diff options
| author | A.Mahfoodh <ab.mahfoodh@gmail.com> | 2013-09-23 14:13:40 -0400 |
|---|---|---|
| committer | Johann <johann.koenig@duck.com> | 2013-09-23 17:23:27 -0700 |
| commit | 13c7715a75e91bcd855e0aa647ecd88d90152e63 (patch) | |
| tree | 63b6923e72084cc78dada44e03b86d49e6fb8db8 /vp9 | |
| parent | 54c87058bfa5124408f0725deb8373ef94c4b98e (diff) | |
| download | libvpx-13c7715a75e91bcd855e0aa647ecd88d90152e63.tar libvpx-13c7715a75e91bcd855e0aa647ecd88d90152e63.tar.gz libvpx-13c7715a75e91bcd855e0aa647ecd88d90152e63.tar.bz2 libvpx-13c7715a75e91bcd855e0aa647ecd88d90152e63.zip | |
Number of instructions in fdct4_1d_sse2 reduced by two.
Mathematically the results are the same.
Change-Id: I1c5126cd3ca64e8515ca6331e0989c6f7dd651a0
Diffstat (limited to 'vp9')
| -rw-r--r-- | vp9/encoder/x86/vp9_dct_sse2.c | 19 |
1 files changed, 9 insertions, 10 deletions
diff --git a/vp9/encoder/x86/vp9_dct_sse2.c b/vp9/encoder/x86/vp9_dct_sse2.c index eb271fef3..ad3d01da9 100644 --- a/vp9/encoder/x86/vp9_dct_sse2.c +++ b/vp9/encoder/x86/vp9_dct_sse2.c @@ -171,22 +171,21 @@ static INLINE void transpose_4x4(__m128i *res) { void fdct4_1d_sse2(__m128i *in) { const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); - const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); - const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); __m128i u[4], v[4]; - u[0] = _mm_add_epi16(in[0], in[3]); - u[1] = _mm_add_epi16(in[1], in[2]); - u[2] = _mm_sub_epi16(in[1], in[2]); - u[3] = _mm_sub_epi16(in[0], in[3]); + u[0]=_mm_unpacklo_epi16(in[0], in[1]); + u[1]=_mm_unpacklo_epi16(in[3], in[2]); + + v[0] = _mm_add_epi16(u[0], u[1]); + v[1] = _mm_sub_epi16(u[0], u[1]); - v[0] = _mm_unpacklo_epi16(u[0], u[1]); - v[1] = _mm_unpacklo_epi16(u[2], u[3]); u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2 - u[2] = _mm_madd_epi16(v[1], k__cospi_p24_p08); // 1 - u[3] = _mm_madd_epi16(v[1], k__cospi_m08_p24); // 3 + u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1 + u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3 v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); |