Optimized dot SIMD implementation with SSE3

glm/detail/func_geometric_simd.inl

@@ -9,10 +9,7 @@ namespace detail
 		GLM_FUNC_QUALIFIER static float call(tvec4<float, P> const& x, tvec4<float, P> const& y)
 		{
 			__m128 const dot0 = glm_dot_ss(x.data, y.data);
-
+			return _mm_cvtss_f32(dot0);
-			float Result = 0;
-			_mm_store_ss(&Result, dot0);
-			return Result;
 		}
 	};
 }//namespace detail

glm/simd/common.h

@@ -3,26 +3,6 @@
 
 #pragma once
 
-#if(GLM_COMPILER & GLM_COMPILER_VC)
-#pragma warning(push)
-#pragma warning(disable : 4510 4512 4610)
-#endif
-
-	union ieee754_QNAN
-	{
-		const float f;
-		struct i
-		{
-			const unsigned int mantissa:23, exp:8, sign:1;
-		};
-
-		ieee754_QNAN() : f(0.0)/*, mantissa(0x7FFFFF), exp(0xFF), sign(0x0)*/ {}
-	};
-
-#if(GLM_COMPILER & GLM_COMPILER_VC)
-#pragma warning(pop)
-#endif
-
 static const __m128 GLM_VAR_USED glm_zero = _mm_setzero_ps();
 static const __m128 GLM_VAR_USED glm_one = _mm_set_ps1(1.0f);
 static const __m128 GLM_VAR_USED glm_half = _mm_set_ps1(0.5f);
@@ -30,15 +10,13 @@ static const __m128 GLM_VAR_USED glm_minus_one = _mm_set_ps1(-1.0f);
 static const __m128 GLM_VAR_USED glm_two = _mm_set_ps1(2.0f);
 static const __m128 GLM_VAR_USED glm_three = _mm_set_ps1(3.0f);
 
-static const ieee754_QNAN glm_abs_mask;
-static const __m128 GLM_VAR_USED glm_abs4_mask = _mm_set_ps1(glm_abs_mask.f);
 static const __m128 GLM_VAR_USED glm_epi32_sign_mask = _mm_castsi128_ps(_mm_set1_epi32(static_cast<int>(0x80000000)));
 static const __m128 GLM_VAR_USED glm_ps_2pow23 = _mm_set_ps1(8388608.0f);
 static const __m128 GLM_VAR_USED glm_ps_1 = _mm_set_ps1(1.0f);
 
 GLM_FUNC_QUALIFIER __m128 glm_abs_ps(__m128 x)
 {
-	return _mm_and_ps(glm_abs4_mask, x);
+	return _mm_and_ps(x, _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF)));
 }
 
 //sign

glm/simd/geometric.h

@@ -9,6 +9,11 @@ GLM_FUNC_QUALIFIER __m128 glm_dot_ps(__m128 v1, __m128 v2)
 {
 #	if GLM_ARCH & GLM_ARCH_AVX
 		return _mm_dp_ps(v1, v2, 0xff);
+#	elif GLM_ARCH & GLM_ARCH_SSE3
+		__m128 const Mul0 = _mm_mul_ps(v1, v2);
+		__m128 const Hadd0 = _mm_hadd_ps(Mul0, Mul0);
+		__m128 const Hadd1 = _mm_hadd_ps(Hadd0, Hadd0);
+		return Hadd1;
 #	else
 		__m128 const mul0 = _mm_mul_ps(v1, v2);
 		__m128 const swp0 = _mm_shuffle_ps(mul0, mul0, _MM_SHUFFLE(2, 3, 0, 1));
@@ -21,12 +26,21 @@ GLM_FUNC_QUALIFIER __m128 glm_dot_ps(__m128 v1, __m128 v2)
 
 GLM_FUNC_QUALIFIER __m128 glm_dot_ss(__m128 v1, __m128 v2)
 {
-	__m128 const mul0 = _mm_mul_ps(v1, v2);
+#	if GLM_ARCH & GLM_ARCH_AVX
-	__m128 const mov0 = _mm_movehl_ps(mul0, mul0);
+		return _mm_dp_ps(v1, v2, 0xff);
-	__m128 const add0 = _mm_add_ps(mov0, mul0);
+#	elif GLM_ARCH & GLM_ARCH_SSE3
-	__m128 const swp1 = _mm_shuffle_ps(add0, add0, 1);
+		__m128 const Mul0 = _mm_mul_ps(v1, v2);
-	__m128 const add1 = _mm_add_ss(add0, swp1);
+		__m128 const Hadd0 = _mm_hadd_ps(Mul0, Mul0);
-	return add1;
+		__m128 const Hadd1 = _mm_hadd_ps(Hadd0, Hadd0);
+		return Hadd1;
+#	else
+		__m128 const mul0 = _mm_mul_ps(v1, v2);
+		__m128 const mov0 = _mm_movehl_ps(mul0, mul0);
+		__m128 const add0 = _mm_add_ps(mov0, mul0);
+		__m128 const swp1 = _mm_shuffle_ps(add0, add0, 1);
+		__m128 const add1 = _mm_add_ss(add0, swp1);
+		return add1;
+#	endif
 }
 
 GLM_FUNC_QUALIFIER __m128 glm_len_ps(__m128 x)