From f4e4db43f9442d3aa11f96292c09b1a627ace3df Mon Sep 17 00:00:00 2001
From: Richard Thier <magosit@outlook.hu>
Date: Sat, 27 Sep 2025 01:43:55 +0200
Subject: [PATCH] 4096-wise thiersort2

---
 thiersort2.h | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/thiersort2.h b/thiersort2.h
index d22e711..bcc93d7 100644
--- a/thiersort2.h
+++ b/thiersort2.h
@@ -42,7 +42,7 @@ static inline uint32_t witch_bucket(uint32_t key) {
 
 	uint32_t witch_base = (key <= 2) ? 0 : (as.u >> 23) - 128; // 0, [127..159] -> [0..31]
 	// return witch_base * 8 + ((as.u >> (23 - 3)) & 7); // 0..255
-	return witch_base * 64 + ((as.u >> (23 - 6)) & 63); // 0..2047
+	return witch_base * 128 + ((as.u >> (23 - 7)) & 127); // 0..2047
 	/* Alternative (but I measure it being worse):
 	return (as.u >> 23);
 	*/
@@ -57,8 +57,8 @@ static inline uint32_t witch_bucket(uint32_t key) {
  * @param rstate Create with sch_rand_state rstate = schwab_rand_state(junk_uint32_t);
  */
 static inline void thiersort2(uint32_t *arr, uint32_t *temparr, int n, sch_rand_state *rstate) {
-	int bucket[2048]; /* Inclusive */
-	int bucket_end[2048]; /* Not inclusive */
+	int bucket[4096]; /* Inclusive */
+	int bucket_end[4096]; /* Not inclusive */
 
 	/* Check if need to sort at all - needed for invariants later */
 	if(n < 2) {
@@ -67,7 +67,7 @@ static inline void thiersort2(uint32_t *arr, uint32_t *temparr, int n, sch_rand_
 
 	/* Count */
 	#pragma GCC unroll 64
-	for(int i = 0; i < 2048; ++i) {
+	for(int i = 0; i < 4096; ++i) {
 		bucket[i] = 0;
 	}
 	#pragma GCC unroll 64
@@ -78,14 +78,14 @@ static inline void thiersort2(uint32_t *arr, uint32_t *temparr, int n, sch_rand_
 	/* Prefix sum (like in Magyarsort) */
 	uint32_t prev = 0;
 	#pragma GCC unroll 4
-	for (int i = 0; i < 2048; i++) {
+	for (int i = 0; i < 4096; i++) {
 		bucket[i] += prev;
 		prev = bucket[i];
 	}
 
 	/* Save end-offsets */
 	#pragma GCC unroll 64
-	for(int i = 0; i < 2048; ++i) {
+	for(int i = 0; i < 4096; ++i) {
 		bucket_end[i] = bucket[i];
 	}
 
@@ -102,7 +102,7 @@ static inline void thiersort2(uint32_t *arr, uint32_t *temparr, int n, sch_rand_
 
 	/* temparr -> arr each bucket and sort them in-place */
 	#pragma GCC unroll 64
-	for(int b = 0; b < 2048; ++b) {
+	for(int b = 0; b < 4096; ++b) {
 		int begin = bucket[b];
 		int end = bucket_end[b];