ref: 22dc946a7e46e9a24e25d61232fb2f4678e9b9f6
parent: eb7acb55242a0421f811a2e96118ba59b48e7db8
author: Scott LaVarnway <slavarnway@google.com>
date: Thu Aug 29 04:30:17 EDT 2013
Improved mb_lpf_horizontal_edge_w_sse2_8 This patch is a reformatted version of optimizations done by engineers at Intel (Erik/Tamar) who have been providing performance feedback for VP9. For the test clips used (720p, 1080p), up to 1.2% performance improvement was seen. Change-Id: Ic1a7149098740079d5453b564da6fbfdd0b2f3d2
--- a/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
+++ b/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
@@ -17,23 +17,11 @@
const unsigned char *_blimit,
const unsigned char *_limit,
const unsigned char *_thresh) {- DECLARE_ALIGNED(16, unsigned char, flat2_op[7][8]);
- DECLARE_ALIGNED(16, unsigned char, flat2_oq[7][8]);
-
- DECLARE_ALIGNED(16, unsigned char, flat_op[3][8]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq[3][8]);
-
- DECLARE_ALIGNED(16, unsigned char, ap[8][8]);
- DECLARE_ALIGNED(16, unsigned char, aq[8][8]);
-
-
__m128i mask, hev, flat, flat2;
const __m128i zero = _mm_set1_epi16(0);
const __m128i one = _mm_set1_epi8(1);
- __m128i p7, p6, p5;
- __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
- __m128i q5, q6, q7;
- int i = 0;
+ __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+ __m128i abs_p1p0;
const unsigned int extended_thresh = _thresh[0] * 0x01010101u;
const unsigned int extended_limit = _limit[0] * 0x01010101u;
const unsigned int extended_blimit = _blimit[0] * 0x01010101u;
@@ -44,41 +32,35 @@
const __m128i blimit =
_mm_shuffle_epi32(_mm_cvtsi32_si128((int)extended_blimit), 0);
- p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
- p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
- p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
- p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
- p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
- q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p));
- q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
- q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p));
- q3 = _mm_loadl_epi64((__m128i *)(s + 3 * p));
- q4 = _mm_loadl_epi64((__m128i *)(s + 4 * p));
+ q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+ q4p4 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q4p4),
+ (__m64 *)(s + 4 * p)));
+ q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+ q3p3 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q3p3),
+ (__m64 *)(s + 3 * p)));
+ q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+ q2p2 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q2p2),
+ (__m64 *)(s + 2 * p)));
+ q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+ q1p1 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q1p1),
+ (__m64 *)(s + 1 * p)));
+ p1q1 = _mm_shuffle_epi32(q1p1, 78);
+ q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+ q0p0 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q0p0),
+ (__m64 *)(s - 0 * p)));
+ p0q0 = _mm_shuffle_epi32(q0p0, 78);
- _mm_storel_epi64((__m128i *)ap[4], p4);
- _mm_storel_epi64((__m128i *)ap[3], p3);
- _mm_storel_epi64((__m128i *)ap[2], p2);
- _mm_storel_epi64((__m128i *)ap[1], p1);
- _mm_storel_epi64((__m128i *)ap[0], p0);
- _mm_storel_epi64((__m128i *)aq[4], q4);
- _mm_storel_epi64((__m128i *)aq[3], q3);
- _mm_storel_epi64((__m128i *)aq[2], q2);
- _mm_storel_epi64((__m128i *)aq[1], q1);
- _mm_storel_epi64((__m128i *)aq[0], q0);
-
-
{- const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
- _mm_subs_epu8(p0, p1));
- const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
- _mm_subs_epu8(q0, q1));
- const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
- _mm_subs_epu8(q0, p0));
- __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
- _mm_subs_epu8(q1, p1));
- __m128i work;
+ __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+ abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
+ _mm_subs_epu8(q0p0, q1p1));
+ abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+ fe = _mm_set1_epi8(0xfe);
+ ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+ abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
+ _mm_subs_epu8(p0q0, q0p0));
+ abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
+ _mm_subs_epu8(p1q1, q1p1));
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
@@ -88,19 +70,16 @@
mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
// mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(flat, mask);
+ mask = _mm_max_epu8(abs_p1p0, mask);
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
- _mm_subs_epu8(p1, p2)),
- _mm_or_si128(_mm_subs_epu8(p3, p2),
- _mm_subs_epu8(p2, p3)));
+
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
+ _mm_subs_epu8(q1p1, q2p2)),
+ _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
+ _mm_subs_epu8(q2p2, q3p3)));
mask = _mm_max_epu8(work, mask);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
- _mm_subs_epu8(q1, q2)),
- _mm_or_si128(_mm_subs_epu8(q3, q2),
