Minor optimization for linear search

Add a small loop unrolling to when BGEN_MAYBELESSEQUAL is set.
tidwall · Dec 20, 2024 · 1e2ae70 · 1e2ae70
1 parent 5fbb5a1
commit 1e2ae70
Show file tree

Hide file tree

Showing 3 changed files with 17 additions and 5 deletions.
diff --git a/bgen.h b/bgen.h
@@ -920,13 +920,24 @@ static int BGEN_SYM(search_bsearch)(BGEN_ITEM *items, int nitems,
     return i;
 }
 #else
+
+#if defined(BGEN_MLESIMD)
+#include <immintrin.h>
+#endif
+
 BGEN_INLINE
 static int BGEN_SYM(search_linear)(BGEN_ITEM *items, int nitems, BGEN_ITEM key,
     void *udata, int *found)
 {
     int i = 0;
     *found = 0;
 #ifdef BGEN_MAYBELESSEQUAL
+    while (nitems-i >= 4) {
+        if (BGEN_SYM(maybelessequal)(key, items[i], udata)){goto compare;}i++;
+        if (BGEN_SYM(maybelessequal)(key, items[i], udata)){goto compare;}i++;
+        if (BGEN_SYM(maybelessequal)(key, items[i], udata)){goto compare;}i++;
+        if (BGEN_SYM(maybelessequal)(key, items[i], udata)){goto compare;}i++;
+    }
     for (; i < nitems; i++) {
         if (BGEN_SYM(maybelessequal)(key, items[i], udata)) {
             goto compare;

diff --git a/tests/bench_s.c b/tests/bench_s.c
@@ -19,7 +19,7 @@ int C = 0;        // -1 = worse-case, 0 = average, +1 = best-case
 // #define USECOMPARE
 
 struct point {
-    uint32_t curve;
+    int32_t curve;
     int id;
     double x;
     double y;
@@ -195,7 +195,7 @@ int main(void) {
         point.x = rand_double()*360.0-180.0;
         point.y = rand_double()*180.0-90.0;
         point.curve = curve_hilbert(point.x, point.y, 
-            (double[4]){-180, -90, 180, 90});
+            (double[4]){-180, -90, 180, 90})/2;
         // point.curve = curve_z(point.y, point.x);
         keys[i] = point;
     }

diff --git a/tests/test_spatial2x.c b/tests/test_spatial2x.c
@@ -5,8 +5,8 @@
 #include "curve.h"
 
 struct point {
-    int id;
     uint32_t curve;
+    int id;
     double x;
     double y;
 };
@@ -40,8 +40,9 @@ static void shuffle_points(struct point *array, size_t numels) {
 #define BGEN_FANOUT     4
 #define BGEN_COUNTED
 #define BGEN_SPATIAL
-#define BGEN_ITEMRECT   { point_rect(item, min, max); }
-#define BGEN_COMPARE    { return point_compare(a, b); }
+#define BGEN_ITEMRECT point_rect(item, min, max);
+#define BGEN_MAYBELESSEQUAL return a.curve <= b.curve;
+#define BGEN_COMPARE return point_compare(a, b);
 #include "../bgen.h"