Add Reciprocal

numpy/_core/code_generators/generate_umath.py

@@ -988,7 +988,7 @@ def english_upper(s):
           docstrings.get('numpy._core.umath.rint'),
           None,
           TD('e', f='rint', astype={'e': 'f'}),
-          TD(inexactvec, dispatch=[('loop_unary_fp', 'fd')]),
+          TD(inexactvec, dispatch=[('loops_unary_fp', 'fd')]),
           TD('fdg' + cmplx, f='rint'),
           TD(P, f='rint'),
           ),

numpy/_core/meson.build

@@ -107,7 +107,8 @@ if use_highway
       # required for hwy::Abort symbol
       'src/highway/hwy/abort.cc',
       'src/highway/hwy/per_target.cc',
-      'src/highway/hwy/targets.cc'
+      'src/highway/hwy/targets.cc',
+      'src/highway/hwy/print.cc'
     ],
     cpp_args: '-DTOOLCHAIN_MISS_ASM_HWCAP_H',
     include_directories: ['src/highway'],
@@ -969,17 +970,7 @@ foreach gen_mtargets : [
   ],
   [
     'loops_unary_fp.dispatch.h',
-    src_file.process('src/umath/loops_unary_fp.dispatch.c.src'),
-    [
-      SSE42, SSE2,
-      VSX2,
-      ASIMD, NEON,
-      VXE, VX
-    ]
-  ],
-  [
-    'loop_unary_fp.dispatch.h',
-    'src/umath/loop_unary_fp.dispatch.cpp',
+    'src/umath/loops_unary_fp.dispatch.cpp',
     [
       SSE42, SSE2,
       VSX2,

numpy/_core/src/umath/loops.h.src

@@ -247,7 +247,6 @@ NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@kind@,
  */
 #ifndef NPY_DISABLE_OPTIMIZATION
     #include "loops_unary_fp.dispatch.h"
-    #include "loop_unary_fp.dispatch.h"
 #endif
 /**begin repeat
  *  #TYPE = FLOAT, DOUBLE#

numpy/_core/src/umath/loop_unary_fp.dispatch.cpp → numpy/_core/src/umath/loops_unary_fp.dispatch.cpp

@@ -90,8 +90,15 @@ namespace hn = hwy::HWY_NAMESPACE;
         hn::ScatterIndex(x, d, output_array + i * sdst, store_index);          \
       }                                                                        \
       if (remainder) {                                                         \
-        const auto in = hn::GatherIndexN(d, input_array + full * ssrc,         \
+        hn::Vec<hn::ScalableTag<T>> in;                                        \
+        if (IS_RECIP) {                                                        \
+          auto one = hn::Set(d, 1);                                            \
+          in = hn::GatherIndexNOr(one, d, input_array + full * ssrc,           \
                                          load_index, remainder);               \
+        } else {                                                               \
+          in = hn::GatherIndexN(d, input_array + full * ssrc,                  \
+                                         load_index, remainder);               \
+        }                                                                      \
         auto x = FUNC(in);                                                     \
         hn::ScatterIndexN(x, d, output_array + full * sdst, store_index,       \
                           remainder);                                          \
@@ -100,6 +107,7 @@ namespace hn = hwy::HWY_NAMESPACE;
   }
 
 #define Square(x) hn::Mul(x, x) 
+#define Reciprocal(x) hn::Div(hn::Set(d, T(1.0)), x)
 //ceil, trunc, sqrt, square, 
 SUPER(Rint, hn::Round, false)
 SUPER(Floor, hn::Floor, false)
@@ -108,6 +116,7 @@ SUPER(Trunc, hn::Trunc, false)
 SUPER(Sqrt, hn::Sqrt, false)
 SUPER(Square, Square, false)
 SUPER(Absolute, hn::Abs, false)
+SUPER(Reciprocal, Reciprocal, true)
 
 extern "C" {
 NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(DOUBLE_rint)
@@ -195,4 +204,16 @@ NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(FLOAT_absolute)
   SuperAbsolute<npy_float>(args, dimensions, steps);
   npy_clear_floatstatus_barrier((char*)dimensions);
 }
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(DOUBLE_reciprocal)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+  SuperReciprocal<npy_double>(args, dimensions, steps);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(FLOAT_reciprocal)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+  SuperReciprocal<npy_float>(args, dimensions, steps);
+}
 }