1/*
2 *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
3 *
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.
9 */
10
11#include <tmmintrin.h>  // SSSE3
12
13#include "./vp9_rtcd.h"
14#include "./vpx_dsp_rtcd.h"
15#include "./vpx_scale_rtcd.h"
16#include "vpx_dsp/x86/convolve_ssse3.h"
17#include "vpx_dsp/x86/mem_sse2.h"
18#include "vpx_dsp/x86/transpose_sse2.h"
19#include "vpx_scale/yv12config.h"
20
21static INLINE __m128i scale_plane_2_to_1_phase_0_kernel(
22    const uint8_t *const src, const __m128i *const mask) {
23  const __m128i a = _mm_loadu_si128((const __m128i *)(&src[0]));
24  const __m128i b = _mm_loadu_si128((const __m128i *)(&src[16]));
25  const __m128i a_and = _mm_and_si128(a, *mask);
26  const __m128i b_and = _mm_and_si128(b, *mask);
27  return _mm_packus_epi16(a_and, b_and);
28}
29
30static void scale_plane_2_to_1_phase_0(const uint8_t *src,
31                                       const ptrdiff_t src_stride, uint8_t *dst,
32                                       const ptrdiff_t dst_stride,
33                                       const int dst_w, const int dst_h) {
34  const int max_width = (dst_w + 15) & ~15;
35  const __m128i mask = _mm_set1_epi16(0x00FF);
36  int y = dst_h;
37
38  do {
39    int x = max_width;
40    do {
41      const __m128i d = scale_plane_2_to_1_phase_0_kernel(src, &mask);
42      _mm_storeu_si128((__m128i *)dst, d);
43      src += 32;
44      dst += 16;
45      x -= 16;
46    } while (x);
47    src += 2 * (src_stride - max_width);
48    dst += dst_stride - max_width;
49  } while (--y);
50}
51
52static void scale_plane_4_to_1_phase_0(const uint8_t *src,
53                                       const ptrdiff_t src_stride, uint8_t *dst,
54                                       const ptrdiff_t dst_stride,
55                                       const int dst_w, const int dst_h) {
56  const int max_width = (dst_w + 15) & ~15;
57  const __m128i mask = _mm_set1_epi32(0x000000FF);
58  int y = dst_h;
59
60  do {
61    int x = max_width;
62    do {
63      const __m128i d0 = scale_plane_2_to_1_phase_0_kernel(&src[0], &mask);
64      const __m128i d1 = scale_plane_2_to_1_phase_0_kernel(&src[32], &mask);
65      const __m128i d2 = _mm_packus_epi16(d0, d1);
66      _mm_storeu_si128((__m128i *)dst, d2);
67      src += 64;
68      dst += 16;
69      x -= 16;
70    } while (x);
71    src += 4 * (src_stride - max_width);
72    dst += dst_stride - max_width;
73  } while (--y);
74}
75
76static INLINE __m128i scale_plane_bilinear_kernel(const __m128i *const s,
77                                                  const __m128i c0c1) {
78  const __m128i k_64 = _mm_set1_epi16(1 << 6);
79  const __m128i t0 = _mm_maddubs_epi16(s[0], c0c1);
80  const __m128i t1 = _mm_maddubs_epi16(s[1], c0c1);
81  // round and shift by 7 bit each 16 bit
82  const __m128i t2 = _mm_adds_epi16(t0, k_64);
83  const __m128i t3 = _mm_adds_epi16(t1, k_64);
84  const __m128i t4 = _mm_srai_epi16(t2, 7);
85  const __m128i t5 = _mm_srai_epi16(t3, 7);
86  return _mm_packus_epi16(t4, t5);
87}
88
89static void scale_plane_2_to_1_bilinear(const uint8_t *src,
90                                        const ptrdiff_t src_stride,
91                                        uint8_t *dst,
92                                        const ptrdiff_t dst_stride,
93                                        const int dst_w, const int dst_h,
94                                        const __m128i c0c1) {
95  const int max_width = (dst_w + 15) & ~15;
96  int y = dst_h;
97
98  do {
99    int x = max_width;
100    do {
101      __m128i s[2], d[2];
102
103      // Horizontal
104      // Even rows
105      s[0] = _mm_loadu_si128((const __m128i *)(src + 0));
106      s[1] = _mm_loadu_si128((const __m128i *)(src + 16));
107      d[0] = scale_plane_bilinear_kernel(s, c0c1);
108
109      // odd rows
110      s[0] = _mm_loadu_si128((const __m128i *)(src + src_stride + 0));
111      s[1] = _mm_loadu_si128((const __m128i *)(src + src_stride + 16));
112      d[1] = scale_plane_bilinear_kernel(s, c0c1);
113
114      // Vertical
115      s[0] = _mm_unpacklo_epi8(d[0], d[1]);
116      s[1] = _mm_unpackhi_epi8(d[0], d[1]);
117      d[0] = scale_plane_bilinear_kernel(s, c0c1);
118
119      _mm_storeu_si128((__m128i *)dst, d[0]);
120      src += 32;
121      dst += 16;
122      x -= 16;
123    } while (x);
124    src += 2 * (src_stride - max_width);
125    dst += dst_stride - max_width;
126  } while (--y);
127}
128
129static void scale_plane_4_to_1_bilinear(const uint8_t *src,
130                                        const ptrdiff_t src_stride,
131                                        uint8_t *dst,
132                                        const ptrdiff_t dst_stride,
133                                        const int dst_w, const int dst_h,
134                                        const __m128i c0c1) {
135  const int max_width = (dst_w + 15) & ~15;
136  int y = dst_h;
137
138  do {
139    int x = max_width;
140    do {
141      __m128i s[8], d[8];
142
143      // Note: Using _mm_packus_epi32() in SSE4.1 could be faster.
