From 1582d0d1f71bb3fb827890de6d863f9a9511ba6d Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Wed, 3 Jun 2026 10:31:55 +0200 Subject: kernels: add neon rope apply path --- qwen_asr_kernels.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/qwen_asr_kernels.c b/qwen_asr_kernels.c index efa606c..db083d4 100644 --- a/qwen_asr_kernels.c +++ b/qwen_asr_kernels.c @@ -8,6 +8,9 @@ #if (defined(__AVX512F__) || defined(__AVX2__)) && (defined(__x86_64__) || defined(__i386__) || defined(_M_X64) || defined(_M_IX86)) #include #endif +#if defined(__ARM_NEON) || defined(__ARM_NEON__) +#include +#endif #ifdef __APPLE__ #include #else @@ -1616,6 +1619,24 @@ void qwen_apply_rope_neox(float *x, const float *cos_vals, const float *sin_vals vec[d] = x1 * c[d] + (-x2) * sn[d]; vec[half + d] = x2 * c[half + d] + x1 * sn[half + d]; } +#elif defined(__ARM_NEON) || defined(__ARM_NEON__) + int d = 0; + for (; d + 4 <= half; d += 4) { + float32x4_t x1 = vld1q_f32(vec + d); + float32x4_t x2 = vld1q_f32(vec + half + d); + float32x4_t cc = vld1q_f32(c + d); + float32x4_t ss = vld1q_f32(sn + d); + float32x4_t new1 = vmlsq_f32(vmulq_f32(x1, cc), x2, ss); + float32x4_t new2 = vfmaq_f32(vmulq_f32(x2, cc), x1, ss); + vst1q_f32(vec + d, new1); + vst1q_f32(vec + half + d, new2); + } + for (; d < half; d++) { + float x1 = vec[d]; + float x2 = vec[half + d]; + vec[d] = x1 * c[d] + (-x2) * sn[d]; + vec[half + d] = x2 * c[half + d] + x1 * sn[half + d]; + } #else for (int d = 0; d < half; d++) { float x1 = vec[d]; /* first half */ -- cgit v1.3.1