Commit 28762341 authored by Linfeng Zhang's avatar Linfeng Zhang Committed by Gerrit Code Review
Browse files

Merge changes Ib9105462,Idfac00ed,If8d8a0e2

* changes:
  cosmetics: NEON scaling code
  Refactor convolve NEON code
  Refactor convolve code
parents a059dc09 7c052972
......@@ -78,7 +78,16 @@ class ScaleTest : public VpxScaleBase,
src_width, src_height, dst_width, dst_height));
ReferenceScaleFrame(filter_type, phase_scaler);
ScaleFrame(filter_type, phase_scaler);
PrintDiff();
if (memcmp(dst_img_.buffer_alloc, ref_img_.buffer_alloc,
ref_img_.frame_size)) {
printf(
"filter_type = %d, phase_scaler = %d, src_width = %4d, "
"src_height = %4d, dst_width = %4d, dst_height = %4d, "
"scale factor = %d:%d\n",
filter_type, phase_scaler, src_width, src_height,
dst_width, dst_height, sf_down, sf_up);
PrintDiff();
}
CompareImages(dst_img_);
DeallocScaleImages();
}
......@@ -134,8 +143,8 @@ TEST_P(ScaleTest, DISABLED_Speed) {
static const int kCountSpeedTestBlock = 100;
static const int kNumScaleFactorsToTest = 4;
static const int kScaleFactors[] = { 1, 2, 3, 4 };
const int src_height = 1280;
const int src_width = 720;
const int src_width = 1280;
const int src_height = 720;
for (INTERP_FILTER filter_type = 2; filter_type < 4; ++filter_type) {
for (int phase_scaler = 0; phase_scaler < 2; ++phase_scaler) {
for (int sf_up_idx = 0; sf_up_idx < kNumScaleFactorsToTest; ++sf_up_idx) {
......
......@@ -170,37 +170,18 @@ static INLINE void scale_plane_4_to_1_bilinear_phase_non_0_neon(
} while (--y);
}
static INLINE uint8x8_t scale_filter(const uint8x8_t *const s,
const int16x8_t filters) {
const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
const int16x8_t filter4 = vdupq_lane_s16(vget_high_s16(filters), 0);
int16x8_t ss[8];
ss[0] = vreinterpretq_s16_u16(vmovl_u8(s[0]));
ss[1] = vreinterpretq_s16_u16(vmovl_u8(s[1]));
ss[2] = vreinterpretq_s16_u16(vmovl_u8(s[2]));
ss[3] = vreinterpretq_s16_u16(vmovl_u8(s[3]));
ss[4] = vreinterpretq_s16_u16(vmovl_u8(s[4]));
ss[5] = vreinterpretq_s16_u16(vmovl_u8(s[5]));
ss[6] = vreinterpretq_s16_u16(vmovl_u8(s[6]));
ss[7] = vreinterpretq_s16_u16(vmovl_u8(s[7]));
return convolve8_8(ss[0], ss[1], ss[2], ss[3], ss[4], ss[5], ss[6], ss[7],
filters, filter3, filter4);
}
static void scale_plane_2_to_1_general_neon(const uint8_t *src,
const int src_stride, uint8_t *dst,
const int dst_stride, const int w,
const int h,
const int16_t *const coef,
uint8_t *const temp_buffer) {
const int max_width_hor = (w + 3) & ~3;
const int max_width_ver = (w + 7) & ~7;
const int max_height_hor = (2 * h + SUBPEL_TAPS - 2 + 7) & ~7;
const int max_height_ver = (h + 3) & ~3;
const int width_hor = (w + 3) & ~3;
const int width_ver = (w + 7) & ~7;
const int height_hor = (2 * h + SUBPEL_TAPS - 2 + 7) & ~7;
const int height_ver = (h + 3) & ~3;
const int16x8_t filters = vld1q_s16(coef);
int x, y = max_height_hor;
int x, y = height_hor;
uint8_t *t = temp_buffer;
uint8x8_t s[14], d[4];
......@@ -208,14 +189,14 @@ static void scale_plane_2_to_1_general_neon(const uint8_t *src,
src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 1;
// horizontal
// horizontal 4x8
// Note: processing 4x8 is about 20% faster than processing row by row using
// vld4_u8().