- _mm_subs_epu8(q2, q3)));
- mask = _mm_max_epu8(work, mask);
+ mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
mask = _mm_subs_epu8(mask, limit);
mask = _mm_cmpeq_epi8(mask, zero);
}
@@ -110,21 +89,19 @@
const __m128i t4 = _mm_set1_epi8(4);
const __m128i t3 = _mm_set1_epi8(3);
const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i te0 = _mm_set1_epi8(0xe0);
- const __m128i t1f = _mm_set1_epi8(0x1f);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i t7f = _mm_set1_epi8(0x7f);
-
- __m128i ps1 = _mm_xor_si128(p1, t80);
- __m128i ps0 = _mm_xor_si128(p0, t80);
- __m128i qs0 = _mm_xor_si128(q0, t80);
- __m128i qs1 = _mm_xor_si128(q1, t80);
+ const __m128i t1 = _mm_set1_epi16(0x1);
+ __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+ __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+ __m128i qs0 = _mm_xor_si128(p0q0, t80);
+ __m128i qs1 = _mm_xor_si128(p1q1, t80);
__m128i filt;
__m128i work_a;
__m128i filter1, filter2;
+ __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+ __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
+ filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, qs0ps0);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
@@ -134,82 +111,60 @@
filter1 = _mm_adds_epi8(filt, t4);
filter2 = _mm_adds_epi8(filt, t3);
+ filter1 = _mm_unpacklo_epi8(zero, filter1);
+ filter1 = _mm_srai_epi16(filter1, 0xB);
+ filter2 = _mm_unpacklo_epi8(zero, filter2);
+ filter2 = _mm_srai_epi16(filter2, 0xB);
+
/* Filter1 >> 3 */
- work_a = _mm_cmpgt_epi8(zero, filter1);
- filter1 = _mm_srli_epi16(filter1, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter1 = _mm_and_si128(filter1, t1f);
- filter1 = _mm_or_si128(filter1, work_a);
- qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+ filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+ qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
- /* Filter2 >> 3 */
- work_a = _mm_cmpgt_epi8(zero, filter2);
- filter2 = _mm_srli_epi16(filter2, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter2 = _mm_and_si128(filter2, t1f);
- filter2 = _mm_or_si128(filter2, work_a);
- ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
-
/* filt >> 1 */
- filt = _mm_adds_epi8(filter1, t1);
- work_a = _mm_cmpgt_epi8(zero, filt);
- filt = _mm_srli_epi16(filt, 1);
- work_a = _mm_and_si128(work_a, t80);
- filt = _mm_and_si128(filt, t7f);
- filt = _mm_or_si128(filt, work_a);
- filt = _mm_andnot_si128(hev, filt);
- ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
- qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+ filt = _mm_adds_epi16(filter1, t1);
+ filt = _mm_srai_epi16(filt, 1);
+ filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
+ filt);
+ filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+ qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
// loopfilter done
{__m128i work;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
- _mm_subs_epu8(p0, p2)),
- _mm_or_si128(_mm_subs_epu8(q2, q0),
- _mm_subs_epu8(q0, q2)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
- _mm_subs_epu8(p0, p3)),
- _mm_or_si128(_mm_subs_epu8(q3, q0),
- _mm_subs_epu8(q0, q3)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p4, p0),
- _mm_subs_epu8(p0, p4)),
- _mm_or_si128(_mm_subs_epu8(q4, q0),
- _mm_subs_epu8(q0, q4)));
+ flat = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
+ _mm_subs_epu8(q0p0, q2p2)),
+ _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
+ _mm_subs_epu8(q0p0, q3p3)));
+ flat = _mm_max_epu8(abs_p1p0, flat);
+ flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
flat = _mm_subs_epu8(flat, one);
flat = _mm_cmpeq_epi8(flat, zero);
flat = _mm_and_si128(flat, mask);
- p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
- q5 = _mm_loadl_epi64((__m128i *)(s + 5 * p));
- flat2 = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p5, p0),
- _mm_subs_epu8(p0, p5)),
- _mm_or_si128(_mm_subs_epu8(q5, q0),
- _mm_subs_epu8(q0, q5)));
- _mm_storel_epi64((__m128i *)ap[5], p5);
- _mm_storel_epi64((__m128i *)aq[5], q5);
- flat2 = _mm_max_epu8(work, flat2);
- p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
- q6 = _mm_loadl_epi64((__m128i *)(s + 6 * p));
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p6, p0),
- _mm_subs_epu8(p0, p6)),
- _mm_or_si128(_mm_subs_epu8(q6, q0),
- _mm_subs_epu8(q0, q6)));
- _mm_storel_epi64((__m128i *)ap[6], p6);
- _mm_storel_epi64((__m128i *)aq[6], q6);
- flat2 = _mm_max_epu8(work, flat2);
+ q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+ q5p5 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q5p5),
+ (__m64 *)(s + 5 * p)));
- p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
- q7 = _mm_loadl_epi64((__m128i *)(s + 7 * p));
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p7, p0),
- _mm_subs_epu8(p0, p7)),
- _mm_or_si128(_mm_subs_epu8(q7, q0),
- _mm_subs_epu8(q0, q7)));
- _mm_storel_epi64((__m128i *)ap[7], p7);
- _mm_storel_epi64((__m128i *)aq[7], q7);
+ q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+ q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6),
