diff options
Diffstat (limited to 'extern/Eigen3/Eigen/src/Core/arch/NEON')
| -rw-r--r-- | extern/Eigen3/Eigen/src/Core/arch/NEON/Complex.h | 10 | ||||
| -rw-r--r-- | extern/Eigen3/Eigen/src/Core/arch/NEON/PacketMath.h | 50 |
2 files changed, 20 insertions, 40 deletions
diff --git a/extern/Eigen3/Eigen/src/Core/arch/NEON/Complex.h b/extern/Eigen3/Eigen/src/Core/arch/NEON/Complex.h index 795b4be7303..f183d31de2a 100644 --- a/extern/Eigen3/Eigen/src/Core/arch/NEON/Complex.h +++ b/extern/Eigen3/Eigen/src/Core/arch/NEON/Complex.h @@ -68,7 +68,6 @@ template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { Packet4f v1, v2; - float32x2_t a_lo, a_hi; // Get the real values of a | a1_re | a1_re | a2_re | a2_re | v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0)); @@ -81,9 +80,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con // Conjugate v2 v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR)); // Swap real/imag elements in v2. - a_lo = vrev64_f32(vget_low_f32(v2)); - a_hi = vrev64_f32(vget_high_f32(v2)); - v2 = vcombine_f32(a_lo, a_hi); + v2 = vrev64q_f32(v2); // Add and return the result return Packet2cf(vaddq_f32(v1, v2)); } @@ -241,13 +238,10 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con // TODO optimize it for AltiVec Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); Packet4f s, rev_s; - float32x2_t a_lo, a_hi; // this computes the norm s = vmulq_f32(b.v, b.v); - a_lo = vrev64_f32(vget_low_f32(s)); - a_hi = vrev64_f32(vget_high_f32(s)); - rev_s = vcombine_f32(a_lo, a_hi); + rev_s = vrev64q_f32(s); return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); } diff --git a/extern/Eigen3/Eigen/src/Core/arch/NEON/PacketMath.h b/extern/Eigen3/Eigen/src/Core/arch/NEON/PacketMath.h index a20250f7c65..163bac215e6 100644 --- a/extern/Eigen3/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/extern/Eigen3/Eigen/src/Core/arch/NEON/PacketMath.h @@ -115,6 +115,9 @@ template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return vnegq_f32(a); } template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return vnegq_s32(a); } +template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } + template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmulq_f32(a,b); } template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmulq_s32(a,b); } @@ -188,15 +191,15 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EI template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) { float32x2_t lo, hi; - lo = vdup_n_f32(*from); - hi = vdup_n_f32(*(from+1)); + lo = vld1_dup_f32(from); + hi = vld1_dup_f32(from+1); return vcombine_f32(lo, hi); } template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) { int32x2_t lo, hi; - lo = vdup_n_s32(*from); - hi = vdup_n_s32(*(from+1)); + lo = vld1_dup_s32(from); + hi = vld1_dup_s32(from+1); return vcombine_s32(lo, hi); } @@ -237,15 +240,12 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) { float32x2_t a_lo, a_hi, sum; - float s[2]; a_lo = vget_low_f32(a); a_hi = vget_high_f32(a); sum = vpadd_f32(a_lo, a_hi); sum = vpadd_f32(sum, sum); - vst1_f32(s, sum); - - return s[0]; + return vget_lane_f32(sum, 0); } template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) @@ -271,15 +271,12 @@ template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) { int32x2_t a_lo, a_hi, sum; - int32_t s[2]; a_lo = vget_low_s32(a); a_hi = vget_high_s32(a); sum = vpadd_s32(a_lo, a_hi); sum = vpadd_s32(sum, sum); - vst1_s32(s, sum); - - return s[0]; + return vget_lane_s32(sum, 0); } template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) @@ -307,7 +304,6 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) { float32x2_t a_lo, a_hi, prod; - float s[2]; // Get a_lo = |a1|a2| and a_hi = |a3|a4| a_lo = vget_low_f32(a); @@ -316,14 +312,12 @@ template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) prod = vmul_f32(a_lo, a_hi); // Multiply prod with its swapped value |a2*a4|a1*a3| prod = vmul_f32(prod, vrev64_f32(prod)); - vst1_f32(s, prod); - return s[0]; + return vget_lane_f32(prod, 0); } template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) { int32x2_t a_lo, a_hi, prod; - int32_t s[2]; // Get a_lo = |a1|a2| and a_hi = |a3|a4| a_lo = vget_low_s32(a); @@ -332,65 +326,57 @@ template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) prod = vmul_s32(a_lo, a_hi); // Multiply prod with its swapped value |a2*a4|a1*a3| prod = vmul_s32(prod, vrev64_s32(prod)); - vst1_s32(s, prod); - return s[0]; + return vget_lane_s32(prod, 0); } // min template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) { float32x2_t a_lo, a_hi, min; - float s[2]; a_lo = vget_low_f32(a); a_hi = vget_high_f32(a); min = vpmin_f32(a_lo, a_hi); min = vpmin_f32(min, min); - vst1_f32(s, min); - return s[0]; + return vget_lane_f32(min, 0); } + template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) { int32x2_t a_lo, a_hi, min; - int32_t s[2]; a_lo = vget_low_s32(a); a_hi = vget_high_s32(a); min = vpmin_s32(a_lo, a_hi); min = vpmin_s32(min, min); - vst1_s32(s, min); - - return s[0]; + + return vget_lane_s32(min, 0); } // max template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) { float32x2_t a_lo, a_hi, max; - float s[2]; a_lo = vget_low_f32(a); a_hi = vget_high_f32(a); max = vpmax_f32(a_lo, a_hi); max = vpmax_f32(max, max); - vst1_f32(s, max); - return s[0]; + return vget_lane_f32(max, 0); } + template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) { int32x2_t a_lo, a_hi, max; - int32_t s[2]; a_lo = vget_low_s32(a); a_hi = vget_high_s32(a); max = vpmax_s32(a_lo, a_hi); - max = vpmax_s32(max, max); - vst1_s32(s, max); - return s[0]; + return vget_lane_s32(max, 0); } // this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors, |