Apply some tricks to make it slightly faster.

This commit also removes a lengthy code block in rice.rs for simplicity.
yotarok · Oct 4, 2023 · e903242 · e903242
1 parent 349f266
commit e903242
Show file tree

Hide file tree

Showing 4 changed files with 13 additions and 33 deletions.
diff --git a/flacenc-bin/Cargo.toml b/flacenc-bin/Cargo.toml
@@ -13,7 +13,8 @@ repository = "https://github.com/yotarok/flacenc-rs/"
 # This is for enabling use of profilers with a release build.
 [profile.release]
 debug = 1
-
+panic = "abort"
+lto = "fat"
 
 [[bin]]
 name = "flacenc"

diff --git a/report/report.md b/report/report.md
@@ -22,15 +22,15 @@ Sources used: wikimedia.i_love_you_california, wikimedia.winter_kiss, wikimedia.
 
 ### Average compression speed (inverse RTF)
   - Reference
-    - opt8lax: 259.31730941656474
-    - opt8: 256.19449736647505
-    - opt5: 498.1760928807853
+    - opt8lax: 260.5208675241677
+    - opt8: 260.4015355192523
+    - opt5: 496.04988453221785
 
   - Ours
-    - default: 166.75753155120208
-    - st: 73.91379462491477
-    - dmse: 111.56016219091809
-    - bsbs: 9.53753985075257
-    - mae: 25.318101362135653
+    - default: 167.2145640416671
+    - st: 83.99637957900106
+    - dmse: 111.79190850520624
+    - bsbs: 11.622540675983512
+    - mae: 26.102371937358594
 
 
diff --git a/src/fakesimd.rs b/src/fakesimd.rs
@@ -225,6 +225,7 @@ macro_rules! def_binop {
             type Output = Self;
             #[allow(clippy::redundant_closure_call)]
             #[inline]
+            #[allow(clippy::redundant_closure_call)]
             fn $fn_name(self, rhs: Self) -> Self::Output {
                 Self(array::from_fn(|i| ($expr)(self.0[i], rhs.0[i])))
             }
@@ -237,6 +238,7 @@ def_binop!(Sub, sub, |x, y| x - y);
 def_binop!(Mul, mul, |x, y| x * y);
 def_binop!(Div, div, |x, y| x / y);
 def_binop!(BitAnd, bitand, |x, y| x & y);
+def_binop!(Shr, shr, |x, y| x >> y);
 
 impl<T, const N: usize> std::convert::AsRef<[T; N]> for Simd<T, N>
 where

diff --git a/src/rice.rs b/src/rice.rs
@@ -73,31 +73,8 @@ impl PrcBitTable {
             simd::u32x16::splat(offset as u32) + simd::u32x16::splat(errors.len() as u32) * INDEX1;
 
         for v in errors {
-            // Below is faster than doing:
-            //   vs = splat(*v) >> INDEX;
-            // or
-            //   vs = simd::u32x16::from_array(std::array::from_fn(
-            //       |i| v >> i));
-            // Perhaps due to smaller memory footprint by avoiding `splat`?
             let v = *v;
-            let vs = simd::u32x16::from_array([
-                v,
-                v >> 1,
-                v >> 2,
-                v >> 3,
-                v >> 4,
-                v >> 5,
-                v >> 6,
-                v >> 7,
-                v >> 8,
-                v >> 9,
-                v >> 10,
-                v >> 11,
-                v >> 12,
-                v >> 13,
-                v >> 14,
-                v >> 15,
-            ]);
+            let vs = simd::u32x16::splat(v) >> INDEX;
             p_to_bits = (vs + p_to_bits) & MAX_P_TO_BITS_VEC;
         }
         self.p_to_bits = p_to_bits;