Add dot4 expression

src/tenvex/expressions/core.h

@@ -87,6 +87,14 @@ vf4 dot3(vf4 l, vf4 r) noexcept
 	return    _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(0, 0, 0, 0));
 }
 
+[[nodiscard]] TNVX_INLINE
+vf4 dot4(vf4 l, vf4 r) noexcept
+{
+	vf4 t = _mm_mul_ps(l, r);
+	vf4 h = _mm_hadd_ps(t, t);
+	return  _mm_hadd_ps(h, h);
+}
+
 [[nodiscard]] TNVX_INLINE
 vf4 magnitude3(vf4 v) noexcept
 {

src/tenvex/expressions/dot4.h

@@ -0,0 +1,43 @@
+#pragma once
+
+#include "core.h"
+#include "expression.h"
+#include "traits.h"
+#include "concepts.h"
+
+namespace tnvx
+{
+
+template<vec_expr L, vec_expr R>
+struct Dot4 : Expr<Dot4<L, R>>
+{
+	TNVX_INLINE
+	Dot4(const L& TNVX_RESTRICT l, const R& TNVX_RESTRICT r) noexcept;
+
+	[[nodiscard]] TNVX_INLINE
+	vf4 eval() const noexcept;
+
+	[[nodiscard]] TNVX_INLINE
+	operator float() const noexcept;
+
+private:
+	tnvx_ref_or_value_t<L> _l;
+	tnvx_ref_or_value_t<R> _r;
+};
+
+template<vec_expr L, vec_expr R> inline constexpr bool is_scalar_expr<Dot4<L, R>> = true;
+
+template<vec_expr L, vec_expr R>
+[[nodiscard]] TNVX_INLINE
+Dot3<L, R> dot4(const L& l, const R& r) noexcept;
+
+template<vec_expr L, vec_expr R>
+TNVX_INLINE
+Dot3<L, R> dot4(const L& l, const R& r) noexcept
+{
+	return { l, r };
+}
+
+}
+
+#include "dot4_impl.hpp"

src/tenvex/expressions/dot4_impl.hpp

@@ -0,0 +1,30 @@
+#pragma once
+
+#include "dot4.h"
+
+namespace tnvx
+{
+
+template<vec_expr L, vec_expr R>
+TNVX_INLINE
+Dot4<L, R>::Dot4(const L& TNVX_RESTRICT l, const R& TNVX_RESTRICT r) noexcept
+	: _l(l)
+	, _r(r)
+{
+}
+
+template<vec_expr L, vec_expr R>
+TNVX_INLINE
+vf4 Dot4<L, R>::eval() const noexcept
+{
+	return detail::dot4(_l.eval(), _r.eval());
+}
+
+template<vec_expr L, vec_expr R>
+TNVX_INLINE
+Dot4<L, R>::operator float() const noexcept
+{
+	return detail::to_float(eval());
+}
+
+}

src/tenvex/tenvex.h

@@ -7,6 +7,7 @@
 #include "expressions/mul.h"
 #include "expressions/div.h"
 #include "expressions/dot3.h"
+#include "expressions/dot4.h"
 #include "expressions/magnitude3.h"
 #include "expressions/norm3.h"
 #include "expressions/cross3.h"