144      //       Here we tried to not use shuffle instructions which would be slow
145      //       on some x86 CPUs.
146
147      // Horizontal
148      // 000 001 xx xx 004 005 xx xx  008 009 xx xx 00C 00D xx xx
149      // 010 011 xx xx 014 015 xx xx  018 019 xx xx 01C 01D xx xx
150      // 020 021 xx xx 024 025 xx xx  028 029 xx xx 02C 02D xx xx
151      // 030 031 xx xx 034 035 xx xx  038 039 xx xx 03C 03D xx xx
152      // 100 101 xx xx 104 105 xx xx  108 109 xx xx 10C 10D xx xx
153      // 110 111 xx xx 114 115 xx xx  118 119 xx xx 11C 11D xx xx
154      // 120 121 xx xx 124 125 xx xx  128 129 xx xx 12C 12D xx xx
155      // 130 131 xx xx 134 135 xx xx  138 139 xx xx 13C 13D xx xx
156      s[0] = _mm_loadu_si128((const __m128i *)(&src[0]));
157      s[1] = _mm_loadu_si128((const __m128i *)(&src[16]));
158      s[2] = _mm_loadu_si128((const __m128i *)(&src[32]));
159      s[3] = _mm_loadu_si128((const __m128i *)(&src[48]));
160      s[4] = _mm_loadu_si128((const __m128i *)(src + src_stride + 0));
161      s[5] = _mm_loadu_si128((const __m128i *)(src + src_stride + 16));
162      s[6] = _mm_loadu_si128((const __m128i *)(src + src_stride + 32));
163      s[7] = _mm_loadu_si128((const __m128i *)(src + src_stride + 48));
164
165      // 000 001 100 101 xx xx xx xx  004 005 104 105 xx xx xx xx
166      // 008 009 108 109 xx xx xx xx  00C 00D 10C 10D xx xx xx xx
167      // 010 011 110 111 xx xx xx xx  014 015 114 115 xx xx xx xx
168      // 018 019 118 119 xx xx xx xx  01C 01D 11C 11D xx xx xx xx
169      // 020 021 120 121 xx xx xx xx  024 025 124 125 xx xx xx xx
170      // 028 029 128 129 xx xx xx xx  02C 02D 12C 12D xx xx xx xx
171      // 030 031 130 131 xx xx xx xx  034 035 134 135 xx xx xx xx
172      // 038 039 138 139 xx xx xx xx  03C 03D 13C 13D xx xx xx xx
173      d[0] = _mm_unpacklo_epi16(s[0], s[4]);
174      d[1] = _mm_unpackhi_epi16(s[0], s[4]);
175      d[2] = _mm_unpacklo_epi16(s[1], s[5]);
176      d[3] = _mm_unpackhi_epi16(s[1], s[5]);
177      d[4] = _mm_unpacklo_epi16(s[2], s[6]);
178      d[5] = _mm_unpackhi_epi16(s[2], s[6]);
179      d[6] = _mm_unpacklo_epi16(s[3], s[7]);
180      d[7] = _mm_unpackhi_epi16(s[3], s[7]);
181
182      // 000 001 100 101 008 009 108 109  xx xx xx xx xx xx xx xx
183      // 004 005 104 105 00C 00D 10C 10D  xx xx xx xx xx xx xx xx
184      // 010 011 110 111 018 019 118 119  xx xx xx xx xx xx xx xx
185      // 014 015 114 115 01C 01D 11C 11D  xx xx xx xx xx xx xx xx
186      // 020 021 120 121 028 029 128 129  xx xx xx xx xx xx xx xx
187      // 024 025 124 125 02C 02D 12C 12D  xx xx xx xx xx xx xx xx
188      // 030 031 130 131 038 039 138 139  xx xx xx xx xx xx xx xx
189      // 034 035 134 135 03C 03D 13C 13D  xx xx xx xx xx xx xx xx
190      s[0] = _mm_unpacklo_epi32(d[0], d[1]);
191      s[1] = _mm_unpackhi_epi32(d[0], d[1]);
192      s[2] = _mm_unpacklo_epi32(d[2], d[3]);
193      s[3] = _mm_unpackhi_epi32(d[2], d[3]);
194      s[4] = _mm_unpacklo_epi32(d[4], d[5]);
195      s[5] = _mm_unpackhi_epi32(d[4], d[5]);
196      s[6] = _mm_unpacklo_epi32(d[6], d[7]);
197      s[7] = _mm_unpackhi_epi32(d[6], d[7]);
198
199      // 000 001 100 101 004 005 104 105  008 009 108 109 00C 00D 10C 10D
200      // 010 011 110 111 014 015 114 115  018 019 118 119 01C 01D 11C 11D
201      // 020 021 120 121 024 025 124 125  028 029 128 129 02C 02D 12C 12D
202      // 030 031 130 131 034 035 134 135  038 039 138 139 03C 03D 13C 13D
203      d[0] = _mm_unpacklo_epi32(s[0], s[1]);
204      d[1] = _mm_unpacklo_epi32(s[2], s[3]);
205      d[2] = _mm_unpacklo_epi32(s[4], s[5]);
206      d[3] = _mm_unpacklo_epi32(s[6], s[7]);
207
208      d[0] = scale_plane_bilinear_kernel(&d[0], c0c1);
209      d[1] = scale_plane_bilinear_kernel(&d[2], c0c1);
210
211      // Vertical
212      d[0] = scale_plane_bilinear_kernel(d, c0c1);
213
214      _mm_storeu_si128((__m128i *)dst, d[0]);
215      src += 64;
216      dst += 16;
217      x -= 16;
218    } while (x);
219    src += 4 * (src_stride - max_width);
220    dst += dst_stride - max_width;
221  } while (--y);
222}
223
224static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
225                                       uint8_t *dst, const int dst_stride,
226                                       const int w, const int h,
227                                       const int16_t *const coef,
228                                       uint8_t *const temp_buffer) {
229  const int width_hor = (w + 3) & ~3;
