From 645bc19f19cf8791bcb483dfeb1e1e75292ab2e1 Mon Sep 17 00:00:00 2001
From: Richard Thier <magosit@outlook.hu>
Date: Fri, 17 Dec 2021 22:48:38 +0100
Subject: [PATCH] Manual occurence unrolling

---
 magyarsort.h | 72 ++++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 70 insertions(+), 2 deletions(-)

diff --git a/magyarsort.h b/magyarsort.h
index ab098a0..5efb391 100644
--- a/magyarsort.h
+++ b/magyarsort.h
@@ -93,12 +93,80 @@ namespace MagyarSort {
 
 	template<typename COUNTER_TYP>
 	static inline void countOccurences(uint32_t arr[], COUNTER_TYP size, COUNTER_TYP *radicsOut) noexcept {
-		#pragma GCC unroll 64
-		for(COUNTER_TYP i = 0; i < size; ++i) {
+		// #pragma GCC unroll 64
+		COUNTER_TYP i = 0;
+		for(; i < size - 64; i += 64) {
 			// Prefetch caches
 			//__builtin_prefetch(&arr[i + 64]);
 			// Creates no object, struct is empty
 			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 1, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 2, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 3, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 4, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 5, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 6, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 7, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 8, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 9, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 10, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 11, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 12, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 13, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 14, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 15, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 16, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 17, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 18, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 19, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 20, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 21, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 22, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 23, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 24, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 25, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 26, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 27, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 28, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 29, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 30, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 31, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 32, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 33, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 34, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 35, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 36, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 37, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 38, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 39, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 40, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 41, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 42, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 43, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 44, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 45, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 46, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 47, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 48, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 49, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 50, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 51, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 52, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 53, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 54, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 55, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 56, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 57, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 58, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 59, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 60, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 61, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 62, radicsOut);
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i + 63, radicsOut);
+		}
+
+		for(; i < size; ++i) {
+			OccurenceMagic<DIGITS - 1, COUNTER_TYP>(arr, i, radicsOut);
 		}
 	}