Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(128)

Issues Search
    My Issues | Starred     Open | Closed | All

Unified Diff: source/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c

Issue 996503002: libvpx: Pull from upstream (Closed) Base URL: https://chromium.googlesource.com/chromium/deps/libvpx.git@master
Patch Set: Created 5 years, 9 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « source/libvpx/vp9/common/vp9_rtcd_defs.pl ('k') | source/libvpx/vp9/decoder/vp9_decodeframe.c » ('j') | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: source/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
diff --git a/source/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c b/source/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
index 7e63f389eade6def507477c56b44b2fabe1c251e..4bd8ac4a315541280a804fa93de052d33cb73a1e 100644
--- a/source/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
+++ b/source/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
@@ -15,24 +15,38 @@
#include "vpx_ports/emmintrin_compat.h"
static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) {
- __m128i ubounded;
- __m128i lbounded;
- __m128i retval;
-
- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi16(1);
- const __m128i t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), bd - 8);
- const __m128i max = _mm_subs_epi16(
- _mm_subs_epi16(_mm_slli_epi16(one, bd), one), t80);
- const __m128i min = _mm_subs_epi16(zero, t80);
- ubounded = _mm_cmpgt_epi16(value, max);
- lbounded = _mm_cmplt_epi16(value, min);
- retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
- ubounded = _mm_and_si128(ubounded, max);
- lbounded = _mm_and_si128(lbounded, min);
- retval = _mm_or_si128(retval, ubounded);
- retval = _mm_or_si128(retval, lbounded);
- return retval;
+ __m128i ubounded;
+ __m128i lbounded;
+ __m128i retval;
+
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi16(1);
+ __m128i t80, max, min;
+
+ if (bd == 8) {
+ t80 = _mm_set1_epi16(0x80);
+ max = _mm_subs_epi16(
+ _mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80);
+ } else if (bd == 10) {
+ t80 = _mm_set1_epi16(0x200);
+ max = _mm_subs_epi16(
+ _mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80);
+ } else { // bd == 12
+ t80 = _mm_set1_epi16(0x800);
+ max = _mm_subs_epi16(
+ _mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80);
+ }
+
+ min = _mm_subs_epi16(zero, t80);
+
+ ubounded = _mm_cmpgt_epi16(value, max);
+ lbounded = _mm_cmplt_epi16(value, min);
+ retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
+ ubounded = _mm_and_si128(ubounded, max);
+ lbounded = _mm_and_si128(lbounded, min);
+ retval = _mm_or_si128(retval, ubounded);
+ retval = _mm_or_si128(retval, lbounded);
+ return retval;
}
// TODO(debargha, peter): Break up large functions into smaller ones
@@ -45,14 +59,7 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
int bd) {
const __m128i zero = _mm_set1_epi16(0);
const __m128i one = _mm_set1_epi16(1);
- 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 q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0;
__m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0;
__m128i ps1, qs1, ps0, qs0;
@@ -68,6 +75,26 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
__m128i t4, t3, t80, t1;
__m128i eight, four;
+ 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);
+ } 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);
+ } 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);
+ }
+
q4 = _mm_load_si128((__m128i *)(s + 4 * p));
p4 = _mm_load_si128((__m128i *)(s - 5 * p));
q3 = _mm_load_si128((__m128i *)(s + 3 * p));
@@ -121,7 +148,13 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
// highbd_filter4
t4 = _mm_set1_epi16(4);
t3 = _mm_set1_epi16(3);
- t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), bd - 8);
+ if (bd == 8)
+ t80 = _mm_set1_epi16(0x80);
+ else if (bd == 10)
+ t80 = _mm_set1_epi16(0x200);
+ else // bd == 12
+ t80 = _mm_set1_epi16(0x800);
+
t1 = _mm_set1_epi16(0x1);
ps1 = _mm_subs_epi16(p1, t80);
@@ -136,7 +169,6 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
filt = _mm_adds_epi16(filt, work_a);
filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
filt = _mm_and_si128(filt, mask);
-
filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
@@ -153,13 +185,13 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
filt = _mm_adds_epi16(filter1, t1);
filt = _mm_srai_epi16(filt, 1);
filt = _mm_andnot_si128(hev, filt);
-
qs1 = _mm_adds_epi16(
signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
t80);
ps1 = _mm_adds_epi16(
signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
t80);
+
// end highbd_filter4
// loopfilter done
@@ -175,7 +207,14 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
flat = _mm_max_epi16(work, flat);
work = _mm_max_epi16(abs_p1p0, abs_q1q0);
flat = _mm_max_epi16(work, 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);
// end flat_mask4
@@ -215,7 +254,13 @@ static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
_mm_subs_epu16(q0, q7)));
flat2 = _mm_max_epi16(work, flat2);
- flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, bd - 8));
+ if (bd == 8)
+ flat2 = _mm_subs_epu16(flat2, one);
+ else if (bd == 10)
+ flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2));
+ else // bd == 12
+ flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4));
+
flat2 = _mm_cmpeq_epi16(flat2, zero);
flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
// end highbd_flat_mask5
@@ -486,15 +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));
@@ -512,18 +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));
@@ -575,7 +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
@@ -706,15 +775,7 @@ void vp9_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
const uint8_t *_thresh,
int count, int bd) {
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_loadu_si128((__m128i *)(s - 4 * p));
__m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
@@ -737,30 +798,63 @@ void vp9_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
__m128i work;
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);
- const __m128i tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), bd - 8);
- const __m128i tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), bd - 8);
- const __m128i t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 16 - bd);
+ __m128i t80;
+ __m128i tff80;
+ __m128i tffe0;
+ __m128i t1f;
// equivalent to shifting 0x1f left by bitdepth - 8
// and setting new bits to 1
const __m128i t1 = _mm_set1_epi16(0x1);
- const __m128i t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 16 - bd);
+ __m128i t7f;
// equivalent to shifting 0x7f left by bitdepth - 8
// and setting new bits to 1
- const __m128i ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)),
- t80);
- const __m128i ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)),
- t80);
- const __m128i qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)),
- t80);
- const __m128i qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)),
- 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);
+ tff80 = _mm_set1_epi16(0xff80);
+ tffe0 = _mm_set1_epi16(0xffe0);
+ t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8);
+ t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8);
+ } 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_slli_epi16(_mm_set1_epi16(0x80), 2);
+ tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2);
+ tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2);
+ t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6);
+ t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6);
+ } 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_slli_epi16(_mm_set1_epi16(0x80), 4);
+ tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4);
+ tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4);
+ t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4);
+ t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4);
+ }
+
+ ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+ ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+ qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+ qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+
// filter_mask and hev_mask
flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu16(flat, thresh);
@@ -796,6 +890,7 @@ void vp9_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
filt = _mm_adds_epi16(filt, work_a);
filt = _mm_adds_epi16(filt, work_a);
filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+
// (vp9_filter + 3 * (qs0 - ps0)) & mask
filt = _mm_and_si128(filt, mask);
« no previous file with comments | « source/libvpx/vp9/common/vp9_rtcd_defs.pl ('k') | source/libvpx/vp9/decoder/vp9_decodeframe.c » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698