From 52a5f9821235e5a9f7e9b3e0198d45d42a1cb267 Mon Sep 17 00:00:00 2001 From: Rasmus Munk Larsen Date: Thu, 24 Jun 2021 15:47:48 -0700 Subject: Get rid of code duplication for conj_helper. For packets where LhsType=RhsType a single generic implementation suffices. For scalars, the generic implementation of pconj automatically forwards to numext::conj, so much of the existing specialization can be avoided. For mixed types we still need specializations. --- Eigen/src/Core/arch/SSE/Complex.h | 119 ++------------------------------------ 1 file changed, 6 insertions(+), 113 deletions(-) (limited to 'Eigen/src/Core/arch/SSE/Complex.h') diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 1cab374c0..8fe22da46 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -165,74 +165,21 @@ template<> EIGEN_STRONG_INLINE std::complex predux_mul(const P return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v)))); } -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(pmul(x,y),c); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const - { - #ifdef EIGEN_VECTORIZE_SSE3 - return internal::pmul(a, pconj(b)); - #else - const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); - return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), - _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), - vec4f_swizzle1(b.v, 1, 0, 3, 2)))); - #endif - } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(pmul(x,y),c); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const - { - #ifdef EIGEN_VECTORIZE_SSE3 - return internal::pmul(pconj(a), b); - #else - const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); - return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), - _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), - vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); - #endif - } -}; - -template<> struct conj_helper +EIGEN_STRONG_INLINE Packet2cf pcplxflip/* */(const Packet2cf& x) { - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(pmul(x,y),c); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const - { - #ifdef EIGEN_VECTORIZE_SSE3 - return pconj(internal::pmul(a, b)); - #else - const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); - return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), - _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), - vec4f_swizzle1(b.v, 1, 0, 3, 2)))); - #endif - } -}; + return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2)); +} EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, const Packet2cf& b) { // TODO optimize it for SSE3 and 4 - Packet2cf res = conj_helper().pmul(a,b); + Packet2cf res = pmul(a, pconj(b)); __m128 s = _mm_mul_ps(b.v,b.v); - return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1))))); + return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,vec4f_swizzle1(s, 1, 0, 3, 2)))); } -EIGEN_STRONG_INLINE Packet2cf pcplxflip/* */(const Packet2cf& x) -{ - return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2)); -} //---------- double ---------- @@ -348,66 +295,12 @@ template<> EIGEN_STRONG_INLINE std::complex predux_mul(const return pfirst(a); } -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(pmul(x,y),c); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const - { - #ifdef EIGEN_VECTORIZE_SSE3 - return internal::pmul(a, pconj(b)); - #else - const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); - return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), - _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), - vec2d_swizzle1(b.v, 1, 0)))); - #endif - } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(pmul(x,y),c); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const - { - #ifdef EIGEN_VECTORIZE_SSE3 - return internal::pmul(pconj(a), b); - #else - const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); - return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), - _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), - vec2d_swizzle1(b.v, 1, 0)), mask))); - #endif - } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(pmul(x,y),c); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const - { - #ifdef EIGEN_VECTORIZE_SSE3 - return pconj(internal::pmul(a, b)); - #else - const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); - return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), - _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), - vec2d_swizzle1(b.v, 1, 0)))); - #endif - } -}; - EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) { // TODO optimize it for SSE3 and 4 - Packet1cd res = conj_helper().pmul(a,b); + Packet1cd res = pmul(a,pconj(b)); __m128d s = _mm_mul_pd(b.v,b.v); return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1)))); } -- cgit v1.2.3