Assembly GEMM kernel NC loop use SUBS instead of CMP+SUBS
PiperOrigin-RevId: 283908353
diff --git a/src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S b/src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S
index 9e460ea..0229b7e 100644
--- a/src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S
+++ b/src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S
@@ -421,7 +421,7 @@
FMAX v31.4s, v31.4s, v5.4s
# Store full 4 x 8
- CMP x1, 8
+ SUBS x1, x1, 8
B.LO 8f
STP q30, q31, [x7]
@@ -436,7 +436,6 @@
SUB x4, x4, x3 // a -= ks
# nc loop
- SUBS x1, x1, 8
B.HI 0b
# Restore d8-d15 from stack