diff options
| author | Max Kellermann <max@duempel.org> | 2014-03-16 08:57:04 +0100 |
|---|---|---|
| committer | Max Kellermann <max@duempel.org> | 2014-03-16 09:05:30 +0100 |
| commit | 6f4775a8eeb30d0fde622f4e89979179af47d219 (patch) | |
| tree | 4bf9185d16cecb79921e4e961c37e477cdb29977 | |
| parent | c7e2f558a0608eda343ab6f4d8801204497bebf5 (diff) | |
| download | mpd-6f4775a8eeb30d0fde622f4e89979179af47d219.tar.gz mpd-6f4775a8eeb30d0fde622f4e89979179af47d219.tar.xz mpd-6f4775a8eeb30d0fde622f4e89979179af47d219.zip | |
pcm/Neon: apply bit shift during float->int conversion
Avoid multiplication. This is a speedup of 20%.
| -rw-r--r-- | src/pcm/Neon.hxx | 9 |
1 files changed, 2 insertions, 7 deletions
diff --git a/src/pcm/Neon.hxx b/src/pcm/Neon.hxx index a6f52db42..2d3a59d2b 100644 --- a/src/pcm/Neon.hxx +++ b/src/pcm/Neon.hxx @@ -68,21 +68,16 @@ struct NeonFloatTo16 { static constexpr size_t BLOCK_SIZE = 16; void Convert(int16_t *dst, const float *src, const size_t n) const { - const float32x4_t factor = - vdupq_n_f32(1 << (DstTraits::BITS - 1)); - for (unsigned i = 0; i < n / BLOCK_SIZE; ++i, src += BLOCK_SIZE, dst += BLOCK_SIZE) { /* load 16 float samples into 4 quad registers */ float32x4x4_t value = vld4q_f32(src); - /* apply factor */ - neon_x4_b(vmulq_f32, value, value, factor); - /* convert to 32 bit integer */ int32x4x4_t ivalue; - neon_x4_u(vcvtq_s32_f32, ivalue, value); + neon_x4_b(vcvtq_n_s32_f32, ivalue, value, + DstTraits::BITS - 1); /* convert to 16 bit integer with saturation */ int16x4x4_t nvalue; |