Use FMA and rsqrt in more places

Author: GabrielMajeri

examples/aobench/src/geometry/vec.rs

@@ -36,7 +36,9 @@ impl V3D {
     #[inline(always)]
     #[must_use]
     pub fn normalized(self) -> Self {
-        self * (1. / self.dot(self).sqrt())
+        let len2 = self.dot(self);
+        let invlen = len2.sqrt().recip();
+        invlen * self
     }
     #[inline(always)]
     #[must_use]

examples/aobench/src/geometry/vecxN.rs

@@ -27,7 +27,9 @@ impl V3DxN {
     #[inline(always)]
     #[must_use]
     pub fn normalized(self) -> Self {
-        (1. / self.dot(self).sqrte()) * self
+        let len2 = self.dot(self);
+        let invlen = len2.rsqrte();
+        invlen * self
     }
 
     pub fn get(&self, idx: usize) -> V3D {
@@ -148,17 +150,15 @@ impl Dot<V3DxN> for V3DxN {
     type Output = f32xN;
     #[inline(always)]
     fn dot(self, o: Self) -> Self::Output {
-        self.x * o.x + self.y * o.y + self.z * o.z
+        self.x.mul_adde(o.x, self.y.mul_adde(o.y, self.z * o.z))
     }
 }
 
 impl Dot<V3D> for V3DxN {
     type Output = f32xN;
     #[inline(always)]
     fn dot(self, o: V3D) -> Self::Output {
-        self.x * f32xN::splat(o.x)
-            + self.y * f32xN::splat(o.y)
-            + self.z * f32xN::splat(o.z)
+        self.x.mul_adde(f32xN::splat(o.x), self.y.mul_adde(f32xN::splat(o.y), self.z * o.z))
     }
 }
 
@@ -204,15 +204,27 @@ impl Mul<V3DxN> for M3x3 {
     #[inline(always)]
     fn mul(self, o: V3DxN) -> Self::Output {
         V3DxN {
-            x: o.x * f32xN::splat(self[0].x)
-                + o.y * f32xN::splat(self[1].x)
-                + o.z * f32xN::splat(self[2].x),
-            y: o.x * f32xN::splat(self[0].y)
-                + o.y * f32xN::splat(self[1].y)
-                + o.z * f32xN::splat(self[2].y),
-            z: o.x * f32xN::splat(self[0].z)
-                + o.y * f32xN::splat(self[1].z)
-                + o.z * f32xN::splat(self[2].z),
+            x: o.x.mul_adde(
+                f32xN::splat(self[0].x),
+                o.y.mul_adde(
+                    f32xN::splat(self[1].x),
+                    o.z * f32xN::splat(self[2].x),
+                ),
+            ),
+            y: o.x.mul_adde(
+                f32xN::splat(self[0].y),
+                o.y.mul_adde(
+                    f32xN::splat(self[1].y),
+                    o.z * f32xN::splat(self[2].y),
+                ),
+            ),
+            z: o.x.mul_adde(
+                f32xN::splat(self[0].z),
+                o.y.mul_adde(
+                    f32xN::splat(self[1].z),
+                    o.z * f32xN::splat(self[2].z),
+                ),
+            ),
         }
     }
 }

examples/aobench/src/intersection/ray_sphere.rs

@@ -40,7 +40,7 @@ impl Intersect<Sphere> for RayxN {
 
         let b = rs.dot(ray.dir);
         let c = rs.dot(rs) - sphere.radius * sphere.radius;
-        let d = b * b - c;
+        let d = b.mul_adde(b, -c);
 
         let old_isect = isect;
 

examples/aobench/src/vector.rs

@@ -23,8 +23,8 @@ fn ao_impl<S: Scene>(scene: &mut S, nsubsamples: usize, img: &mut ::Image) {
                         (x as f32, y as f32, h as f32, w as f32);
 
                     let dir = V3D {
-                        x: (x + du - (w / 2.)) / (w / 2.) * w / h,
-                        y: -(y + dv - (h / 2.)) / (h / 2.),
+                        x: (x + du - (w * 0.5)) / (w * 0.5) * w / h,
+                        y: -(y + dv - (h * 0.5)) / (h * 0.5),
                         z: -1.,
                     };
                     let dir = dir.normalized();
@@ -72,7 +72,7 @@ cfg_if! {
             ao_impl(scene, nsubsamples, img);
         }
 
-        #[target_feature(enable = "avx2")]
+        #[target_feature(enable = "avx2,fma")]
         unsafe fn ao_avx2<S: Scene>(scene: &mut S, nsubsamples: usize,
                                     img: &mut ::Image) {
             ao_impl(scene, nsubsamples, img);
@@ -81,7 +81,7 @@ cfg_if! {
         pub fn ao<S: Scene>(scene: &mut S, nsubsamples: usize,
                             img: &mut ::Image) {
             unsafe {
-                if is_x86_feature_detected!("avx2") {
+                if is_x86_feature_detected!("avx2") && is_x86_feature_detected!("fma") {
                     ao_avx2(scene, nsubsamples, img);
                 } else if is_x86_feature_detected!("avx") {
                     ao_avx(scene, nsubsamples, img);