From 85aaf4b1a10745f87aed916cc3563cab81d684f5 Mon Sep 17 00:00:00 2001
From: Richard Thier <magosit@outlook.hu>
Date: Fri, 9 May 2025 01:10:12 +0200
Subject: [PATCH] testing schwab_sort

---
 top2_partition.c | 83 ++++++++++++++++++++++++++++++++++++++++++++++++
 ypsu.cpp         | 15 ++++++++-
 2 files changed, 97 insertions(+), 1 deletion(-)
 create mode 100644 top2_partition.c

diff --git a/top2_partition.c b/top2_partition.c
new file mode 100644
index 0000000..6eb4699
--- /dev/null
+++ b/top2_partition.c
@@ -0,0 +1,83 @@
+#include <stdio.h>
+#include <stdint.h>
+
+#define TOP2(x) ((x) >> 30)
+
+void swap(uint32_t* a, uint32_t* b) {
+    uint32_t tmp = *a;
+    *a = *b;
+    *b = tmp;
+}
+
+// TODO: instead of swaps, we need a single tmp var + only memcpys!!!
+// TODO: we can do "ILP-memcpy": key from b2->b3, value from b2->b3, key from b1->b2, value from b1...
+// Rem.: The latter is faster likely than calling memcpy function even though its simd-optimized...
+void partition_top2(uint32_t* arr, int n) {
+    int b0 = 0, b1 = 0, b2 = 0, b3 = 0;
+
+    while (b3 < n) {
+        uint32_t val = arr[b3];
+        uint32_t top = TOP2(val);
+
+        if (top == 0) {
+            // Cascade: swap into b2, b1, b0
+            swap(&arr[b3], &arr[b2]);
+            swap(&arr[b2], &arr[b1]);
+            swap(&arr[b1], &arr[b0]);
+            b0++; b1++; b2++;
+        } else if (top == 1) {
+            // Cascade: swap into b2, b1
+            swap(&arr[b3], &arr[b2]);
+            swap(&arr[b2], &arr[b1]);
+            b1++; b2++;
+        } else if (top == 2) {
+            // Swap into b2
+            swap(&arr[b3], &arr[b2]);
+            b2++;
+        }
+        // else (top == 3), do nothing
+
+        b3++;
+    }
+}
+
+int main() {
+    uint32_t arr[] = {
+        0x40000001, // top 2 bits = 01
+        0x00000002, // 00
+        0xC0000003, // 11
+        0x80000004, // 10
+        0x00000005, // 00
+        0x40000006, // 01
+        0xC0000007, // 11
+        0x80000008, // 10
+    };
+
+    int n = sizeof(arr) / sizeof(arr[0]);
+
+    printf("Before:\n");
+    for (int i = 0; i < n; ++i)
+        printf("0x%08X ", arr[i]);
+    printf("\n");
+
+    // Optional: Show top 2 bits for verification
+    printf("\nTop 2 bits:\n");
+    for (int i = 0; i < n; ++i)
+        printf("%u ", TOP2(arr[i]));
+    printf("\n");
+
+    partition_top2(arr, n);
+
+    printf("\nAfter:\n");
+    for (int i = 0; i < n; ++i)
+        printf("0x%08X ", arr[i]);
+    printf("\n");
+
+    // Optional: Show top 2 bits for verification
+    printf("\nTop 2 bits:\n");
+    for (int i = 0; i < n; ++i)
+        printf("%u ", TOP2(arr[i]));
+    printf("\n");
+
+    return 0;
+}
diff --git a/ypsu.cpp b/ypsu.cpp
index a1fd6f2..880d40c 100644
--- a/ypsu.cpp
+++ b/ypsu.cpp
@@ -18,6 +18,7 @@
 #include "thiersort.h"
 #include "qsort/qsort.h"
 #include "qsort/zssort.h"
+#include "qsort/schwab_sort.h"
 #include "qsort/chatgpt_qs.h"
 
 // #define MAGYAR_SORT_DEFAULT_REUSE
@@ -199,6 +200,14 @@ static inline void do_neoqs(uint32_t *a, int n) noexcept {
 	neoqs(a, 0, n - 1, &state);
 }
 
+/** schwab */
+static inline void do_schwab(uint32_t *a, int n) noexcept {
+	assert(n * uint32_t(sizeof(a[0])) <= INT_MAX);
+	uint32_t junk;
+	sch_rand_state state = schwab_rand_state(junk);
+	schwab_sort(a, 0, n - 1, &state);
+}
+
 // mormord — Today at 2:27 AM
 // 1 2 2 2 3
 // 
@@ -924,10 +933,10 @@ int main(void) {
     w = v;
     measure(inputtype, "zsr", [&] { do_zsr(&w[0], w.size()); });
     assert(w == expected);
-    */
     w = v;
     measure(inputtype, "qsr3", [&] { do_qsr3(&w[0], w.size()); });
     assert(w == expected);
+    */
     w = v;
     measure(inputtype, "zsr3", [&] { do_zsr3(&w[0], w.size()); });
     assert(w == expected);
@@ -950,6 +959,10 @@ int main(void) {
     measure(inputtype, "neoqs", [&] { do_neoqs(&w[0], w.size()); });
     assert(w == expected);
 
+    w = v;
+    measure(inputtype, "schwab", [&] { do_schwab(&w[0], w.size()); });
+    assert(w == expected);
+
     /*
     w = v;
     measure(inputtype, "magbuck", [&] { magyar_bucket_sort(&w[0], w.size()); });