ref: 49aa9b1f12af5faf07e450d1a546ce978e652023
parent: 7aa27bd62fd69ed3c530b1b1a61e574f9ec7aa32
author: Linfeng Zhang <linfengz@google.com>
date: Wed Oct 5 06:45:03 EDT 2016
[vpx highbd lpf NEON 2/6] vertical 4 BUG=webm:1300 Change-Id: Ia33a9f2d6c7e2e6b3497ad6f1a09439a85b33983
--- a/test/lpf_test.cc
+++ b/test/lpf_test.cc
@@ -521,7 +521,13 @@
make_tuple(&vpx_highbd_lpf_horizontal_4_neon,
&vpx_highbd_lpf_horizontal_4_c, 10),
make_tuple(&vpx_highbd_lpf_horizontal_4_neon,
- &vpx_highbd_lpf_horizontal_4_c, 12)));
+ &vpx_highbd_lpf_horizontal_4_c, 12),
+ make_tuple(&vpx_highbd_lpf_vertical_4_neon,
+ &vpx_highbd_lpf_vertical_4_c, 8),
+ make_tuple(&vpx_highbd_lpf_vertical_4_neon,
+ &vpx_highbd_lpf_vertical_4_c, 10),
+ make_tuple(&vpx_highbd_lpf_vertical_4_neon,
+ &vpx_highbd_lpf_vertical_4_c, 12)));
INSTANTIATE_TEST_CASE_P(
NEON, Loop8Test9Param,
::testing::Values(make_tuple(&vpx_highbd_lpf_horizontal_4_dual_neon,
@@ -529,7 +535,13 @@
make_tuple(&vpx_highbd_lpf_horizontal_4_dual_neon,
&vpx_highbd_lpf_horizontal_4_dual_c, 10),
make_tuple(&vpx_highbd_lpf_horizontal_4_dual_neon,
- &vpx_highbd_lpf_horizontal_4_dual_c, 12)));
+ &vpx_highbd_lpf_horizontal_4_dual_c, 12),
+ make_tuple(&vpx_highbd_lpf_vertical_4_dual_neon,
+ &vpx_highbd_lpf_vertical_4_dual_c, 8),
+ make_tuple(&vpx_highbd_lpf_vertical_4_dual_neon,
+ &vpx_highbd_lpf_vertical_4_dual_c, 10),
+ make_tuple(&vpx_highbd_lpf_vertical_4_dual_neon,
+ &vpx_highbd_lpf_vertical_4_dual_c, 12)));
#else
INSTANTIATE_TEST_CASE_P(
NEON, Loop8Test6Param,
--- a/vpx_dsp/arm/highbd_loopfilter_neon.c
+++ b/vpx_dsp/arm/highbd_loopfilter_neon.c
@@ -157,6 +157,32 @@
vst1q_u16(s, s3);
}
+static INLINE void store_4x8(uint16_t *s, const int p, const uint16x8_t p1,
+ const uint16x8_t p0, const uint16x8_t q0,
+ const uint16x8_t q1) {+ uint16x8x4_t o;
+
+ o.val[0] = p1;
+ o.val[1] = p0;
+ o.val[2] = q0;
+ o.val[3] = q1;
+ vst4q_lane_u16(s, o, 0);
+ s += p;
+ vst4q_lane_u16(s, o, 1);
+ s += p;
+ vst4q_lane_u16(s, o, 2);
+ s += p;
+ vst4q_lane_u16(s, o, 3);
+ s += p;
+ vst4q_lane_u16(s, o, 4);
+ s += p;
+ vst4q_lane_u16(s, o, 5);
+ s += p;
+ vst4q_lane_u16(s, o, 6);
+ s += p;
+ vst4q_lane_u16(s, o, 7);
+}
+
void vpx_highbd_lpf_horizontal_4_neon(uint16_t *s, int p, const uint8_t *blimit,
const uint8_t *limit,
const uint8_t *thresh, int bd) {@@ -177,4 +203,29 @@
const uint8_t *thresh1, int bd) {vpx_highbd_lpf_horizontal_4_neon(s, p, blimit0, limit0, thresh0, bd);
vpx_highbd_lpf_horizontal_4_neon(s + 8, p, blimit1, limit1, thresh1, bd);
+}
+
+void vpx_highbd_lpf_vertical_4_neon(uint16_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int bd) {+ uint16x8_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+ mask, hev;
+
+ load_8x8(s - 4, p, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+ transpose_s16_8x8((int16x8_t *)&p3, (int16x8_t *)&p2, (int16x8_t *)&p1,
+ (int16x8_t *)&p0, (int16x8_t *)&q0, (int16x8_t *)&q1,
+ (int16x8_t *)&q2, (int16x8_t *)&q3);
+ load_thresh(blimit, limit, thresh, &blimit_vec, &limit_vec, &thresh_vec, bd);
+ filter_hev_mask4(limit_vec, blimit_vec, thresh_vec, p3, p2, p1, p0, q0, q1,
+ q2, q3, &hev, &mask);
+ filter4(mask, hev, p1, p0, q0, q1, &p1, &p0, &q0, &q1, bd);
+ store_4x8(s - 2, p, p1, p0, q0, q1);
+}
+
+void vpx_highbd_lpf_vertical_4_dual_neon(
+ uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+ const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+ const uint8_t *thresh1, int bd) {+ vpx_highbd_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0, bd);
+ vpx_highbd_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1, bd);
}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -469,10 +469,10 @@
specialize qw/vpx_highbd_lpf_vertical_8_dual sse2/;
add_proto qw/void vpx_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
- specialize qw/vpx_highbd_lpf_vertical_4 sse2/;
+ specialize qw/vpx_highbd_lpf_vertical_4 sse2 neon/;
add_proto qw/void vpx_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
- specialize qw/vpx_highbd_lpf_vertical_4_dual sse2/;
+ specialize qw/vpx_highbd_lpf_vertical_4_dual sse2 neon/;
add_proto qw/void vpx_highbd_lpf_horizontal_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
specialize qw/vpx_highbd_lpf_horizontal_16 sse2/;