add sse2 tests and benchmarks

2 files changed, 60 insertions, 0 deletions

diff --git a/c/blake3_c_rust_bindings/benches/bench.rs b/c/blake3_c_rust_bindings/benches/bench.rs
index 4335a26..c6e821e 100644
--- a/c/blake3_c_rust_bindings/benches/bench.rs
+++ b/c/blake3_c_rust_bindings/benches/bench.rs
@@ -71,6 +71,18 @@ fn bench_single_compression_portable(b: &mut Bencher) {
 
 #[bench]
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn bench_single_compression_sse2(b: &mut Bencher) {
+    if !blake3_c_rust_bindings::sse2_detected() {
+        return;
+    }
+    bench_single_compression_fn(
+        b,
+        blake3_c_rust_bindings::ffi::x86::blake3_compress_in_place_sse2,
+    );
+}
+
+#[bench]
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 fn bench_single_compression_sse41(b: &mut Bencher) {
     if !blake3_c_rust_bindings::sse41_detected() {
         return;
@@ -136,6 +148,19 @@ fn bench_many_chunks_fn(b: &mut Bencher, f: HashManyFn, degree: usize) {
 
 #[bench]
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn bench_many_chunks_sse2(b: &mut Bencher) {
+    if !blake3_c_rust_bindings::sse2_detected() {
+        return;
+    }
+    bench_many_chunks_fn(
+        b,
+        blake3_c_rust_bindings::ffi::x86::blake3_hash_many_sse2,
+        4,
+    );
+}
+
+#[bench]
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 fn bench_many_chunks_sse41(b: &mut Bencher) {
     if !blake3_c_rust_bindings::sse41_detected() {
         return;
@@ -215,6 +240,20 @@ fn bench_many_parents_fn(b: &mut Bencher, f: HashManyFn, degree: usize) {
 
 #[bench]
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn bench_many_parents_sse2(b: &mut Bencher) {
+    if !blake3_c_rust_bindings::sse2_detected() {
+        return;
+    }
+    bench_many_parents_fn(
+        b,
+        blake3_c_rust_bindings::ffi::x86::blake3_hash_many_sse2,
+        4,
+    );
+}
+
+
+#[bench]
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 fn bench_many_parents_sse41(b: &mut Bencher) {
     if !blake3_c_rust_bindings::sse41_detected() {
         return;
diff --git a/c/blake3_c_rust_bindings/src/test.rs b/c/blake3_c_rust_bindings/src/test.rs
index 9f46730..62a7ee1 100644
--- a/c/blake3_c_rust_bindings/src/test.rs
+++ b/c/blake3_c_rust_bindings/src/test.rs
@@ -145,6 +145,18 @@ fn test_compress_portable() {
 
 #[test]
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn test_compress_sse2() {
+    if !crate::sse2_detected() {
+        return;
+    }
+    test_compress_fn(
+        crate::ffi::x86::blake3_compress_in_place_sse2,
+        crate::ffi::x86::blake3_compress_xof_sse2,
+    );
+}
+
+#[test]
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 fn test_compress_sse41() {
     if !crate::sse41_detected() {
         return;
@@ -286,6 +298,15 @@ fn test_hash_many_portable() {
 
 #[test]
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn test_hash_many_sse2() {
+    if !crate::sse2_detected() {
+        return;
+    }
+    test_hash_many_fn(crate::ffi::x86::blake3_hash_many_sse2);
+}
+
+#[test]
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 fn test_hash_many_sse41() {
     if !crate::sse41_detected() {
         return;