230  const int width_ver = (w + 7) & ~7;
231  const int height_hor = (2 * h + SUBPEL_TAPS - 2 + 7) & ~7;
232  const int height_ver = (h + 3) & ~3;
233  int x, y = height_hor;
234  uint8_t *t = temp_buffer;
235  __m128i s[11], d[4];
236  __m128i f[4];
237
238  assert(w && h);
239
240  shuffle_filter_ssse3(coef, f);
241  src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 1;
242
243  // horizontal 4x8
244  do {
245    load_8bit_8x8(src + 2, src_stride, s);
246    // 00 01 10 11 20 21 30 31  40 41 50 51 60 61 70 71
247    // 02 03 12 13 22 23 32 33  42 43 52 53 62 63 72 73
248    // 04 05 14 15 24 25 34 35  44 45 54 55 64 65 74 75
249    // 06 07 16 17 26 27 36 37  46 47 56 57 66 67 76 77 (overlapped)
250    transpose_16bit_4x8(s, s);
251    x = width_hor;
252
253    do {
254      src += 8;
255      load_8bit_8x8(src, src_stride, &s[3]);
256      // 06 07 16 17 26 27 36 37  46 47 56 57 66 67 76 77
257      // 08 09 18 19 28 29 38 39  48 49 58 59 68 69 78 79
258      // 0A 0B 1A 1B 2A 2B 3A 3B  4A 4B 5A 5B 6A 6B 7A 7B
259      // 0C 0D 1C 1D 2C 2D 3C 3D  4C 4D 5C 5D 6C 6D 7C 7D
260      transpose_16bit_4x8(&s[3], &s[3]);
261
262      d[0] = convolve8_8_ssse3(&s[0], f);  // 00 10 20 30 40 50 60 70
263      d[1] = convolve8_8_ssse3(&s[1], f);  // 01 11 21 31 41 51 61 71
264      d[2] = convolve8_8_ssse3(&s[2], f);  // 02 12 22 32 42 52 62 72
265      d[3] = convolve8_8_ssse3(&s[3], f);  // 03 13 23 33 43 53 63 73
266
267      // 00 10 20 30 40 50 60 70  02 12 22 32 42 52 62 72
268      // 01 11 21 31 41 51 61 71  03 13 23 33 43 53 63 73
269      d[0] = _mm_packus_epi16(d[0], d[2]);
270      d[1] = _mm_packus_epi16(d[1], d[3]);
271      // 00 10 01 11 20 30 21 31  40 50 41 51 60 70 61 71
272      // 02 12 03 13 22 32 23 33  42 52 43 53 62 72 63 73
273      d[2] = _mm_unpacklo_epi16(d[0], d[1]);
274      d[3] = _mm_unpackhi_epi16(d[0], d[1]);
275      // 00 10 01 11 02 12 03 13  20 30 21 31 22 32 23 33
276      // 40 50 41 51 42 52 43 53  60 70 61 71 62 72 63 73
277      d[0] = _mm_unpacklo_epi32(d[2], d[3]);
278      d[1] = _mm_unpackhi_epi32(d[2], d[3]);
279      store_8bit_8x4_from_16x2(d, t, 2 * width_hor);
280
281      s[0] = s[4];
282      s[1] = s[5];
283      s[2] = s[6];
284
285      t += 8;
286      x -= 4;
287    } while (x);
288    src += 8 * src_stride - 2 * width_hor;
289    t += 6 * width_hor;
290    y -= 8;
291  } while (y);
292
293  // vertical 8x4
294  x = width_ver;
295  t = temp_buffer;
296  do {
297    // 00 10 01 11 02 12 03 13  04 14 05 15 06 16 07 17
298    // 20 30 21 31 22 32 23 33  24 34 25 35 26 36 27 37
299    // 40 50 41 51 42 52 43 53  44 54 45 55 46 56 47 57
300    s[0] = _mm_loadu_si128((const __m128i *)(t + 0 * width_hor));
301    s[1] = _mm_loadu_si128((const __m128i *)(t + 2 * width_hor));
302    s[2] = _mm_loadu_si128((const __m128i *)(t + 4 * width_hor));
303    t += 6 * width_hor;
304    y = height_ver;
305
306    do {
307      // 60 70 61 71 62 72 63 73  64 74 65 75 66 76 67 77
308      // 80 90 81 91 82 92 83 93  84 94 85 95 86 96 87 77
309      // A0 B0 A1 B1 A2 B2 A3 B3  A4 B4 A5 B5 A6 B6 A7 77
310      // C0 D0 C1 D1 C2 D2 C3 D3  C4 D4 C5 D5 C6 D6 C7 77
311      loadu_8bit_16x4(t, 2 * width_hor, &s[3]);
312      t += 8 * width_hor;
313
314      d[0] = convolve8_8_ssse3(&s[0], f);  // 00 01 02 03 04 05 06 07
315      d[1] = convolve8_8_ssse3(&s[1], f);  // 10 11 12 13 14 15 16 17
316      d[2] = convolve8_8_ssse3(&s[2], f);  // 20 21 22 23 24 25 26 27
317      d[3] = convolve8_8_ssse3(&s[3], f);  // 30 31 32 33 34 35 36 37
318
319      // 00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17
320      // 20 21 22 23 24 25 26 27  30 31 32 33 34 35 36 37
321      d[0] = _mm_packus_epi16(d[0], d[1]);
322      d[1] = _mm_packus_epi16(d[2], d[3]);
323      store_8bit_8x4_from_16x2(d, dst, dst_stride);
324
325      s[0] = s[4];
326      s[1] = s[5];
327      s[2] = s[6];
328
329      dst += 4 * dst_stride;
330      y -= 4;
331    } while (y);
332    t -= width_hor * (2 * height_ver + 6);
333    t += 16;
334    dst -= height_ver * dst_stride;
335    dst += 8;
336    x -= 8;
337  } while (x);
338}
339
340static void scale_plane_4_to_1_general(const uint8_t *src, const int src_stride,
341                                       uint8_t *dst, const int dst_stride,
342                                       const int w, const int h,
343                                       const int16_t *const coef,
