VamuveTV · July 22, 2020 12:14 · Apr 20, 2015 · Apr 20, 2015 · Apr 20, 2015 · Apr 20, 2015
diff --git a/SSE Sort b/SSE Sort
@@ -1,19 +1,8 @@
-static inline __m128 vselect(const __m128 a, const __m128 b, const __m128 mask)
-{
-	return _mm_or_ps(_mm_and_ps(mask, a), _mm_andnot_ps(mask, b));
-}
-
-static inline __m128i vselect(const __m128i a, const __m128i b, const __m128i mask)
-{
-	return _mm_or_si128(_mm_and_si128(mask, a), _mm_andnot_si128(mask, b));
-}
-
 /*
 Sort 4 floats in SSE vector using sorting network and return indices of moved values
 */
 inline __m128i v4_sort(__m128& v)
 {
-	// @todo Replace by constants?
 	auto i = _mm_castsi128_ps(_mm_set_epi32(3, 2, 1, 0));
 	const auto mask = _mm_castsi128_ps(_mm_set1_epi32(0xFFFFFFFC));
 
@@ -26,24 +15,26 @@ inline __m128i v4_sort(__m128& v)
 	auto cmp = _mm_cmplt_ps(v, temp);
 
 	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(1, 0, 1, 0));
-	auto temp2 = vselect(v, temp, cmp);
+	auto temp2 = _mm_blendv_ps(temp, v, cmp);
 
 	// Second pass
 	temp = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(2, 3, 0, 1));
 	cmp = _mm_cmplt_ps(temp2, temp);
 
 	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(2, 0, 2, 0));
 
-	temp2 = vselect(temp2, temp, cmp);
+	temp2 = _mm_blendv_ps(temp, temp2, cmp);
 
 	// Third pass
 	temp = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(3, 1, 2, 0));
 	cmp = _mm_cmplt_ps(temp2, temp);
 
 	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(3, 1, 1, 0));
 
-	v = vselect(temp2, temp, cmp);
+	v = _mm_blendv_ps(temp, temp2, cmp);
 
 	// Remove indices from input and return them
-	return _mm_castps_si128(_mm_andnot_ps(mask, v));
+	auto ret = _mm_castps_si128(_mm_andnot_ps(mask, v));
+
+	return ret;
 }
diff --git a/SSE Sort b/SSE Sort
@@ -13,6 +13,7 @@ Sort 4 floats in SSE vector using sorting network and return indices of moved va
 */
 inline __m128i v4_sort(__m128& v)
 {
+	// @todo Replace by constants?
 	auto i = _mm_castsi128_ps(_mm_set_epi32(3, 2, 1, 0));
 	const auto mask = _mm_castsi128_ps(_mm_set1_epi32(0xFFFFFFFC));
 
@@ -44,8 +45,5 @@ inline __m128i v4_sort(__m128& v)
 	v = vselect(temp2, temp, cmp);
 
 	// Remove indices from input and return them
-	auto ret = _mm_castps_si128(_mm_andnot_ps(mask, v));
-	v = _mm_and_ps(mask, v);
-
-	return ret;
+	return _mm_castps_si128(_mm_andnot_ps(mask, v));
 }
diff --git a/SSE Sort b/SSE Sort
@@ -9,38 +9,43 @@ static inline __m128i vselect(const __m128i a, const __m128i b, const __m128i ma
 }
 