do {
load_u8_8x8(src + 2, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
&s[6], &s[7]);
transpose_u8_8x8(&s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6], &s[7]);
x = max_width_hor;
x = width_hor;
do {
src += 8;
......@@ -224,31 +205,31 @@ static void scale_plane_2_to_1_general_neon(const uint8_t *src,
transpose_u8_8x8(&s[6], &s[7], &s[8], &s[9], &s[10], &s[11], &s[12],
&s[13]);
d[0] = scale_filter(&s[0], filters); // 00 10 20 30 40 50 60 70
d[1] = scale_filter(&s[2], filters); // 01 11 21 31 41 51 61 71
d[2] = scale_filter(&s[4], filters); // 02 12 22 32 42 52 62 72
d[3] = scale_filter(&s[6], filters); // 03 13 23 33 43 53 63 73
d[0] = scale_filter_8(&s[0], filters); // 00 10 20 30 40 50 60 70
d[1] = scale_filter_8(&s[2], filters); // 01 11 21 31 41 51 61 71
d[2] = scale_filter_8(&s[4], filters); // 02 12 22 32 42 52 62 72
d[3] = scale_filter_8(&s[6], filters); // 03 13 23 33 43 53 63 73
// 00 01 02 03 40 41 42 43
// 10 11 12 13 50 51 52 53
// 20 21 22 23 60 61 62 63
// 30 31 32 33 70 71 72 73
transpose_u8_8x4(&d[0], &d[1], &d[2], &d[3]);
vst1_lane_u32((uint32_t *)(t + 0 * max_width_hor),
vreinterpret_u32_u8(d[0]), 0);
vst1_lane_u32((uint32_t *)(t + 1 * max_width_hor),
vreinterpret_u32_u8(d[1]), 0);
vst1_lane_u32((uint32_t *)(t + 2 * max_width_hor),
vreinterpret_u32_u8(d[2]), 0);
vst1_lane_u32((uint32_t *)(t + 3 * max_width_hor),
vreinterpret_u32_u8(d[3]), 0);
vst1_lane_u32((uint32_t *)(t + 4 * max_width_hor),
vreinterpret_u32_u8(d[0]), 1);
vst1_lane_u32((uint32_t *)(t + 5 * max_width_hor),
vreinterpret_u32_u8(d[1]), 1);
vst1_lane_u32((uint32_t *)(t + 6 * max_width_hor),
vreinterpret_u32_u8(d[2]), 1);
vst1_lane_u32((uint32_t *)(t + 7 * max_width_hor),
vreinterpret_u32_u8(d[3]), 1);
vst1_lane_u32((uint32_t *)(t + 0 * width_hor), vreinterpret_u32_u8(d[0]),
0);
vst1_lane_u32((uint32_t *)(t + 1 * width_hor), vreinterpret_u32_u8(d[1]),
0);
vst1_lane_u32((uint32_t *)(t + 2 * width_hor), vreinterpret_u32_u8(d[2]),
0);
vst1_lane_u32((uint32_t *)(t + 3 * width_hor), vreinterpret_u32_u8(d[3]),
0);
vst1_lane_u32((uint32_t *)(t + 4 * width_hor), vreinterpret_u32_u8(d[0]),
1);
vst1_lane_u32((uint32_t *)(t + 5 * width_hor), vreinterpret_u32_u8(d[1]),
1);
vst1_lane_u32((uint32_t *)(t + 6 * width_hor), vreinterpret_u32_u8(d[2]),
1);
vst1_lane_u32((uint32_t *)(t + 7 * width_hor), vreinterpret_u32_u8(d[3]),
1);
s[0] = s[8];
s[1] = s[9];
......@@ -260,29 +241,29 @@ static void scale_plane_2_to_1_general_neon(const uint8_t *src,
