2 * Copyright (c) 2016 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
11 #include <tmmintrin.h> // SSSE3
13 #include "./vp9_rtcd.h"
14 #include "./vpx_dsp_rtcd.h"
15 #include "./vpx_scale_rtcd.h"
16 #include "vpx_scale/yv12config.h"
18 extern void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
19 YV12_BUFFER_CONFIG *dst,
20 uint8_t filter_type, int phase_scaler);
22 static void downsample_2_to_1_ssse3(const uint8_t *src, ptrdiff_t src_stride,
23 uint8_t *dst, ptrdiff_t dst_stride, int w,
25 const __m128i mask = _mm_set1_epi16(0x00FF);
26 const int max_width = w & ~15;
28 for (y = 0; y < h; ++y) {
30 for (x = 0; x < max_width; x += 16) {
31 const __m128i a = _mm_loadu_si128((const __m128i *)(src + x * 2 + 0));
32 const __m128i b = _mm_loadu_si128((const __m128i *)(src + x * 2 + 16));
33 const __m128i a_and = _mm_and_si128(a, mask);
34 const __m128i b_and = _mm_and_si128(b, mask);
35 const __m128i c = _mm_packus_epi16(a_and, b_and);
36 _mm_storeu_si128((__m128i *)(dst + x), c);
38 for (; x < w; ++x) dst[x] = src[x * 2];
39 src += src_stride * 2;
44 static INLINE __m128i filter(const __m128i *const a, const __m128i *const b,
45 const __m128i *const c, const __m128i *const d,
46 const __m128i *const e, const __m128i *const f,
47 const __m128i *const g, const __m128i *const h) {
48 const __m128i coeffs_ab =
49 _mm_set_epi8(6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1);
50 const __m128i coeffs_cd = _mm_set_epi8(78, -19, 78, -19, 78, -19, 78, -19, 78,
51 -19, 78, -19, 78, -19, 78, -19);
52 const __m128i const64_x16 = _mm_set1_epi16(64);
53 const __m128i ab = _mm_unpacklo_epi8(*a, *b);
54 const __m128i cd = _mm_unpacklo_epi8(*c, *d);
55 const __m128i fe = _mm_unpacklo_epi8(*f, *e);
56 const __m128i hg = _mm_unpacklo_epi8(*h, *g);
57 const __m128i ab_terms = _mm_maddubs_epi16(ab, coeffs_ab);
58 const __m128i cd_terms = _mm_maddubs_epi16(cd, coeffs_cd);
59 const __m128i fe_terms = _mm_maddubs_epi16(fe, coeffs_cd);
60 const __m128i hg_terms = _mm_maddubs_epi16(hg, coeffs_ab);
62 const __m128i abcd_terms = _mm_add_epi16(ab_terms, cd_terms);
64 const __m128i fehg_terms = _mm_add_epi16(fe_terms, hg_terms);
65 // can overflow, use saturating add
66 const __m128i terms = _mm_adds_epi16(abcd_terms, fehg_terms);
67 const __m128i round = _mm_adds_epi16(terms, const64_x16);
68 const __m128i shift = _mm_srai_epi16(round, 7);
69 return _mm_packus_epi16(shift, shift);
72 static void eight_tap_row_ssse3(const uint8_t *src, uint8_t *dst, int w) {
73 const int max_width = w & ~7;
75 for (; x < max_width; x += 8) {
76 const __m128i a = _mm_loadl_epi64((const __m128i *)(src + x + 0));
77 const __m128i b = _mm_loadl_epi64((const __m128i *)(src + x + 1));
78 const __m128i c = _mm_loadl_epi64((const __m128i *)(src + x + 2));
79 const __m128i d = _mm_loadl_epi64((const __m128i *)(src + x + 3));
80 const __m128i e = _mm_loadl_epi64((const __m128i *)(src + x + 4));
81 const __m128i f = _mm_loadl_epi64((const __m128i *)(src + x + 5));
82 const __m128i g = _mm_loadl_epi64((const __m128i *)(src + x + 6));
83 const __m128i h = _mm_loadl_epi64((const __m128i *)(src + x + 7));
84 const __m128i pack = filter(&a, &b, &c, &d, &e, &f, &g, &h);
85 _mm_storel_epi64((__m128i *)(dst + x), pack);
89 static void upsample_1_to_2_ssse3(const uint8_t *src, ptrdiff_t src_stride,
90 uint8_t *dst, ptrdiff_t dst_stride, int dst_w,
95 DECLARE_ALIGNED(16, uint8_t, tmp[1920 * 8]);
96 uint8_t *tmp0 = tmp + dst_w * 0;
97 uint8_t *tmp1 = tmp + dst_w * 1;
98 uint8_t *tmp2 = tmp + dst_w * 2;
99 uint8_t *tmp3 = tmp + dst_w * 3;
100 uint8_t *tmp4 = tmp + dst_w * 4;
101 uint8_t *tmp5 = tmp + dst_w * 5;
102 uint8_t *tmp6 = tmp + dst_w * 6;
103 uint8_t *tmp7 = tmp + dst_w * 7;