344                                       uint8_t *const temp_buffer) {
345  const int width_hor = (w + 1) & ~1;
346  const int width_ver = (w + 7) & ~7;
347  const int height_hor = (4 * h + SUBPEL_TAPS - 2 + 7) & ~7;
348  const int height_ver = (h + 1) & ~1;
349  int x, y = height_hor;
350  uint8_t *t = temp_buffer;
351  __m128i s[11], d[4];
352  __m128i f[4];
353
354  assert(w && h);
355
356  shuffle_filter_ssse3(coef, f);
357  src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 3;
358
359  // horizontal 2x8
360  do {
361    load_8bit_8x8(src + 4, src_stride, s);
362    // 00 01 10 11 20 21 30 31  40 41 50 51 60 61 70 71
363    // 02 03 12 13 22 23 32 33  42 43 52 53 62 63 72 73
364    // 04 05 14 15 24 25 34 35  44 45 54 55 64 65 74 75 (overlapped)
365    // 06 07 16 17 26 27 36 37  46 47 56 57 66 67 76 77 (overlapped)
366    transpose_16bit_4x8(s, s);
367    x = width_hor;
368
369    do {
370      src += 8;
371      load_8bit_8x8(src, src_stride, &s[2]);
372      // 04 05 14 15 24 25 34 35  44 45 54 55 64 65 74 75
373      // 06 07 16 17 26 27 36 37  46 47 56 57 66 67 76 77
374      // 08 09 18 19 28 29 38 39  48 49 58 59 68 69 78 79
375      // 0A 0B 1A 1B 2A 2B 3A 3B  4A 4B 5A 5B 6A 6B 7A 7B
376      transpose_16bit_4x8(&s[2], &s[2]);
377
378      d[0] = convolve8_8_ssse3(&s[0], f);  // 00 10 20 30 40 50 60 70
379      d[1] = convolve8_8_ssse3(&s[2], f);  // 01 11 21 31 41 51 61 71
380
381      // 00 10 20 30 40 50 60 70  xx xx xx xx xx xx xx xx
382      // 01 11 21 31 41 51 61 71  xx xx xx xx xx xx xx xx
383      d[0] = _mm_packus_epi16(d[0], d[0]);
384      d[1] = _mm_packus_epi16(d[1], d[1]);
385      // 00 10 01 11 20 30 21 31  40 50 41 51 60 70 61 71
386      d[0] = _mm_unpacklo_epi16(d[0], d[1]);
387      store_8bit_4x4_sse2(d[0], t, 2 * width_hor);
388
389      s[0] = s[4];
390      s[1] = s[5];
391
392      t += 4;
393      x -= 2;
394    } while (x);
395    src += 8 * src_stride - 4 * width_hor;
396    t += 6 * width_hor;
397    y -= 8;
398  } while (y);
399
400  // vertical 8x2
401  x = width_ver;
402  t = temp_buffer;
403  do {
404    // 00 10 01 11 02 12 03 13  04 14 05 15 06 16 07 17
405    // 20 30 21 31 22 32 23 33  24 34 25 35 26 36 27 37
406    s[0] = _mm_loadu_si128((const __m128i *)(t + 0 * width_hor));
407    s[1] = _mm_loadu_si128((const __m128i *)(t + 2 * width_hor));
408    t += 4 * width_hor;
409    y = height_ver;
410
411    do {
412      // 40 50 41 51 42 52 43 53  44 54 45 55 46 56 47 57
413      // 60 70 61 71 62 72 63 73  64 74 65 75 66 76 67 77
414      // 80 90 81 91 82 92 83 93  84 94 85 95 86 96 87 77
415      // A0 B0 A1 B1 A2 B2 A3 B3  A4 B4 A5 B5 A6 B6 A7 77
416      loadu_8bit_16x4(t, 2 * width_hor, &s[2]);
417      t += 8 * width_hor;
418
419      d[0] = convolve8_8_ssse3(&s[0], f);  // 00 01 02 03 04 05 06 07
420      d[1] = convolve8_8_ssse3(&s[2], f);  // 10 11 12 13 14 15 16 17
421
422      // 00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17
423      d[0] = _mm_packus_epi16(d[0], d[1]);
424      _mm_storel_epi64((__m128i *)(dst + 0 * dst_stride), d[0]);
425      _mm_storeh_epi64((__m128i *)(dst + 1 * dst_stride), d[0]);
426
427      s[0] = s[4];
428      s[1] = s[5];
429
430      dst += 2 * dst_stride;
431      y -= 2;
432    } while (y);
433    t -= width_hor * (4 * height_ver + 4);
434    t += 16;
435    dst -= height_ver * dst_stride;
436    dst += 8;
437    x -= 8;
438  } while (x);
439}
440
441typedef void (*shuffle_filter_funcs)(const int16_t *const filter,
442                                     __m128i *const f);
443
444typedef __m128i (*convolve8_funcs)(const __m128i *const s,
445                                   const __m128i *const f);
446
447static void scale_plane_4_to_3_general(const uint8_t *src, const int src_stride,
448                                       uint8_t *dst, const int dst_stride,
449                                       const int w, const int h,
450                                       const InterpKernel *const coef,
451                                       const int phase_scaler,
452                                       uint8_t *const temp_buffer) {
453  static const int step_q4 = 16 * 4 / 3;
454  const int width_hor = (w + 5) - ((w + 5) % 6);
455  const int stride_hor = 2 * width_hor + 4;  // store 4 extra pixels
456  const int width_ver = (w + 7) & ~7;
457  // We need (SUBPEL_TAPS - 1) extra rows: (SUBPEL_TAPS / 2 - 1) extra rows
458  // above and (SUBPEL_TAPS / 2) extra rows below.