+ (__m64 *)(s + 6 * p)));
+
+ flat2 = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q4p4, q0p0),
+ _mm_subs_epu8(q0p0, q4p4)),
+ _mm_or_si128(_mm_subs_epu8(q5p5, q0p0),
+ _mm_subs_epu8(q0p0, q5p5)));
+
+ q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+ q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7),
+ (__m64 *)(s + 7 * p)));
+
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q6p6, q0p0),
+ _mm_subs_epu8(q0p0, q6p6)),
+ _mm_or_si128(_mm_subs_epu8(q7p7, q0p0),
+ _mm_subs_epu8(q0p0, q7p7)));
+
flat2 = _mm_max_epu8(work, flat2);
+ flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
flat2 = _mm_subs_epu8(flat2, one);
flat2 = _mm_cmpeq_epi8(flat2, zero);
flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
@@ -220,260 +175,198 @@
{const __m128i eight = _mm_set1_epi16(8);
const __m128i four = _mm_set1_epi16(4);
- {- __m128i workp_shft;
- __m128i a, b, c;
+ __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+ __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+ __m128i pixelFilter_p, pixelFilter_q;
+ __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+ __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
- p7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[7])), zero);
- p6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[6])), zero);
- p5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[5])), zero);
- p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[4])), zero);
- p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[3])), zero);
- p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[2])), zero);
- p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[1])), zero);
- p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[0])), zero);
- q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[0])), zero);
- q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[1])), zero);
- q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[2])), zero);
- q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[3])), zero);
- q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[4])), zero);
- q5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[5])), zero);
- q6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[6])), zero);
- q7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[7])), zero);
+ p7_16 = _mm_unpacklo_epi8(q7p7, zero);;
+ p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+ p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+ p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+ p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+ p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+ p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+ p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+ q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+ q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+ q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+ q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+ q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+ q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+ q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+ q7_16 = _mm_unpackhi_epi8(q7p7, zero);
- c = _mm_sub_epi16(_mm_slli_epi16(p7, 3), p7); // p7 * 7
- c = _mm_add_epi16(_mm_slli_epi16(p6, 1), _mm_add_epi16(p4, c));
+ pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+ _mm_add_epi16(p4_16, p3_16));
+ pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+ _mm_add_epi16(q4_16, q3_16));
- b = _mm_add_epi16(_mm_add_epi16(p3, four), _mm_add_epi16(p3, p2));
- a = _mm_add_epi16(p3, _mm_add_epi16(p2, p1));
- a = _mm_add_epi16(_mm_add_epi16(p0, q0), a);
+ pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+ pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
- _mm_storel_epi64((__m128i *)&flat_op[2][i*8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+ pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+ pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+ pixelFilter_q));
+ pixetFilter_p2p1p0 = _mm_add_epi16(four,
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ pixetFilter_q2q1q0));
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(p7_16, p0_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(q7_16, q0_16)), 4);
+ flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(p3_16, p0_16)), 3);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(q3_16, q0_16)), 3);
- c = _mm_add_epi16(_mm_add_epi16(p5, eight), c);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[6][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ flat_q0p0 = _mm_packus_epi16(res_p, res_q);
- a = _mm_add_epi16(q1, a);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p2)), p1);
- _mm_storel_epi64((__m128i *)&flat_op[1][i*8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ sum_p7 = _mm_add_epi16(p7_16, p7_16);
+ sum_q7 = _mm_add_epi16(q7_16, q7_16);
+ sum_p3 = _mm_add_epi16(p3_16, p3_16);
