add benchmarks for AVX-512 and NEON

benches/bench.rs

@@ -80,6 +80,15 @@ fn bench_single_compression_sse41(b: &mut Bencher) {
     bench_single_compression_fn(b, blake3::sse41::compress);
 }
 
+#[bench]
+#[cfg(feature = "c_avx512")]
+fn bench_single_compression_avx512(b: &mut Bencher) {
+    if !blake3::platform::avx512_detected() {
+        return;
+    }
+    bench_single_compression_fn(b, blake3::c_avx512::compress);
+}
+
 type HashManyFn<A> = unsafe fn(
     inputs: &[&A],
     key: &[u8; blake3::KEY_LEN],
@@ -136,6 +145,22 @@ fn bench_many_chunks_avx2(b: &mut Bencher) {
     bench_many_chunks_fn(b, blake3::avx2::hash_many, blake3::avx2::DEGREE);
 }
 
+#[bench]
+#[cfg(feature = "c_avx512")]
+fn bench_many_chunks_avx512(b: &mut Bencher) {
+    if !blake3::platform::avx512_detected() {
+        return;
+    }
+    bench_many_chunks_fn(b, blake3::c_avx512::hash_many, blake3::c_avx512::DEGREE);
+}
+
+#[bench]
+#[cfg(feature = "c_neon")]
+fn bench_many_chunks_neon(b: &mut Bencher) {
+    // When "c_neon" is on, NEON support is assumed.
+    bench_many_chunks_fn(b, blake3::c_neon::hash_many, blake3::c_neon::DEGREE);
+}
+
 // TODO: When we get const generics we can unify this with the chunks code.
 fn bench_many_parents_fn(b: &mut Bencher, f: HashManyFn<[u8; BLOCK_LEN]>, degree: usize) {
     let mut inputs = Vec::new();
@@ -182,6 +207,22 @@ fn bench_many_parents_avx2(b: &mut Bencher) {
     bench_many_parents_fn(b, blake3::avx2::hash_many, blake3::avx2::DEGREE);
 }
 
+#[bench]
+#[cfg(feature = "c_avx512")]
+fn bench_many_parents_avx512(b: &mut Bencher) {
+    if !blake3::platform::avx512_detected() {
+        return;
+    }
+    bench_many_parents_fn(b, blake3::c_avx512::hash_many, blake3::c_avx512::DEGREE);
+}
+
+#[bench]
+#[cfg(feature = "c_neon")]
+fn bench_many_parents_neon(b: &mut Bencher) {
+    // When "c_neon" is on, NEON support is assumed.
+    bench_many_parents_fn(b, blake3::c_neon::hash_many, blake3::c_neon::DEGREE);
+}
+
 fn bench_atonce(b: &mut Bencher, len: usize) {
     let mut input = RandomInput::new(b, len);
     b.iter(|| blake3::hash(input.get()));

src/c_avx512.rs

@@ -1,5 +1,7 @@
 use crate::{OffsetDeltas, BLOCK_LEN, KEY_LEN, OUT_LEN};
 
+pub const DEGREE: usize = 16;
+
 // Unsafe because this may only be called on platforms supporting AVX-512.
 pub unsafe fn compress(
     cv: &[u8; 32],

src/c_neon.rs

@@ -1,5 +1,7 @@
 use crate::{OffsetDeltas, BLOCK_LEN, KEY_LEN, OUT_LEN};
 
+pub const DEGREE: usize = 4;
+
 // Unsafe because this may only be called on platforms supporting NEON.
 pub unsafe fn hash_many<A: arrayvec::Array<Item = u8>>(
     inputs: &[&A],