Unify NEON's pexp with generic implementation

2 files changed, 25 insertions, 64 deletions

diff --git a/Eigen/src/Core/arch/NEON/MathFunctions.h b/Eigen/src/Core/arch/NEON/MathFunctions.h
index 8a70b771c..f3f6c279e 100644
--- a/Eigen/src/Core/arch/NEON/MathFunctions.h
+++ b/Eigen/src/Core/arch/NEON/MathFunctions.h
@@ -21,69 +21,7 @@ namespace internal {
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f pexp<Packet4f>(const Packet4f& _x)
 {
-  Packet4f x = _x;
-  Packet4f tmp, fx;
-
-  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
-  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
-  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
-  _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
-  _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
-  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
-
-  x = vminq_f32(x, p4f_exp_hi);
-  x = vmaxq_f32(x, p4f_exp_lo);
-
-  /* express exp(x) as exp(g + n*log(2)) */
-  fx = vmlaq_f32(p4f_half, x, p4f_cephes_LOG2EF);
-
-  /* perform a floorf */
-  tmp = vcvtq_f32_s32(vcvtq_s32_f32(fx));
-
-  /* if greater, substract 1 */
-  Packet4ui mask = vcgtq_f32(tmp, fx);
-  mask = vandq_u32(mask, vreinterpretq_u32_f32(p4f_1));
-
-  fx = vsubq_f32(tmp, vreinterpretq_f32_u32(mask));
-
-  tmp = vmulq_f32(fx, p4f_cephes_exp_C1);
-  Packet4f z = vmulq_f32(fx, p4f_cephes_exp_C2);
-  x = vsubq_f32(x, tmp);
-  x = vsubq_f32(x, z);
-
-  Packet4f y = vmulq_f32(p4f_cephes_exp_p0, x);
-  z = vmulq_f32(x, x);
-  y = vaddq_f32(y, p4f_cephes_exp_p1);
-  y = vmulq_f32(y, x);
-  y = vaddq_f32(y, p4f_cephes_exp_p2);
-  y = vmulq_f32(y, x);
-  y = vaddq_f32(y, p4f_cephes_exp_p3);
-  y = vmulq_f32(y, x);
-  y = vaddq_f32(y, p4f_cephes_exp_p4);
-  y = vmulq_f32(y, x);
-  y = vaddq_f32(y, p4f_cephes_exp_p5);
-
-  y = vmulq_f32(y, z);
-  y = vaddq_f32(y, x);
-  y = vaddq_f32(y, p4f_1);
-
-  /* build 2^n */
-  int32x4_t mm;
-  mm = vcvtq_s32_f32(fx);
-  mm = vaddq_s32(mm, p4i_0x7f);
-  mm = vshlq_n_s32(mm, 23);
-  Packet4f pow2n = vreinterpretq_f32_s32(mm);
-
-  y = vmulq_f32(y, pow2n);
-  return y;
+  return pexp_float(_x);
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 72f076e50..dc432f0d2 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -108,7 +108,8 @@ template<> struct packet_traits<float>  : default_packet_traits
     size = 4,
     HasHalfPacket=0, // Packet2f intrinsics not implemented yet
    
-    HasDiv  = 1,
+    HasDiv   = 1,
+    HasFloor = 1,
     // FIXME check the Has*
     HasSin  = 0,
     HasCos  = 0,
@@ -256,6 +257,18 @@ template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4
 template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return vreinterpretq_f32_u32(vceqq_f32(a,b)); }
 template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return vreinterpretq_f32_u32(vmvnq_u32(vcgeq_f32(a,b))); }
 
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
+{
+  const Packet4f cst_1 = pset1<Packet4f>(1.0f);
+  /* perform a floorf */
+  Packet4f tmp = vcvtq_f32_s32(vcvtq_s32_f32(a));
+
+  /* if greater, substract 1 */
+  Packet4ui mask = vcgtq_f32(tmp, a);
+  mask = vandq_u32(mask, vreinterpretq_u32_f32(cst_1));
+  return vsubq_f32(tmp, vreinterpretq_f32_u32(mask));
+}
+
 // Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
 {
@@ -379,6 +392,16 @@ template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Pack
   return pfrexp_float(a,exponent);
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f pcast_and_shiftleft<Packet4f>(Packet4f v, int n)
+{
+  Packet4i vi = vcvtq_s32_f32(v);
+  return vreinterpretq_f32_s32(vshlq_n_s32(vi, n));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) {
+  return pldexp_float(a,exponent);
+}
+
 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
 {
   float32x2_t a_lo, a_hi, sum;