+ sum_q3 = _mm_add_epi16(q3_16, q3_16);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p6)), p5);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[5][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p1_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q1_16)), 4);
+ flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
- a = _mm_add_epi16(q2, a);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p1)), p0);
- _mm_storel_epi64((__m128i *)&flat_op[0][i*8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p1_16)), 3);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q1_16)), 3);
+ flat_q1p1 = _mm_packus_epi16(res_p, res_q);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p5)), p4);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[4][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+ sum_q3 = _mm_add_epi16(sum_q3, q3_16);
- a = _mm_add_epi16(q3, a);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p0)), q0);
- _mm_storel_epi64((__m128i *)&flat_oq[0][i*8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p2_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q2_16)), 4);
+ flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p4)), p3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[3][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
- b = _mm_add_epi16(q3, b);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p2, q0)), q1);
- _mm_storel_epi64((__m128i *)&flat_oq[1][i*8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p2_16)), 3);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q2_16)), 3);
+ flat_q2p2 = _mm_packus_epi16(res_p, res_q);
- c = _mm_add_epi16(q4, c);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p3)), p2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[2][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p3_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q3_16)), 4);
+ flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
- b = _mm_add_epi16(q3, b);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p1, q1)), q2);
- _mm_storel_epi64((__m128i *)&flat_oq[2][i*8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
- a = _mm_add_epi16(q5, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p2)), p1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[1][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p4_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q4_16)), 4);
+ flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
- a = _mm_add_epi16(q6, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p1)), p0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[0][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p5_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q5_16)), 4);
+ flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p0)), q0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[0][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p6, q0)), q1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[1][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p5, q1)), q2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[2][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p4, q2)), q3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[3][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p3, q3)), q4);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[4][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p2, q4)), q5);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[5][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p1, q5)), q6);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[6][i*8],
- _mm_packus_epi16(workp_shft, workp_shft));
- }
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p6_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q6_16)), 4);
+ flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
}
// wide flat
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- work_a = _mm_loadl_epi64((__m128i *)ap[2]);
- p2 = _mm_loadl_epi64((__m128i *)flat_op[2]);
- work_a = _mm_andnot_si128(flat, work_a);
- p2 = _mm_and_si128(flat, p2);
- p2 = _mm_or_si128(work_a, p2);
- _mm_storel_epi64((__m128i *)flat_op[2], p2);
+ flat = _mm_shuffle_epi32(flat, 68);
+ flat2 = _mm_shuffle_epi32(flat2, 68);
- p1 = _mm_loadl_epi64((__m128i *)flat_op[1]);
- work_a = _mm_andnot_si128(flat, ps1);
- p1 = _mm_and_si128(flat, p1);
- p1 = _mm_or_si128(work_a, p1);
- _mm_storel_epi64((__m128i *)flat_op[1], p1);
+ q2p2 = _mm_andnot_si128(flat, q2p2);
+ flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+ q2p2 = _mm_or_si128(q2p2, flat_q2p2);
- p0 = _mm_loadl_epi64((__m128i *)flat_op[0]);
- work_a = _mm_andnot_si128(flat, ps0);
- p0 = _mm_and_si128(flat, p0);
- p0 = _mm_or_si128(work_a, p0);
- _mm_storel_epi64((__m128i *)flat_op[0], p0);
+ qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+ flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+ q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
- q0 = _mm_loadl_epi64((__m128i *)flat_oq[0]);
- work_a = _mm_andnot_si128(flat, qs0);
- q0 = _mm_and_si128(flat, q0);
- q0 = _mm_or_si128(work_a, q0);
