perf: reduce encode int buffer u128 -> u64 to boost parallel perf

.github/workflows/audit.yml

@@ -1,8 +1,5 @@
 on:
   pull_request:
-    paths:
-      - '**/Cargo.toml'
-      - '**/Cargo.lock'
   push:
     branches:
       - main

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "rbase64"
-version = "1.3.1"
+version = "1.4.0"
 edition = "2021"
 description = "A simple base64 encoder / decoder CLI tool made in Rust"
 authors = ["Marcel Riera <marcel.riera@outlook.com>"]

src/common.rs

@@ -2,17 +2,18 @@ pub const ENCODE_MAP: &[u8; 64] =
     b"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
 pub const DECODE_MAP: &[u8; 256] = &construct_decode_map();
 
-pub const SIX_BIT_MASK: u128 = 0x3f;
+pub const SIX_BIT_MASK: u64 = 0x3f;
 pub const BYTE_MASK: u64 = 0xff;
 pub const INVALID_BYTE: u8 = 0x40;
 
-pub const ENC_CHUNK_SIZE: usize = 4;
+pub const ENC_CHUNK_SIZE: usize = 2;
 pub const DEC_CHUNK_SIZE: usize = 2;
-pub const PARALLEL_THRESHOLD_BYTES: usize = 2 << 17; // 256 KiB
 
 pub const ENC_U128_OFFSET: usize = (ENC_CHUNK_SIZE * 3 - 1) * 8;
 pub const DEC_U64_OFFSET: usize = (DEC_CHUNK_SIZE * 4 - 1) * 6;
 
+pub const PARALLEL_THRESHOLD_BYTES: usize = 2 << 16; // 128 KiB
+
 const fn construct_decode_map() -> [u8; 256] {
     let mut map = [INVALID_BYTE; 256];
     let mut index = 0;

src/encode.rs

@@ -4,7 +4,7 @@ use std::cmp::min;
 use std::iter::zip;
 
 #[inline(always)]
-pub(crate) fn encode_u128_chunks(input: &[u8], buffer: &mut [u8]) {
+pub(crate) fn encode_u64_chunks(input: &[u8], buffer: &mut [u8]) {
     let in_chunks = input.chunks_exact(ENC_CHUNK_SIZE * 3);
     let out_chunks = buffer.chunks_exact_mut(ENC_CHUNK_SIZE * 4);
 
@@ -14,7 +14,7 @@ pub(crate) fn encode_u128_chunks(input: &[u8], buffer: &mut [u8]) {
 }
 
 #[inline(always)]
-pub(crate) fn encode_u128_chunks_parallel(input: &[u8], buffer: &mut [u8]) {
+pub(crate) fn encode_u64_chunks_parallel(input: &[u8], buffer: &mut [u8]) {
     let in_chunks = input.par_chunks_exact(ENC_CHUNK_SIZE * 3);
     let out_chunks = buffer.par_chunks_exact_mut(ENC_CHUNK_SIZE * 4);
 
@@ -57,13 +57,13 @@ fn encode_u128(input: &[u8], buffer: &mut [u8]) {
 }
 
 #[inline(always)]
-fn read_u128_partial(bytes: &[u8]) -> u128 {
-    let size = min(bytes.len(), 16);
-    let mut buffer = [0u8; 16];
+fn read_u128_partial(bytes: &[u8]) -> u64 {
+    let mut buffer = [0u8; 8];
+    let size = min(bytes.len(), 8);
 
-    buffer[16 - size..].copy_from_slice(&bytes[..size]);
+    buffer[8 - size..].copy_from_slice(&bytes[..size]);
 
-    u128::from_be_bytes(buffer)
+    u64::from_be_bytes(buffer)
 }
 
 #[inline(always)]

src/lib.rs

@@ -11,9 +11,9 @@ pub fn encode(input: &[u8]) -> String {
     let total_chunks = input.len() / (ENC_CHUNK_SIZE * 3);
 
     if input.len() < PARALLEL_THRESHOLD_BYTES {
-        encode::encode_u128_chunks(input, &mut buffer);
+        encode::encode_u64_chunks(input, &mut buffer);
     } else {
-        encode::encode_u128_chunks_parallel(input, &mut buffer);
+        encode::encode_u64_chunks_parallel(input, &mut buffer);
     };
 
     let bytes_rem = encode_u128_remainder(