From ae0a749d08cf598b7d5978eb14a3b0a2717c62ba Mon Sep 17 00:00:00 2001
From: Ahmet Inan <inan@aicodix.de>
Date: Tue, 5 Mar 2024 11:46:33 +0100
Subject: [PATCH] added insertion sort and regression test

---
 sort.hh                       | 119 +++++++---------------------------
 tests/sort_regression_test.cc |  44 +++++++++++++
 2 files changed, 66 insertions(+), 97 deletions(-)
 create mode 100644 tests/sort_regression_test.cc

diff --git a/sort.hh b/sort.hh
index c36623d..1303658 100644
--- a/sort.hh
+++ b/sort.hh
@@ -6,107 +6,32 @@ Copyright 2024 Ahmet Inan <inan@aicodix.de>
 
 #pragma once
 
-#include <algorithm>
-
-template <int WIDTH>
-static inline SIMD<uint32_t, WIDTH> vorder(SIMD<float, WIDTH> a)
+template <typename TYPE, int WIDTH>
+static inline SIMD<typename SIMD<TYPE, WIDTH>::uint_type, WIDTH> vorder(SIMD<TYPE, WIDTH> a)
 {
-	SIMD<uint32_t, WIDTH> tmp;
-	for (int i = 0; i < WIDTH; ++i)
-		tmp.v[i] = i;
-	std::stable_sort(tmp.v, tmp.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
-	return tmp;
+	SIMD<typename SIMD<TYPE, WIDTH>::uint_type, WIDTH> p;
+	p.v[0] = 0;
+	for (int i = 1, j; i < WIDTH; ++i) {
+		TYPE t = a.v[i];
+		for (j = i; j > 0 && a.v[j-1] > t; --j) {
+			a.v[j] = a.v[j-1];
+			p.v[j] = p.v[j-1];
+		}
+		a.v[j] = t;
+		p.v[j] = i;
+	}
+	return p;
 }
 
-template <int WIDTH>
-static inline SIMD<uint64_t, WIDTH> vorder(SIMD<double, WIDTH> a)
+template <typename TYPE, int WIDTH>
+static inline SIMD<TYPE, WIDTH> vsort(SIMD<TYPE, WIDTH> a)
 {
-	SIMD<uint64_t, WIDTH> tmp;
-	for (int i = 0; i < WIDTH; ++i)
-		tmp.v[i] = i;
-	std::stable_sort(tmp.v, tmp.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
-	return tmp;
-}
-
-template <int WIDTH>
-static inline SIMD<uint8_t, WIDTH> vorder(SIMD<int8_t, WIDTH> a)
-{
-	SIMD<uint8_t, WIDTH> tmp;
-	for (int i = 0; i < WIDTH; ++i)
-		tmp.v[i] = i;
-	std::stable_sort(tmp.v, tmp.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
-	return tmp;
-}
-
-template <int WIDTH>
-static inline SIMD<uint16_t, WIDTH> vorder(SIMD<int16_t, WIDTH> a)
-{
-	SIMD<uint16_t, WIDTH> tmp;
-	for (int i = 0; i < WIDTH; ++i)
-		tmp.v[i] = i;
-	std::stable_sort(tmp.v, tmp.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
-	return tmp;
-}
-
-template <int WIDTH>
-static inline SIMD<uint32_t, WIDTH> vorder(SIMD<int32_t, WIDTH> a)
-{
-	SIMD<uint32_t, WIDTH> tmp;
-	for (int i = 0; i < WIDTH; ++i)
-		tmp.v[i] = i;
-	std::stable_sort(tmp.v, tmp.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
-	return tmp;
-}
-
-template <int WIDTH>
-static inline SIMD<uint64_t, WIDTH> vorder(SIMD<int64_t, WIDTH> a)
-{
-	SIMD<uint64_t, WIDTH> tmp;
-	for (int i = 0; i < WIDTH; ++i)
-		tmp.v[i] = i;
-	std::stable_sort(tmp.v, tmp.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
-	return tmp;
-}
-
-template <int WIDTH>
-static inline SIMD<float, WIDTH> vsort(SIMD<float, WIDTH> a)
-{
-	std::sort(a.v, a.v+WIDTH);
-	return a;
-}
-
-template <int WIDTH>
-static inline SIMD<double, WIDTH> vsort(SIMD<double, WIDTH> a)
-{
-	std::sort(a.v, a.v+WIDTH);
-	return a;
-}
-
-template <int WIDTH>
-static inline SIMD<int8_t, WIDTH> vsort(SIMD<int8_t, WIDTH> a)
-{
-	std::sort(a.v, a.v+WIDTH);
-	return a;
-}
-
-template <int WIDTH>
-static inline SIMD<int16_t, WIDTH> vsort(SIMD<int16_t, WIDTH> a)
-{
-	std::sort(a.v, a.v+WIDTH);
-	return a;
-}
-
-template <int WIDTH>
-static inline SIMD<int32_t, WIDTH> vsort(SIMD<int32_t, WIDTH> a)
-{
-	std::sort(a.v, a.v+WIDTH);
-	return a;
-}
-
-template <int WIDTH>
-static inline SIMD<int64_t, WIDTH> vsort(SIMD<int64_t, WIDTH> a)
-{
-	std::sort(a.v, a.v+WIDTH);
+	for (int i = 1, j; i < WIDTH; ++i) {
+		TYPE t = a.v[i];
+		for (j = i; j > 0 && a.v[j-1] > t; --j)
+			a.v[j] = a.v[j-1];
+		a.v[j] = t;
+	}
 	return a;
 }
 
diff --git a/tests/sort_regression_test.cc b/tests/sort_regression_test.cc
new file mode 100644
index 0000000..1b18a2b
--- /dev/null
+++ b/tests/sort_regression_test.cc
@@ -0,0 +1,44 @@
+/*
+Regression Test for the SIMD wrappers sort
+
+Copyright 2024 Ahmet Inan <inan@aicodix.de>
+*/
+
+#include <random>
+#include <cassert>
+#include <iostream>
+#include <algorithm>
+#include <functional>
+#include "simd.hh"
+#include "sort.hh"
+
+int main()
+{
+	const int WIDTH = 32;
+	typedef int32_t TYPE;
+	typedef SIMD<TYPE, WIDTH>::uint_type UINT;
+	std::random_device rd;
+	typedef std::default_random_engine generator;
+	typedef std::uniform_int_distribution<int> distribution;
+	auto rand = std::bind(distribution(0, WIDTH), generator(rd()));
+	for (int loop = 0; loop < 1000000; ++loop) {
+		SIMD<TYPE, WIDTH> a, b, c;
+		for (int i = 0; i < WIDTH; ++i)
+			a.v[i] = rand();
+		b = vsort(a);
+		c = a;
+		std::sort(c.v, c.v+WIDTH);
+		for (int i = 0; i < WIDTH; ++i)
+			assert(b.v[i] == c.v[i]);
+		SIMD<UINT, WIDTH> d, e;
+		d = vorder(a);
+		for (int i = 0; i < WIDTH; ++i)
+			e.v[i] = i;
+		std::stable_sort(e.v, e.v+WIDTH, [a](int i, int j){ return a.v[i] < a.v[j]; });
+		for (int i = 0; i < WIDTH; ++i)
+			assert(d.v[i] == e.v[i]);
+	}
+	std::cerr << "Sort regression test passed!" << std::endl;
+	return 0;
+}
+