From 7ddb5daa866984caa78e3fa4b5cd4869f4ee94cf Mon Sep 17 00:00:00 2001
From: Sayed Adel <seiko@imavr.com>
Date: Tue, 7 Feb 2023 21:04:27 +0200
Subject: ENH, SIMD: removes #NPY_ENABLE_AVX512_QSORT and use #directives
 instead

---
 numpy/core/setup.py                                  | 19 ++-----------------
 numpy/core/src/npysort/quicksort.cpp                 | 11 +++++++++++
 numpy/core/src/npysort/simd_qsort.dispatch.cpp       |  4 ++--
 numpy/core/src/npysort/simd_qsort_16bit.dispatch.cpp |  4 ++--
 4 files changed, 17 insertions(+), 21 deletions(-)

(limited to 'numpy/core')

diff --git a/numpy/core/setup.py b/numpy/core/setup.py
index cfae34e31..d6117f02d 100644
--- a/numpy/core/setup.py
+++ b/numpy/core/setup.py
@@ -68,14 +68,6 @@ class CallOnceOnly:
             out = copy.deepcopy(pickle.loads(self._check_complex))
         return out
 
-# Temporarily disable AVX512 sorting on WIN32 until we can figure
-# out why it has test failures
-def enable_avx512_qsort():
-    enable = True
-    if "win32" in sysconfig.get_platform():
-        enable = False
-    return enable
-
 def can_link_svml():
     """SVML library is supported only on x86_64 architecture and currently
     only on linux
@@ -492,9 +484,6 @@ def configuration(parent_package='',top_path=None):
             if can_link_svml():
                 moredefs.append(('NPY_CAN_LINK_SVML', 1))
 
-            if enable_avx512_qsort():
-                moredefs.append(('NPY_ENABLE_AVX512_QSORT', 1))
-
             # Use bogus stride debug aid to flush out bugs where users use
             # strides of dimensions with length 1 to index a full contiguous
             # array.
@@ -975,14 +964,10 @@ def configuration(parent_package='',top_path=None):
             # links to the arm64 npymath library,
             # see gh-22673
             join('src', 'npymath', 'arm64_exports.c'),
+            join('src', 'npysort', 'simd_qsort.dispatch.cpp'),
+            join('src', 'npysort', 'simd_qsort_16bit.dispatch.cpp'),
             ]
 
-    if enable_avx512_qsort():
-        multiarray_src += [
-                join('src', 'npysort', 'simd_qsort.dispatch.cpp'),
-                join('src', 'npysort', 'simd_qsort_16bit.dispatch.cpp'),
-                ]
-
     #######################################################################
     #             _multiarray_umath module - umath part                   #
     #######################################################################
diff --git a/numpy/core/src/npysort/quicksort.cpp b/numpy/core/src/npysort/quicksort.cpp
index 0e65dc9bc..625fdebbb 100644
--- a/numpy/core/src/npysort/quicksort.cpp
+++ b/numpy/core/src/npysort/quicksort.cpp
@@ -69,6 +69,15 @@
 #define SMALL_MERGESORT 20
 #define SMALL_STRING 16
 
+// Temporarily disable AVX512 sorting on WIN32 until we can figure
+//  out why it has test failures
+#ifdef _MSC_VER
+template<typename T>
+inline bool quicksort_dispatch(T*, npy_intp)
+{
+    return false;
+}
+#else
 template<typename T>
 inline bool quicksort_dispatch(T *start, npy_intp num)
 {
@@ -92,6 +101,8 @@ inline bool quicksort_dispatch(T *start, npy_intp num)
     }
     return false;
 }
+#endif // _MSC_VER
+
 /*
  *****************************************************************************
  **                            NUMERIC SORTS                                **
diff --git a/numpy/core/src/npysort/simd_qsort.dispatch.cpp b/numpy/core/src/npysort/simd_qsort.dispatch.cpp
index 36b5d799c..c2ac5a2ae 100644
--- a/numpy/core/src/npysort/simd_qsort.dispatch.cpp
+++ b/numpy/core/src/npysort/simd_qsort.dispatch.cpp
@@ -7,14 +7,14 @@
 
 #include "simd_qsort.hpp"
 
-#ifdef NPY_HAVE_AVX512_SKX
+#if defined(NPY_HAVE_AVX512_SKX) && !defined(_MSC_VER)
     #include "avx512-32bit-qsort.hpp"
     #include "avx512-64bit-qsort.hpp"
 #endif
 
 namespace np { namespace qsort_simd {
 
-#ifdef NPY_HAVE_AVX512_SKX
+#if defined(NPY_HAVE_AVX512_SKX) && !defined(_MSC_VER)
 template<> void NPY_CPU_DISPATCH_CURFX(QSort)(int32_t *arr, intptr_t size)
 {
     avx512_qsort(arr, size);
diff --git a/numpy/core/src/npysort/simd_qsort_16bit.dispatch.cpp b/numpy/core/src/npysort/simd_qsort_16bit.dispatch.cpp
index a816b8781..673a2f81e 100644
--- a/numpy/core/src/npysort/simd_qsort_16bit.dispatch.cpp
+++ b/numpy/core/src/npysort/simd_qsort_16bit.dispatch.cpp
@@ -7,13 +7,13 @@
 
 #include "simd_qsort.hpp"
 
-#ifdef NPY_HAVE_AVX512_ICL
+#if defined(NPY_HAVE_AVX512_ICL) && !defined(_MSC_VER)
     #include "avx512-16bit-qsort.hpp"
 #endif
 
 namespace np { namespace qsort_simd {
 
-#ifdef NPY_HAVE_AVX512_ICL
+#if defined(NPY_HAVE_AVX512_ICL) && !defined(_MSC_VER)
 template<> void NPY_CPU_DISPATCH_CURFX(QSort)(Half *arr, intptr_t size)
 {
     avx512_qsort_fp16(reinterpret_cast<uint16_t*>(arr), size);
-- 
cgit v1.2.1