diff options
Diffstat (limited to 'media/libaom/src/aom_dsp/x86/fwd_txfm_sse2.h')
| -rw-r--r-- | media/libaom/src/aom_dsp/x86/fwd_txfm_sse2.h | 25 |
1 files changed, 15 insertions, 10 deletions
diff --git a/media/libaom/src/aom_dsp/x86/fwd_txfm_sse2.h b/media/libaom/src/aom_dsp/x86/fwd_txfm_sse2.h index 260d8dd58e..ab3cd91557 100644 --- a/media/libaom/src/aom_dsp/x86/fwd_txfm_sse2.h +++ b/media/libaom/src/aom_dsp/x86/fwd_txfm_sse2.h @@ -136,16 +136,21 @@ static INLINE int check_epi16_overflow_x32( } static INLINE void store_output(const __m128i *poutput, tran_low_t *dst_ptr) { - if (sizeof(tran_low_t) == 4) { - const __m128i zero = _mm_setzero_si128(); - const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); - __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); - __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); - _mm_store_si128((__m128i *)(dst_ptr), out0); - _mm_store_si128((__m128i *)(dst_ptr + 4), out1); - } else { - _mm_store_si128((__m128i *)(dst_ptr), *poutput); - } + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + _mm_store_si128((__m128i *)(dst_ptr), out0); + _mm_store_si128((__m128i *)(dst_ptr + 4), out1); +} + +static INLINE void storeu_output(const __m128i *poutput, tran_low_t *dst_ptr) { + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + _mm_storeu_si128((__m128i *)(dst_ptr), out0); + _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1); } #ifdef __cplusplus |