add a benchmark

2 files changed, 11 insertions, 1 deletions

diff --git a/benches/bench.rs b/benches/bench.rs
index 5efb9e6..95b67e1 100644
--- a/benches/bench.rs
+++ b/benches/bench.rs
@@ -151,6 +151,15 @@ fn bench_many_chunks_neon(b: &mut Bencher) {
     }
 }
 
+#[bench]
+fn bench_many_chunks_kernel(b: &mut Bencher) {
+    let mut input = RandomInput::new(b, 16 * CHUNK_LEN);
+    let mut out = [blake3::kernel::Words16([0; 16]); 8];
+    b.iter(|| unsafe {
+        blake3::kernel::chunks16(input.get().try_into().unwrap(), &[0; 8], 0, 0, &mut out);
+    });
+}
+
 // TODO: When we get const generics we can unify this with the chunks code.
 fn bench_many_parents_fn(b: &mut Bencher, platform: Platform) {
     let degree = platform.simd_degree();
diff --git a/src/kernel.rs b/src/kernel.rs
index a9c58ab..4ac76eb 100644
--- a/src/kernel.rs
+++ b/src/kernel.rs
@@ -964,6 +964,7 @@ global_asm!(
     // CVs in zmm0-zmm7 are written to the out pointer.
     // --------------------------------------------------------------------------------------------
     "blake3_avx512_chunks_16:",
+    // TODO: Prefetches
     // Broadcast the key into zmm0-zmm7. Use ecx as scratch.
     "mov ecx, dword ptr [rsi + 0 * 4]",
     "vpbroadcastd zmm0, ecx",
@@ -1038,7 +1039,7 @@ global_asm!(
 
 #[repr(C, align(64))]
 #[derive(Copy, Clone, Debug)]
-pub struct Words16([u32; 16]);
+pub struct Words16(pub [u32; 16]);
 
 #[no_mangle]
 static BLAKE3_IV0_16: Words16 = Words16([crate::IV[0]; 16]);