From 20cf68679c8b8bbbbb053cfb1b90115af1d28b56 Mon Sep 17 00:00:00 2001
From: Christophe Riccio <christopheri@unity3d.com>
Date: Sun, 29 May 2016 02:57:53 +0200
Subject: [PATCH] SIMD sqrt optimizations including lowp

---
 glm/detail/func_exponential.inl      | 11 +++++++-
 glm/detail/func_exponential_simd.inl | 26 ++++++++++++++++++
 glm/detail/func_geometric_simd.inl   |  3 +++
 glm/simd/common.h                    | 26 ------------------
 glm/simd/exponential.h               | 17 ++++++++++++
 test/core/core_func_exponential.cpp  | 40 +++++++++-------------------
 6 files changed, 68 insertions(+), 55 deletions(-)

diff --git a/glm/detail/func_exponential.inl b/glm/detail/func_exponential.inl
index 5cec61c7..d597c55b 100644
--- a/glm/detail/func_exponential.inl
+++ b/glm/detail/func_exponential.inl
@@ -29,6 +29,15 @@ namespace detail
 		}
 	};
 
+	template <template <class, precision> class vecType, typename T, precision P>
+	struct compute_sqrt
+	{
+		GLM_FUNC_QUALIFIER static vecType<T, P> call(vecType<T, P> const & x)
+		{
+			return detail::functor1<T, T, P, vecType>::call(std::sqrt, x);
+		}
+	};
+
 	template <template <class, precision> class vecType, typename T, precision P>
 	struct compute_inversesqrt
 	{
@@ -113,7 +122,7 @@ namespace detail
 	GLM_FUNC_QUALIFIER vecType<T, P> sqrt(vecType<T, P> const & x)
 	{
 		GLM_STATIC_ASSERT(std::numeric_limits<T>::is_iec559, "'sqrt' only accept floating-point inputs");
-		return detail::functor1<T, T, P, vecType>::call(sqrt, x);
+		return detail::compute_sqrt<vecType, T, P>::call(x);
 	}
 
 	// inversesqrt
diff --git a/glm/detail/func_exponential_simd.inl b/glm/detail/func_exponential_simd.inl
index 7358c15f..9f593dea 100644
--- a/glm/detail/func_exponential_simd.inl
+++ b/glm/detail/func_exponential_simd.inl
@@ -1,9 +1,35 @@
 /// @ref core
 /// @file glm/detail/func_exponential_simd.inl
 
+#include "../simd/exponential.h"
+
+#if GLM_ARCH & GLM_ARCH_SSE2_BIT
+
 namespace glm{
 namespace detail
 {
+	template <precision P>
+	struct compute_sqrt<tvec4, float, P>
+	{
+		GLM_FUNC_QUALIFIER static tvec4<float, P> call(tvec4<float, P> const & v)
+		{
+			tvec4<float, P> result(uninitialize);
+			result.data = _mm_sqrt_ps(v.data);
+			return result;
+		}
+	};
 
+	template <>
+	struct compute_sqrt<tvec4, float, lowp>
+	{
+		GLM_FUNC_QUALIFIER static tvec4<float, lowp> call(tvec4<float, lowp> const & v)
+		{
+			tvec4<float, lowp> result(uninitialize);
+			result.data = glm_f32v4_sqrt_lowp(v.data);
+			return result;
+		}
+	};
 }//namespace detail
 }//namespace glm
+
+#endif//GLM_ARCH & GLM_ARCH_SSE2_BIT
diff --git a/glm/detail/func_geometric_simd.inl b/glm/detail/func_geometric_simd.inl
index e2e4d4ef..c99276f9 100644
--- a/glm/detail/func_geometric_simd.inl
+++ b/glm/detail/func_geometric_simd.inl
@@ -1,3 +1,6 @@
+/// @ref core
+/// @file glm/detail/func_geometric_simd.inl
+
 #include "../simd/geometric.h"
 
 #if GLM_ARCH & GLM_ARCH_SSE2_BIT
diff --git a/glm/simd/common.h b/glm/simd/common.h
index 6bd82499..90cb5cf4 100644
--- a/glm/simd/common.h
+++ b/glm/simd/common.h
@@ -178,30 +178,4 @@ GLM_FUNC_QUALIFIER __m128 glm_f32v4_inf(__m128 x)
 	return _mm_castsi128_ps(_mm_cmpeq_epi32(t2, _mm_set1_epi32(0xFF000000)));		// exponent is all 1s, fraction is 0
 }
 
