ENH: unroll vector minmax loop

Improves speed by 10% on intel cpus.
Simplify code by moving the fenv support check into the dispatcher. fenv works
on all common platforms (including windows), the fallback is not worth it for
the exotic platforms where it might not be available.

1 files changed, 22 insertions, 26 deletions

diff --git a/numpy/core/src/umath/simd.inc.src b/numpy/core/src/umath/simd.inc.src
index e274e0596..e5c4e9335 100644
--- a/numpy/core/src/umath/simd.inc.src
+++ b/numpy/core/src/umath/simd.inc.src
@@ -138,6 +138,7 @@ static const npy_int32 fanout_4[] = {
  * #func = sqrt, absolute, minimum, maximum#
  * #check = IS_BLOCKABLE_UNARY, IS_BLOCKABLE_UNARY, IS_BLOCKABLE_REDUCE, IS_BLOCKABLE_REDUCE#
  * #name = unary, unary, unary_reduce, unary_reduce#
+ * #minmax = 0, 0, 1, 1#
  */
 
 #if @vector@ && defined NPY_HAVE_SSE2_INTRINSICS
@@ -151,6 +152,9 @@ sse2_@func@_@TYPE@(@type@ *, @type@ *, const npy_intp n);
 static NPY_INLINE int
 run_@name@_simd_@func@_@TYPE@(char **args, npy_intp *dimensions, npy_intp *steps)
 {
+#if @minmax@ && (defined NO_FLOATING_POINT_SUPPORT)
+    return 0;
+#else
 #if @vector@ && defined NPY_HAVE_SSE2_INTRINSICS
     if (@check@(sizeof(@type@), 16)) {
         sse2_@func@_@TYPE@((@type@*)args[1], (@type@*)args[0], dimensions[0]);
@@ -158,6 +162,7 @@ run_@name@_simd_@func@_@TYPE@(char **args, npy_intp *dimensions, npy_intp *steps
     }
 #endif
     return 0;
+#endif
 }
 
 /**end repeat1**/
@@ -660,42 +665,33 @@ sse2_absolute_@TYPE@(@type@ * op, @type@ * ip, const npy_intp n)
 static void
 sse2_@kind@_@TYPE@(@type@ * ip, @type@ * op, const npy_intp n)
 {
+    const size_t stride = 16 / sizeof(@type@);
     LOOP_BLOCK_ALIGN_VAR(ip, @type@, 16) {
         *op = (*op @OP@ ip[i] || npy_isnan(*op)) ? *op : ip[i];
     }
-    assert(n < (16 / sizeof(@type@)) || npy_is_aligned(&ip[i], 16));
-    if (i + 2 * 16 / sizeof(@type@) <= n) {
+    assert(n < (stride) || npy_is_aligned(&ip[i], 16));
+    if (i + 3 * stride <= n) {
         /* load the first elements */
-        @vtype@ c = @vpre@_load_@vsuf@((@type@*)&ip[i]);
-#ifdef NO_FLOATING_POINT_SUPPORT
-        @vtype@ cnan = @vpre@_cmpneq_@vsuf@(c, c);
-#else
+        @vtype@ c1 = @vpre@_load_@vsuf@((@type@*)&ip[i]);
+        @vtype@ c2 = @vpre@_load_@vsuf@((@type@*)&ip[i + stride]);
+        i += 2 * stride;
+
         /* minps/minpd will set invalid flag if nan is encountered */
         PyUFunc_clearfperr();
-#endif
-        i += 16 / sizeof(@type@);
-
-        LOOP_BLOCKED(@type@, 16) {
-            @vtype@ v = @vpre@_load_@vsuf@((@type@*)&ip[i]);
-            c = @vpre@_@VOP@_@vsuf@(c, v);
-#ifdef NO_FLOATING_POINT_SUPPORT
-            /* check for nan, breaking the loop makes non nan case slow */
-            cnan = @vpre@_or_@vsuf@(@vpre@_cmpneq_@vsuf@(v, v), cnan);
+        LOOP_BLOCKED(@type@, 32) {
+            @vtype@ v1 = @vpre@_load_@vsuf@((@type@*)&ip[i]);
+            @vtype@ v2 = @vpre@_load_@vsuf@((@type@*)&ip[i + stride]);
+            c1 = @vpre@_@VOP@_@vsuf@(c1, v1);
+            c2 = @vpre@_@VOP@_@vsuf@(c2, v2);
         }
+        c1 = @vpre@_@VOP@_@vsuf@(c1, c2);
 
-        if (@vpre@_movemask_@vsuf@(cnan)) {
+        if (PyUFunc_getfperr() & UFUNC_FPE_INVALID) {
             *op = @nan@;
-            return;
-        }
-#else
         }
-#endif
-        {
-            @type@ tmp = sse2_horizontal_@VOP@_@vtype@(c);
-            if (PyUFunc_getfperr() & UFUNC_FPE_INVALID)
-                *op = @nan@;
-            else
-                *op  = (*op @OP@ tmp || npy_isnan(*op)) ? *op : tmp;
+        else {
+            @type@ tmp = sse2_horizontal_@VOP@_@vtype@(c1);
+            *op  = (*op @OP@ tmp || npy_isnan(*op)) ? *op : tmp;
         }
     }
     LOOP_BLOCKED_END {