diff options
| author | Jingning Han <jingning@google.com> | 2015-02-27 09:49:00 -0800 |
|---|---|---|
| committer | Gerrit Code Review <gerrit@gerrit.golo.chromium.org> | 2015-02-27 09:49:00 -0800 |
| commit | 43bb97f7d0b72eff2a00d34ecde98b9492638701 (patch) | |
| tree | 6b1c541dfe99f767ce4ca8bc7d2aa23bbdca41ef /vp9/common/x86 | |
| parent | 4800b0e80d36680fd035720bdaf29121be4a695c (diff) | |
| parent | 8ec22296b34283ba01f27fad7c1ad4b474389cce (diff) | |
| download | libvpx-43bb97f7d0b72eff2a00d34ecde98b9492638701.tar libvpx-43bb97f7d0b72eff2a00d34ecde98b9492638701.tar.gz libvpx-43bb97f7d0b72eff2a00d34ecde98b9492638701.tar.bz2 libvpx-43bb97f7d0b72eff2a00d34ecde98b9492638701.zip | |
Merge "Fix high bit-depth loop-filter sse2 compiling issue - part 3"
Diffstat (limited to 'vp9/common/x86')
| -rw-r--r-- | vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c | 54 |
1 files changed, 39 insertions, 15 deletions
diff --git a/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c b/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c index 26e85a3d6..04823ba9c 100644 --- a/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c +++ b/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c @@ -531,15 +531,7 @@ void vp9_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p, DECLARE_ALIGNED_ARRAY(16, uint16_t, flat_oq1, 16); DECLARE_ALIGNED_ARRAY(16, uint16_t, flat_oq0, 16); const __m128i zero = _mm_set1_epi16(0); - const __m128i blimit = _mm_slli_epi16( - _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), - bd - 8); - const __m128i limit = _mm_slli_epi16( - _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), - bd - 8); - const __m128i thresh = _mm_slli_epi16( - _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), - bd - 8); + __m128i blimit, limit, thresh; __m128i mask, hev, flat; __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p)); __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p)); @@ -557,18 +549,43 @@ void vp9_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p, const __m128i t4 = _mm_set1_epi16(4); const __m128i t3 = _mm_set1_epi16(3); - const __m128i t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), bd - 8); + __m128i t80; const __m128i t1 = _mm_set1_epi16(0x1); - const __m128i ps1 = _mm_subs_epi16(p1, t80); - const __m128i ps0 = _mm_subs_epi16(p0, t80); - const __m128i qs0 = _mm_subs_epi16(q0, t80); - const __m128i qs1 = _mm_subs_epi16(q1, t80); + __m128i ps1, ps0, qs0, qs1; __m128i filt; __m128i work_a; __m128i filter1, filter2; (void)count; + if (bd == 8) { + blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); + limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); + thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); + t80 = _mm_set1_epi16(0x80); + } else if (bd == 10) { + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); + t80 = _mm_set1_epi16(0x200); + } else { // bd == 12 + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); + t80 = _mm_set1_epi16(0x800); + } + + ps1 = _mm_subs_epi16(p1, t80); + ps0 = _mm_subs_epi16(p0, t80); + qs0 = _mm_subs_epi16(q0, t80); + qs1 = _mm_subs_epi16(q1, t80); + // filter_mask and hev_mask abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); @@ -620,7 +637,14 @@ void vp9_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p, flat = _mm_max_epi16(work, flat); flat = _mm_max_epi16(abs_p1p0, flat); flat = _mm_max_epi16(abs_q1q0, flat); - flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8)); + + if (bd == 8) + flat = _mm_subs_epu16(flat, one); + else if (bd == 10) + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2)); + else // bd == 12 + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4)); + flat = _mm_cmpeq_epi16(flat, zero); flat = _mm_and_si128(flat, mask); // flat & mask |