generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Add 4 to 1 scaling x86 optimization

Browse Source 
Change-Id: I51c190f0a88685867df36912522e67bdae58a673
This commit is contained in:
Linfeng Zhang 2017-10-04 09:55:56 -07:00
parent dcfae2cc64
commit 16166bfdaa
2 changed files with 299 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

310
vp9/encoder/x86/vp9_frame_scale_ssse3.c
View File

@ -18,6 +18,15 @@
#include "vpx_dsp/x86/transpose_sse2.h"
#include "vpx_scale/yv12config.h"
static INLINE __m128i scale_plane_2_to_1_phase_0_kernel(
const uint8_t *const src, const __m128i *const mask) {
const __m128i a = _mm_loadu_si128((const __m128i *)(&src[0]));
const __m128i b = _mm_loadu_si128((const __m128i *)(&src[16]));
const __m128i a_and = _mm_and_si128(a, *mask);
const __m128i b_and = _mm_and_si128(b, *mask);
return _mm_packus_epi16(a_and, b_and);
}
static void scale_plane_2_to_1_phase_0(const uint8_t *src,
const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride,
@ -29,12 +38,8 @@ static void scale_plane_2_to_1_phase_0(const uint8_t *src,
do {
int x = max_width;
do {
const __m128i a = _mm_loadu_si128((const __m128i *)(src + 0));
const __m128i b = _mm_loadu_si128((const __m128i *)(src + 16));
const __m128i a_and = _mm_and_si128(a, mask);
const __m128i b_and = _mm_and_si128(b, mask);
const __m128i c = _mm_packus_epi16(a_and, b_and);
_mm_storeu_si128((__m128i *)dst, c);
const __m128i d = scale_plane_2_to_1_phase_0_kernel(src, &mask);
_mm_storeu_si128((__m128i *)dst, d);
src += 32;
dst += 16;
x -= 16;
@ -44,6 +49,30 @@ static void scale_plane_2_to_1_phase_0(const uint8_t *src,
} while (--y);
}
static void scale_plane_4_to_1_phase_0(const uint8_t *src,
const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride,
const int dst_w, const int dst_h) {
const int max_width = (dst_w + 15) & ~15;
const __m128i mask = _mm_set1_epi32(0x000000FF);
int y = dst_h;
do {
int x = max_width;
do {
const __m128i d0 = scale_plane_2_to_1_phase_0_kernel(&src[0], &mask);
const __m128i d1 = scale_plane_2_to_1_phase_0_kernel(&src[32], &mask);
const __m128i d2 = _mm_packus_epi16(d0, d1);
_mm_storeu_si128((__m128i *)dst, d2);
src += 64;
dst += 16;
x -= 16;
} while (x);
src += 4 * (src_stride - max_width);
dst += dst_stride - max_width;
} while (--y);
}
static INLINE __m128i scale_plane_bilinear_kernel(const __m128i *const s,
const __m128i c0c1) {
const __m128i k_64 = _mm_set1_epi16(1 << 6);
@ -97,6 +126,101 @@ static void scale_plane_2_to_1_bilinear(const uint8_t *src,
} while (--y);
}
static void scale_plane_4_to_1_bilinear(const uint8_t *src,
const ptrdiff_t src_stride,
uint8_t *dst,
const ptrdiff_t dst_stride,
const int dst_w, const int dst_h,
const __m128i c0c1) {
const int max_width = (dst_w + 15) & ~15;
int y = dst_h;
do {
int x = max_width;
do {
__m128i s[8], d[8];
// Note: Using _mm_packus_epi32() in SSE4.1 could be faster.
// Here we tried to not use shuffle instructions which would be slow
// on some x86 CPUs.
