Support more SSE and AVX vector types

8 files changed, 1460 insertions, 103 deletions

diff --git a/.gitignore b/.gitignore
index c58f8d1..b69d47d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,3 +5,4 @@ example
 old
 test
 .ninja_log
+OUT
diff --git a/mathfuncs_sqrt.h b/mathfuncs_sqrt.h
index c1ea3f7..050fc91 100644
--- a/mathfuncs_sqrt.h
+++ b/mathfuncs_sqrt.h
@@ -57,6 +57,14 @@ namespace vecmathlib {
   template<typename realvec_t>
   realvec_t mathfuncs<realvec_t>::vml_rsqrt(realvec_t x)
   {
+#if 0
+    double const x_2 = 0.5 * x;   // x/2
+    // Newton method:
+    // Solve   f(r) = 0   for   f(r) = x - 1/r^2
+    //    r <- r - f(r) / f'(r)
+    //    r <- (3 r - r^3 x) / 2
+    return r * (1.5 - r*r * x_2);
+#endif
     return rcp(sqrt(x));
   }
   
diff --git a/test.cc b/test.cc
index 9d45069..c61d0ae 100644
--- a/test.cc
+++ b/test.cc
@@ -40,7 +40,7 @@ struct vecmathlib_test {
   
   
   // Test each function with this many random values
-  static int const imax = 100000;
+  static int const imax = 10000;
   // Require that 3/4 of the digits are correct
   static real_t constexpr accuracy = pow(realvec_t::epsilon(), R(0.75));
   
@@ -380,7 +380,11 @@ int main(int argc, char** argv)
   cout << "Testing math functions:\n";
   
   vecmathlib_test<realvec<float,1>>::test();
+  vecmathlib_test<realvec<float,4>>::test();
+  vecmathlib_test<realvec<float,8>>::test();
+  
   vecmathlib_test<realvec<double,1>>::test();
+  vecmathlib_test<realvec<double,2>>::test();
   vecmathlib_test<realvec<double,4>>::test();
   
   cout << "\n";
diff --git a/vec_double_avx.h b/vec_double_avx.h
index f311095..081045f 100644
--- a/vec_double_avx.h
+++ b/vec_double_avx.h
@@ -171,34 +171,27 @@ namespace vecmathlib {
     // Note: not all arithmetic operations are supported!
     
     intvec operator+() const { return *this; }
-    intvec operator-() const
-    {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      v01 = _mm_sub_epi64(_mm_set1_epi64x(0), v01);
-      v23 = _mm_sub_epi64(_mm_set1_epi64x(0), v23);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
-    }
+    intvec operator-() const { return IV(I(0)) - *this; }
     
     intvec operator+(intvec x) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      __m128i xv01 = _mm256_castsi256_si128(x.v);
-      __m128i xv23 = _mm256_extractf128_si256(x.v, 1);
-      v01 = _mm_add_epi64(v01, xv01);
-      v23 = _mm_add_epi64(v23, xv23);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i xvlo = _mm256_castsi256_si128(x.v);
+      __m128i xvhi = _mm256_extractf128_si256(x.v, 1);
+      vlo = _mm_add_epi64(vlo, xvlo);
+      vhi = _mm_add_epi64(vhi, xvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec operator-(intvec x) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      __m128i xv01 = _mm256_castsi256_si128(x.v);
-      __m128i xv23 = _mm256_extractf128_si256(x.v, 1);
-      v01 = _mm_sub_epi64(v01, xv01);
-      v23 = _mm_sub_epi64(v23, xv23);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i xvlo = _mm256_castsi256_si128(x.v);
+      __m128i xvhi = _mm256_extractf128_si256(x.v, 1);
+      vlo = _mm_sub_epi64(vlo, xvlo);
+      vhi = _mm_sub_epi64(vhi, xvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     
     intvec& operator+=(intvec const& x) { return *this=*this+x; }
@@ -206,11 +199,7 @@ namespace vecmathlib {
     
     
     
-    intvec operator~() const
-    {
-      return _mm256_castpd_si256(_mm256_xor_pd(_mm256_castsi256_pd(IV(~U(0))),
-                                               _mm256_castsi256_pd(v)));
-    }
+    intvec operator~() const { return IV(~U(0)) ^ *this; }
     
     intvec operator&(intvec x) const
     {
@@ -234,118 +223,109 @@ namespace vecmathlib {
     
     
     
-    // SSE2 shift instructions potentially relevant for 64-bit
-    // operations:
-    //    _mm_slli_epi64
-    //    _mm_srli_epi64
-    //    _mm_sll_epi64
-    //    _mm_srl_epi64
-    //    _mm_srai_epi32
-    //    _mm_sra_epi32
-    //    _mm_srli_si128
-    //    _mm_slli_si128
     intvec lsr(int_t n) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      v01 = _mm_srli_epi64(v01, n);
-      v23 = _mm_srli_epi64(v23, n);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      vlo = _mm_srli_epi64(vlo, n);
+      vhi = _mm_srli_epi64(vhi, n);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec operator>>(int_t n) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      // There is no _mm_srai_epi64. To emulate it, add 0x8000000
-      // before shifting, and subtracted the shifted 0x80000000 after shifting
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      // There is no _mm_srai_epi64. To emulate it, add 0x80000000
+      // before shifting, and subtract the shifted 0x80000000 after
+      // shifting
 #if 0
       __m128i signmask01 = _mm_sub_epi64(_mm_set1_epi64x(0),
-                                         _mm_srli_epi64(v01, 63));
+                                         _mm_srli_epi64(vlo, 63));
       __m128i signmask23 = _mm_sub_epi64(_mm_set1_epi64x(0),
-                                         _mm_srli_epi64(v23, 63));
-      v01 = _mm_xor_si128(signmask01, v01);
-      v23 = _mm_xor_si128(signmask23, v23);
-      v01 = _mm_srli_epi64(v01, n);
-      v23 = _mm_srli_epi64(v23, n);
-      v01 = _mm_xor_si128(signmask01, v01);
-      v23 = _mm_xor_si128(signmask23, v23);
+                                         _mm_srli_epi64(vhi, 63));
+      vlo = _mm_xor_si128(signmask01, vlo);
+      vhi = _mm_xor_si128(signmask23, vhi);
+      vlo = _mm_srli_epi64(vlo, n);
+      vhi = _mm_srli_epi64(vhi, n);
+      vlo = _mm_xor_si128(signmask01, vlo);
+      vhi = _mm_xor_si128(signmask23, vhi);
 #else
       // Convert signed to unsiged
-      v01 += _mm_set1_epi64x(U(1) << (bits-1));
-      v23 += _mm_set1_epi64x(U(1) << (bits-1));
+      vlo += _mm_set1_epi64x(U(1) << (bits-1));
+      vhi += _mm_set1_epi64x(U(1) << (bits-1));
       // Shift
-      v01 = _mm_srli_epi64(v01, n);
-      v23 = _mm_srli_epi64(v23, n);
+      vlo = _mm_srli_epi64(vlo, n);
+      vhi = _mm_srli_epi64(vhi, n);
       // Undo conversion
-      v01 -= _mm_set1_epi64x(U(1) << (bits-n));
-      v23 -= _mm_set1_epi64x(U(1) << (bits-n));
+      vlo -= _mm_set1_epi64x(U(1) << (bits-n));
+      vhi -= _mm_set1_epi64x(U(1) << (bits-n));
 #endif
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec operator<<(int_t n) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      v01 = _mm_slli_epi64(v01, n);
-      v23 = _mm_slli_epi64(v23, n);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      vlo = _mm_slli_epi64(vlo, n);
+      vhi = _mm_slli_epi64(vhi, n);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec& operator>>=(int_t n) { return *this=*this>>n; }
     intvec& operator<<=(int_t n) { return *this=*this<<n; }
     