459  const int height_hor = (4 * h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
460  const int height_ver = (h + 5) - ((h + 5) % 6);
461  int x, y = height_hor;
462  uint8_t *t = temp_buffer;
463  __m128i s[12], d[6], dd[4];
464  __m128i f0[4], f1[5], f2[5];
465  // The offset of the first row is always less than 1 pixel.
466  const int offset1_q4 = phase_scaler + 1 * step_q4;
467  const int offset2_q4 = phase_scaler + 2 * step_q4;
468  // offset_idxx indicates the pixel offset is even (0) or odd (1).
469  // It's used to choose the src offset and filter coefficient offset.
470  const int offset_idx1 = (offset1_q4 >> 4) & 1;
471  const int offset_idx2 = (offset2_q4 >> 4) & 1;
472  static const shuffle_filter_funcs shuffle_filter_funcs[2] = {
473    shuffle_filter_ssse3, shuffle_filter_odd_ssse3
474  };
475  static const convolve8_funcs convolve8_funcs[2] = {
476    convolve8_8_even_offset_ssse3, convolve8_8_odd_offset_ssse3
477  };
478
479  assert(w && h);
480
481  shuffle_filter_ssse3(coef[(phase_scaler + 0 * step_q4) & SUBPEL_MASK], f0);
482  shuffle_filter_funcs[offset_idx1](coef[offset1_q4 & SUBPEL_MASK], f1);
483  shuffle_filter_funcs[offset_idx2](coef[offset2_q4 & SUBPEL_MASK], f2);
484
485  // Sub 64 to avoid overflow.
486  // Coef 128 would be treated as -128 in PMADDUBSW. Sub 64 here.
487  // Coef 128 is in either fx[1] or fx[2] depending on the phase idx.
488  // When filter phase idx is 1, the two biggest coefficients are shuffled
489  // together, and the sum of them are always no less than 128. Sub 64 here.
490  // After the subtraction, when the sum of all positive coefficients are no
491  // larger than 128, and the sum of all negative coefficients are no
492  // less than -128, there will be no overflow in the convolve8 functions.
493  f0[1] = _mm_sub_epi8(f0[1], _mm_set1_epi8(64));
494  f1[1 + offset_idx1] = _mm_sub_epi8(f1[1 + offset_idx1], _mm_set1_epi8(64));
495  f2[1 + offset_idx2] = _mm_sub_epi8(f2[1 + offset_idx2], _mm_set1_epi8(64));
496
497  src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 - 1;
498
499  // horizontal 6x8
500  do {
501    load_8bit_8x8(src, src_stride, s);
502    // 00 01 10 11 20 21 30 31  40 41 50 51 60 61 70 71
503    // 02 03 12 13 22 23 32 33  42 43 52 53 62 63 72 73
504    // 04 05 14 15 24 25 34 35  44 45 54 55 64 65 74 75
505    // 06 07 16 17 26 27 36 37  46 47 56 57 66 67 76 77
506    transpose_16bit_4x8(s, s);
507    x = width_hor;
508
509    do {
510      src += 8;
511      load_8bit_8x8(src, src_stride, &s[4]);
512      // 08 09 18 19 28 29 38 39  48 49 58 59 68 69 78 79
513      // 0A 0B 1A 1B 2A 2B 3A 3B  4A 4B 5A 5B 6A 6B 7A 7B
514      // OC 0D 1C 1D 2C 2D 3C 3D  4C 4D 5C 5D 6C 6D 7C 7D
515      // 0E 0F 1E 1F 2E 2F 3E 3F  4E 4F 5E 5F 6E 6F 7E 7F
516      transpose_16bit_4x8(&s[4], &s[4]);
517
518      // 00 10 20 30 40 50 60 70
519      // 01 11 21 31 41 51 61 71
520      // 02 12 22 32 42 52 62 72
521      // 03 13 23 33 43 53 63 73
522      // 04 14 24 34 44 54 64 74
523      // 05 15 25 35 45 55 65 75
524      d[0] = convolve8_8_even_offset_ssse3(&s[0], f0);
525      d[1] = convolve8_funcs[offset_idx1](&s[offset1_q4 >> 5], f1);
526      d[2] = convolve8_funcs[offset_idx2](&s[offset2_q4 >> 5], f2);
527      d[3] = convolve8_8_even_offset_ssse3(&s[2], f0);
528      d[4] = convolve8_funcs[offset_idx1](&s[2 + (offset1_q4 >> 5)], f1);
529      d[5] = convolve8_funcs[offset_idx2](&s[2 + (offset2_q4 >> 5)], f2);
530
531      // 00 10 20 30 40 50 60 70  02 12 22 32 42 52 62 72
532      // 01 11 21 31 41 51 61 71  03 13 23 33 43 53 63 73
533      // 04 14 24 34 44 54 64 74  xx xx xx xx xx xx xx xx
534      // 05 15 25 35 45 55 65 75  xx xx xx xx xx xx xx xx
535      dd[0] = _mm_packus_epi16(d[0], d[2]);
536      dd[1] = _mm_packus_epi16(d[1], d[3]);
537      dd[2] = _mm_packus_epi16(d[4], d[4]);
538      dd[3] = _mm_packus_epi16(d[5], d[5]);
539
540      // 00 10 01 11 20 30 21 31  40 50 41 51 60 70 61 71
541      // 02 12 03 13 22 32 23 33  42 52 43 53 62 72 63 73
542      // 04 14 05 15 24 34 25 35  44 54 45 55 64 74 65 75
543      d[0] = _mm_unpacklo_epi16(dd[0], dd[1]);