// Horizontal
// 000 001 xx xx 004 005 xx xx 008 009 xx xx 00C 00D xx xx
// 010 011 xx xx 014 015 xx xx 018 019 xx xx 01C 01D xx xx
// 020 021 xx xx 024 025 xx xx 028 029 xx xx 02C 02D xx xx
// 030 031 xx xx 034 035 xx xx 038 039 xx xx 03C 03D xx xx
// 100 101 xx xx 104 105 xx xx 108 109 xx xx 10C 10D xx xx
// 110 111 xx xx 114 115 xx xx 118 119 xx xx 11C 11D xx xx
// 120 121 xx xx 124 125 xx xx 128 129 xx xx 12C 12D xx xx
// 130 131 xx xx 134 135 xx xx 138 139 xx xx 13C 13D xx xx
s[0] = _mm_loadu_si128((const __m128i *)(&src[0]));
s[1] = _mm_loadu_si128((const __m128i *)(&src[16]));
s[2] = _mm_loadu_si128((const __m128i *)(&src[32]));
s[3] = _mm_loadu_si128((const __m128i *)(&src[48]));
s[4] = _mm_loadu_si128((const __m128i *)(src + src_stride + 0));
s[5] = _mm_loadu_si128((const __m128i *)(src + src_stride + 16));
s[6] = _mm_loadu_si128((const __m128i *)(src + src_stride + 32));
s[7] = _mm_loadu_si128((const __m128i *)(src + src_stride + 48));
// 000 001 100 101 xx xx xx xx 004 005 104 105 xx xx xx xx
// 008 009 108 109 xx xx xx xx 00C 00D 10C 10D xx xx xx xx
// 010 011 110 111 xx xx xx xx 014 015 114 115 xx xx xx xx
// 018 019 118 119 xx xx xx xx 01C 01D 11C 11D xx xx xx xx
// 020 021 120 121 xx xx xx xx 024 025 124 125 xx xx xx xx
// 028 029 128 129 xx xx xx xx 02C 02D 12C 12D xx xx xx xx
// 030 031 130 131 xx xx xx xx 034 035 134 135 xx xx xx xx
// 038 039 138 139 xx xx xx xx 03C 03D 13C 13D xx xx xx xx
d[0] = _mm_unpacklo_epi16(s[0], s[4]);
d[1] = _mm_unpackhi_epi16(s[0], s[4]);
d[2] = _mm_unpacklo_epi16(s[1], s[5]);
d[3] = _mm_unpackhi_epi16(s[1], s[5]);
d[4] = _mm_unpacklo_epi16(s[2], s[6]);
d[5] = _mm_unpackhi_epi16(s[2], s[6]);
d[6] = _mm_unpacklo_epi16(s[3], s[7]);
d[7] = _mm_unpackhi_epi16(s[3], s[7]);
// 000 001 100 101 008 009 108 109 xx xx xx xx xx xx xx xx
// 004 005 104 105 00C 00D 10C 10D xx xx xx xx xx xx xx xx
// 010 011 110 111 018 019 118 119 xx xx xx xx xx xx xx xx
// 014 015 114 115 01C 01D 11C 11D xx xx xx xx xx xx xx xx
// 020 021 120 121 028 029 128 129 xx xx xx xx xx xx xx xx
// 024 025 124 125 02C 02D 12C 12D xx xx xx xx xx xx xx xx
// 030 031 130 131 038 039 138 139 xx xx xx xx xx xx xx xx
// 034 035 134 135 03C 03D 13C 13D xx xx xx xx xx xx xx xx
s[0] = _mm_unpacklo_epi32(d[0], d[1]);
s[1] = _mm_unpackhi_epi32(d[0], d[1]);
s[2] = _mm_unpacklo_epi32(d[2], d[3]);
s[3] = _mm_unpackhi_epi32(d[2], d[3]);
s[4] = _mm_unpacklo_epi32(d[4], d[5]);
s[5] = _mm_unpackhi_epi32(d[4], d[5]);
s[6] = _mm_unpacklo_epi32(d[6], d[7]);
s[7] = _mm_unpackhi_epi32(d[6], d[7]);
// 000 001 100 101 004 005 104 105 008 009 108 109 00C 00D 10C 10D
// 010 011 110 111 014 015 114 115 018 019 118 119 01C 01D 11C 11D
// 020 021 120 121 024 025 124 125 028 029 128 129 02C 02D 12C 12D
// 030 031 130 131 034 035 134 135 038 039 138 139 03C 03D 13C 13D
d[0] = _mm_unpacklo_epi32(s[0], s[1]);
d[1] = _mm_unpacklo_epi32(s[2], s[3]);
d[2] = _mm_unpacklo_epi32(s[4], s[5]);
d[3] = _mm_unpacklo_epi32(s[6], s[7]);
d[0] = scale_plane_bilinear_kernel(&d[0], c0c1);
d[1] = scale_plane_bilinear_kernel(&d[2], c0c1);
// Vertical
d[0] = scale_plane_bilinear_kernel(d, c0c1);
_mm_storeu_si128((__m128i *)dst, d[0]);
src += 64;
dst += 16;
x -= 16;
} while (x);
src += 4 * (src_stride - max_width);
dst += dst_stride - max_width;
} while (--y);
}
static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
uint8_t *dst, const int dst_stride,
const int w, const int h,
@ -148,8 +272,11 @@ static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
// 02 12 03 13 22 32 23 33 42 52 43 53 62 72 63 73
d[2] = _mm_unpacklo_epi16(d[0], d[1]);
d[3] = _mm_unpackhi_epi16(d[0], d[1]);
store_8bit_4x4_sse2(d[2], t + 0, 2 * width_hor);
store_8bit_4x4_sse2(d[3], t + 4, 2 * width_hor);
// 00 10 01 11 02 12 03 13 20 30 21 31 22 32 23 33
// 40 50 41 51 42 52 43 53 60 70 61 71 62 72 63 73
d[0] = _mm_unpacklo_epi32(d[2], d[3]);
d[1] = _mm_unpackhi_epi32(d[2], d[3]);