     intvec lsr(intvec n) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      __m128i nv01 = _mm256_castsi256_si128(n.v);
-      __m128i nv23 = _mm256_extractf128_si256(n.v, 1);
-      v01 = _mm_srl_epi64(v01, nv01);
-      v23 = _mm_srl_epi64(v23, nv23);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i nvlo = _mm256_castsi256_si128(n.v);
+      __m128i nvhi = _mm256_extractf128_si256(n.v, 1);
+      vlo = _mm_srl_epi64(vlo, nvlo);
+      vhi = _mm_srl_epi64(vhi, nvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec operator>>(intvec n) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      __m128i nv01 = _mm256_castsi256_si128(n.v);
-      __m128i nv23 = _mm256_extractf128_si256(n.v, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i nvlo = _mm256_castsi256_si128(n.v);
+      __m128i nvhi = _mm256_extractf128_si256(n.v, 1);
 #if 0
       // There is no _mm_srai_epi64. To emulate it, invert all bits
       // before and after shifting if the sign bit is set.
       __m128i signmask01 = _mm_sub_epi64(_mm_set1_epi64x(0),
-                                         _mm_srli_epi64(v01, 63));
+                                         _mm_srli_epi64(vlo, 63));
       __m128i signmask23 = _mm_sub_epi64(_mm_set1_epi64x(0),
-                                         _mm_srli_epi64(v23, 63));
-      v01 = _mm_xor_si128(signmask01, v01);
-      v23 = _mm_xor_si128(signmask23, v23);
-      v01 = _mm_srl_epi64(v01, nv01);
-      v23 = _mm_srl_epi64(v23, nv23);
-      v01 = _mm_xor_si128(signmask01, v01);
-      v23 = _mm_xor_si128(signmask23, v23);
+                                         _mm_srli_epi64(vhi, 63));
+      vlo = _mm_xor_si128(signmask01, vlo);
+      vhi = _mm_xor_si128(signmask23, vhi);
+      vlo = _mm_srl_epi64(vlo, nvlo);
+      vhi = _mm_srl_epi64(vhi, nvhi);
+      vlo = _mm_xor_si128(signmask01, vlo);
+      vhi = _mm_xor_si128(signmask23, vhi);
 #else
       // Convert signed to unsiged
-      v01 += _mm_set1_epi64x(U(1) << (bits-1));
-      v23 += _mm_set1_epi64x(U(1) << (bits-1));
+      vlo += _mm_set1_epi64x(U(1) << (bits-1));
+      vhi += _mm_set1_epi64x(U(1) << (bits-1));
       // Shift
-      v01 = _mm_srl_epi64(v01, nv01);
-      v23 = _mm_srl_epi64(v23, nv23);
+      vlo = _mm_srl_epi64(vlo, nvlo);
+      vhi = _mm_srl_epi64(vhi, nvhi);
       // Undo conversion
-      v01 -= _mm_sll_epi64(_mm_set1_epi64x(1),
-                           _mm_sub_epi64(_mm_set1_epi64x(bits), nv01));
-      v23 -= _mm_sll_epi64(_mm_set1_epi64x(1),
-                           _mm_sub_epi64(_mm_set1_epi64x(bits), nv23));
+      vlo -= _mm_sll_epi64(_mm_set1_epi64x(1),
+                           _mm_sub_epi64(_mm_set1_epi64x(bits), nvlo));
+      vhi -= _mm_sll_epi64(_mm_set1_epi64x(1),
+                           _mm_sub_epi64(_mm_set1_epi64x(bits), nvhi));
 #endif
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec operator<<(intvec n) const
     {
-      __m128i v01 = _mm256_castsi256_si128(v);
-      __m128i v23 = _mm256_extractf128_si256(v, 1);
-      __m128i nv01 = _mm256_castsi256_si128(n.v);
-      __m128i nv23 = _mm256_extractf128_si256(n.v, 1);
-      v01 = _mm_sll_epi64(v01, nv01);
-      v23 = _mm_sll_epi64(v23, nv23);
-      return _mm256_insertf128_si256(_mm256_castsi128_si256(v01), v23, 1);
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i nvlo = _mm256_castsi256_si128(n.v);
+      __m128i nvhi = _mm256_extractf128_si256(n.v, 1);
+      vlo = _mm_sll_epi64(vlo, nvlo);
+      vhi = _mm_sll_epi64(vhi, nvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
     }
     intvec& operator>>=(intvec n) { return *this=*this>>n; }
     intvec& operator<<=(intvec n) { return *this=*this<<n; }
@@ -403,7 +383,7 @@ namespace vecmathlib {
     
     
     realvec operator+() const { return *this; }
-    realvec operator-() const { return _mm256_sub_pd(_mm256_set1_pd(0.0), v); }
+    realvec operator-() const { return RV(0.0) - *this; }
     