t += 4;
x -= 4;
} while (x);
src += 8 * src_stride - 2 * max_width_hor;
t += 7 * max_width_hor;
src += 8 * src_stride - 2 * width_hor;
t += 7 * width_hor;
y -= 8;
} while (y);
// vertical
x = max_width_ver;
// vertical 8x4
x = width_ver;
t = temp_buffer;
do {
load_u8_8x8(t, max_width_hor, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
&s[6], &s[7]);
t += 6 * max_width_hor;
y = max_height_ver;
load_u8_8x8(t, width_hor, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
&s[7]);
t += 6 * width_hor;
y = height_ver;
do {
load_u8_8x8(t, max_width_hor, &s[6], &s[7], &s[8], &s[9], &s[10], &s[11],
load_u8_8x8(t, width_hor, &s[6], &s[7], &s[8], &s[9], &s[10], &s[11],
&s[12], &s[13]);
t += 8 * max_width_hor;
t += 8 * width_hor;
d[0] = scale_filter(&s[0], filters);
d[1] = scale_filter(&s[2], filters);
d[2] = scale_filter(&s[4], filters);
d[3] = scale_filter(&s[6], filters);
d[0] = scale_filter_8(&s[0], filters);
d[1] = scale_filter_8(&s[2], filters);
d[2] = scale_filter_8(&s[4], filters);
d[3] = scale_filter_8(&s[6], filters);
vst1_u8(dst + 0 * dst_stride, d[0]);
vst1_u8(dst + 1 * dst_stride, d[1]);
vst1_u8(dst + 2 * dst_stride, d[2]);
......@@ -298,9 +279,9 @@ static void scale_plane_2_to_1_general_neon(const uint8_t *src,
dst += 4 * dst_stride;
y -= 4;
} while (y);
t -= max_width_hor * (2 * max_height_ver + 6);
t -= width_hor * (2 * height_ver + 6);
t += 8;
dst -= max_height_ver * dst_stride;
dst -= height_ver * dst_stride;
dst += 8;
x -= 8;
} while (x);
......@@ -312,12 +293,12 @@ static void scale_plane_4_to_1_general_neon(const uint8_t *src,
const int h,
const int16_t *const coef,
uint8_t *const temp_buffer) {
const int max_width_hor = (w + 1) & ~1;
const int max_width_ver = (w + 7) & ~7;
const int max_height_hor = (4 * h + SUBPEL_TAPS - 2 + 7) & ~7;
const int max_height_ver = (h + 1) & ~1;
const int width_hor = (w + 1) & ~1;
const int width_ver = (w + 7) & ~7;
const int height_hor = (4 * h + SUBPEL_TAPS - 2 + 7) & ~7;
const int height_ver = (h + 1) & ~1;
const int16x8_t filters = vld1q_s16(coef);
int x, y = max_height_hor;
int x, y = height_hor;
uint8_t *t = temp_buffer;
uint8x8_t s[12], d[2];
......@@ -325,14 +306,14 @@ static void scale_plane_4_to_1_general_neon(const uint8_t *src,
src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 3;
// horizontal
// horizontal 2x8
// Note: processing 2x8 is about 20% faster than processing row by row using
// vld4_u8().