 /*
-  Sort 4 floats in SSE vector using sorting network and return indices of moved values
+Sort 4 floats in SSE vector using sorting network and return indices of moved values
 */
 inline __m128i v4_sort(__m128& v)
 {
-	auto i = _mm_set_epi32(3, 2, 1, 0);
-
+	auto i = _mm_castsi128_ps(_mm_set_epi32(3, 2, 1, 0));
+	const auto mask = _mm_castsi128_ps(_mm_set1_epi32(0xFFFFFFFC));
+
+	// Place indices in lower 2 bits of mantissa
+	v = _mm_or_ps(_mm_and_ps(v, mask), i);
+
 	// Simple sorting network for n=4
 	// First pass
 	auto temp = _mm_shuffle_ps(v, v, _MM_SHUFFLE(1, 0, 3, 2));
 	auto cmp = _mm_cmplt_ps(v, temp);
-	
+
 	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(1, 0, 1, 0));
 	auto temp2 = vselect(v, temp, cmp);
-	// Swap indices
-	i = vselect(i, _mm_shuffle_epi32(i, _MM_SHUFFLE(1, 0, 3, 2)), _mm_castps_si128(cmp));
-
+
 	// Second pass
 	temp = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(2, 3, 0, 1));
 	cmp = _mm_cmplt_ps(temp2, temp);
-	
+
 	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(2, 0, 2, 0));
-	
+
 	temp2 = vselect(temp2, temp, cmp);
-	i = vselect(i, _mm_shuffle_epi32(i, _MM_SHUFFLE(2, 3, 0, 1)), _mm_castps_si128(cmp));
-
+
 	// Third pass
 	temp = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(3, 1, 2, 0));
 	cmp = _mm_cmplt_ps(temp2, temp);
-	
+
 	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(3, 1, 1, 0));
-	
+
 	v = vselect(temp2, temp, cmp);
-
-	return vselect(i, _mm_shuffle_epi32(i, _MM_SHUFFLE(3, 1, 2, 0)), _mm_castps_si128(cmp));
+
+	// Remove indices from input and return them
+	auto ret = _mm_castps_si128(_mm_andnot_ps(mask, v));
+	v = _mm_and_ps(mask, v);
+
+	return ret;
 }
diff --git a/SSE Sort b/SSE Sort
@@ -0,0 +1,46 @@
+static inline __m128 vselect(const __m128 a, const __m128 b, const __m128 mask)
+{
+	return _mm_or_ps(_mm_and_ps(mask, a), _mm_andnot_ps(mask, b));
+}
+
+static inline __m128i vselect(const __m128i a, const __m128i b, const __m128i mask)
+{
+	return _mm_or_si128(_mm_and_si128(mask, a), _mm_andnot_si128(mask, b));
+}
+
+/*
+  Sort 4 floats in SSE vector using sorting network and return indices of moved values
+*/
+inline __m128i v4_sort(__m128& v)
+{
+	auto i = _mm_set_epi32(3, 2, 1, 0);
+
+	// Simple sorting network for n=4
+	// First pass
+	auto temp = _mm_shuffle_ps(v, v, _MM_SHUFFLE(1, 0, 3, 2));
+	auto cmp = _mm_cmplt_ps(v, temp);
+
+	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(1, 0, 1, 0));
+	auto temp2 = vselect(v, temp, cmp);
+	// Swap indices
+	i = vselect(i, _mm_shuffle_epi32(i, _MM_SHUFFLE(1, 0, 3, 2)), _mm_castps_si128(cmp));
+
+	// Second pass
+	temp = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(2, 3, 0, 1));
+	cmp = _mm_cmplt_ps(temp2, temp);
+
+	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(2, 0, 2, 0));
+
+	temp2 = vselect(temp2, temp, cmp);
+	i = vselect(i, _mm_shuffle_epi32(i, _MM_SHUFFLE(2, 3, 0, 1)), _mm_castps_si128(cmp));
+
+	// Third pass
+	temp = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(3, 1, 2, 0));
+	cmp = _mm_cmplt_ps(temp2, temp);
+
+	cmp = _mm_shuffle_ps(cmp, cmp, _MM_SHUFFLE(3, 1, 1, 0));
+
+	v = vselect(temp2, temp, cmp);
+
+	return vselect(i, _mm_shuffle_epi32(i, _MM_SHUFFLE(3, 1, 2, 0)), _mm_castps_si128(cmp));
+}