     realvec operator+(realvec x) const { return _mm256_add_pd(v, x.v); }
     realvec operator-(realvec x) const { return _mm256_sub_pd(v, x.v); }
diff --git a/vec_double_sse2.h b/vec_double_sse2.h
new file mode 100644
index 0000000..1c6648c
--- /dev/null
+++ b/vec_double_sse2.h
@@ -0,0 +1,433 @@
+// -*-C++-*-
+
+#ifndef VEC_DOUBLE_SSE2_H
+#define VEC_DOUBLE_SSE2_H
+
+#include "floatprops.h"
+#include "mathfuncs.h"
+#include "vec_base.h"
+
+#include <cmath>
+
+// SSE2 intrinsics
+#include <emmintrin.h>
+#ifdef __SSE4_1__               // Intel's SSE 4.1
+#  include <smmintrin.h>
+#endif
+#ifdef __SSE4A__                // AMD's SSE 4a
+#  include <ammintrin.h>
+#endif
+
+
+
+namespace vecmathlib {
+  
+  template<> struct boolvec<double,2>;
+  template<> struct intvec<double,2>;
+  template<> struct realvec<double,2>;
+  
+  
+  
+  template<>
+  struct boolvec<double,2>: floatprops<double>
+  {
+    static int const size = 2;
+    typedef bool scalar_t;
+    typedef __m128d bvector_t;
+    
+    static_assert(size * sizeof(real_t) == sizeof(bvector_t),
+                  "vector size is wrong");
+    
+  private:
+    // true values have the sign bit set, false values have it unset
+    static uint_t from_bool(bool a) { return - uint_t(a); }
+    static bool to_bool(uint_t a) { return int_t(a) < int_t(0); }
+  public:
+    
+    typedef boolvec boolvec_t;
+    typedef intvec<real_t, size> intvec_t;
+    typedef realvec<real_t, size> realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    bvector_t v;
+    
+    boolvec() {}
+    boolvec(boolvec const& x): v(x.v) {}
+    boolvec& operator=(boolvec const& x) { return v=x.v, *this; }
+    boolvec(bvector_t x): v(x) {}
+    boolvec(bool a):
+    v(_mm_castsi128_pd(_mm_set1_epi64x(from_bool(a)))) {}
+    boolvec(bool const* as):
+    v(_mm_castsi128_pd(_mm_set_epi64x(from_bool(as[1]), from_bool(as[0])))) {}
+    
+    operator bvector_t() const { return v; }
+    bool operator[](int n) const { return to_bool(((uint_t const*)&v)[n]); }
+    boolvec& set_elt(int n, bool a)
+    {
+      return ((int_t*)&v)[n] = from_bool(a), *this;
+    }
+    
+    
+    
+    auto as_int() const -> intvec_t;      // defined after intvec
+    auto convert_int() const -> intvec_t; // defined after intvec
+    
+    
+    
+    boolvec operator!() const { return _mm_xor_pd(boolvec(true), v); }
+    
+    boolvec operator&&(boolvec x) const { return _mm_and_pd(v, x.v); }
+    boolvec operator||(boolvec x) const { return _mm_or_pd(v, x.v); }
+    boolvec operator==(boolvec x) const { return !(*this==x); }
+    boolvec operator!=(boolvec x) const { return _mm_xor_pd(v, x.v); }
+    
+    bool all() const
+    {
+      return (*this)[0] && (*this)[1];
+    }
+    bool any() const
+    {
+      return (*this)[0] || (*this)[1];
+    }
+    
+    
+    
+    // ifthen(condition, then-value, else-value)
+    auto ifthen(intvec_t x,
+                intvec_t y) const -> intvec_t; // defined after intvec
+    auto ifthen(realvec_t x,
+                realvec_t y) const -> realvec_t; // defined after realvec
+    
+  };
+  
+  
+  
+  template<>
+  struct intvec<double,2>: floatprops<double>
+  {
+    static int const size = 2;
+    typedef int_t scalar_t;
+    typedef __m128i ivector_t;
+    
+    static_assert(size * sizeof(real_t) == sizeof(ivector_t),
+                  "vector size is wrong");
+    
+    typedef boolvec<real_t, size> boolvec_t;
+    typedef intvec intvec_t;
+    typedef realvec<real_t, size> realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    ivector_t v;
+    
+    intvec() {}
+    intvec(intvec const& x): v(x.v) {}
+    intvec& operator=(intvec const& x) { return v=x.v, *this; }
+    intvec(ivector_t x): v(x) {}
+    intvec(int_t a): v(_mm_set1_epi64x(a)) {}
+    intvec(int_t const* as): v(_mm_set_epi64x(as[1], as[0])) {}
+    
+    operator ivector_t() const { return v; }
+    int_t operator[](int n) const { return ((int_t const*)&v)[n]; }
+    intvec& set_elt(int n, int_t a) { return ((int_t*)&v)[n]=a, *this; }
+    
+    
+    
+    auto as_bool() const -> boolvec_t { return _mm_castsi128_pd(v); }
+    auto convert_bool() const -> boolvec_t
+    {
+      // Result: convert_bool(0)=false, convert_bool(else)=true
+      // There is no intrinsic to compare with zero. Instead, we check
+      // whether x is positive and x-1 is negative.
+      intvec x = *this;
+      intvec xm1 = x - 1;
+      // We know that boolvec values depend only on the sign bit
+      // return (~xm1 | x).as_bool();
+      return x.as_bool() || !xm1.as_bool();
+    }
+    auto as_float() const -> realvec_t; // defined after realvec
+    auto convert_float() const -> realvec_t; // defined after realvec
+    
+    
+    
+    // Note: not all arithmetic operations are supported!
+    
+    intvec operator+() const { return *this; }
+    intvec operator-() const { return IV(I(0)) - *this; }
+    
+    intvec operator+(intvec x) const { return _mm_add_epi64(v, x.v); }
+    intvec operator-(intvec x) const { return _mm_sub_epi64(v, x.v); }
+    
+    intvec& operator+=(intvec const& x) { return *this=*this+x; }
+    intvec& operator-=(intvec const& x) { return *this=*this-x; }
+    
+    
+    
+    intvec operator~() const { return IV(~U(0)) ^ *this; }
+    
+    intvec operator&(intvec x) const
+    {
+      return _mm_castpd_si128(_mm_and_pd(_mm_castsi128_pd(v),
+                                         _mm_castsi128_pd(x.v)));
+    }
+    intvec operator|(intvec x) const
+    {
+      return _mm_castpd_si128(_mm_or_pd(_mm_castsi128_pd(v),
+                                        _mm_castsi128_pd(x.v)));
+    }
+    intvec operator^(intvec x) const
+    {
+      return _mm_castpd_si128(_mm_xor_pd(_mm_castsi128_pd(v),
+                                         _mm_castsi128_pd(x.v)));
+    }
+    
+    intvec& operator&=(intvec const& x) { return *this=*this&x; }
+    intvec& operator|=(intvec const& x) { return *this=*this|x; }
+    intvec& operator^=(intvec const& x) { return *this=*this^x; }
+    
+    
+    
+    intvec lsr(int_t n) const { return _mm_srli_epi64(v, n); }
+    intvec operator>>(int_t n) const
+    {
+      // There is no _mm_srai_epi64. To emulate it, add 0x80000000
+      // before shifting, and subtract the shifted 0x80000000 after
+      // shifting
+      intvec x = *this;
+      // Convert signed to unsiged
+      x += U(1) << (bits-1);
+      // Shift
+      x = x.lsr(x);
+      // Undo conversion
+      x -= U(1) << (bits-n);
+      return x;
+    }
+    intvec operator<<(int_t n) const { return _mm_slli_epi64(v, n); }
+    intvec& operator>>=(int_t n) { return *this=*this>>n; }
+    intvec& operator<<=(int_t n) { return *this=*this<<n; }
+    
+    intvec lsr(intvec n) const { return _mm_srl_epi64(v, n.v); }
+    intvec operator>>(intvec n) const
+    {
+      // There is no _mm_srai_epi64. To emulate it, add 0x80000000
+      // before shifting, and subtract the shifted 0x80000000 after
+      // shifting
+      intvec x = *this;
+      // Convert signed to unsiged
