diff options
| author | lvqcl <lvqcl.mail@gmail.com> | 2018-09-19 20:03:37 +0300 |
|---|---|---|
| committer | Erik de Castro Lopo <erikd@mega-nerd.com> | 2018-09-20 07:20:13 +1000 |
| commit | 421961f00b505dcad41603305b47eda0b2ddfe92 (patch) | |
| tree | 051990fb5c29ee092201be9562fa6479834a25f3 /src/libFLAC/stream_encoder_intrin_sse2.c | |
| parent | faafa4c82c31e5aed7bc7c0e87a379825372c6ac (diff) | |
| download | flac-421961f00b505dcad41603305b47eda0b2ddfe92.tar.gz | |
Replace hadd with shuffle + add
Diffstat (limited to 'src/libFLAC/stream_encoder_intrin_sse2.c')
| -rw-r--r-- | src/libFLAC/stream_encoder_intrin_sse2.c | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/src/libFLAC/stream_encoder_intrin_sse2.c b/src/libFLAC/stream_encoder_intrin_sse2.c index ed94ec3e..44ee4d35 100644 --- a/src/libFLAC/stream_encoder_intrin_sse2.c +++ b/src/libFLAC/stream_encoder_intrin_sse2.c @@ -97,8 +97,8 @@ void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual mm_sum = _mm_add_epi32(mm_sum, mm_res); } - mm_sum = _mm_add_epi32(mm_sum, _mm_srli_si128(mm_sum, 8)); - mm_sum = _mm_add_epi32(mm_sum, _mm_srli_si128(mm_sum, 4)); + mm_sum = _mm_add_epi32(mm_sum, _mm_shuffle_epi32(mm_sum, _MM_SHUFFLE(1,0,3,2))); + mm_sum = _mm_add_epi32(mm_sum, _mm_shufflelo_epi16(mm_sum, _MM_SHUFFLE(1,0,3,2))); abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(mm_sum); /* workaround for MSVC bugs (at least versions 2015 and 2017 are affected) */ #if (defined _MSC_VER) && (defined FLAC__CPU_X86_64) |