Fix swapped scalar and simd implementations of strip

The logic for choosing whether to use the scalar or simd version of the strip kernel was backwards.

This makes a pretty small performance difference; it just isn't a large part of the total time.

Also optimize clamping behavior to take advantage of saturation in conversion operations.

Author: raphlinus

cpu-sparse/src/fine.rs

@@ -57,7 +57,7 @@ impl<'a> Fine<'a> {
             for i in 0..WIDE_TILE_WIDTH {
                 let mut rgba_f32 = [0.0; 4];
                 rgba_f32.copy_from_slice(&self.scratch[(i * STRIP_HEIGHT + j) * 4..][..4]);
-                let rgba_u8 = rgba_f32.map(|x| (x.clamp(0., 1.) * 255.0).round() as u8);
+                let rgba_u8 = rgba_f32.map(|x| (x * 255.0).round() as u8);
                 self.out_buf[line_ix + i * 4..][..4].copy_from_slice(&rgba_u8);
             }
         }

cpu-sparse/src/simd.rs

@@ -44,11 +44,11 @@ impl<'a> Fine<'a> {
 
     pub(crate) fn strip(&mut self, x: usize, width: usize, alphas: &[u32], color: [f32; 4]) {
         if self.use_simd {
-            self.strip_scalar(x, width, alphas, color);
-        } else {
             unsafe {
                 self.strip_simd(x, width, alphas, color);
             }
+        } else {
+            self.strip_scalar(x, width, alphas, color);
         }
     }
 }

cpu-sparse/src/simd/neon.rs

@@ -23,7 +23,7 @@ impl<'a> Fine<'a> {
 
     pub fn pack_simd(&mut self, x: usize, y: usize) {
         unsafe fn cvt(v: float32x4_t) -> uint8x16_t {
-            let clamped = vminq_f32(vmaxq_f32(v, vdupq_n_f32(0.0)), vdupq_n_f32(1.0));
+            let clamped = vminq_f32(v, vdupq_n_f32(1.0));
             let scaled = vmulq_f32(clamped, vdupq_n_f32(255.0));
             vreinterpretq_u8_u32(vcvtnq_u32_f32(scaled))
         }