+      x += U(1) << (bits-1);
+      // Shift
+      x = x.lsr(n);
+      // Undo conversion
+      x -= IV(1) << (IV(bits) - n);
+      return x;
+    }
+    intvec operator<<(intvec n) const { return _mm_sll_epi64(v, n.v); }
+    intvec& operator>>=(intvec n) { return *this=*this>>n; }
+    intvec& operator<<=(intvec n) { return *this=*this<<n; }
+  };
+  
+  
+  
+  template<>
+  struct realvec<double,2>: floatprops<double>
+  {
+    static int const size = 2;
+    typedef real_t scalar_t;
+    typedef __m128d vector_t;
+    
+    static constexpr char const* const name = "<SSE2:2*double>";
+    inline void barrier() { asm("": "+x" (v)); }
+    
+    static_assert(size * sizeof(real_t) == sizeof(vector_t),
+                  "vector size is wrong");
+    
+    typedef boolvec<real_t, size> boolvec_t;
+    typedef intvec<real_t, size> intvec_t;
+    typedef realvec realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    vector_t v;
+    
+    realvec() {}
+    realvec(realvec const& x): v(x.v) {}
+    realvec& operator=(realvec const& x) { return v=x.v, *this; }
+    realvec(vector_t x): v(x) {}
+    realvec(real_t a): v(_mm_set1_pd(a)) {}
+    realvec(real_t const* as): v(_mm_set_pd(as[1], as[0])) {}
+    
+    operator vector_t() const { return v; }
+    real_t operator[](int n) const { return ((real_t const*)&v)[n]; }
+    realvec& set_elt(int n, real_t a) { return ((real_t*)&v)[n]=a, *this; }
+    
+    
+    
+    intvec_t as_int() const { return _mm_castpd_si128(v); }
+    intvec_t convert_int() const { return MF::vml_convert_int(*this); }
+    
+    
+    
+    realvec operator+() const { return *this; }
+    realvec operator-() const { return RV(0.0) - *this; }
+    
+    realvec operator+(realvec x) const { return _mm_add_pd(v, x.v); }
+    realvec operator-(realvec x) const { return _mm_sub_pd(v, x.v); }
+    realvec operator*(realvec x) const { return _mm_mul_pd(v, x.v); }
+    realvec operator/(realvec x) const { return _mm_div_pd(v, x.v); }
+    
+    realvec& operator+=(realvec const& x) { return *this=*this+x; }
+    realvec& operator-=(realvec const& x) { return *this=*this-x; }
+    realvec& operator*=(realvec const& x) { return *this=*this*x; }
+    realvec& operator/=(realvec const& x) { return *this=*this/x; }
+    
+    real_t prod() const
+    {
+      return (*this)[0] * (*this)[1];
+    }
+    real_t sum() const
+    {
+      return (*this)[0] + (*this)[1];
+    }
+    
+    
+    
+    boolvec_t operator==(realvec const& x) const
+    {
+      return _mm_cmpeq_pd(v, x.v);
+    }
+    boolvec_t operator!=(realvec const& x) const
+    {
+      return _mm_cmpneq_pd(v, x.v);
+    }
+    boolvec_t operator<(realvec const& x) const
+    {
+      return _mm_cmplt_pd(v, x.v);
+    }
+    boolvec_t operator<=(realvec const& x) const
+    {
+      return _mm_cmple_pd(v, x.v);
+    }
+    boolvec_t operator>(realvec const& x) const
+    {
+      return _mm_cmpgt_pd(v, x.v);
+    }
+    boolvec_t operator>=(realvec const& x) const
+    {
+      return _mm_cmpge_pd(v, x.v);
+    }
+    
+    
+    
+    realvec acos() const { return MF::vml_acos(*this); }
+    realvec acosh() const { return MF::vml_acosh(*this); }
+    realvec asin() const { return MF::vml_asin(*this); }
+    realvec asinh() const { return MF::vml_asinh(*this); }
+    realvec atan() const { return MF::vml_atan(*this); }
+    realvec atanh() const { return MF::vml_atanh(*this); }
+    realvec ceil() const { return _mm_ceil_pd(v); }
+    realvec copysign(realvec y) const { return MF::vml_copysign(*this, y); }
+    realvec cos() const { return MF::vml_cos(*this); }
+    realvec cosh() const { return MF::vml_cosh(*this); }
+    realvec exp() const { return MF::vml_exp(*this); }
+    realvec exp10() const { return MF::vml_exp10(*this); }
+    realvec exp2() const { return MF::vml_exp2(*this); }
+    realvec expm1() const { return MF::vml_expm1(*this); }
+    realvec fabs() const { return MF::vml_fabs(*this); }
+    realvec floor() const { return _mm_floor_pd(v); }
+    realvec fmod(realvec y) const { return MF::vml_fmod(*this, y); }
+    intvec_t ilogb() const { return MF::vml_ilogb(*this); }
+    realvec log() const { return MF::vml_log(*this); }
+    realvec log10() const { return MF::vml_log10(*this); }
+    realvec log1p() const { return MF::vml_log1p(*this); }
+    realvec log2() const { return MF::vml_log2(*this); }
+    realvec pow(realvec y) const { return MF::vml_pow(*this, y); }
+    realvec rcp() const { return _mm_div_pd(_mm_set1_pd(1.0), v); }
+    realvec remainder(realvec y) const { return MF::vml_remainder(*this, y); }
+    realvec round() const { return _mm_round_pd(v, _MM_FROUND_NINT); }
+    realvec rsqrt() const { return MF::vml_rsqrt(*this); }
+    realvec scalbn(intvec_t n) const { return MF::vml_scalbn(*this, n); }
+    boolvec_t signbit() const { return v; }
+    realvec sin() const { return MF::vml_sin(*this); }
+    realvec sinh() const { return MF::vml_sinh(*this); }
+    realvec sqrt() const { return _mm_sqrt_pd(v); }
+    realvec tan() const { return MF::vml_tan(*this); }
+    realvec tanh() const { return MF::vml_tanh(*this); }
+  };
+  
+  
+  
+  // boolvec definitions
+  
+  inline
+  auto boolvec<double,2>::as_int() const -> intvec_t
+  {
+    return _mm_castpd_si128(v);
+  }
+  
+  inline
+  auto boolvec<double,2>::convert_int() const -> intvec_t
+  {
+    //return ifthen(v, U(1), U(0));
+    return lsr(as_int(), bits-1);
+  }
+  
+  inline
+  auto boolvec<double,2>::ifthen(intvec_t x, intvec_t y) const -> intvec_t
+  {
+    return ifthen(x.as_float(), y.as_float()).as_int();
+  }
+  
+  inline
+  auto boolvec<double,2>::ifthen(realvec_t x, realvec_t y) const -> realvec_t
+  {
+    return _mm_blendv_pd(y.v, x.v, v);
+  }
+
+  
+  
+  // intvec definitions
+  
+  inline auto intvec<double,2>::as_float() const -> realvec_t
+  {
+    return _mm_castsi128_pd(v);
+  }
+  
+  inline auto intvec<double,2>::convert_float() const -> realvec_t
+  {
+    return MF::vml_convert_float(*this);
+  }
+  
+} // namespace vecmathlib
+
+#endif  // #ifndef VEC_DOUBLE_SSE2_H
diff --git a/vec_float_avx.h b/vec_float_avx.h
new file mode 100644
index 0000000..4a19673
--- /dev/null
+++ b/vec_float_avx.h
@@ -0,0 +1,506 @@
+// -*-C++-*-
+
+#ifndef VEC_FLOAT_AVX_H
+#define VEC_FLOAT_AVX_H
+
+#include "floatprops.h"
+#include "mathfuncs.h"
+#include "vec_base.h"
+
+#include <cmath>
+
+// AVX intrinsics
+#include <immintrin.h>
+
+
+
+namespace vecmathlib {
+  
+  template<> struct boolvec<float,8>;
+  template<> struct intvec<float,8>;
+  template<> struct realvec<float,8>;
+  
+  
+  
+  template<>
+  struct boolvec<float,8>: floatprops<float>
+  {
+    static int const size = 8;
+    typedef bool scalar_t;
+    typedef __m256 bvector_t;
+    
+    static_assert(size * sizeof(real_t) == sizeof(bvector_t),
+                  "vector size is wrong");
+    
+  private:
+    // true values have the sign bit set, false values have it unset
+    static uint_t from_bool(bool a) { return - uint_t(a); }