do {
load_u8_8x8(src + 4, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
&s[6], &s[7]);
transpose_u8_4x8(&s[0], &s[1], &s[2], &s[3], s[4], s[5], s[6], s[7]);
x = max_width_hor;
x = width_hor;
do {
uint8x8x2_t dd;
......@@ -342,26 +323,26 @@ static void scale_plane_4_to_1_general_neon(const uint8_t *src,
transpose_u8_8x8(&s[4], &s[5], &s[6], &s[7], &s[8], &s[9], &s[10],
&s[11]);
d[0] = scale_filter(&s[0], filters); // 00 10 20 30 40 50 60 70
d[1] = scale_filter(&s[4], filters); // 01 11 21 31 41 51 61 71
d[0] = scale_filter_8(&s[0], filters); // 00 10 20 30 40 50 60 70
d[1] = scale_filter_8(&s[4], filters); // 01 11 21 31 41 51 61 71
// dd.val[0]: 00 01 20 21 40 41 60 61
// dd.val[1]: 10 11 30 31 50 51 70 71
dd = vtrn_u8(d[0], d[1]);
vst1_lane_u16((uint16_t *)(t + 0 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 0 * width_hor),
vreinterpret_u16_u8(dd.val[0]), 0);
vst1_lane_u16((uint16_t *)(t + 1 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 1 * width_hor),
vreinterpret_u16_u8(dd.val[1]), 0);
vst1_lane_u16((uint16_t *)(t + 2 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 2 * width_hor),
vreinterpret_u16_u8(dd.val[0]), 1);
vst1_lane_u16((uint16_t *)(t + 3 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 3 * width_hor),
vreinterpret_u16_u8(dd.val[1]), 1);
vst1_lane_u16((uint16_t *)(t + 4 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 4 * width_hor),
vreinterpret_u16_u8(dd.val[0]), 2);
vst1_lane_u16((uint16_t *)(t + 5 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 5 * width_hor),
vreinterpret_u16_u8(dd.val[1]), 2);
vst1_lane_u16((uint16_t *)(t + 6 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 6 * width_hor),
vreinterpret_u16_u8(dd.val[0]), 3);
vst1_lane_u16((uint16_t *)(t + 7 * max_width_hor),
vst1_lane_u16((uint16_t *)(t + 7 * width_hor),
vreinterpret_u16_u8(dd.val[1]), 3);
s[0] = s[8];
......@@ -372,26 +353,26 @@ static void scale_plane_4_to_1_general_neon(const uint8_t *src,
t += 2;
x -= 2;
} while (x);
src += 8 * src_stride - 4 * max_width_hor;
t += 7 * max_width_hor;
src += 8 * src_stride - 4 * width_hor;
t += 7 * width_hor;
y -= 8;
} while (y);
// vertical
x = max_width_ver;
// vertical 8x2
x = width_ver;
t = temp_buffer;
do {
load_u8_8x4(t, max_width_hor, &s[0], &s[1], &s[2], &s[3]);
t += 4 * max_width_hor;
y = max_height_ver;
load_u8_8x4(t, width_hor, &s[0], &s[1], &s[2], &s[3]);
t += 4 * width_hor;
y = height_ver;
do {
load_u8_8x8(t, max_width_hor, &s[4], &s[5], &s[6], &s[7], &s[8], &s[9],
load_u8_8x8(t, width_hor, &s[4], &s[5], &s[6], &s[7], &s[8], &s[9],
&s[10], &s[11]);
t += 8 * max_width_hor;
t += 8 * width_hor;
d[0] = scale_filter(&s[0], filters);
d[1] = scale_filter(&s[4], filters);
d[0] = scale_filter_8(&s[0], filters);
d[1] = scale_filter_8(&s[4], filters);
vst1_u8(dst + 0 * dst_stride, d[0]);
vst1_u8(dst + 1 * dst_stride, d[1]);
......@@ -403,9 +384,9 @@ static void scale_plane_4_to_1_general_neon(const uint8_t *src,
dst += 2 * dst_stride;
y -= 2;
} while (y);
t -= max_width_hor * (4 * max_height_ver + 4);
t -= width_hor * (4 * height_ver + 4);
t += 8;
dst -= max_height_ver * dst_stride;
dst -= height_ver * dst_stride;
dst += 8;
x -= 8;
} while (x);
......@@ -423,7 +404,11 @@ void vp9_scale_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
const int dst_uv_h = dst_h / 2;
int scaled = 0;
// phase_scaler is usually 0 or 8.
