Commit ed2dc59c authored by Jingning Han's avatar Jingning Han

Integral projection based motion estimation

This commit introduces a new block match motion estimation
using integral projection measurement. The 2-D block and the nearby
region is projected onto the horizontal and vertical 1-D vectors,
respectively. It then runs vector match, instead of block match,
over the two separate 1-D vectors to locate the motion compensated
reference block.

This process is run per 64x64 block to align the reference before
choosing partitioning in speed 6. The overall CPU cycle cost due
to this additional 64x64 block match (SSE2 version) takes around 2%
at low bit-rate rtc speed 6. When strong motion activities exist in
the video sequence, it substantially improves the partition
selection accuracy, thereby achieving better compression performance
and lower CPU cycles.

The experiments were tested in RTC speed -6 setting:
cloud 1080p 500 kbps
17006 b/f, 37.086 dB, 5386 ms ->
16669 b/f, 37.970 dB, 5085 ms (>0.9dB gain and 6% faster)

pedestrian_area 1080p 500 kbps
53537 b/f, 36.771 dB, 18706 ms ->
51897 b/f, 36.792 dB, 18585 ms (4% bit-rate savings)

blue_sky 1080p 500 kbps
70214 b/f, 33.600 dB, 13979 ms ->
53885 b/f, 33.645 dB, 10878 ms (30% bit-rate savings, 25% faster)

jimred 400 kbps
13380 b/f, 36.014 dB, 5723 ms ->
13377 b/f, 36.087 dB, 5831 ms  (2% bit-rate savings, 2% slower)