+    static bool to_bool(uint_t a) { return int_t(a) < int_t(0); }
+  public:
+    
+    typedef boolvec boolvec_t;
+    typedef intvec<real_t, size> intvec_t;
+    typedef realvec<real_t, size> realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    bvector_t v;
+    
+    boolvec() {}
+    boolvec(boolvec const& x): v(x.v) {}
+    boolvec& operator=(boolvec const& x) { return v=x.v, *this; }
+    boolvec(bvector_t x): v(x) {}
+    boolvec(bool a):
+    v(_mm256_castsi256_ps(_mm256_set1_epi32(from_bool(a)))) {}
+    boolvec(bool const* as):
+    v(_mm256_castsi256_ps(_mm256_set_epi32(from_bool(as[7]),
+                                           from_bool(as[6]),
+                                           from_bool(as[5]),
+                                           from_bool(as[4]),
+                                           from_bool(as[3]),
+                                           from_bool(as[2]),
+                                           from_bool(as[1]),
+                                           from_bool(as[0])))) {}
+    
+    operator bvector_t() const { return v; }
+    bool operator[](int n) const { return to_bool(((uint_t const*)&v)[n]); }
+    boolvec& set_elt(int n, bool a)
+    {
+      return ((int_t*)&v)[n] = from_bool(a), *this;
+    }
+    
+    
+    
+    auto as_int() const -> intvec_t;      // defined after intvec
+    auto convert_int() const -> intvec_t; // defined after intvec
+    
+    
+    
+    boolvec operator!() const { return _mm256_xor_ps(boolvec(true), v); }
+    
+    boolvec operator&&(boolvec x) const { return _mm256_and_ps(v, x.v); }
+    boolvec operator||(boolvec x) const { return _mm256_or_ps(v, x.v); }
+    boolvec operator==(boolvec x) const { return !(*this==x); }
+    boolvec operator!=(boolvec x) const { return _mm256_xor_ps(v, x.v); }
+    
+    bool all() const
+    {
+      return (*this)[0] && (*this)[1] && (*this)[2] && (*this)[3];
+    }
+    bool any() const
+    {
+      return (*this)[0] || (*this)[1] || (*this)[2] || (*this)[3];
+    }
+    
+    
+    
+    // ifthen(condition, then-value, else-value)
+    auto ifthen(intvec_t x,
+                intvec_t y) const -> intvec_t; // defined after intvec
+    auto ifthen(realvec_t x,
+                realvec_t y) const -> realvec_t; // defined after realvec
+    
+  };
+  
+  
+  
+  template<>
+  struct intvec<float,8>: floatprops<float>
+  {
+    static int const size = 8;
+    typedef int_t scalar_t;
+    typedef __m256i ivector_t;
+    
+    static_assert(size * sizeof(real_t) == sizeof(ivector_t),
+                  "vector size is wrong");
+    
+    typedef boolvec<real_t, size> boolvec_t;
+    typedef intvec intvec_t;
+    typedef realvec<real_t, size> realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    ivector_t v;
+    
+    intvec() {}
+    intvec(intvec const& x): v(x.v) {}
+    intvec& operator=(intvec const& x) { return v=x.v, *this; }
+    intvec(ivector_t x): v(x) {}
+    intvec(int_t a): v(_mm256_set1_epi32(a)) {}
+    intvec(int_t const* as): v(_mm256_set_epi32(as[7], as[6], as[5], as[4],
+                                                as[3], as[2], as[1], as[0])) {}
+    
+    operator ivector_t() const { return v; }
+    int_t operator[](int n) const { return ((int_t const*)&v)[n]; }
+    intvec& set_elt(int n, int_t a) { return ((int_t*)&v)[n]=a, *this; }
+    
+    
+    
+    auto as_bool() const -> boolvec_t { return _mm256_castsi256_ps(v); }
+    auto convert_bool() const -> boolvec_t
+    {
+      // Result: convert_bool(0)=false, convert_bool(else)=true
+      // There is no intrinsic to compare with zero. Instead, we check
+      // whether x is positive and x-1 is negative.
+      intvec x = *this;
+      intvec xm1 = x - 1;
+      // We know that boolvec values depend only on the sign bit
+      // return (~xm1 | x).as_bool();
+      return x.as_bool() || !xm1.as_bool();
+    }
+    auto as_float() const -> realvec_t; // defined after realvec
+    auto convert_float() const -> realvec_t; // defined after realvec
+    
+    
+    
+    // Note: not all arithmetic operations are supported!
+    
+    intvec operator+() const { return *this; }
+    intvec operator-() const { return IV(0) - *this; }
+    
+    intvec operator+(intvec x) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i xvlo = _mm256_castsi256_si128(x.v);
+      __m128i xvhi = _mm256_extractf128_si256(x.v, 1);
+      vlo = _mm_add_epi32(vlo, xvlo);
+      vhi = _mm_add_epi32(vhi, xvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec operator-(intvec x) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i xvlo = _mm256_castsi256_si128(x.v);
+      __m128i xvhi = _mm256_extractf128_si256(x.v, 1);
+      vlo = _mm_sub_epi32(vlo, xvlo);
+      vhi = _mm_sub_epi32(vhi, xvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    
+    intvec& operator+=(intvec const& x) { return *this=*this+x; }
+    intvec& operator-=(intvec const& x) { return *this=*this-x; }
+    
+    
+    
+    intvec operator~() const { return IV(~U(0)) ^ *this; }
+    
+    intvec operator&(intvec x) const
+    {
+      return _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(v),
+                                               _mm256_castsi256_ps(x.v)));
+    }
+    intvec operator|(intvec x) const
+    {
+      return _mm256_castps_si256(_mm256_or_ps(_mm256_castsi256_ps(v),
+                                              _mm256_castsi256_ps(x.v)));
+    }
+    intvec operator^(intvec x) const
+    {
+      return _mm256_castps_si256(_mm256_xor_ps(_mm256_castsi256_ps(v),
+                                               _mm256_castsi256_ps(x.v)));
+    }
+    
+    intvec& operator&=(intvec const& x) { return *this=*this&x; }
+    intvec& operator|=(intvec const& x) { return *this=*this|x; }
+    intvec& operator^=(intvec const& x) { return *this=*this^x; }
+    
+    
+    
+    intvec lsr(int_t n) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      vlo = _mm_srli_epi32(vlo, n);
+      vhi = _mm_srli_epi32(vhi, n);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec operator>>(int_t n) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      // There is no _mm_srai_epi32. To emulate it, add 0x80000000
+      // before shifting, and subtract the shifted 0x80000000 after
+      // shifting
+      // Convert signed to unsiged
+      vlo += _mm_set1_epi32(U(1) << (bits-1));
+      vhi += _mm_set1_epi32(U(1) << (bits-1));
+      // Shift
+      vlo = _mm_srli_epi32(vlo, n);
+      vhi = _mm_srli_epi32(vhi, n);
+      // Undo conversion
+      vlo -= _mm_set1_epi32(U(1) << (bits-n));
+      vhi -= _mm_set1_epi32(U(1) << (bits-n));
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec operator<<(int_t n) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      vlo = _mm_slli_epi32(vlo, n);
+      vhi = _mm_slli_epi32(vhi, n);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec& operator>>=(int_t n) { return *this=*this>>n; }
+    intvec& operator<<=(int_t n) { return *this=*this<<n; }
+    
