ref: 127864deb3d72aa8ec90f161b1a0dc7511bcffd7
parent: b809442521bc05cc838b45d0e4ccf1c8ba3e71d4
author: Linfeng Zhang <linfengz@google.com>
date: Tue Oct 3 12:09:19 EDT 2017
Generalize 2:1 vp9_scale_and_extend_frame_ssse3() Change-Id: I882da3a04884d5fabd4cd591c28682cbb2d76aa5
--- a/vp9/encoder/x86/vp9_frame_scale_ssse3.c
+++ b/vp9/encoder/x86/vp9_frame_scale_ssse3.c
@@ -15,6 +15,7 @@
#include "./vpx_scale_rtcd.h"
#include "vpx_dsp/x86/convolve_ssse3.h"
#include "vpx_dsp/x86/mem_sse2.h"
+#include "vpx_dsp/x86/transpose_sse2.h"
#include "vpx_scale/yv12config.h"
static void scale_plane_2_to_1_phase_0(const uint8_t *src,
@@ -21,8 +22,8 @@
const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride,
const int dst_w, const int dst_h) {- const __m128i mask = _mm_set1_epi16(0x00FF);
const int max_width = (dst_w + 15) & ~15;
+ const __m128i mask = _mm_set1_epi16(0x00FF);
int y = dst_h;
do {@@ -43,6 +44,174 @@
} 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);
+ const __m128i t0 = _mm_maddubs_epi16(s[0], c0c1);
+ const __m128i t1 = _mm_maddubs_epi16(s[1], c0c1);
+ // round and shift by 7 bit each 16 bit
+ const __m128i t2 = _mm_adds_epi16(t0, k_64);
+ const __m128i t3 = _mm_adds_epi16(t1, k_64);
+ const __m128i t4 = _mm_srai_epi16(t2, 7);
+ const __m128i t5 = _mm_srai_epi16(t3, 7);
+ return _mm_packus_epi16(t4, t5);
+}
+
+static void scale_plane_2_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[2], d[2];
+
+ // Horizontal
+ // Even rows
+ s[0] = _mm_loadu_si128((const __m128i *)(src + 0));
+ s[1] = _mm_loadu_si128((const __m128i *)(src + 16));
+ d[0] = scale_plane_bilinear_kernel(s, c0c1);
+
+ // odd rows
+ s[0] = _mm_loadu_si128((const __m128i *)(src + src_stride + 0));
+ s[1] = _mm_loadu_si128((const __m128i *)(src + src_stride + 16));
+ d[1] = scale_plane_bilinear_kernel(s, c0c1);
+
+ // Vertical
+ s[0] = _mm_unpacklo_epi8(d[0], d[1]);
+ s[1] = _mm_unpackhi_epi8(d[0], d[1]);
+ d[0] = scale_plane_bilinear_kernel(s, c0c1);
+
+ _mm_storeu_si128((__m128i *)dst, d[0]);
+ src += 32;
+ dst += 16;
+ x -= 16;
+ } while (x);
+ src += 2 * (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,
+ const int16_t *const coef,
+ uint8_t *const temp_buffer) {+ 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;
+ 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 + 1;
+
+ // horizontal 4x8
+ do {+ load_8bit_8x8(src + 2, 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
+ // 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[3]);
+ // 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
+ // 0C 0D 1C 1D 2C 2D 3C 3D 4C 4D 5C 5D 6C 6D 7C 7D
+ transpose_16bit_4x8(&s[3], &s[3]);
+
+ d[0] = convolve8_8_ssse3(&s[0], f); // 00 10 20 30 40 50 60 70
+ d[1] = convolve8_8_ssse3(&s[1], f); // 01 11 21 31 41 51 61 71
+ d[2] = convolve8_8_ssse3(&s[2], f); // 02 12 22 32 42 52 62 72
+ d[3] = convolve8_8_ssse3(&s[3], f); // 03 13 23 33 43 53 63 73
+
+ // 00 10 20 30 40 50 60 70 02 12 22 32 42 52 62 72
+ // 01 11 21 31 41 51 61 71 03 13 23 33 43 53 63 73
+ d[0] = _mm_packus_epi16(d[0], d[2]);
+ d[1] = _mm_packus_epi16(d[1], d[3]);
+ // 00 10 01 11 20 30 21 31 40 50 41 51 60 70 61 71
+ // 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);
+
+ s[0] = s[4];
+ s[1] = s[5];
+ s[2] = s[6];
+
+ t += 8;
+ x -= 4;
+ } while (x);
+ src += 8 * src_stride - 2 * width_hor;
+ t += 6 * width_hor;
+ y -= 8;
+ } while (y);
+
+ // vertical 8x4
+ 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
+ // 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);
+ t += 6 * width_hor;
+ y = height_ver;
+
+ do {+ // 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
+ // C0 D0 C1 D1 C2 D2 C3 D3 C4 D4 C5 D5 C6 D6 C7 77
+ 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);
+
+ // 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]);
+
+ s[0] = s[4];
+ s[1] = s[5];
+ s[2] = s[6];
+
+ dst += 4 * dst_stride;
+ y -= 4;
+ } while (y);
+ t -= width_hor * (2 * height_ver + 6);
+ 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;
@@ -165,17 +334,54 @@
const int dst_h = dst->y_crop_height;
int scaled = 0;
- if (dst_w * 2 == src_w && dst_h * 2 == src_h && phase_scaler == 0) {+ // phase_scaler is usually 0 or 8.