Change-Id: Iffdb6ea5b16b77016bfa3dd3904d284168ae649c
parent 5041aa0f
......@@ -1109,6 +1109,15 @@ specialize qw/vp9_avg_8x8 sse2 neon/;
add_proto qw/unsigned int vp9_avg_4x4/, "const uint8_t *, int p";
specialize qw/vp9_avg_4x4 sse2/;
add_proto qw/void vp9_int_pro_row/, "int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height";
specialize qw/vp9_int_pro_row sse2/;
add_proto qw/int16_t vp9_int_pro_col/, "uint8_t const *ref, const int width";
specialize qw/vp9_int_pro_col sse2/;
add_proto qw/int vp9_vector_sad/, "int16_t const *ref, int16_t const *src, const int width";
specialize qw/vp9_vector_sad sse2/;
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/unsigned int vp9_highbd_avg_8x8/, "const uint8_t *, int p";
specialize qw/vp9_highbd_avg_8x8/;
......
......@@ -28,6 +28,36 @@ unsigned int vp9_avg_4x4_c(const uint8_t *s, int p) {
return (sum + 8) >> 4;
}
// Integer projection onto row vectors.
void vp9_int_pro_row_c(int16_t *hbuf, uint8_t const *ref,
const int ref_stride, const int height) {
int idx;
for (idx = 0; idx < 16; ++idx) {
int i;
hbuf[idx] = 0;
for (i = 0; i < height; ++i)
hbuf[idx] += ref[i * ref_stride];
++ref;
}
}
int16_t vp9_int_pro_col_c(uint8_t const *ref, const int width) {
int idx;
int16_t sum = 0;
for (idx = 0; idx < width; ++idx)
sum += ref[idx];
return sum;
}
int vp9_vector_sad_c(int16_t const *ref, int16_t const *src,
const int width) {
int i;
int this_sad = 0;
for (i = 0; i < width; ++i)
this_sad += abs(ref[i] - src[i]);
return this_sad;
}
#if CONFIG_VP9_HIGHBITDEPTH
unsigned int vp9_highbd_avg_8x8_c(const uint8_t *s8, int p) {
int i, j;
......
......@@ -507,6 +507,119 @@ void vp9_set_vbp_thresholds(VP9_COMP *cpi, int q) {
}
}
#if CONFIG_VP9_HIGHBITDEPTH
#define GLOBAL_MOTION 0
#else
#define GLOBAL_MOTION 1
#endif
#if GLOBAL_MOTION
static int vector_match(int16_t *ref, int16_t *src) {
int best_sad = INT_MAX;
int this_sad;
int d;
int center, offset = 0;
for (d = 0; d <= 64; d += 16) {
this_sad = vp9_vector_sad(&ref[d], src, 64);
if (this_sad < best_sad) {
best_sad = this_sad;
offset = d;
}
}
center = offset;
for (d = -8; d <= 8; d += 4) {
int this_pos = offset + d;
// check limit
if (this_pos < 0 || this_pos > 64 || this_pos == 32)
continue;
this_sad = vp9_vector_sad(&ref[this_pos], src, 64);
if (this_sad < best_sad) {
best_sad = this_sad;
center = this_pos;
}
}
offset = center;
for (d = -4; d <= 4; d += 2) {
int this_pos = offset + d;
// check limit
if (this_pos < 0 || this_pos > 64 || this_pos == 32)
continue;
this_sad = vp9_vector_sad(&ref[this_pos], src, 64);
if (this_sad < best_sad) {
best_sad = this_sad;
center = this_pos;
}
}
offset = center;
for (d = -2; d <= 2; d += 1) {
int this_pos = offset + d;
// check limit
if (this_pos < 0 || this_pos > 64 || this_pos == 32)
continue;
this_sad = vp9_vector_sad(&ref[this_pos], src, 64);
if (this_sad < best_sad) {
best_sad = this_sad;
center = this_pos;
}
}
return (center - 32);
}
static void motion_estimation(MACROBLOCK *x) {
MACROBLOCKD *xd = &x->e_mbd;
DECLARE_ALIGNED(16, int16_t, hbuf[128]);
DECLARE_ALIGNED(16, int16_t, vbuf[128]);
DECLARE_ALIGNED(16, int16_t, src_hbuf[64]);
DECLARE_ALIGNED(16, int16_t, src_vbuf[64]);
int idx;
const int stride = 64;
const int search_width = 128;
const int search_height = 128;
const int src_stride = x->plane[0].src.stride;
const int ref_stride = xd->plane[0].pre[0].stride;
uint8_t const *ref_buf, *src_buf;
MV *tmp_mv = &xd->mi[0].src_mi->mbmi.mv[0].as_mv;
// Set up prediction 1-D reference set
ref_buf = xd->plane[0].pre[0].buf + (-32);
for (idx = 0; idx < search_width; idx += 16) {
vp9_int_pro_row(&hbuf[idx], ref_buf, ref_stride, 64);
ref_buf += 16;
}
ref_buf = xd->plane[0].pre[0].buf + (-32) * ref_stride;
for (idx = 0; idx < search_height; ++idx) {
vbuf[idx] = vp9_int_pro_col(ref_buf, 64);
ref_buf += ref_stride;
}
// Set up src 1-D reference set
for (idx = 0; idx < stride; idx += 16) {
src_buf = x->plane[0].src.buf + idx;