+    intvec lsr(intvec n) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i nvlo = _mm256_castsi256_si128(n.v);
+      __m128i nvhi = _mm256_extractf128_si256(n.v, 1);
+      vlo = _mm_srl_epi32(vlo, nvlo);
+      vhi = _mm_srl_epi32(vhi, nvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec operator>>(intvec n) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i nvlo = _mm256_castsi256_si128(n.v);
+      __m128i nvhi = _mm256_extractf128_si256(n.v, 1);
+      // Convert signed to unsiged
+      vlo += _mm_set1_epi32(U(1) << (bits-1));
+      vhi += _mm_set1_epi32(U(1) << (bits-1));
+      // Shift
+      vlo = _mm_srl_epi32(vlo, nvlo);
+      vhi = _mm_srl_epi32(vhi, nvhi);
+      // Undo conversion
+      vlo -= _mm_sll_epi32(_mm_set1_epi32(1),
+                           _mm_sub_epi32(_mm_set1_epi32(bits), nvlo));
+      vhi -= _mm_sll_epi32(_mm_set1_epi32(1),
+                           _mm_sub_epi32(_mm_set1_epi32(bits), nvhi));
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec operator<<(intvec n) const
+    {
+      __m128i vlo = _mm256_castsi256_si128(v);
+      __m128i vhi = _mm256_extractf128_si256(v, 1);
+      __m128i nvlo = _mm256_castsi256_si128(n.v);
+      __m128i nvhi = _mm256_extractf128_si256(n.v, 1);
+      vlo = _mm_sll_epi32(vlo, nvlo);
+      vhi = _mm_sll_epi32(vhi, nvhi);
+      return _mm256_insertf128_si256(_mm256_castsi128_si256(vlo), vhi, 1);
+    }
+    intvec& operator>>=(intvec n) { return *this=*this>>n; }
+    intvec& operator<<=(intvec n) { return *this=*this<<n; }
+  };
+  
+  
+  
+  template<>
+  struct realvec<float,8>: floatprops<float>
+  {
+    static int const size = 8;
+    typedef real_t scalar_t;
+    typedef __m256 vector_t;
+    
+    static constexpr char const* const name = "<AVX:8*float>";
+    inline void barrier() { asm("": "+x" (v)); }
+    
+    static_assert(size * sizeof(real_t) == sizeof(vector_t),
+                  "vector size is wrong");
+    
+    typedef boolvec<real_t, size> boolvec_t;
+    typedef intvec<real_t, size> intvec_t;
+    typedef realvec realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    vector_t v;
+    
+    realvec() {}
+    realvec(realvec const& x): v(x.v) {}
+    realvec& operator=(realvec const& x) { return v=x.v, *this; }
+    realvec(vector_t x): v(x) {}
+    realvec(real_t a): v(_mm256_set1_ps(a)) {}
+    realvec(real_t const* as): v(_mm256_set_ps(as[7], as[6], as[5], as[4],
+                                               as[3], as[2], as[1], as[0])) {}
+    
+    operator vector_t() const { return v; }
+    real_t operator[](int n) const { return ((real_t const*)&v)[n]; }
+    realvec& set_elt(int n, real_t a) { return ((real_t*)&v)[n]=a, *this; }
+    
+    
+    
+    intvec_t as_int() const { return _mm256_castps_si256(v); }
+    intvec_t convert_int() const { return _mm256_cvtps_epi32(v); }
+    
+    
+    
+    realvec operator+() const { return *this; }
+    realvec operator-() const { return RV(0.0) - *this; }
+    
+    realvec operator+(realvec x) const { return _mm256_add_ps(v, x.v); }
+    realvec operator-(realvec x) const { return _mm256_sub_ps(v, x.v); }
+    realvec operator*(realvec x) const { return _mm256_mul_ps(v, x.v); }
+    realvec operator/(realvec x) const { return _mm256_div_ps(v, x.v); }
+    
+    realvec& operator+=(realvec const& x) { return *this=*this+x; }
+    realvec& operator-=(realvec const& x) { return *this=*this-x; }
+    realvec& operator*=(realvec const& x) { return *this=*this*x; }
+    realvec& operator/=(realvec const& x) { return *this=*this/x; }
+    
+    real_t prod() const
+    {
+      return (*this)[0] * (*this)[1] * (*this)[2] * (*this)[3];
+    }
+    real_t sum() const
+    {
+      return (*this)[0] + (*this)[1] + (*this)[2] + (*this)[3];
+    }
+    
+    
+    
+    boolvec_t operator==(realvec const& x) const
+    {
+      return _mm256_cmp_ps(v, x.v, _CMP_EQ_OQ);
+    }
+    boolvec_t operator!=(realvec const& x) const
+    {
+      return _mm256_cmp_ps(v, x.v, _CMP_NEQ_OQ);
+    }
+    boolvec_t operator<(realvec const& x) const
+    {
+      return _mm256_cmp_ps(v, x.v, _CMP_LT_OQ);
+    }
+    boolvec_t operator<=(realvec const& x) const
+    {
+      return _mm256_cmp_ps(v, x.v, _CMP_LE_OQ);
+    }
+    boolvec_t operator>(realvec const& x) const
+    {
+      return _mm256_cmp_ps(v, x.v, _CMP_GT_OQ);
+    }
+    boolvec_t operator>=(realvec const& x) const
+    {
+      return _mm256_cmp_ps(v, x.v, _CMP_GE_OQ);
+    }
+    
+    
+    
+    realvec acos() const { return MF::vml_acos(*this); }
+    realvec acosh() const { return MF::vml_acosh(*this); }
+    realvec asin() const { return MF::vml_asin(*this); }
+    realvec asinh() const { return MF::vml_asinh(*this); }
+    realvec atan() const { return MF::vml_atan(*this); }
+    realvec atanh() const { return MF::vml_atanh(*this); }
+    realvec ceil() const { return _mm256_ceil_ps(v); }
+    realvec copysign(realvec y) const { return MF::vml_copysign(*this, y); }
+    realvec cos() const { return MF::vml_cos(*this); }
+    realvec cosh() const { return MF::vml_cosh(*this); }
+    realvec exp() const { return MF::vml_exp(*this); }
+    realvec exp10() const { return MF::vml_exp10(*this); }
+    realvec exp2() const { return MF::vml_exp2(*this); }
+    realvec expm1() const { return MF::vml_expm1(*this); }
+    realvec fabs() const { return MF::vml_fabs(*this); }
+    realvec floor() const { return _mm256_floor_ps(v); }
+    realvec fmod(realvec y) const { return MF::vml_fmod(*this, y); }
+    intvec_t ilogb() const { return MF::vml_ilogb(*this); }
+    realvec log() const { return MF::vml_log(*this); }
+    realvec log10() const { return MF::vml_log10(*this); }
+    realvec log1p() const { return MF::vml_log1p(*this); }
+    realvec log2() const { return MF::vml_log2(*this); }
+    realvec pow(realvec y) const { return MF::vml_pow(*this, y); }
+    realvec rcp() const
+    {
+      realvec x = *this;
+      realvec r = _mm256_rcp_ps(x); // this is only an approximation
+      r *= RV(2.0) - r*x;        // one Newton iteration (see vml_rcp)
+      return r;
+    }
+    realvec remainder(realvec y) const { return MF::vml_remainder(*this, y); }
+    realvec round() const { return _mm256_round_ps(v, _MM_FROUND_NINT); }
+    realvec rsqrt() const
+    {
+      realvec x = *this;
+      realvec r = _mm256_rsqrt_ps(x);    // this is only an approximation
+      r *= RV(1.5) - RV(0.5)*x * r*r; // one Newton iteration (see vml_rsqrt)
+      return r;
+    }
+    realvec scalbn(intvec_t n) const { return MF::vml_scalbn(*this, n); }
+    boolvec_t signbit() const { return v; }
+    realvec sin() const { return MF::vml_sin(*this); }
+    realvec sinh() const { return MF::vml_sinh(*this); }
+    realvec sqrt() const { return _mm256_sqrt_ps(v); }
+    realvec tan() const { return MF::vml_tan(*this); }
+    realvec tanh() const { return MF::vml_tanh(*this); }
+  };
+  
+  
+  
+  // boolvec definitions
+  
+  inline
+  auto boolvec<float,8>::as_int() const -> intvec_t
+  {
+    return _mm256_castps_si256(v);
+  }
+  
+  inline
