From 6629d96d6f72e62027ded7f8dc89dde7a864b2b3 Mon Sep 17 00:00:00 2001 From: Krzysztof Rymski Date: Fri, 26 Jun 2026 02:10:04 -0700 Subject: [PATCH] Add attention kernels optimized for arm's i8mm instruction. They give about 8x higher throughput compared to previous i8 implementation. PiperOrigin-RevId: 938467018 --- hwy/ops/set_macros-inl.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/hwy/ops/set_macros-inl.h b/hwy/ops/set_macros-inl.h index 4db9e9ab0b..4d540837cc 100644 --- a/hwy/ops/set_macros-inl.h +++ b/hwy/ops/set_macros-inl.h @@ -637,7 +637,7 @@ #define HWY_HAVE_FLOAT64 1 #define HWY_MEM_OPS_MIGHT_FAULT 0 #define HWY_NATIVE_FMA 1 -#if HWY_SVE_HAVE_BF16_FEATURE || HWY_TARGET == HWY_SVE2_128 +#if HWY_SVE_HAVE_BF16_FEATURE #define HWY_NATIVE_DOT_BF16 1 #else #define HWY_NATIVE_DOT_BF16 0