104 uint8_t *tmp8 = NULL;
105 const int max_width = dst_w & ~7;
107 eight_tap_row_ssse3(src - src_stride * 3 - 3, tmp0, dst_w);
108 eight_tap_row_ssse3(src - src_stride * 2 - 3, tmp1, dst_w);
109 eight_tap_row_ssse3(src - src_stride * 1 - 3, tmp2, dst_w);
110 eight_tap_row_ssse3(src + src_stride * 0 - 3, tmp3, dst_w);
111 eight_tap_row_ssse3(src + src_stride * 1 - 3, tmp4, dst_w);
112 eight_tap_row_ssse3(src + src_stride * 2 - 3, tmp5, dst_w);
113 eight_tap_row_ssse3(src + src_stride * 3 - 3, tmp6, dst_w);
114 for (y = 0; y < dst_h; y++) {
116 eight_tap_row_ssse3(src + src_stride * 4 - 3, tmp7, dst_w);
117 for (x = 0; x < max_width; x += 8) {
118 const __m128i A = _mm_loadl_epi64((const __m128i *)(src + x));
119 const __m128i B = _mm_loadl_epi64((const __m128i *)(tmp3 + x));
120 const __m128i AB = _mm_unpacklo_epi8(A, B);
122 _mm_storeu_si128((__m128i *)(dst + x * 2), AB);
125 _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 3));
127 _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 2));
129 _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 1));
131 _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 0));
133 _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 1));
135 _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 2));
137 _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 3));
139 _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 4));
140 C = filter(&a, &b, &c, &d, &e, &f, &g, &h);
143 const __m128i a = _mm_loadl_epi64((const __m128i *)(tmp0 + x));
144 const __m128i b = _mm_loadl_epi64((const __m128i *)(tmp1 + x));
145 const __m128i c = _mm_loadl_epi64((const __m128i *)(tmp2 + x));
146 const __m128i d = _mm_loadl_epi64((const __m128i *)(tmp3 + x));
147 const __m128i e = _mm_loadl_epi64((const __m128i *)(tmp4 + x));
148 const __m128i f = _mm_loadl_epi64((const __m128i *)(tmp5 + x));
149 const __m128i g = _mm_loadl_epi64((const __m128i *)(tmp6 + x));
150 const __m128i h = _mm_loadl_epi64((const __m128i *)(tmp7 + x));
151 D = filter(&a, &b, &c, &d, &e, &f, &g, &h);
153 CD = _mm_unpacklo_epi8(C, D);
154 _mm_storeu_si128((__m128i *)(dst + x * 2 + dst_stride), CD);
157 dst += dst_stride * 2;
171 void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
172 YV12_BUFFER_CONFIG *dst,
173 uint8_t filter_type, int phase_scaler) {
174 const int src_w = src->y_crop_width;
175 const int src_h = src->y_crop_height;
176 const int dst_w = dst->y_crop_width;
177 const int dst_h = dst->y_crop_height;
178 const int dst_uv_w = dst_w / 2;
179 const int dst_uv_h = dst_h / 2;
181 if (dst_w * 2 == src_w && dst_h * 2 == src_h && phase_scaler == 0) {
182 downsample_2_to_1_ssse3(src->y_buffer, src->y_stride, dst->y_buffer,
183 dst->y_stride, dst_w, dst_h);
184 downsample_2_to_1_ssse3(src->u_buffer, src->uv_stride, dst->u_buffer,
185 dst->uv_stride, dst_uv_w, dst_uv_h);
186 downsample_2_to_1_ssse3(src->v_buffer, src->uv_stride, dst->v_buffer,
187 dst->uv_stride, dst_uv_w, dst_uv_h);
188 vpx_extend_frame_borders(dst);
189 } else if (dst_w == src_w * 2 && dst_h == src_h * 2 && phase_scaler == 0) {
190 // The upsample() supports widths up to 1920 * 2. If greater, fall back
191 // to vp9_scale_and_extend_frame_c().
192 if (dst_w / 2 <= 1920) {
193 upsample_1_to_2_ssse3(src->y_buffer, src->y_stride, dst->y_buffer,
194 dst->y_stride, dst_w, dst_h);
195 upsample_1_to_2_ssse3(src->u_buffer, src->uv_stride, dst->u_buffer,
196 dst->uv_stride, dst_uv_w, dst_uv_h);
197 upsample_1_to_2_ssse3(src->v_buffer, src->uv_stride, dst->v_buffer,
198 dst->uv_stride, dst_uv_w, dst_uv_h);
199 vpx_extend_frame_borders(dst);
201 vp9_scale_and_extend_frame_c(src, dst, filter_type, phase_scaler);
204 vp9_scale_and_extend_frame_c(src, dst, filter_type, phase_scaler);