- _mm_storel_epi64((__m128i *)flat_oq[0], q0);
+ qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+ flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+ q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
- q1 = _mm_loadl_epi64((__m128i *)flat_oq[1]);
- work_a = _mm_andnot_si128(flat, qs1);
- q1 = _mm_and_si128(flat, q1);
- q1 = _mm_or_si128(work_a, q1);
- _mm_storel_epi64((__m128i *)flat_oq[1], q1);
+ q6p6 = _mm_andnot_si128(flat2, q6p6);
+ flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+ q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+ _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
+ _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
- work_a = _mm_loadl_epi64((__m128i *)aq[2]);
- q2 = _mm_loadl_epi64((__m128i *)flat_oq[2]);
- work_a = _mm_andnot_si128(flat, work_a);
- q2 = _mm_and_si128(flat, q2);
- q2 = _mm_or_si128(work_a, q2);
- _mm_storel_epi64((__m128i *)flat_oq[2], q2);
+ q5p5 = _mm_andnot_si128(flat2, q5p5);
+ flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+ q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+ _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+ _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
- // write out op6 - op3
- {- unsigned char *dst = (s - 7 * p);
- for (i = 6; i > 2; i--) {- __m128i flat2_output;
- work_a = _mm_loadl_epi64((__m128i *)ap[i]);
- flat2_output = _mm_loadl_epi64((__m128i *)flat2_op[i]);
- work_a = _mm_andnot_si128(flat2, work_a);
- flat2_output = _mm_and_si128(flat2, flat2_output);
- work_a = _mm_or_si128(work_a, flat2_output);
- _mm_storel_epi64((__m128i *)dst, work_a);
- dst += p;
- }
- }
+ q4p4 = _mm_andnot_si128(flat2, q4p4);
+ flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+ q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+ _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+ _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
- work_a = _mm_loadl_epi64((__m128i *)flat_op[2]);
- p2 = _mm_loadl_epi64((__m128i *)flat2_op[2]);
- work_a = _mm_andnot_si128(flat2, work_a);
- p2 = _mm_and_si128(flat2, p2);
- p2 = _mm_or_si128(work_a, p2);
- _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+ q3p3 = _mm_andnot_si128(flat2, q3p3);
+ flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+ q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+ _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+ _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
- work_a = _mm_loadl_epi64((__m128i *)flat_op[1]);
- p1 = _mm_loadl_epi64((__m128i *)flat2_op[1]);
- work_a = _mm_andnot_si128(flat2, work_a);
- p1 = _mm_and_si128(flat2, p1);
- p1 = _mm_or_si128(work_a, p1);
- _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
+ q2p2 = _mm_andnot_si128(flat2, q2p2);
+ flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+ q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+ _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+ _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
- work_a = _mm_loadl_epi64((__m128i *)flat_op[0]);
- p0 = _mm_loadl_epi64((__m128i *)flat2_op[0]);
- work_a = _mm_andnot_si128(flat2, work_a);
- p0 = _mm_and_si128(flat2, p0);
- p0 = _mm_or_si128(work_a, p0);
- _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
+ q1p1 = _mm_andnot_si128(flat2, q1p1);
+ flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+ q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+ _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+ _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
- work_a = _mm_loadl_epi64((__m128i *)flat_oq[0]);
- q0 = _mm_loadl_epi64((__m128i *)flat2_oq[0]);
- work_a = _mm_andnot_si128(flat2, work_a);
- q0 = _mm_and_si128(flat2, q0);
- q0 = _mm_or_si128(work_a, q0);
- _mm_storel_epi64((__m128i *)(s - 0 * p), q0);
-
- work_a = _mm_loadl_epi64((__m128i *)flat_oq[1]);
- q1 = _mm_loadl_epi64((__m128i *)flat2_oq[1]);
- work_a = _mm_andnot_si128(flat2, work_a);
- q1 = _mm_and_si128(flat2, q1);
- q1 = _mm_or_si128(work_a, q1);
- _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
-
- work_a = _mm_loadl_epi64((__m128i *)flat_oq[2]);
- q2 = _mm_loadl_epi64((__m128i *)flat2_oq[2]);
- work_a = _mm_andnot_si128(flat2, work_a);
- q2 = _mm_and_si128(flat2, q2);
- q2 = _mm_or_si128(work_a, q2);
- _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
-
- // write out oq3 - oq7
- {- unsigned char *dst = (s + 3 * p);
- for (i = 3; i < 7; i++) {- __m128i flat2_output;
- work_a = _mm_loadl_epi64((__m128i *)aq[i]);
- flat2_output = _mm_loadl_epi64((__m128i *)flat2_oq[i]);
- work_a = _mm_andnot_si128(flat2, work_a);
- flat2_output = _mm_and_si128(flat2, flat2_output);
- work_a = _mm_or_si128(work_a, flat2_output);
- _mm_storel_epi64((__m128i *)dst, work_a);
- dst += p;
- }
- }
+ q0p0 = _mm_andnot_si128(flat2, q0p0);
+ flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+ q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+ _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+ _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
}
}