-// SSE scalar reciprocal sqrt using rsqrt op, plus one Newton-Rhaphson iteration
-// By Elan Ruskin, http://assemblyrequired.crashworks.org/
-GLM_FUNC_QUALIFIER __m128 glm_f32v1_sqrt_wip(__m128 x)
-{
-	__m128 const Rcp0 = _mm_rsqrt_ss(x);  // "estimate" opcode
-	__m128 const Mul0 = _mm_mul_ss(_mm_set1_ps(0.5f), Rcp0);
-	__m128 const Mul1 = _mm_mul_ss(Rcp0, Rcp0);
-	__m128 const Mul2 = _mm_mul_ss(x, Mul1);
-	__m128 const Sub0 = _mm_sub_ss(_mm_set1_ps(3.0f), Mul2);
-	__m128 const Mul3 = _mm_mul_ss(Mul0, Sub0);
-	return Mul3;
-}
-
-// SSE scalar reciprocal sqrt using rsqrt op, plus one Newton-Rhaphson iteration
-// By Elan Ruskin, http://assemblyrequired.crashworks.org/
-GLM_FUNC_QUALIFIER __m128 glm_f32v4_sqrt_wip(__m128 x)
-{
-	__m128 const Rcp0 = _mm_rsqrt_ps(x);  // "estimate" opcode
-	__m128 const Mul0 = _mm_mul_ps(_mm_set1_ps(0.5f), Rcp0);
-	__m128 const Mul1 = _mm_mul_ps(Mul0, Mul0);
-	__m128 const Mul2 = _mm_mul_ps(x, Mul1);
-	__m128 const Sub0 = _mm_sub_ps(_mm_set1_ps(3.0f), Mul2);
-	__m128 const Mul3 = _mm_mul_ps(Mul0, Sub0);
-	return Mul3;
-}
-
 #endif//GLM_ARCH & GLM_ARCH_SSE2_BIT
diff --git a/glm/simd/exponential.h b/glm/simd/exponential.h
index 49be6cb4..bc7f8cf3 100644
--- a/glm/simd/exponential.h
+++ b/glm/simd/exponential.h
@@ -3,3 +3,20 @@
 
 #pragma once
 
+#if GLM_ARCH & GLM_ARCH_SSE2_BIT
+
+// SSE scalar reciprocal sqrt using rsqrt op, plus one Newton-Rhaphson iteration
+// By Elan Ruskin, http://assemblyrequired.crashworks.org/
+GLM_FUNC_QUALIFIER __m128 glm_f32v1_sqrt_lowp(__m128 x)
+{
+	return _mm_mul_ss(_mm_rsqrt_ss(x), x);
+}
+
+// SSE scalar reciprocal sqrt using rsqrt op, plus one Newton-Rhaphson iteration
+// By Elan Ruskin, http://assemblyrequired.crashworks.org/
+GLM_FUNC_QUALIFIER __m128 glm_f32v4_sqrt_lowp(__m128 x)
+{
+	return _mm_mul_ps(_mm_rsqrt_ps(x), x);
+}
+
+#endif//GLM_ARCH & GLM_ARCH_SSE2_BIT
diff --git a/test/core/core_func_exponential.cpp b/test/core/core_func_exponential.cpp
index b06090f0..b2f73da5 100644
--- a/test/core/core_func_exponential.cpp
+++ b/test/core/core_func_exponential.cpp
@@ -1,33 +1,5 @@
-///////////////////////////////////////////////////////////////////////////////////
-/// OpenGL Mathematics (glm.g-truc.net)
-///
-/// Copyright (c) 2005 - 2015 G-Truc Creation (www.g-truc.net)
-/// Permission is hereby granted, free of charge, to any person obtaining a copy
-/// of this software and associated documentation files (the "Software"), to deal
-/// in the Software without restriction, including without limitation the rights
-/// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-/// copies of the Software, and to permit persons to whom the Software is
-/// furnished to do so, subject to the following conditions:
-/// 
-/// The above copyright notice and this permission notice shall be included in
-/// all copies or substantial portions of the Software.
-/// 
-/// Restrictions:
-///		By making use of the Software for military purposes, you choose to make
-///		a Bunny unhappy.
-/// 
-/// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-/// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-/// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-/// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-/// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-/// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-/// THE SOFTWARE.
-///
 /// @file test/core/func_exponential.cpp
 /// @date 2011-01-15 / 2011-09-13
-/// @author Christophe Riccio
-///////////////////////////////////////////////////////////////////////////////////
 
 #include <glm/common.hpp>
 #include <glm/exponential.hpp>
@@ -103,6 +75,18 @@ int test_sqrt()
 {
 	int Error(0);
 
+#	if GLM_ARCH & GLM_ARCH_SSE2_BIT
+	for(float f = 0.1f; f < 30.0f; f += 0.1f)
+	{
+		float q = _mm_cvtss_f32(_mm_sqrt_ps(_mm_set1_ps(f)));
+		float r = _mm_cvtss_f32(glm_f32v4_sqrt_lowp(_mm_set1_ps(f)));
+		float s = std::sqrt(f);
+		Error += glm::abs(q - s) < 0.01f ? 0 : 1;
+		Error += glm::abs(r - s) < 0.01f ? 0 : 1;
+		assert(!Error);
+	}
+#	endif//GLM_ARCH & GLM_ARCH_SSE2_BIT
+
 	float A = glm::sqrt(10.f);
 	glm::vec1 B = glm::sqrt(glm::vec1(10.f));
 	glm::vec2 C = glm::sqrt(glm::vec2(10.f));