From a1c3f5d8372f6c217d1ef26f11ba6085d9a45ea6 Mon Sep 17 00:00:00 2001
From: Artem Zyktin <zyktin.a.a@gmail.com>
Date: Tue, 2 Jun 2026 05:12:21 +0800
Subject: [PATCH] Optimize dot3

---
 src/tenvex/expressions/core.h | 10 ++++++++--
 src/tenvex/expressions/vec4.h |  3 +++
 2 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/src/tenvex/expressions/core.h b/src/tenvex/expressions/core.h
index 813e718..50ef0d7 100644
--- a/src/tenvex/expressions/core.h
+++ b/src/tenvex/expressions/core.h
@@ -66,13 +66,19 @@ vf4 div(vf4 l, vf4 r) noexcept
 [[nodiscard]] TNVX_INLINE
 vf4 dot3(vf4 l, vf4 r) noexcept
 {
-	return _mm_dp_ps(l, r, 0b01111111);
+	//return _mm_dp_ps(l, r, 0b01111111);
+	vf4 t   = _mm_mul_ps(l, r);
+	vf4 ty  = _mm_shuffle_ps(t, t, _MM_SHUFFLE(1, 1, 1, 1));
+	vf4 xy  = _mm_add_ss(t, ty);
+	vf4 tz  = _mm_movehl_ps(t, t);
+	vf4 xyz = _mm_add_ss(xy, tz);
+	return    _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(0, 0, 0, 0));
 }
 
 [[nodiscard]] TNVX_INLINE
 vf4 magnitude3(vf4 v) noexcept
 {
-	return _mm_sqrt_ps(_mm_dp_ps(v, v, 0b01111111));
+	return _mm_sqrt_ps(dot3(v, v));
 }
 
 [[nodiscard]] TNVX_INLINE
diff --git a/src/tenvex/expressions/vec4.h b/src/tenvex/expressions/vec4.h
index 30be4e7..82960ec 100644
--- a/src/tenvex/expressions/vec4.h
+++ b/src/tenvex/expressions/vec4.h
@@ -35,6 +35,9 @@ struct alignas(16) vec4 : Expr<vec4>
 	[[nodiscard]] TNVX_INLINE
 	vf4 eval() const noexcept;
 
+	[[nodiscard]] TNVX_INLINE
+	vec4 operator-() const noexcept;
+
 private:
 
 	vf4 _data;