sorry for the confusion.

[antonwolfy]

It seems the changelog needs to be updated with that change

[ndgrigorian]

LGTM!

[AndresGuzman-Ballen]

By the way, will we need to increment the patch version number for mkl_umath after this is merged?

[ndgrigorian]

By the way, will we need to increment the patch version number for mkl_umath after this is merged?

Do you mean the package version? I figured we would cut a new version around the time of the release, but if numpy builds need it, I can go ahead and do it whenever.

[AndresGuzman-Ballen]

@ndgrigorian good point! I'll keep an eye out for new versions around release time then :) thanks!