mewin/glm
1
0
Fork 0
Code Releases Activity

SIMD sqrt optimizations including lowp

Browse Source
This commit is contained in:
Christophe Riccio
2016-05-29 02:57:53 +02:00
parent 449c7ccedf
commit 20cf68679c
6 changed files with 68 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
glm/detail/func_exponential.inl
View File

@@ -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

26
glm/detail/func_exponential_simd.inl
View File

@@ -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

3
glm/detail/func_geometric_simd.inl
View File

@@ -1,3 +1,6 @@
/// @ref core
/// @file glm/detail/func_geometric_simd.inl
#include "../simd/geometric.h"
#if GLM_ARCH & GLM_ARCH_SSE2_BIT

26
glm/simd/common.h
View File

@@ -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

17
glm/simd/exponential.h
View File

@@ -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