544      d[1] = _mm_unpackhi_epi16(dd[0], dd[1]);
545      d[2] = _mm_unpacklo_epi16(dd[2], dd[3]);
546
547      // 00 10 01 11 02 12 03 13  20 30 21 31 22 32 23 33
548      // 40 50 41 51 42 52 43 53  60 70 61 71 62 72 63 73
549      // 04 14 05 15 xx xx xx xx  24 34 25 35 xx xx xx xx
550      // 44 54 45 55 xx xx xx xx  64 74 65 75 xx xx xx xx
551      dd[0] = _mm_unpacklo_epi32(d[0], d[1]);
552      dd[1] = _mm_unpackhi_epi32(d[0], d[1]);
553      dd[2] = _mm_unpacklo_epi32(d[2], d[2]);
554      dd[3] = _mm_unpackhi_epi32(d[2], d[2]);
555
556      // 00 10 01 11 02 12 03 13  04 14 05 15 xx xx xx xx
557      // 20 30 21 31 22 32 23 33  24 34 25 35 xx xx xx xx
558      // 40 50 41 51 42 52 43 53  44 54 45 55 xx xx xx xx
559      // 60 70 61 71 62 72 63 73  64 74 65 75 xx xx xx xx
560      d[0] = _mm_unpacklo_epi64(dd[0], dd[2]);
561      d[1] = _mm_unpackhi_epi64(dd[0], dd[2]);
562      d[2] = _mm_unpacklo_epi64(dd[1], dd[3]);
563      d[3] = _mm_unpackhi_epi64(dd[1], dd[3]);
564
565      // store 4 extra pixels
566      storeu_8bit_16x4(d, t, stride_hor);
567
568      s[0] = s[4];
569      s[1] = s[5];
570      s[2] = s[6];
571      s[3] = s[7];
572
573      t += 12;
574      x -= 6;
575    } while (x);
576    src += 8 * src_stride - 4 * width_hor / 3;
577    t += 3 * stride_hor + 4;
578    y -= 8;
579  } while (y);
580
581  // vertical 8x6
582  x = width_ver;
583  t = temp_buffer;
584  do {
585    // 00 10 01 11 02 12 03 13  04 14 05 15 06 16 07 17
586    // 20 30 21 31 22 32 23 33  24 34 25 35 26 36 27 37
587    // 40 50 41 51 42 52 43 53  44 54 45 55 46 56 47 57
588    // 60 70 61 71 62 72 63 73  64 74 65 75 66 76 67 77
589    loadu_8bit_16x4(t, stride_hor, s);
590    y = height_ver;
591
592    do {
593      // 80 90 81 91 82 92 83 93  84 94 85 95 86 96 87 97
594      // A0 B0 A1 B1 A2 B2 A3 B3  A4 B4 A5 B5 A6 B6 A7 B7
595      // C0 D0 C1 D1 C2 D2 C3 D3  C4 D4 C5 D5 C6 D6 C7 D7
596      // E0 F0 E1 F1 E2 F2 E3 F3  E4 F4 E5 F5 E6 F6 E7 F7
597      t += 4 * stride_hor;
598      loadu_8bit_16x4(t, stride_hor, &s[4]);
599
600      d[0] = convolve8_8_even_offset_ssse3(&s[0], f0);
601      d[1] = convolve8_funcs[offset_idx1](&s[offset1_q4 >> 5], f1);
602      d[2] = convolve8_funcs[offset_idx2](&s[offset2_q4 >> 5], f2);
603      d[3] = convolve8_8_even_offset_ssse3(&s[2], f0);
604      d[4] = convolve8_funcs[offset_idx1](&s[2 + (offset1_q4 >> 5)], f1);
605      d[5] = convolve8_funcs[offset_idx2](&s[2 + (offset2_q4 >> 5)], f2);
606
607      // 00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17
608      // 20 21 22 23 24 25 26 27  30 31 32 33 34 35 36 37
609      // 40 41 42 43 44 45 46 47  50 51 52 53 54 55 56 57
610      d[0] = _mm_packus_epi16(d[0], d[1]);
611      d[2] = _mm_packus_epi16(d[2], d[3]);
612      d[4] = _mm_packus_epi16(d[4], d[5]);
613
614      _mm_storel_epi64((__m128i *)(dst + 0 * dst_stride), d[0]);
615      _mm_storeh_epi64((__m128i *)(dst + 1 * dst_stride), d[0]);
616      _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), d[2]);
617      _mm_storeh_epi64((__m128i *)(dst + 3 * dst_stride), d[2]);
618      _mm_storel_epi64((__m128i *)(dst + 4 * dst_stride), d[4]);
619      _mm_storeh_epi64((__m128i *)(dst + 5 * dst_stride), d[4]);
620
621      s[0] = s[4];
622      s[1] = s[5];
623      s[2] = s[6];
624      s[3] = s[7];
625
626      dst += 6 * dst_stride;
627      y -= 6;
628    } while (y);
629    t -= stride_hor * 2 * height_ver / 3;
630    t += 16;
631    dst -= height_ver * dst_stride;
632    dst += 8;
633    x -= 8;
634  } while (x);
635}
636
637static INLINE __m128i scale_1_to_2_phase_0_kernel(const __m128i *const s,
638                                                  const __m128i *const f) {
639  __m128i ss[4], temp;
640
641  ss[0] = _mm_unpacklo_epi8(s[0], s[1]);
642  ss[1] = _mm_unpacklo_epi8(s[2], s[3]);
643  ss[2] = _mm_unpacklo_epi8(s[4], s[5]);
644  ss[3] = _mm_unpacklo_epi8(s[6], s[7]);
645  temp = convolve8_8_ssse3(ss, f);
646  return _mm_packus_epi16(temp, temp);
647}
648
649// Only calculate odd columns since even columns are just src pixels' copies.