store_8bit_8x4_from_16x2(d, t, 2 * width_hor);
s[0] = s[4];
s[1] = s[5];
@ -170,8 +297,9 @@ static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
// 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
// 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
// 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
// 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 (overlapped)
loadu_8bit_16x4(t, 2 * width_hor, s);
s[0] = _mm_loadu_si128((const __m128i *)(t + 0 * width_hor));
s[1] = _mm_loadu_si128((const __m128i *)(t + 2 * width_hor));
s[2] = _mm_loadu_si128((const __m128i *)(t + 4 * width_hor));
t += 6 * width_hor;
y = height_ver;
@ -183,19 +311,16 @@ static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
loadu_8bit_16x4(t, 2 * width_hor, &s[3]);
t += 8 * width_hor;
d[0] = convolve8_8_ssse3(&s[0], f);
d[1] = convolve8_8_ssse3(&s[1], f);
d[2] = convolve8_8_ssse3(&s[2], f);
d[3] = convolve8_8_ssse3(&s[3], f);
d[0] = convolve8_8_ssse3(&s[0], f); // 00 01 02 03 04 05 06 07
d[1] = convolve8_8_ssse3(&s[1], f); // 10 11 12 13 14 15 16 17
d[2] = convolve8_8_ssse3(&s[2], f); // 20 21 22 23 24 25 26 27
d[3] = convolve8_8_ssse3(&s[3], f); // 30 31 32 33 34 35 36 37
// 00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17
// 20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37
d[0] = _mm_packus_epi16(d[0], d[1]);
d[1] = _mm_packus_epi16(d[2], d[3]);
_mm_storel_epi64((__m128i *)(dst + 0 * dst_stride), d[0]);
_mm_storeh_epi64((__m128i *)(dst + 1 * dst_stride), d[0]);
_mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), d[1]);
_mm_storeh_epi64((__m128i *)(dst + 3 * dst_stride), d[1]);
store_8bit_8x4_from_16x2(d, dst, dst_stride);
s[0] = s[4];
s[1] = s[5];
@ -212,6 +337,107 @@ static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
} while (x);
}
static void scale_plane_4_to_1_general(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 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;
int x, y = height_hor;
uint8_t *t = temp_buffer;
__m128i s[11], d[4];
__m128i f[4];
assert(w && h);
shuffle_filter_ssse3(coef, f);
src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 3;
// horizontal 2x8
do {
load_8bit_8x8(src + 4, src_stride, s);
// 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71
// 02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73
// 04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75 (overlapped)
// 06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77 (overlapped)
transpose_16bit_4x8(s, s);
x = width_hor;
do {
src += 8;
load_8bit_8x8(src, src_stride, &s[2]);
// 04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75
// 06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77
// 08 09 18 19 28 29 38 39 48 49 58 59 68 69 78 79
// 0A 0B 1A 1B 2A 2B 3A 3B 4A 4B 5A 5B 6A 6B 7A 7B
transpose_16bit_4x8(&s[2], &s[2]);
d[0] = convolve8_8_ssse3(&s[0], f); // 00 10 20 30 40 50 60 70
d[1] = convolve8_8_ssse3(&s[2], f); // 01 11 21 31 41 51 61 71
// 00 10 20 30 40 50 60 70 xx xx xx xx xx xx xx xx
// 01 11 21 31 41 51 61 71 xx xx xx xx xx xx xx xx
d[0] = _mm_packus_epi16(d[0], d[0]);
d[1] = _mm_packus_epi16(d[1], d[1]);
// 00 10 01 11 20 30 21 31 40 50 41 51 60 70 61 71
d[0] = _mm_unpacklo_epi16(d[0], d[1]);
store_8bit_4x4_sse2(d[0], t, 2 * width_hor);
s[0] = s[4];
s[1] = s[5];
t += 4;
x -= 2;
} while (x);
src += 8 * src_stride - 4 * width_hor;
t += 6 * width_hor;
y -= 8;
} while (y);
// vertical 8x2
x = width_ver;
t = temp_buffer;
do {
// 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
// 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
s[0] = _mm_loadu_si128((const __m128i *)(t + 0 * width_hor));
s[1] = _mm_loadu_si128((const __m128i *)(t + 2 * width_hor));
t += 4 * width_hor;
y = height_ver;
do {
// 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
// 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
// 80 90 81 91 82 92 83 93 84 94 85 95 86 96 87 77
