diff options
Diffstat (limited to 'extern/Eigen3/Eigen/src/Core/arch/SSE/MathFunctions.h')
-rw-r--r-- | extern/Eigen3/Eigen/src/Core/arch/SSE/MathFunctions.h | 14 |
1 files changed, 7 insertions, 7 deletions
diff --git a/extern/Eigen3/Eigen/src/Core/arch/SSE/MathFunctions.h b/extern/Eigen3/Eigen/src/Core/arch/SSE/MathFunctions.h index 99cbd0d95be..2b07168ae4f 100644 --- a/extern/Eigen3/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/extern/Eigen3/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -52,7 +52,7 @@ Packet4f plog<Packet4f>(const Packet4f& _x) Packet4i emm0; - Packet4f invalid_mask = _mm_cmplt_ps(x, _mm_setzero_ps()); + Packet4f invalid_mask = _mm_cmpnge_ps(x, _mm_setzero_ps()); // not greater equal is true if x is NaN Packet4f iszero_mask = _mm_cmpeq_ps(x, _mm_setzero_ps()); x = pmax(x, p4f_min_norm_pos); /* cut off denormalized stuff */ @@ -126,7 +126,7 @@ Packet4f pexp<Packet4f>(const Packet4f& _x) _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); - Packet4f tmp = _mm_setzero_ps(), fx; + Packet4f tmp, fx; Packet4i emm0; // clamp x @@ -166,7 +166,7 @@ Packet4f pexp<Packet4f>(const Packet4f& _x) emm0 = _mm_cvttps_epi32(fx); emm0 = _mm_add_epi32(emm0, p4i_0x7f); emm0 = _mm_slli_epi32(emm0, 23); - return pmul(y, _mm_castsi128_ps(emm0)); + return pmax(pmul(y, Packet4f(_mm_castsi128_ps(emm0))), _x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d pexp<Packet2d>(const Packet2d& _x) @@ -195,7 +195,7 @@ Packet2d pexp<Packet2d>(const Packet2d& _x) _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6); static const __m128i p4i_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0); - Packet2d tmp = _mm_setzero_pd(), fx; + Packet2d tmp, fx; Packet4i emm0; // clamp x @@ -239,7 +239,7 @@ Packet2d pexp<Packet2d>(const Packet2d& _x) emm0 = _mm_add_epi32(emm0, p4i_1023_0); emm0 = _mm_slli_epi32(emm0, 20); emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3)); - return pmul(x, _mm_castsi128_pd(emm0)); + return pmax(pmul(x, Packet2d(_mm_castsi128_pd(emm0))), _x); } /* evaluation of 4 sines at onces, using SSE2 intrinsics. @@ -279,7 +279,7 @@ Packet4f psin<Packet4f>(const Packet4f& _x) _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI - Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y; + Packet4f xmm1, xmm2, xmm3, sign_bit, y; Packet4i emm0, emm2; sign_bit = x; @@ -378,7 +378,7 @@ Packet4f pcos<Packet4f>(const Packet4f& _x) _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI - Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, y; + Packet4f xmm1, xmm2, xmm3, y; Packet4i emm0, emm2; x = pabs(x); |