+ assert(phase_scaler >= 0 && phase_scaler < 16);
+
+ 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;
- scale_plane_2_to_1_phase_0(src->y_buffer, src->y_stride, dst->y_buffer,
- dst->y_stride, dst_w, dst_h);
- scale_plane_2_to_1_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
- dst->uv_stride, dst_uv_w, dst_uv_h);
- scale_plane_2_to_1_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
- dst->uv_stride, dst_uv_w, dst_uv_h);
+
+ if (phase_scaler == 0) {+ scale_plane_2_to_1_phase_0(src->y_buffer, src->y_stride, dst->y_buffer,
+ dst->y_stride, dst_w, dst_h);
+ scale_plane_2_to_1_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
+ dst->uv_stride, dst_uv_w, dst_uv_h);
+ scale_plane_2_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_2_to_1_bilinear(src->y_buffer, src->y_stride, dst->y_buffer,
+ dst->y_stride, dst_w, dst_h, c0c1);
+ scale_plane_2_to_1_bilinear(src->u_buffer, src->uv_stride, dst->u_buffer,
+ dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
+ scale_plane_2_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 + 3) & ~3;
+ const int buffer_height = (2 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
+ uint8_t *const temp_buffer =
+ (uint8_t *)malloc(buffer_stride * buffer_height);
+ if (temp_buffer) {+ scale_plane_2_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_2_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_2_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));
--- a/vpx_dsp/x86/mem_sse2.h
+++ b/vpx_dsp/x86/mem_sse2.h
@@ -55,18 +55,24 @@
d[7] = _mm_load_si128((const __m128i *)(s + 7 * stride));
}
-static INLINE void loadu_8bit_16x8(const uint8_t *const s,
+static INLINE void loadu_8bit_16x4(const uint8_t *const s,
const ptrdiff_t stride, __m128i *const d) {d[0] = _mm_loadu_si128((const __m128i *)(s + 0 * stride));
d[1] = _mm_loadu_si128((const __m128i *)(s + 1 * stride));
d[2] = _mm_loadu_si128((const __m128i *)(s + 2 * stride));
d[3] = _mm_loadu_si128((const __m128i *)(s + 3 * stride));
- d[4] = _mm_loadu_si128((const __m128i *)(s + 4 * stride));
- d[5] = _mm_loadu_si128((const __m128i *)(s + 5 * stride));
- d[6] = _mm_loadu_si128((const __m128i *)(s + 6 * stride));
- d[7] = _mm_loadu_si128((const __m128i *)(s + 7 * stride));
}
+static INLINE void loadu_8bit_16x8(const uint8_t *const s,
+ const ptrdiff_t stride, __m128i *const d) {+ loadu_8bit_16x4(s + 0 * stride, stride, &d[0]);
+ loadu_8bit_16x4(s + 4 * stride, stride, &d[4]);
+}
+
+static INLINE void _mm_storeh_epi64(__m128i *const d, const __m128i s) {+ _mm_storeh_pi((__m64 *)d, _mm_castsi128_ps(s));
+}
+
static INLINE void store_8bit_4x4(const __m128i *const s, uint8_t *const d,
const ptrdiff_t stride) {*(int *)(d + 0 * stride) = _mm_cvtsi128_si32(s[0]);
@@ -73,6 +79,17 @@
*(int *)(d + 1 * stride) = _mm_cvtsi128_si32(s[1]);
*(int *)(d + 2 * stride) = _mm_cvtsi128_si32(s[2]);
*(int *)(d + 3 * stride) = _mm_cvtsi128_si32(s[3]);
+}
+
+static INLINE void store_8bit_4x4_sse2(const __m128i s, uint8_t *const d,
+ const ptrdiff_t stride) {+ __m128i ss[4];
+
+ ss[0] = s;
+ ss[1] = _mm_srli_si128(s, 4);
+ ss[2] = _mm_srli_si128(s, 8);
+ ss[3] = _mm_srli_si128(s, 12);
+ store_8bit_4x4(ss, d, stride);
}
static INLINE void store_8bit_8x8(const __m128i *const s, uint8_t *const d,