GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-06-04 18:54:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

slight optimization of SSE base integer mul (thanks to Rohit Garg)

Browse Source
This commit is contained in:
Gael Guennebaud 2009-03-08 10:14:07 +00:00
parent f9790a649c
commit 7718a8ed83
2 changed files with 26 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@ -29,6 +29,18 @@
#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 16 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 16
#endif #endif
#define ei_vec4f_swizzle1(v,p,q,r,s) \
(_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p)))))
#define ei_vec4i_swizzle1(v,p,q,r,s) \
(_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p))))
#define ei_vec4f_swizzle2(a,b,p,q,r,s) \
(_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p))))
#define ei_vec4i_swizzle2(a,b,p,q,r,s) \
(_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p))))))
template<> struct ei_packet_traits<float> { typedef __m128 type; enum {size=4}; }; template<> struct ei_packet_traits<float> { typedef __m128 type; enum {size=4}; };
template<> struct ei_packet_traits<double> { typedef __m128d type; enum {size=2}; }; template<> struct ei_packet_traits<double> { typedef __m128d type; enum {size=2}; };
template<> struct ei_packet_traits<int> { typedef __m128i type; enum {size=4}; }; template<> struct ei_packet_traits<int> { typedef __m128i type; enum {size=4}; };
@ -54,14 +66,13 @@ template<> EIGEN_STRONG_INLINE __m128d ei_pmul<__m128d>(const __m128d& a, const
template<> EIGEN_STRONG_INLINE __m128i ei_pmul<__m128i>(const __m128i& a, const __m128i& b) template<> EIGEN_STRONG_INLINE __m128i ei_pmul<__m128i>(const __m128i& a, const __m128i& b)
{ {
// this version is very slightly faster than 4 scalar products // this version is very slightly faster than 4 scalar products
return _mm_or_si128( return ei_vec4i_swizzle1(
_mm_and_si128( ei_vec4i_swizzle2(
_mm_mul_epu32(a,b), _mm_mul_epu32(a,b),
_mm_setr_epi32(0xffffffff,0,0xffffffff,0)), _mm_mul_epu32(ei_vec4i_swizzle1(a,1,0,3,2),
_mm_slli_si128( ei_vec4i_swizzle1(b,1,0,3,2)),
_mm_and_si128( 0,2,0,2),
_mm_mul_epu32(_mm_srli_si128(a,4),_mm_srli_si128(b,4)), 0,2,1,3);
_mm_setr_epi32(0xffffffff,0,0xffffffff,0)), 4));
} }
template<> EIGEN_STRONG_INLINE __m128 ei_pdiv<__m128>(const __m128& a, const __m128& b) { return _mm_div_ps(a,b); } template<> EIGEN_STRONG_INLINE __m128 ei_pdiv<__m128>(const __m128& a, const __m128& b) { return _mm_div_ps(a,b); }

17
Eigen/src/Geometry/arch/Geometry_SSE.h
View File

@ -26,9 +26,6 @@
#ifndef EIGEN_GEOMETRY_SSE_H #ifndef EIGEN_GEOMETRY_SSE_H
#define EIGEN_GEOMETRY_SSE_H #define EIGEN_GEOMETRY_SSE_H
#define vec4f_swizzle(v,p,q,r,s) (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), \
((s)<<6|(r)<<4|(q)<<2|(p)))))
template<> inline Quaternion<float> template<> inline Quaternion<float>
ei_quaternion_product<EiArch_SSE,float>(const Quaternion<float>& _a, const Quaternion<float>& _b) ei_quaternion_product<EiArch_SSE,float>(const Quaternion<float>& _a, const Quaternion<float>& _b)
{ {
@ -36,14 +33,14 @@ ei_quaternion_product<EiArch_SSE,float>(const Quaternion<float>& _a, const Quate
Quaternion<float> res; Quaternion<float> res;
__m128 a = _a.coeffs().packet<Aligned>(0); __m128 a = _a.coeffs().packet<Aligned>(0);
__m128 b = _b.coeffs().packet<Aligned>(0); __m128 b = _b.coeffs().packet<Aligned>(0);
__m128 flip1 = _mm_xor_ps(_mm_mul_ps(vec4f_swizzle(a,1,2,0,2), __m128 flip1 = _mm_xor_ps(_mm_mul_ps(ei_vec4f_swizzle1(a,1,2,0,2),
vec4f_swizzle(b,2,0,1,2)),mask); ei_vec4f_swizzle1(b,2,0,1,2)),mask);
__m128 flip2 = _mm_xor_ps(_mm_mul_ps(vec4f_swizzle(a,3,3,3,1), __m128 flip2 = _mm_xor_ps(_mm_mul_ps(ei_vec4f_swizzle1(a,3,3,3,1),
vec4f_swizzle(b,0,1,2,1)),mask); ei_vec4f_swizzle1(b,0,1,2,1)),mask);
ei_pstore(&res.x(), ei_pstore(&res.x(),
_mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle(b,3,3,3,3)), _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,ei_vec4f_swizzle1(b,3,3,3,3)),
_mm_mul_ps(vec4f_swizzle(a,2,0,1,0), _mm_mul_ps(ei_vec4f_swizzle1(a,2,0,1,0),
vec4f_swizzle(b,1,2,0,0))), ei_vec4f_swizzle1(b,1,2,0,0))),
_mm_add_ps(flip1,flip2))); _mm_add_ps(flip1,flip2)));
return res; return res;
} }