benches

Cargo.toml

@@ -2,6 +2,6 @@
 members = ["crates/*", "benches", "analyze"]
 resolver = "2"
 
-[profile.release]
+[profile.bench]
 lto = "fat"
 codegen-units = 1

benches/Cargo.toml

@@ -41,6 +41,7 @@ detect = [
     "uuid-simd/detect",
     "radix64/simd",
 ]
+unstable = ["criterion/real_blackbox"]
 
 [dependencies.base32-simd]
 path = "../crates/base32-simd"
@@ -83,4 +84,4 @@ base32ct = "0.1.0"
 data-encoding = "2.3.2"
 
 [dev-dependencies]
-criterion = { version = "0.4.0", features = ["real_blackbox"] }
+criterion = "0.4.0"

benches/quick.rs

@@ -1,12 +1,14 @@
 use std::time::Instant;
 
+#[inline]
 fn time(f: impl FnOnce()) -> u128 {
     let t0 = Instant::now();
     f();
     let t1 = Instant::now();
     (t1 - t0).as_nanos()
 }
 
+#[inline(never)]
 fn bench_base64(n: usize, src: &str) -> u128 {
     let encoded = base64_simd::STANDARD.encode_type::<Box<[u8]>>(src.as_bytes());
 
@@ -19,6 +21,7 @@ fn bench_base64(n: usize, src: &str) -> u128 {
     })
 }
 
+#[inline(never)]
 fn bench_ascii(n: usize, data: &[u8]) -> u128 {
     time(|| {
         for _ in 0..n {
@@ -27,6 +30,7 @@ fn bench_ascii(n: usize, data: &[u8]) -> u128 {
     })
 }
 
+#[inline(never)]
 fn bench_utf32(n: usize, data: &[u32]) -> u128 {
     time(|| {
         for _ in 0..n {
@@ -35,26 +39,35 @@ fn bench_utf32(n: usize, data: &[u32]) -> u128 {
     })
 }
 
+/// GiB/s
+fn throughput(bytes: u128, ns: u128) -> f64 {
+    let time_sec = ns as f64 / 1e9;
+    let gib = u128::pow(1024, 3);
+    bytes as f64 / gib as f64 / time_sec
+}
+
 fn main() {
     println!("simd-benches quick mode\n");
 
     {
         println!("base64-simd forgiving_decode_inplace");
 
         {
-            let src = "helloworld".repeat(100_000);
+            let data = "helloworld".repeat(100_000);
             let n = 100;
-            let time = bench_base64(n, &src);
+            let time = bench_base64(n, &data);
             let time_per_op = (time / n as u128) as f64;
-            println!("long  | n = {n:<8} time = {time_per_op:>8}ns");
+            let throughput = throughput(n as u128 * data.len() as u128, time);
+            println!("long  | n = {n:<8} time = {time_per_op:>8}ns, throughout = {throughput:.6}GiB/s");
         }
 
         {
-            let src = "123";
+            let data = "123";
             let n = 1_000_000;
-            let time = bench_base64(n, src);
+            let time = bench_base64(n, data);
             let time_per_op = time / n as u128;
-            println!("short | n = {n:<8} time = {time_per_op:>8}ns");
+            let throughput = throughput(n as u128 * data.len() as u128, time);
+            println!("short | n = {n:<8} time = {time_per_op:>8}ns, throughout = {throughput:.6}GiB/s");
         }
         println!();
     }
@@ -65,12 +78,7 @@ fn main() {
         let n = 10000;
         let time = bench_ascii(n, data.as_bytes());
         let time_per_op = time / n as u128;
-        let throughput = {
-            let total = n as u128 * data.len() as u128;
-            let time_sec = time as f64 / 1e9;
-            let gib = u128::pow(1024, 3);
-            total as f64 / gib as f64 / time_sec
-        };
+        let throughput = throughput(n as u128 * data.len() as u128, time);
         println!(
             "long  | n = {n:<8} time = {:>8}ns, throughout = {:.6}GiB/s",
             time_per_op, throughput
@@ -84,12 +92,7 @@ fn main() {
         let n = 1000;
         let time = bench_utf32(n, &data);
         let time_per_op = time / n as u128;
-        let throughput = {
-            let total = n as u128 * data.len() as u128 * 4;
-            let time_sec = time as f64 / 1e9;
-            let gib = u128::pow(1024, 3);
-            total as f64 / gib as f64 / time_sec
-        };
+        let throughput = throughput(n as u128 * data.len() as u128 * 4, time);
         println!(
             "long  | n = {n:<8} time = {:>8}ns, throughput = {:.6}GiB/s",
             time_per_op, throughput

justfile

@@ -14,10 +14,10 @@ doc pkg="vsimd":
     cargo doc --no-deps --all-features
     cargo doc --open -p {{pkg}}
 
-x86-bench dispatch *ARGS:
+bench dispatch *ARGS:
     #!/bin/bash -ex
     cd {{justfile_directory()}}
-    mkdir -p target/x86-bench
+    mkdir -p target/simd-benches
     COMMIT_HASH=`git rev-parse --short HEAD`
 
     case "{{dispatch}}" in
@@ -39,21 +39,21 @@ x86-bench dispatch *ARGS:
             ;;
     esac
 
-    NAME=target/x86-bench/$COMMIT_HASH-{{dispatch}}
+    NAME=target/simd-benches/$COMMIT_HASH-{{dispatch}}
 
     time cargo criterion -p simd-benches --history-id $COMMIT_HASH --message-format json --features "$FEATURES" {{ARGS}} > $NAME.jsonl
     just analyze $COMMIT_HASH {{dispatch}} > $NAME.md
-    bat $NAME.md
+    bat --paging=never $NAME.md
 
 bench-all:
-    just x86-bench static
-    just x86-bench dynamic
-    just x86-bench fallback
+    just bench static
+    just bench dynamic
+    just bench fallback
 
 analyze commit dispatch:
     #!/bin/bash -ex
     cd {{justfile_directory()}}
-    cargo run -q -p analyze -- target/x86-bench/{{commit}}-{{dispatch}}.jsonl
+    cargo run -q -p analyze -- target/simd-benches/{{commit}}-{{dispatch}}.jsonl
 
 js-bench:
     #!/bin/bash -e