diff --git a/randominus.h b/randominus.h
index 67709ae..1708045 100644
--- a/randominus.h
+++ b/randominus.h
@@ -3,6 +3,7 @@
 /* To randomize an array - hopefully as fast as possible */
 
 #include "fastrand/fastrand.h"
+#define RD_REL_WIDTH 16
 
 /** swap */
 static void inline rd_swap(uint32_t *a, uint32_t *b) {
@@ -24,23 +25,24 @@ static inline void randominus(uint32_t *a, int n, uint32_t seed) {
 	}
 
 	/** Go over the array and randomly swap stuff - hand unrolled with ILP random get! */
-	for(int i = 0; i < (n - 8); i += 8) {
-		uint32_t to0 = fastmodlike(lcg_ilp(&rsi, A), n);
-		uint32_t to1 = fastmodlike(lcg_ilp(&rsi, B), n);
-		uint32_t to2 = fastmodlike(lcg_ilp(&rsi, C), n);
-		uint32_t to3 = fastmodlike(lcg_ilp(&rsi, D), n);
-		uint32_t to4 = fastmodlike(lcg_ilp(&rsi, E), n);
-		uint32_t to5 = fastmodlike(lcg_ilp(&rsi, F), n);
-		uint32_t to6 = fastmodlike(lcg_ilp(&rsi, G), n);
-		uint32_t to7 = fastmodlike(lcg_ilp(&rsi, H), n);
-		rd_swap(&a[i], &a[to0]);
-		rd_swap(&a[i + 1], &a[to1]);
-		rd_swap(&a[i + 2], &a[to2]);
-		rd_swap(&a[i + 3], &a[to3]);
-		rd_swap(&a[i + 4], &a[to4]);
-		rd_swap(&a[i + 5], &a[to5]);
-		rd_swap(&a[i + 6], &a[to6]);
-		rd_swap(&a[i + 7], &a[to7]);
+	#pragma GCC unroll 16
+	for(int i = (RD_REL_WIDTH/2 + 1); i < (n - 8 - (RD_REL_WIDTH/2 + 1)); i += 8) {
+		uint32_t to = fastmodlike(lcg_ilp(&rsi, A), n);
+		uint32_t relto1 = fastmodlike(lcg_ilp(&rsi, B), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		uint32_t relto2 = fastmodlike(lcg_ilp(&rsi, C), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		uint32_t relto3 = fastmodlike(lcg_ilp(&rsi, D), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		uint32_t relto4 = fastmodlike(lcg_ilp(&rsi, E), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		uint32_t relto5 = fastmodlike(lcg_ilp(&rsi, F), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		uint32_t relto6 = fastmodlike(lcg_ilp(&rsi, G), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		uint32_t relto7 = fastmodlike(lcg_ilp(&rsi, H), RD_REL_WIDTH) - (RD_REL_WIDTH / 2);
+		rd_swap(&a[i], &a[to]);
+		rd_swap(&a[i + 1], &a[i + relto1]);
+		rd_swap(&a[i + 2], &a[i + relto2]);
+		rd_swap(&a[i + 3], &a[i + relto3]);
+		rd_swap(&a[i + 4], &a[i + relto4]);
+		rd_swap(&a[i + 5], &a[i + relto5]);
+		rd_swap(&a[i + 6], &a[i + relto6]);
+		rd_swap(&a[i + 7], &a[i + relto7]);
 	}
 }