+  auto boolvec<float,8>::convert_int() const -> intvec_t
+  {
+    return lsr(as_int(), bits-1);
+  }
+  
+  inline
+  auto boolvec<float,8>::ifthen(intvec_t x, intvec_t y) const -> intvec_t
+  {
+    return ifthen(x.as_float(), y.as_float()).as_int();
+  }
+  
+  inline
+  auto boolvec<float,8>::ifthen(realvec_t x, realvec_t y) const -> realvec_t
+  {
+    return _mm256_blendv_ps(y.v, x.v, v);
+  }
+
+  
+  
+  // intvec definitions
+  
+  inline auto intvec<float,8>::as_float() const -> realvec_t
+  {
+    return _mm256_castsi256_ps(v);
+  }
+  
+  inline auto intvec<float,8>::convert_float() const -> realvec_t
+  {
+    return _mm256_cvtepi32_ps(v);
+  }
+  
+} // namespace vecmathlib
+
+#endif  // #ifndef VEC_FLOAT_AVX_H
diff --git a/vec_float_sse2.h b/vec_float_sse2.h
new file mode 100644
index 0000000..f99eae8
--- /dev/null
+++ b/vec_float_sse2.h
@@ -0,0 +1,415 @@
+// -*-C++-*-
+
+#ifndef VEC_FLOAT_SSE2_H
+#define VEC_FLOAT_SSE2_H
+
+#include "floatprops.h"
+#include "mathfuncs.h"
+#include "vec_base.h"
+
+#include <cmath>
+
+// SSE2 intrinsics
+#include <xmmintrin.h>
+#if defined __SSE4_1__          // Intel's SSE 4.1
+#  include <smmintrin.h>
+#endif
+#if defined __SSE4A__           // AMD's SSE 4a
+#  include <ammintrin.h>
+#endif
+
+
+
+namespace vecmathlib {
+  
+  template<> struct boolvec<float,4>;
+  template<> struct intvec<float,4>;
+  template<> struct realvec<float,4>;
+  
+  
+  
+  template<>
+  struct boolvec<float,4>: floatprops<float>
+  {
+    static int const size = 4;
+    typedef bool scalar_t;
+    typedef __m128 bvector_t;
+    
+    static_assert(size * sizeof(real_t) == sizeof(bvector_t),
+                  "vector size is wrong");
+    
+  private:
+    // true values have the sign bit set, false values have it unset
+    static uint_t from_bool(bool a) { return - uint_t(a); }
+    static bool to_bool(uint_t a) { return int_t(a) < int_t(0); }
+  public:
+    
+    typedef boolvec boolvec_t;
+    typedef intvec<real_t, size> intvec_t;
+    typedef realvec<real_t, size> realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    bvector_t v;
+    
+    boolvec() {}
+    boolvec(boolvec const& x): v(x.v) {}
+    boolvec& operator=(boolvec const& x) { return v=x.v, *this; }
+    boolvec(bvector_t x): v(x) {}
+    boolvec(bool a):
+    v(_mm_castsi128_ps(_mm_set1_epi32(from_bool(a)))) {}
+    boolvec(bool const* as):
+    v(_mm_castsi128_ps(_mm_set_epi32(from_bool(as[3]),
+                                     from_bool(as[2]),
+                                     from_bool(as[1]),
+                                     from_bool(as[0])))) {}
+    
+    operator bvector_t() const { return v; }
+    bool operator[](int n) const { return to_bool(((uint_t const*)&v)[n]); }
+    boolvec& set_elt(int n, bool a)
+    {
+      return ((int_t*)&v)[n] = from_bool(a), *this;
+    }
+    
+    
+    
+    auto as_int() const -> intvec_t;      // defined after intvec
+    auto convert_int() const -> intvec_t; // defined after intvec
+    
+    
+    
+    boolvec operator!() const { return _mm_xor_ps(boolvec(true), v); }
+    
+    boolvec operator&&(boolvec x) const { return _mm_and_ps(v, x.v); }
+    boolvec operator||(boolvec x) const { return _mm_or_ps(v, x.v); }
+    boolvec operator==(boolvec x) const { return !(*this==x); }
+    boolvec operator!=(boolvec x) const { return _mm_xor_ps(v, x.v); }
+    
+    bool all() const
+    {
+      return (*this)[0] && (*this)[1] && (*this)[2] && (*this)[3];
+    }
+    bool any() const
+    {
+      return (*this)[0] || (*this)[1] || (*this)[2] || (*this)[3];
+    }
+    
+    
+    
+    // ifthen(condition, then-value, else-value)
+    auto ifthen(intvec_t x,
+                intvec_t y) const -> intvec_t; // defined after intvec
+    auto ifthen(realvec_t x,
+                realvec_t y) const -> realvec_t; // defined after realvec
+    
+  };
+  
+  
+  
+  template<>
+  struct intvec<float,4>: floatprops<float>
+  {
+    static int const size = 4;
+    typedef int_t scalar_t;
+    typedef __m128i ivector_t;
+    
+    static_assert(size * sizeof(real_t) == sizeof(ivector_t),
+                  "vector size is wrong");
+    
+    typedef boolvec<real_t, size> boolvec_t;
+    typedef intvec intvec_t;
+    typedef realvec<real_t, size> realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    ivector_t v;
+    
+    intvec() {}
+    intvec(intvec const& x): v(x.v) {}
+    intvec& operator=(intvec const& x) { return v=x.v, *this; }
+    intvec(ivector_t x): v(x) {}
+    intvec(int_t a): v(_mm_set1_epi32(a)) {}
+    intvec(int_t const* as): v(_mm_set_epi32(as[3], as[2], as[1], as[0])) {}
+    
+    operator ivector_t() const { return v; }
+    int_t operator[](int n) const { return ((int_t const*)&v)[n]; }
+    intvec& set_elt(int n, int_t a) { return ((int_t*)&v)[n]=a, *this; }
+    
+    
+    
+    auto as_bool() const -> boolvec_t { return _mm_castsi128_ps(v); }
+    auto convert_bool() const -> boolvec_t
+    {
+      // Result: convert_bool(0)=false, convert_bool(else)=true
+      return ! IV(_mm_cmpeq_epi32(v, IV(0))).as_bool();
+    }
+    auto as_float() const -> realvec_t; // defined after realvec
+    auto convert_float() const -> realvec_t; // defined after realvec
+    
+    
+    
+    // Note: not all arithmetic operations are supported!
+    
+    intvec operator+() const { return *this; }
+    intvec operator-() const { return IV(0) - *this; }
+    
+    intvec operator+(intvec x) const { return _mm_add_epi32(v, x.v); }
+    intvec operator-(intvec x) const { return _mm_sub_epi32(v, x.v); }
+    
+    intvec& operator+=(intvec const& x) { return *this=*this+x; }
+    intvec& operator-=(intvec const& x) { return *this=*this-x; }
+    
+    
+    
+    intvec operator~() const { return IV(~U(0)) ^ *this; }
+    
+    intvec operator&(intvec x) const
+    {
+      return _mm_castps_si128(_mm_and_ps(_mm_castsi128_ps(v),
+                                         _mm_castsi128_ps(x.v)));
+    }
+    intvec operator|(intvec x) const
+    {
+      return _mm_castps_si128(_mm_or_ps(_mm_castsi128_ps(v),
+                                        _mm_castsi128_ps(x.v)));
+    }
+    intvec operator^(intvec x) const
+    {
+      return _mm_castps_si128(_mm_xor_ps(_mm_castsi128_ps(v),
+                                         _mm_castsi128_ps(x.v)));
+    }
+    
+    intvec& operator&=(intvec const& x) { return *this=*this&x; }
+    intvec& operator|=(intvec const& x) { return *this=*this|x; }
+    intvec& operator^=(intvec const& x) { return *this=*this^x; }
+    
+    
+    
+    intvec lsr(int_t n) const { return _mm_srli_epi32(v, n); }
+    intvec operator>>(int_t n) const { return _mm_srai_epi32(v, n); }
+    intvec operator<<(int_t n) const { return _mm_slli_epi32(v, n); }
+    intvec& operator>>=(int_t n) { return *this=*this>>n; }
+    intvec& operator<<=(int_t n) { return *this=*this<<n; }
+    
+    intvec lsr(intvec n) const { return _mm_srl_epi32(v, n.v); }
+    intvec operator>>(intvec n) const { return _mm_sra_epi32(v, n.v); }
