workaround bug around inline asm

sonos · Jan 31, 2024 · c057192 · c057192
1 parent d3991af
commit c057192
Showing 1 changed file with 8 additions and 12 deletions.
diff --git a/linalg/src/arm64/arm64fp16/max.rs b/linalg/src/arm64/arm64fp16/max.rs
@@ -15,8 +15,9 @@ reduce_impl_wrap!(
         unsafe fn run(buf: &[f16]) -> f16 {
             let len = buf.len();
             let ptr = buf.as_ptr();
-            let mut out: u16 = f16::MIN.to_bits();
-                /*
+            let mut out: u16;
+            std::arch::asm!("
+                ins v0.h[0], {min:w}
                 dup v0.8h, v0.h[0]
                 dup v1.8h, v0.h[0]
                 dup v2.8h, v0.h[0]
@@ -36,17 +37,12 @@ reduce_impl_wrap!(
                 fmax v2.8h, v2.8h, v3.8h
                 fmax v0.8h, v0.8h, v2.8h
                 fmaxv h0, v0.8h
-                */
-            std::arch::asm!("
-                dup v0.8h, v0.h[0]
-                2:
-
-                    subs {len}, {len}, 32
-                    bne 2b
                 ",
+                // using v0 as inout triggers https://github.com/rust-lang/rust/issues/120374
+                min = in(reg) f16::MIN.to_bits(),
+                ptr = inout(reg) ptr => _,
                 len = inout(reg) len => _,
-//                ptr = inout(reg) ptr => _,
-                inout("v0") out, out("v1") _, out("v2") _, out("v3") _,
+                out("v0") out, out("v1") _, out("v2") _, out("v3") _,
                 out("v4") _, out("v5") _, out("v6") _, out("v7") _,);
             f16::from_bits(out)
         }
@@ -59,7 +55,7 @@ reduce_impl_wrap!(
 );
 
 #[cfg(test)]
-mod test_arm64simd_max_f16_32n {
+mod test_arm64fp16_max_f16_32n {
     use super::*;
     max_frame_tests!(crate::arm64::has_fp16(), f16, arm64fp16_max_f16_32n);
 }