650static void scale_1_to_2_phase_0_row(const uint8_t *src, uint8_t *dst,
651                                     const int w, const __m128i *const f) {
652  int x = w;
653
654  do {
655    __m128i s[8], temp;
656    s[0] = _mm_loadl_epi64((const __m128i *)(src + 0));
657    s[1] = _mm_loadl_epi64((const __m128i *)(src + 1));
658    s[2] = _mm_loadl_epi64((const __m128i *)(src + 2));
659    s[3] = _mm_loadl_epi64((const __m128i *)(src + 3));
660    s[4] = _mm_loadl_epi64((const __m128i *)(src + 4));
661    s[5] = _mm_loadl_epi64((const __m128i *)(src + 5));
662    s[6] = _mm_loadl_epi64((const __m128i *)(src + 6));
663    s[7] = _mm_loadl_epi64((const __m128i *)(src + 7));
664    temp = scale_1_to_2_phase_0_kernel(s, f);
665    _mm_storel_epi64((__m128i *)dst, temp);
666    src += 8;
667    dst += 8;
668    x -= 8;
669  } while (x);
670}
671
672static void scale_plane_1_to_2_phase_0(const uint8_t *src,
673                                       const ptrdiff_t src_stride, uint8_t *dst,
674                                       const ptrdiff_t dst_stride,
675                                       const int src_w, const int src_h,
676                                       const int16_t *const coef,
677                                       uint8_t *const temp_buffer) {
678  int max_width;
679  int y;
680  uint8_t *tmp[9];
681  __m128i f[4];
682
683  max_width = (src_w + 7) & ~7;
684  tmp[0] = temp_buffer + 0 * max_width;
685  tmp[1] = temp_buffer + 1 * max_width;
686  tmp[2] = temp_buffer + 2 * max_width;
687  tmp[3] = temp_buffer + 3 * max_width;
688  tmp[4] = temp_buffer + 4 * max_width;
689  tmp[5] = temp_buffer + 5 * max_width;
690  tmp[6] = temp_buffer + 6 * max_width;
691  tmp[7] = temp_buffer + 7 * max_width;
692
693  shuffle_filter_ssse3(coef, f);
694
695  scale_1_to_2_phase_0_row(src - 3 * src_stride - 3, tmp[0], max_width, f);
696  scale_1_to_2_phase_0_row(src - 2 * src_stride - 3, tmp[1], max_width, f);
697  scale_1_to_2_phase_0_row(src - 1 * src_stride - 3, tmp[2], max_width, f);
698  scale_1_to_2_phase_0_row(src + 0 * src_stride - 3, tmp[3], max_width, f);
699  scale_1_to_2_phase_0_row(src + 1 * src_stride - 3, tmp[4], max_width, f);
700  scale_1_to_2_phase_0_row(src + 2 * src_stride - 3, tmp[5], max_width, f);
701  scale_1_to_2_phase_0_row(src + 3 * src_stride - 3, tmp[6], max_width, f);
702
703  y = src_h;
704  do {
705    int x;
706    scale_1_to_2_phase_0_row(src + 4 * src_stride - 3, tmp[7], max_width, f);
707    for (x = 0; x < max_width; x += 8) {
708      __m128i s[8], C, D, CD;
709
710      // Even rows
711      const __m128i a = _mm_loadl_epi64((const __m128i *)(src + x));
712      const __m128i b = _mm_loadl_epi64((const __m128i *)(tmp[3] + x));
713      const __m128i ab = _mm_unpacklo_epi8(a, b);
714      _mm_storeu_si128((__m128i *)(dst + 2 * x), ab);
715
716      // Odd rows
717      // Even columns
718      load_8bit_8x8(src + x - 3 * src_stride, src_stride, s);
719      C = scale_1_to_2_phase_0_kernel(s, f);
720
721      // Odd columns
722      s[0] = _mm_loadl_epi64((const __m128i *)(tmp[0] + x));
723      s[1] = _mm_loadl_epi64((const __m128i *)(tmp[1] + x));
724      s[2] = _mm_loadl_epi64((const __m128i *)(tmp[2] + x));
725      s[3] = _mm_loadl_epi64((const __m128i *)(tmp[3] + x));
726      s[4] = _mm_loadl_epi64((const __m128i *)(tmp[4] + x));
727      s[5] = _mm_loadl_epi64((const __m128i *)(tmp[5] + x));
728      s[6] = _mm_loadl_epi64((const __m128i *)(tmp[6] + x));
729      s[7] = _mm_loadl_epi64((const __m128i *)(tmp[7] + x));
730      D = scale_1_to_2_phase_0_kernel(s, f);
731
732      CD = _mm_unpacklo_epi8(C, D);
733      _mm_storeu_si128((__m128i *)(dst + dst_stride + 2 * x), CD);
734    }
735
736    src += src_stride;
737    dst += 2 * dst_stride;
738    tmp[8] = tmp[0];
739    tmp[0] = tmp[1];
740    tmp[1] = tmp[2];
741    tmp[2] = tmp[3];
742    tmp[3] = tmp[4];
743    tmp[4] = tmp[5];
744    tmp[5] = tmp[6];
745    tmp[6] = tmp[7];
746    tmp[7] = tmp[8];
747  } while (--y);
748}
749
750void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
751                                      YV12_BUFFER_CONFIG *dst,
752                                      uint8_t filter_type, int phase_scaler) {