+    intvec operator<<(intvec n) const { return _mm_sll_epi32(v, n.v); }
+    intvec& operator>>=(intvec n) { return *this=*this>>n; }
+    intvec& operator<<=(intvec n) { return *this=*this<<n; }
+  };
+  
+  
+  
+  template<>
+  struct realvec<float,4>: floatprops<float>
+  {
+    static int const size = 4;
+    typedef real_t scalar_t;
+    typedef __m128 vector_t;
+    
+    static constexpr char const* const name = "<SSE2:4*float>";
+    inline void barrier() { asm("": "+x" (v)); }
+    
+    static_assert(size * sizeof(real_t) == sizeof(vector_t),
+                  "vector size is wrong");
+    
+    typedef boolvec<real_t, size> boolvec_t;
+    typedef intvec<real_t, size> intvec_t;
+    typedef realvec realvec_t;
+    
+    // Short names for type casts
+    typedef real_t R;
+    typedef int_t I;
+    typedef uint_t U;
+    typedef realvec_t RV;
+    typedef intvec_t IV;
+    typedef boolvec_t BV;
+    typedef floatprops<real_t> FP;
+    typedef mathfuncs<realvec_t> MF;
+    
+    
+    
+    vector_t v;
+    
+    realvec() {}
+    realvec(realvec const& x): v(x.v) {}
+    realvec& operator=(realvec const& x) { return v=x.v, *this; }
+    realvec(vector_t x): v(x) {}
+    realvec(real_t a): v(_mm_set1_ps(a)) {}
+    realvec(real_t const* as): v(_mm_set_ps(as[3], as[2], as[1], as[0])) {}
+    
+    operator vector_t() const { return v; }
+    real_t operator[](int n) const { return ((real_t const*)&v)[n]; }
+    realvec& set_elt(int n, real_t a) { return ((real_t*)&v)[n]=a, *this; }
+    
+    
+    
+    intvec_t as_int() const { return _mm_castps_si128(v); }
+    intvec_t convert_int() const { return _mm_cvtps_epi32(v); }
+    
+    
+    
+    realvec operator+() const { return *this; }
+    realvec operator-() const { return RV(0.0) - *this; }
+    
+    realvec operator+(realvec x) const { return _mm_add_ps(v, x.v); }
+    realvec operator-(realvec x) const { return _mm_sub_ps(v, x.v); }
+    realvec operator*(realvec x) const { return _mm_mul_ps(v, x.v); }
+    realvec operator/(realvec x) const { return _mm_div_ps(v, x.v); }
+    
+    realvec& operator+=(realvec const& x) { return *this=*this+x; }
+    realvec& operator-=(realvec const& x) { return *this=*this-x; }
+    realvec& operator*=(realvec const& x) { return *this=*this*x; }
+    realvec& operator/=(realvec const& x) { return *this=*this/x; }
+    
+    real_t prod() const
+    {
+      return (*this)[0] * (*this)[1] * (*this)[2] * (*this)[3];
+    }
+    real_t sum() const
+    {
+      return (*this)[0] + (*this)[1] + (*this)[2] + (*this)[3];
+    }
+    
+    
+    
+    boolvec_t operator==(realvec const& x) const
+    {
+      return _mm_cmpeq_ps(v, x.v);
+    }
+    boolvec_t operator!=(realvec const& x) const
+    {
+      return _mm_cmpneq_ps(v, x.v);
+    }
+    boolvec_t operator<(realvec const& x) const
+    {
+      return _mm_cmplt_ps(v, x.v);
+    }
+    boolvec_t operator<=(realvec const& x) const
+    {
+      return _mm_cmple_ps(v, x.v);
+    }
+    boolvec_t operator>(realvec const& x) const
+    {
+      return _mm_cmpgt_ps(v, x.v);
+    }
+    boolvec_t operator>=(realvec const& x) const
+    {
+      return _mm_cmpge_ps(v, x.v);
+    }
+    
+    
+    
+    realvec acos() const { return MF::vml_acos(*this); }
+    realvec acosh() const { return MF::vml_acosh(*this); }
+    realvec asin() const { return MF::vml_asin(*this); }
+    realvec asinh() const { return MF::vml_asinh(*this); }
+    realvec atan() const { return MF::vml_atan(*this); }
+    realvec atanh() const { return MF::vml_atanh(*this); }
+    realvec ceil() const { return _mm_ceil_ps(v); }
+    realvec copysign(realvec y) const { return MF::vml_copysign(*this, y); }
+    realvec cos() const { return MF::vml_cos(*this); }
+    realvec cosh() const { return MF::vml_cosh(*this); }
+    realvec exp() const { return MF::vml_exp(*this); }
+    realvec exp10() const { return MF::vml_exp10(*this); }
+    realvec exp2() const { return MF::vml_exp2(*this); }
+    realvec expm1() const { return MF::vml_expm1(*this); }
+    realvec fabs() const { return MF::vml_fabs(*this); }
+    realvec floor() const { return _mm_floor_ps(v); }
+    realvec fmod(realvec y) const { return MF::vml_fmod(*this, y); }
+    intvec_t ilogb() const { return MF::vml_ilogb(*this); }
+    realvec log() const { return MF::vml_log(*this); }
+    realvec log10() const { return MF::vml_log10(*this); }
+    realvec log1p() const { return MF::vml_log1p(*this); }
+    realvec log2() const { return MF::vml_log2(*this); }
+    realvec pow(realvec y) const { return MF::vml_pow(*this, y); }
+    realvec rcp() const
+    {
+      realvec x = *this;
+      realvec r = _mm_rcp_ps(x); // this is only an approximation
+      r *= RV(2.0) - r*x;        // one Newton iteration (see vml_rcp)
+      return r;
+    }
+    realvec remainder(realvec y) const { return MF::vml_remainder(*this, y); }
+    realvec round() const { return _mm_round_ps(v, _MM_FROUND_NINT); }
+    realvec rsqrt() const
+    {
+      realvec x = *this;
+      realvec r = _mm_rsqrt_ps(x);    // this is only an approximation
+      r *= RV(1.5) - RV(0.5)*x * r*r; // one Newton iteration (see vml_rsqrt)
+      return r;
+    }
+    realvec scalbn(intvec_t n) const { return MF::vml_scalbn(*this, n); }
+    boolvec_t signbit() const { return v; }
+    realvec sin() const { return MF::vml_sin(*this); }
+    realvec sinh() const { return MF::vml_sinh(*this); }
+    realvec sqrt() const { return _mm_sqrt_ps(v); }
+    realvec tan() const { return MF::vml_tan(*this); }
+    realvec tanh() const { return MF::vml_tanh(*this); }
+  };
+  
+  
+  
+  // boolvec definitions
+  
+  inline
+  auto boolvec<float,4>::as_int() const -> intvec_t
+  {
+    return _mm_castps_si128(v);
+  }
+  
+  inline
+  auto boolvec<float,4>::convert_int() const -> intvec_t
+  {
+    return lsr(as_int(), bits-1);
+  }
+  
+  inline
+  auto boolvec<float,4>::ifthen(intvec_t x, intvec_t y) const -> intvec_t
+  {
+    return ifthen(x.as_float(), y.as_float()).as_int();
+  }
+  
+  inline
+  auto boolvec<float,4>::ifthen(realvec_t x, realvec_t y) const -> realvec_t
+  {
+    return _mm_blendv_ps(y.v, x.v, v);
+  }
+
+  
+  
+  // intvec definitions
+  
+  inline auto intvec<float,4>::as_float() const -> realvec_t
+  {
+    return _mm_castsi128_ps(v);
+  }
+  
+  inline auto intvec<float,4>::convert_float() const -> realvec_t
+  {
+    return _mm_cvtepi32_ps(v);
+  }
+  
+} // namespace vecmathlib
+
+#endif  // #ifndef VEC_FLOAT_SSE2_H
diff --git a/vecmathlib.h b/vecmathlib.h
index b5703bb..83396ad 100644
--- a/vecmathlib.h
+++ b/vecmathlib.h
@@ -4,7 +4,17 @@
 #define VECMATHLIB_H
 
 #include "vec_float.h"
+#if defined __SSE2__            // Intel SSE 2
+#  include "vec_float_sse2.h"
+#endif
+#if defined __AVX__             // Intel AVX
+#  include "vec_float_avx.h"
+#endif
+
 #include "vec_double.h"
+#if defined __SSE2__            // Intel SSE 2
+#  include "vec_double_sse2.h"
+#endif
 #if defined __AVX__             // Intel AVX
 #  include "vec_double_avx.h"
 #endif