assert(phase_scaler >= 0 && phase_scaler < 16);
if (2 * dst_w == src_w && 2 * dst_h == src_h) {
// 2 to 1
scaled = 1;
if (phase_scaler == 0) {
scale_plane_2_to_1_phase_0_neon(src->y_buffer, src->y_stride,
......@@ -448,10 +433,10 @@ void vp9_scale_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
dst_uv_w, dst_uv_h, c0, c1);
} else {
const int buffer_max_width = (dst_w + 3) & ~3;
const int buffer_max_height = (2 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
const int buffer_stride = (dst_w + 3) & ~3;
const int buffer_height = (2 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
uint8_t *const temp_buffer =
(uint8_t *)malloc(buffer_max_width * buffer_max_height);
(uint8_t *)malloc(buffer_stride * buffer_height);
if (temp_buffer) {
scale_plane_2_to_1_general_neon(
src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
......@@ -470,6 +455,7 @@ void vp9_scale_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
}
}
} else if (4 * dst_w == src_w && 4 * dst_h == src_h) {
// 4 to 1
scaled = 1;
if (phase_scaler == 0) {
scale_plane_4_to_1_phase_0_neon(src->y_buffer, src->y_stride,
......@@ -494,10 +480,10 @@ void vp9_scale_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
dst_uv_w, dst_uv_h, c0, c1);
} else {
const int buffer_max_width = (dst_w + 1) & ~1;
const int buffer_max_height = (4 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
const int buffer_stride = (dst_w + 1) & ~1;
const int buffer_height = (4 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
uint8_t *const temp_buffer =
(uint8_t *)malloc(buffer_max_width * buffer_max_height);
(uint8_t *)malloc(buffer_stride * buffer_height);
if (temp_buffer) {
scale_plane_4_to_1_general_neon(
src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
......@@ -520,6 +506,7 @@ void vp9_scale_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
if (scaled) {
vpx_extend_frame_borders(dst);
} else {
// Call c version for all other scaling ratios.
vp9_scale_and_extend_frame_c(src, dst, filter_type, phase_scaler);
}
}
......@@ -85,11 +85,10 @@ static INLINE void store_8x8(uint16_t *s, const ptrdiff_t p,
vst1q_u16(s, s7);
}
static INLINE int32x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
const int16x4_t s2, const int16x4_t s3,
const int16x4_t s4, const int16x4_t s5,
const int16x4_t s6, const int16x4_t s7,
const int16x8_t filters) {
static INLINE int32x4_t highbd_convolve8_4(
const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
const int16x4_t s6, const int16x4_t s7, const int16x8_t filters) {
const int16x4_t filters_lo = vget_low_s16(filters);
const int16x4_t filters_hi = vget_high_s16(filters);
int32x4_t sum;
......@@ -105,12 +104,11 @@ static INLINE int32x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
return sum;
}
static INLINE uint16x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
const int16x8_t s2, const int16x8_t s3,
const int16x8_t s4, const int16x8_t s5,
const int16x8_t s6, const int16x8_t s7,
const int16x8_t filters,
const uint16x8_t max) {
static INLINE uint16x8_t
highbd_convolve8_8(const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
const int16x8_t s6, const int16x8_t s7,
const int16x8_t filters, const uint16x8_t max) {
const int16x4_t filters_lo = vget_low_s16(filters);
const int16x4_t filters_hi = vget_high_s16(filters);
int32x4_t sum0, sum1;
......@@ -183,10 +181,10 @@ void vpx_highbd_convolve8_horiz_neon(const uint16_t *src, ptrdiff_t src_stride,
load_4x4((const int16_t *)src, src_stride, &s7, &s8, &s9, &s10);
transpose_s16_4x4d(&s7, &s8, &s9, &s10);
d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
d0 = highbd_convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
d1 = highbd_convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
d2 = highbd_convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
d3 = highbd_convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
d01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
d23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
......@@ -242,10 +240,11 @@ void vpx_highbd_convolve8_horiz_neon(const uint16_t *src, ptrdiff_t src_stride,
__builtin_prefetch(src + 5 * src_stride);
__builtin_prefetch(src + 6 * src_stride);
__builtin_prefetch(src + 7 * src_stride);
d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
d0 = highbd_convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
d1 = highbd_convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
d2 = highbd_convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
d3 =
highbd_convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
transpose_u16_8x4(&d0, &d1, &d2, &d3);
vst1_u16(dst, vget_low_u16(d0));
......@@ -303,14 +302,22 @@ void vpx_highbd_convolve8_horiz_neon(const uint16_t *src, ptrdiff_t src_stride,
&s12, &s13, &s14);
transpose_s16_8x8(&s7, &s8, &s9, &s10, &s11, &s12, &s13, &s14);
