diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-03 10:59:02 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-14 08:27:03 +0200 |
| commit | 283352665bbbf57fef2433ca49c9ceb99ff6503a (patch) | |
| tree | 4975ede788a998e476910c838ac73705f06e6ab7 | |
| parent | dd2ace08d85fb3245ab7791ea4114618930e419e (diff) | |
| download | ffwd-283352665bbbf57fef2433ca49c9ceb99ff6503a.tar.gz ffwd-283352665bbbf57fef2433ca49c9ceb99ff6503a.zip | |
kernels: vectorize bf16 widening on arm
| -rw-r--r-- | qwen_asr_kernels.c | 14 |
1 files changed, 14 insertions, 0 deletions
diff --git a/qwen_asr_kernels.c b/qwen_asr_kernels.c index 2c932ff..c948c03 100644 --- a/qwen_asr_kernels.c +++ b/qwen_asr_kernels.c @@ -234,9 +234,23 @@ void qwen_linear_nobias(float *y, const float *x, const float *W, /* Convert bf16 buffer to f32 buffer */ static void bf16_to_f32_buf(float *dst, const uint16_t *src, size_t n) { +#if defined(__ARM_NEON) || defined(__ARM_NEON__) + size_t i = 0; + uint32_t *d = (uint32_t *)(void *)dst; + for (; i + 8 <= n; i += 8) { + uint16x8_t v = vld1q_u16(src + i); + uint32x4_t lo = vshlq_n_u32(vmovl_u16(vget_low_u16(v)), 16); + uint32x4_t hi = vshlq_n_u32(vmovl_u16(vget_high_u16(v)), 16); + vst1q_u32(d + i, lo); + vst1q_u32(d + i + 4, hi); + } + for (; i < n; i++) + d[i] = ((uint32_t)src[i]) << 16; +#else uint32_t *d = (uint32_t *)(void *)dst; for (size_t i = 0; i < n; i++) d[i] = ((uint32_t)src[i]) << 16; +#endif } /* Reusable scratch buffer for bf16->f32 conversion */ |
