commit 7ffbc9d0fd2c0e93de9befde73122654c3fac185
author Cherma Rajan A <cherma.rajan@ittiam.com> Wed Jun 13 15:09:29 2018 +0530
committer Harish Mahendrakar <harish.mahendrakar@ittiam.com> Wed Jun 13 17:08:34 2018 +0000
tree 0ac3bf56565e2042fda59629637e3db10232c1a8
parent 0ba3c471eaa44bc1396817bf16a6f5b22dc868ed

Add NEON optimization of lpf_vertical_8

Intrinsic optimization and unit test changes of vertical
filter 8 added.

Performance gain w.r.t. C,
lpf_vertical_8  ~3.1x

Change-Id: Icbb2b43867c8a14c39af1e24d08a20662ed39937

aom_dsp/aom_dsp_rtcd_defs.pl

diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 15fac17..dfbf31c 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -379,7 +379,7 @@
 specialize qw/aom_lpf_vertical_6 sse2/;
 
 add_proto qw/void aom_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/aom_lpf_vertical_8 sse2/;
+specialize qw/aom_lpf_vertical_8 sse2 neon/;
 
 add_proto qw/void aom_lpf_vertical_8_dual/, "uint8_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";
 specialize qw/aom_lpf_vertical_8_dual sse2/;

aom_dsp/arm/loopfilter_neon.c

diff --git a/aom_dsp/arm/loopfilter_neon.c b/aom_dsp/arm/loopfilter_neon.c
index dfbbab8..4e39175 100644
--- a/aom_dsp/arm/loopfilter_neon.c
+++ b/aom_dsp/arm/loopfilter_neon.c
@@ -449,6 +449,49 @@
   store_u8_8x16(src - 8, stride, row0, row1, row2, row3);
 }
 
+void aom_lpf_vertical_8_neon(uint8_t *src, int stride, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  uint32x2x2_t p2q2_p1q1, p3q3_p0q0;
+  uint32x2_t pq_rev;
+  uint8x8_t p3q0, p2q1, p1q2, p0q3;
+  uint8x8_t p0q0, p1q1, p2q2, p3q3;
+
+  // row0: p3 p2 p1 p0 | q0 q1 q2 q3
+  // row1: p3 p2 p1 p0 | q0 q1 q2 q3
+  // row2: p3 p2 p1 p0 | q0 q1 q2 q3
+  // row3: p3 p2 p1 p0 | q0 q1 q2 q3
+  load_u8_8x4(src - 4, stride, &p3q0, &p2q1, &p1q2, &p0q3);
+
+  transpose_u8_8x4(&p3q0, &p2q1, &p1q2, &p0q3);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p0q3));
+  p3q3_p0q0 = vtrn_u32(vreinterpret_u32_u8(p3q0), pq_rev);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q2));
+  p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q1), pq_rev);
+
+  p0q0 = vreinterpret_u8_u32(vrev64_u32(p3q3_p0q0.val[1]));
+  p1q1 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1]));
+  p2q2 = vreinterpret_u8_u32(p2q2_p1q1.val[0]);
+  p3q3 = vreinterpret_u8_u32(p3q3_p0q0.val[0]);
+
+  lpf_8_neon(&p3q3, &p2q2, &p1q1, &p0q0, *blimit, *limit, *thresh);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p0q0));
+  p3q3_p0q0 = vtrn_u32(vreinterpret_u32_u8(p3q3), pq_rev);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q1));
+  p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q2), pq_rev);
+
+  p0q3 = vreinterpret_u8_u32(vrev64_u32(p3q3_p0q0.val[1]));
+  p1q2 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1]));
+  p2q1 = vreinterpret_u8_u32(p2q2_p1q1.val[0]);
+  p3q0 = vreinterpret_u8_u32(p3q3_p0q0.val[0]);
+  transpose_u8_8x4(&p3q0, &p2q1, &p1q2, &p0q3);
+
+  store_u8_8x4(src - 4, stride, p3q0, p2q1, p1q2, p0q3);
+}
+
 void aom_lpf_horizontal_8_neon(uint8_t *src, int stride, const uint8_t *blimit,
                                const uint8_t *limit, const uint8_t *thresh) {
   uint8x8_t p0q0, p1q1, p2q2, p3q3;

test/lpf_test.cc

diff --git a/test/lpf_test.cc b/test/lpf_test.cc
index 4ee14c8..c31276e 100644
--- a/test/lpf_test.cc
+++ b/test/lpf_test.cc
@@ -566,6 +566,7 @@
 #if HAVE_NEON
 const loop_param_t kLoop8Test6[] = {
   make_tuple(&aom_lpf_vertical_14_neon, &aom_lpf_vertical_14_c, 8),
+  make_tuple(&aom_lpf_vertical_8_neon, &aom_lpf_vertical_8_c, 8),
   make_tuple(&aom_lpf_horizontal_8_neon, &aom_lpf_horizontal_8_c, 8)
 };