d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters, max);
d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters, max);
d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters, max);
d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters, max);
d0 = highbd_convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters,
max);
d1 = highbd_convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters,
max);
d2 = highbd_convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters,
max);
d3 = highbd_convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters,
max);
d4 = highbd_convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters,
max);
d5 = highbd_convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters,
max);
d6 = highbd_convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters,
max);
d7 = highbd_convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14,
filters, max);
transpose_u16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
store_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
......@@ -383,10 +390,10 @@ void vpx_highbd_convolve8_avg_horiz_neon(const uint16_t *src,
load_4x4((const int16_t *)src, src_stride, &s7, &s8, &s9, &s10);
transpose_s16_4x4d(&s7, &s8, &s9, &s10);
d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
d0 = highbd_convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
d1 = highbd_convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
d2 = highbd_convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
d3 = highbd_convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
t01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
t23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
......@@ -449,10 +456,11 @@ void vpx_highbd_convolve8_avg_horiz_neon(const uint16_t *src,
__builtin_prefetch(src + 5 * src_stride);
__builtin_prefetch(src + 6 * src_stride);
__builtin_prefetch(src + 7 * src_stride);
t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
t0 = highbd_convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
t1 = highbd_convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
t2 = highbd_convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
t3 =
highbd_convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
transpose_u16_8x4(&t0, &t1, &t2, &t3);
d0 = vcombine_u16(vld1_u16(dst + 0 * dst_stride),
......@@ -523,14 +531,22 @@ void vpx_highbd_convolve8_avg_horiz_neon(const uint16_t *src,
&s12, &s13, &s14);
transpose_s16_8x8(&s7, &s8, &s9, &s10, &s11, &s12, &s13, &s14);
d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters, max);
d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters, max);
d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters, max);
d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters, max);
d0 = highbd_convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters,
max);
d1 = highbd_convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters,
max);
d2 = highbd_convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters,
max);
d3 = highbd_convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters,
max);
d4 = highbd_convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters,
max);
d5 = highbd_convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters,
max);
d6 = highbd_convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters,
max);
d7 = highbd_convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14,
filters, max);
transpose_u16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
......@@ -620,10 +636,10 @@ void vpx_highbd_convolve8_vert_neon(const uint16_t *src, ptrdiff_t src_stride,
__builtin_prefetch(src + 1 * src_stride);
__builtin_prefetch(src + 2 * src_stride);
__builtin_prefetch(src + 3 * src_stride);
d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
d0 = highbd_convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
d1 = highbd_convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
d2 = highbd_convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
d3 = highbd_convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
d01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
d23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
......@@ -698,10 +714,11 @@ void vpx_highbd_convolve8_vert_neon(const uint16_t *src, ptrdiff_t src_stride,
__builtin_prefetch(s + 1 * src_stride);
__builtin_prefetch(s + 2 * src_stride);
__builtin_prefetch(s + 3 * src_stride);
d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
d0 = highbd_convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
d1 = highbd_convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
d2 = highbd_convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
d3 =
highbd_convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
vst1q_u16(d, d0);