// A0 B0 A1 B1 A2 B2 A3 B3 A4 B4 A5 B5 A6 B6 A7 77
loadu_8bit_16x4(t, 2 * width_hor, &s[2]);
t += 8 * width_hor;
d[0] = convolve8_8_ssse3(&s[0], f); // 00 01 02 03 04 05 06 07
d[1] = convolve8_8_ssse3(&s[2], f); // 10 11 12 13 14 15 16 17
// 00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17
d[0] = _mm_packus_epi16(d[0], d[1]);
_mm_storel_epi64((__m128i *)(dst + 0 * dst_stride), d[0]);
_mm_storeh_epi64((__m128i *)(dst + 1 * dst_stride), d[0]);
s[0] = s[4];
s[1] = s[5];
dst += 2 * dst_stride;
y -= 2;
} while (y);
t -= width_hor * (4 * height_ver + 4);
t += 16;
dst -= height_ver * dst_stride;
dst += 8;
x -= 8;
} while (x);
}
static INLINE __m128i scale_1_to_2_phase_0_kernel(const __m128i *const s,
const __m128i *const f) {
__m128i ss[4], temp;
@ -332,6 +558,8 @@ void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
const int src_h = src->y_crop_height;
const int dst_w = dst->y_crop_width;
const int dst_h = dst->y_crop_height;
const int dst_uv_w = dst_w / 2;
const int dst_uv_h = dst_h / 2;
int scaled = 0;
// phase_scaler is usually 0 or 8.
@ -339,8 +567,6 @@ void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
if (dst_w * 2 == src_w && dst_h * 2 == src_h) {
// 2 to 1
const int dst_uv_w = dst_w / 2;
const int dst_uv_h = dst_h / 2;
scaled = 1;
if (phase_scaler == 0) {
@ -382,6 +608,50 @@ void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
scaled = 0;
}
}
} 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(src->y_buffer, src->y_stride, dst->y_buffer,
dst->y_stride, dst_w, dst_h);
scale_plane_4_to_1_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
dst->uv_stride, dst_uv_w, dst_uv_h);
scale_plane_4_to_1_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
dst->uv_stride, dst_uv_w, dst_uv_h);
} else if (filter_type == BILINEAR) {
const int16_t c0 = vp9_filter_kernels[BILINEAR][phase_scaler][3];
const int16_t c1 = vp9_filter_kernels[BILINEAR][phase_scaler][4];
const __m128i c0c1 = _mm_set1_epi16(c0 | (c1 << 8)); // c0 and c1 >= 0
scale_plane_4_to_1_bilinear(src->y_buffer, src->y_stride, dst->y_buffer,
dst->y_stride, dst_w, dst_h, c0c1);
scale_plane_4_to_1_bilinear(src->u_buffer, src->uv_stride, dst->u_buffer,
dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
scale_plane_4_to_1_bilinear(src->v_buffer, src->uv_stride, dst->v_buffer,
dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
} else {
const int buffer_stride = (dst_w + 1) & ~1;
const int buffer_height = (4 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
// When dst_w is 1 or 2, we need extra padding to avoid heap read overflow
const int extra_padding = 16;
uint8_t *const temp_buffer =
(uint8_t *)malloc(buffer_stride * buffer_height + extra_padding);
if (temp_buffer) {
scale_plane_4_to_1_general(
src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
dst_h, vp9_filter_kernels[filter_type][phase_scaler], temp_buffer);
scale_plane_4_to_1_general(
src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
temp_buffer);
scale_plane_4_to_1_general(
src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
temp_buffer);
free(temp_buffer);
} else {
scaled = 0;
}
}
} else if (dst_w == src_w * 2 && dst_h == src_h * 2 && phase_scaler == 0) {
// 1 to 2
uint8_t *const temp_buffer = (uint8_t *)malloc(8 * ((src_w + 7) & ~7));

9
vpx_dsp/x86/mem_sse2.h
View File

@ -92,6 +92,15 @@ static INLINE void store_8bit_4x4_sse2(const __m128i s, uint8_t *const d,
store_8bit_4x4(ss, d, stride);
}
static INLINE void store_8bit_8x4_from_16x2(const __m128i *const s,
uint8_t *const d,
const ptrdiff_t stride) {
_mm_storel_epi64((__m128i *)(d + 0 * stride), s[0]);
_mm_storeh_epi64((__m128i *)(d + 1 * stride), s[0]);
_mm_storel_epi64((__m128i *)(d + 2 * stride), s[1]);
_mm_storeh_epi64((__m128i *)(d + 3 * stride), s[1]);
}
static INLINE void store_8bit_8x8(const __m128i *const s, uint8_t *const d,
const ptrdiff_t stride) {
_mm_storel_epi64((__m128i *)(d + 0 * stride), s[0]);