From d84f2b23074dd6faf4a8e4c49e033204a6464808 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Branimir=20Karad=C5=BEi=C4=87?= <branimirkaradzic@gmail.com>
Date: Thu, 15 Oct 2015 16:35:21 -0700
Subject: [PATCH] Added radix sort just for keys.

---
 include/bx/radixsort.h | 126 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 124 insertions(+), 2 deletions(-)

diff --git a/include/bx/radixsort.h b/include/bx/radixsort.h
index 9ed4011..d28f651 100644
--- a/include/bx/radixsort.h
+++ b/include/bx/radixsort.h
@@ -14,8 +14,69 @@ namespace bx
 #define BX_RADIXSORT_HISTOGRAM_SIZE (1<<BX_RADIXSORT_BITS)
 #define BX_RADIXSORT_BIT_MASK (BX_RADIXSORT_HISTOGRAM_SIZE-1)
 
+	inline void radixSort32(uint32_t* __restrict _keys, uint32_t* __restrict _tempKeys, uint32_t _size)
+	{
+		uint32_t* __restrict keys = _keys;
+		uint32_t* __restrict tempKeys = _tempKeys;
+
+		uint32_t histogram[BX_RADIXSORT_HISTOGRAM_SIZE];
+		uint16_t shift = 0;
+		uint32_t pass = 0;
+		for (; pass < 3; ++pass)
+		{
+			memset(histogram, 0, sizeof(uint32_t)*BX_RADIXSORT_HISTOGRAM_SIZE);
+
+			bool sorted = true;
+			{
+				uint32_t key = keys[0];
+				uint32_t prevKey = key;
+				for (uint32_t ii = 0; ii < _size; ++ii, prevKey = key)
+				{
+					key = keys[ii];
+					uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
+					++histogram[index];
+					sorted &= prevKey <= key;
+				}
+			}
+
+			if (sorted)
+			{
+				goto done;
+			}
+
+			uint32_t offset = 0;
+			for (uint32_t ii = 0; ii < BX_RADIXSORT_HISTOGRAM_SIZE; ++ii)
+			{
+				uint32_t count = histogram[ii];
+				histogram[ii] = offset;
+				offset += count;
+			}
+
+			for (uint32_t ii = 0; ii < _size; ++ii)
+			{
+				uint32_t key = keys[ii];
+				uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
+				uint32_t dest = histogram[index]++;
+				tempKeys[dest] = key;
+			}
+
+			uint32_t* swapKeys = tempKeys;
+			tempKeys = keys;
+			keys = swapKeys;
+
+			shift += BX_RADIXSORT_BITS;
+		}
+
+done:
+		if (0 != (pass&1) )
+		{
+			// Odd number of passes needs to do copy to the destination.
+			memcpy(_keys, _tempKeys, _size*sizeof(uint32_t) );
+		}
+	}
+
 	template <typename Ty>
-	void radixSort32(uint32_t* __restrict _keys, uint32_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
+	inline void radixSort32(uint32_t* __restrict _keys, uint32_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
 	{
 		uint32_t* __restrict keys = _keys;
 		uint32_t* __restrict tempKeys = _tempKeys;
@@ -87,8 +148,69 @@ done:
 		}
 	}
 
+	inline void radixSort64(uint64_t* __restrict _keys, uint64_t* __restrict _tempKeys, uint32_t _size)
+	{
+		uint64_t* __restrict keys = _keys;
+		uint64_t* __restrict tempKeys = _tempKeys;
+
+		uint32_t histogram[BX_RADIXSORT_HISTOGRAM_SIZE];
+		uint16_t shift = 0;
+		uint32_t pass = 0;
+		for (; pass < 6; ++pass)
+		{
+			memset(histogram, 0, sizeof(uint32_t)*BX_RADIXSORT_HISTOGRAM_SIZE);
+
+			bool sorted = true;
+			{
+				uint64_t key = keys[0];
+				uint64_t prevKey = key;
+				for (uint32_t ii = 0; ii < _size; ++ii, prevKey = key)
+				{
+					key = keys[ii];
+					uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
+					++histogram[index];
+					sorted &= prevKey <= key;
+				}
+			}
+
+			if (sorted)
+			{
+				goto done;
+			}
+
+			uint32_t offset = 0;
+			for (uint32_t ii = 0; ii < BX_RADIXSORT_HISTOGRAM_SIZE; ++ii)
+			{
+				uint32_t count = histogram[ii];
+				histogram[ii] = offset;
+				offset += count;
+			}
+
+			for (uint32_t ii = 0; ii < _size; ++ii)
+			{
+				uint64_t key = keys[ii];
+				uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
+				uint32_t dest = histogram[index]++;
+				tempKeys[dest] = key;
+			}
+
+			uint64_t* swapKeys = tempKeys;
+			tempKeys = keys;
+			keys = swapKeys;
+
+			shift += BX_RADIXSORT_BITS;
+		}
+
+done:
+		if (0 != (pass&1) )
+		{
+			// Odd number of passes needs to do copy to the destination.
+			memcpy(_keys, _tempKeys, _size*sizeof(uint64_t) );
+		}
+	}
+
 	template <typename Ty>
-	void radixSort64(uint64_t* __restrict _keys, uint64_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
+	inline void radixSort64(uint64_t* __restrict _keys, uint64_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
 	{
 		uint64_t* __restrict keys = _keys;
 		uint64_t* __restrict tempKeys = _tempKeys;