diff options
| author | Josh Blum <josh@joshknows.com> | 2010-06-25 15:57:08 -0700 |
|---|---|---|
| committer | Josh Blum <josh@joshknows.com> | 2010-06-28 11:12:19 -0700 |
| commit | a9a47d0a69419e862657567a4228e8de0f4b8342 (patch) | |
| tree | 93b7a605bea8bf2c349fd2291831989354649e09 | |
| parent | b46f9f53c9252f69c0ab60a948373431da4fc54f (diff) | |
| download | uhd-a9a47d0a69419e862657567a4228e8de0f4b8342.tar.gz uhd-a9a47d0a69419e862657567a4228e8de0f4b8342.tar.bz2 uhd-a9a47d0a69419e862657567a4228e8de0f4b8342.zip | |
uhd: convert types simd, unpack with zero constant for lower half
| -rw-r--r-- | host/lib/transport/convert_types_impl.hpp | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/host/lib/transport/convert_types_impl.hpp b/host/lib/transport/convert_types_impl.hpp index 58cc69eb0..ca1be175c 100644 --- a/host/lib/transport/convert_types_impl.hpp +++ b/host/lib/transport/convert_types_impl.hpp @@ -160,6 +160,7 @@ static UHD_INLINE void item32_to_fc32_bswap( const item32_t *input, fc32_t *output, size_t nsamps ){ __m128 scalar = _mm_set_ps1(floats_per_short/(1 << 16)); + __m128i zeroi = _mm_setzero_si128(); //convert blocks of samples with intrinsics size_t i = 0; for (; i < nsamps/4; i+=4){ @@ -168,8 +169,8 @@ static UHD_INLINE void item32_to_fc32_bswap( //byteswap + unpack -> byteswap 32 bit words tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); - __m128i tmpilo = _mm_unpacklo_epi16(tmpi, tmpi); - __m128i tmpihi = _mm_unpackhi_epi16(tmpi, tmpi); + __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits + __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); //convert and scale __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); |