tanh: give aarch64 cores without FEAT_FP16 an f32-roundtrip f16 kernel

linalg/Cargo.toml

@@ -189,3 +189,7 @@ harness = false
 [[bench]]
 name = "avxvnni_i32"
 harness = false
+
+[[bench]]
+name = "tanh_f16_arm64"
+harness = false

linalg/benches/tanh_f16_arm64.rs

@@ -0,0 +1,25 @@
+// f16 tanh on aarch64 cores without FEAT_FP16: generic scalar vs f32 roundtrip.
+use criterion::*;
+use tract_data::prelude::*;
+use tract_linalg::element_wise::ElementWiseKer;
+
+fn bench(c: &mut Criterion) {
+    for n in [1024usize, 65536] {
+        let mut t = unsafe { Tensor::uninitialized_aligned::<f16>(&[n], 16).unwrap() };
+        let input = unsafe { t.as_slice_mut_unchecked::<f16>() };
+        for (i, x) in input.iter_mut().enumerate() {
+            *x = f16::from_f32((i as f32 / 10.0).sin() * 5.0);
+        }
+        let mut g = c.benchmark_group(format!("tanh_f16/{n}"));
+        g.throughput(Throughput::Elements(n as u64));
+        g.bench_function("generic", |b| b.iter(|| tract_linalg::generic::HTanh8::run(input, ())));
+        #[cfg(target_arch = "aarch64")]
+        g.bench_function("f32-roundtrip", |b| {
+            b.iter(|| tract_linalg::arm64::arm64simd_tanh_f16_4n::run(input, ()))
+        });
+        g.finish();
+    }
+}
+
+criterion_group!(benches, bench);
+criterion_main!(benches);

linalg/src/arm64.rs

@@ -510,9 +510,12 @@ pub fn plug(ops: &mut Ops) {
         // TODO: Change this SiLU kernel once we have a native-FP16 one
         ops.silu_f16 = Box::new(|| arm64simd_silu_f16_4n::ew());
     } else {
-        log::info!("No native fp16 support; f32-roundtrip NEON sigmoid_f16 and silu_f16 activated");
+        log::info!(
+            "No native fp16 support; f32-roundtrip NEON sigmoid_f16, silu_f16 and tanh_f16 activated"
+        );
         ops.sigmoid_f16 = Box::new(|| arm64simd_sigmoid_f16_4n::ew());
         ops.silu_f16 = Box::new(|| arm64simd_silu_f16_4n::ew());
+        ops.tanh_f16 = Box::new(|| arm64simd_tanh_f16_4n::ew());
     }
     #[cfg(any(target_os = "macos", all(target_os = "ios", feature = "apple-amx-ios")))]
     {

linalg/src/arm64/arm64simd.rs

@@ -16,6 +16,7 @@ mod unicast;
 
 pub use act_f16::arm64simd_sigmoid_f16_4n;
 pub use act_f16::arm64simd_silu_f16_4n;
+pub use act_f16::arm64simd_tanh_f16_4n;
 pub use by_scalar::*;
 pub use gelu::arm64simd_gelu_f32_4n;
 pub use gelu_fused::arm64simd_gelu_f32_4n_fused;

linalg/src/arm64/arm64simd/act_f16.rs

@@ -173,3 +173,21 @@ pub mod test_arm64simd_silu_f16_4n {
     use super::*;
     silu_frame_tests!(true, f16, arm64simd_silu_f16_4n);
 }
+
+// f32-roundtrip f16 tanh for arm64 cores without FEAT_FP16.
+ew_impl_f16_via_f32!(
+    arm64simd_tanh_f16_4n,
+    4,
+    4,
+    CHUNK,
+    16,
+    cvt_f16_to_f32,
+    cvt_f32_to_f16,
+    super::arm64simd_tanh_f32_4n
+);
+
+#[cfg(test)]
+pub mod test_arm64simd_tanh_f16_4n {
+    use super::*;
+    tanh_frame_tests!(true, f16, arm64simd_tanh_f16_4n);
+}