vp9_int_pro_row(&src_hbuf[idx], src_buf, src_stride, 64);
}
src_buf = x->plane[0].src.buf;
for (idx = 0; idx < stride; ++idx) {
src_vbuf[idx] = vp9_int_pro_col(src_buf, 64);
src_buf += src_stride;
}
// Find the best match per 1-D search
tmp_mv->col = vector_match(hbuf, src_hbuf);
tmp_mv->row = vector_match(vbuf, src_vbuf);
tmp_mv->row *= 8;
tmp_mv->col *= 8;
x->pred_mv[LAST_FRAME] = *tmp_mv;
}
#endif
// This function chooses partitioning based on the variance between source and
// reconstructed last, where variance is computed for downs-sampled inputs.
......@@ -551,6 +664,11 @@ static void choose_partitioning(VP9_COMP *cpi,
mbmi->ref_frame[1] = NONE;
mbmi->sb_type = BLOCK_64X64;
mbmi->mv[0].as_int = 0;
#if GLOBAL_MOTION
motion_estimation(x);
#endif
vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
for (i = 1; i <= 2; ++i) {
......
......@@ -56,3 +56,117 @@ unsigned int vp9_avg_4x4_sse2(const uint8_t *s, int p) {
avg = _mm_extract_epi16(s0, 0);
return (avg + 8) >> 4;
}
void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
const int ref_stride, const int height) {
int idx;
__m128i zero = _mm_setzero_si128();
__m128i src_line = _mm_load_si128((const __m128i *)ref);
__m128i s0 = _mm_unpacklo_epi8(src_line, zero);
__m128i s1 = _mm_unpackhi_epi8(src_line, zero);
__m128i t0, t1;
int height_1 = height - 1;
ref += ref_stride;
for (idx = 1; idx < height_1; idx += 2) {
src_line = _mm_load_si128((const __m128i *)ref);
t0 = _mm_unpacklo_epi8(src_line, zero);
t1 = _mm_unpackhi_epi8(src_line, zero);
s0 = _mm_adds_epu16(s0, t0);
s1 = _mm_adds_epu16(s1, t1);
ref += ref_stride;
src_line = _mm_load_si128((const __m128i *)ref);
t0 = _mm_unpacklo_epi8(src_line, zero);
t1 = _mm_unpackhi_epi8(src_line, zero);
s0 = _mm_adds_epu16(s0, t0);
s1 = _mm_adds_epu16(s1, t1);
ref += ref_stride;
}
src_line = _mm_load_si128((const __m128i *)ref);
t0 = _mm_unpacklo_epi8(src_line, zero);
t1 = _mm_unpackhi_epi8(src_line, zero);
s0 = _mm_adds_epu16(s0, t0);
s1 = _mm_adds_epu16(s1, t1);
_mm_store_si128((__m128i *)hbuf, s0);
hbuf += 8;
_mm_store_si128((__m128i *)hbuf, s1);
}
int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {
__m128i zero = _mm_setzero_si128();
__m128i src_line = _mm_load_si128((const __m128i *)ref);
__m128i s0 = _mm_sad_epu8(src_line, zero);
__m128i s1;
(void) width; // width = 64
ref += 16;
src_line = _mm_load_si128((const __m128i *)ref);
s1 = _mm_sad_epu8(src_line, zero);
s0 = _mm_adds_epu16(s0, s1);
ref += 16;
src_line = _mm_load_si128((const __m128i *)ref);
s1 = _mm_sad_epu8(src_line, zero);
s0 = _mm_adds_epu16(s0, s1);
ref += 16;
src_line = _mm_load_si128((const __m128i *)ref);
s1 = _mm_sad_epu8(src_line, zero);
s0 = _mm_adds_epu16(s0, s1);
s1 = _mm_srli_si128(s0, 8);
s0 = _mm_adds_epu16(s0, s1);
return _mm_extract_epi16(s0, 0);
}
int vp9_vector_sad_sse2(int16_t const *ref, int16_t const *src,
const int width) {
int idx;
__m128i zero = _mm_setzero_si128();
__m128i sum;
__m128i v0 = _mm_loadu_si128((const __m128i *)ref);
__m128i v1 = _mm_load_si128((const __m128i *)src);
__m128i diff = _mm_subs_epi16(v0, v1);
__m128i sign = _mm_srai_epi16(diff, 15);
diff = _mm_xor_si128(diff, sign);
sum = _mm_sub_epi16(diff, sign);
(void) width; // width = 64;
ref += 8;
src += 8;
v0 = _mm_unpacklo_epi16(sum, zero);
v1 = _mm_unpackhi_epi16(sum, zero);
sum = _mm_add_epi32(v0, v1);
for (idx = 1; idx < 8; ++idx) {
v0 = _mm_loadu_si128((const __m128i *)ref);
v1 = _mm_load_si128((const __m128i *)src);
diff = _mm_subs_epi16(v0, v1);
sign = _mm_srai_epi16(diff, 15);
diff = _mm_xor_si128(diff, sign);
diff = _mm_sub_epi16(diff, sign);
v0 = _mm_unpacklo_epi16(diff, zero);
v1 = _mm_unpackhi_epi16(diff, zero);
sum = _mm_add_epi32(sum, v0);
sum = _mm_add_epi32(sum, v1);
ref += 8;
src += 8;
}
v0 = _mm_srli_si128(sum, 8);
sum = _mm_add_epi32(sum, v0);
v0 = _mm_srli_epi64(sum, 32);
sum = _mm_add_epi32(sum, v0);
return _mm_cvtsi128_si32(sum);
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment