Merge pull request numpy#25781 from r-devulap/sincos-hwy

ENH: Convert fp32 sin/cos from C universal intrinsics to C++ using Highway

Author: Mousius

meson_cpu/x86/meson.build

@@ -59,13 +59,14 @@ FMA3 = mod_features.new(
   'FMA3', 24, implies: F16C, args: '-mfma',
   test_code: files(source_root + '/numpy/distutils/checks/cpu_fma3.c')[0]
 )
+# match this to HWY_AVX2
 AVX2 = mod_features.new(
-  'AVX2', 25, implies: F16C, args: '-mavx2',
+  'AVX2', 25, implies: FMA3, args: ['-mavx2', '-maes', '-mpclmul', '-mbmi', '-mbmi2'],
   test_code: files(source_root + '/numpy/distutils/checks/cpu_avx2.c')[0]
 )
 # 25-40 left as margin for any extra features
 AVX512F = mod_features.new(
-  'AVX512F', 40, implies: [FMA3, AVX2],
+  'AVX512F', 40, implies: [AVX2],
   # Disables mmx because of stack corruption that may happen during mask
   # conversions.
   # TODO (seiko2plus): provide more clarification

numpy/_core/meson.build

@@ -944,12 +944,10 @@ foreach gen_mtargets : [
   ],
   [
     'loops_trigonometric.dispatch.h',
-    src_file.process('src/umath/loops_trigonometric.dispatch.c.src'),
+    'src/umath/loops_trigonometric.dispatch.cpp',
     [
-      AVX512F, [AVX2, FMA3],
-      VSX4, VSX3, VSX2,
+      AVX512_SKX, [AVX2, FMA3],
       NEON_VFPV4,
-      VXE2, VXE
     ]
   ],
   [
@@ -1020,7 +1018,8 @@ foreach gen_mtargets : [
       'src/common',
       'src/multiarray',
       'src/npymath',
-      'src/umath'
+      'src/umath',
+      'src/highway',
     ]
   )
   if not is_variable('multiarray_umath_mtargets')

numpy/_core/src/common/npy_cpu_features.c

@@ -474,6 +474,8 @@ npy__cpu_init_features(void)
     // third call to the cpuid to get extended AVX2 & AVX512 feature bits
     npy__cpu_cpuid(reg, 7);
     npy__cpu_have[NPY_CPU_FEATURE_AVX2]   = (reg[1] & (1 << 5))  != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_AVX2]   = npy__cpu_have[NPY_CPU_FEATURE_AVX2] &&
+                                            npy__cpu_have[NPY_CPU_FEATURE_FMA3];
     if (!npy__cpu_have[NPY_CPU_FEATURE_AVX2])
         return;
     // detect AVX2 & FMA3
@@ -641,7 +643,7 @@ static void
 npy__cpu_init_features(void)
 {
     memset(npy__cpu_have, 0, sizeof(npy__cpu_have[0]) * NPY_CPU_FEATURE_MAX);
-    
+
     unsigned int hwcap = getauxval(AT_HWCAP);
     if ((hwcap & HWCAP_S390_VX) == 0) {
         return;
@@ -653,7 +655,7 @@ npy__cpu_init_features(void)
        npy__cpu_have[NPY_CPU_FEATURE_VXE2] = 1;
        return;
     }
-    
+
     npy__cpu_have[NPY_CPU_FEATURE_VXE] = (hwcap & HWCAP_S390_VXE) != 0;
 
     npy__cpu_have[NPY_CPU_FEATURE_VX]  = 1;

numpy/_core/src/highway

@@ -10,6 +10,10 @@
     #define NPY_NO_EXPORT NPY_VISIBILITY_HIDDEN
 #endif
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /*
  *****************************************************************************
  **                             BOOLEAN LOOPS                               **
@@ -875,5 +879,7 @@ PyUFunc_OOO_O(char **args, npy_intp const *dimensions, npy_intp const *steps, vo
  **                              END LOOPS                                  **
  *****************************************************************************
  */
-
+#ifdef __cplusplus
+}
+#endif
 #endif