753  const int src_w = src->y_crop_width;
754  const int src_h = src->y_crop_height;
755  const int dst_w = dst->y_crop_width;
756  const int dst_h = dst->y_crop_height;
757  const int dst_uv_w = dst_w / 2;
758  const int dst_uv_h = dst_h / 2;
759  int scaled = 0;
760
761  // phase_scaler is usually 0 or 8.
762  assert(phase_scaler >= 0 && phase_scaler < 16);
763
764  if (dst_w * 2 == src_w && dst_h * 2 == src_h) {
765    // 2 to 1
766    scaled = 1;
767
768    if (phase_scaler == 0) {
769      scale_plane_2_to_1_phase_0(src->y_buffer, src->y_stride, dst->y_buffer,
770                                 dst->y_stride, dst_w, dst_h);
771      scale_plane_2_to_1_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
772                                 dst->uv_stride, dst_uv_w, dst_uv_h);
773      scale_plane_2_to_1_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
774                                 dst->uv_stride, dst_uv_w, dst_uv_h);
775    } else if (filter_type == BILINEAR) {
776      const int16_t c0 = vp9_filter_kernels[BILINEAR][phase_scaler][3];
777      const int16_t c1 = vp9_filter_kernels[BILINEAR][phase_scaler][4];
778      const __m128i c0c1 = _mm_set1_epi16(c0 | (c1 << 8));  // c0 and c1 >= 0
779      scale_plane_2_to_1_bilinear(src->y_buffer, src->y_stride, dst->y_buffer,
780                                  dst->y_stride, dst_w, dst_h, c0c1);
781      scale_plane_2_to_1_bilinear(src->u_buffer, src->uv_stride, dst->u_buffer,
782                                  dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
783      scale_plane_2_to_1_bilinear(src->v_buffer, src->uv_stride, dst->v_buffer,
784                                  dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
785    } else {
786      const int buffer_stride = (dst_w + 3) & ~3;
787      const int buffer_height = (2 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
788      uint8_t *const temp_buffer =
789          (uint8_t *)malloc(buffer_stride * buffer_height);
790      if (temp_buffer) {
791        scale_plane_2_to_1_general(
792            src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
793            dst_h, vp9_filter_kernels[filter_type][phase_scaler], temp_buffer);
794        scale_plane_2_to_1_general(
795            src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
796            dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
797            temp_buffer);
798        scale_plane_2_to_1_general(
799            src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
800            dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
801            temp_buffer);
802        free(temp_buffer);
803      } else {
804        scaled = 0;
805      }
806    }
807  } else if (4 * dst_w == src_w && 4 * dst_h == src_h) {
808    // 4 to 1
809    scaled = 1;
810    if (phase_scaler == 0) {
811      scale_plane_4_to_1_phase_0(src->y_buffer, src->y_stride, dst->y_buffer,
812                                 dst->y_stride, dst_w, dst_h);
813      scale_plane_4_to_1_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
814                                 dst->uv_stride, dst_uv_w, dst_uv_h);
815      scale_plane_4_to_1_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
816                                 dst->uv_stride, dst_uv_w, dst_uv_h);
817    } else if (filter_type == BILINEAR) {
818      const int16_t c0 = vp9_filter_kernels[BILINEAR][phase_scaler][3];
819      const int16_t c1 = vp9_filter_kernels[BILINEAR][phase_scaler][4];
820      const __m128i c0c1 = _mm_set1_epi16(c0 | (c1 << 8));  // c0 and c1 >= 0
821      scale_plane_4_to_1_bilinear(src->y_buffer, src->y_stride, dst->y_buffer,
822                                  dst->y_stride, dst_w, dst_h, c0c1);
823      scale_plane_4_to_1_bilinear(src->u_buffer, src->uv_stride, dst->u_buffer,
824                                  dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
825      scale_plane_4_to_1_bilinear(src->v_buffer, src->uv_stride, dst->v_buffer,
826                                  dst->uv_stride, dst_uv_w, dst_uv_h, c0c1);
827    } else {
828      const int buffer_stride = (dst_w + 1) & ~1;
829      const int buffer_height = (4 * dst_h + SUBPEL_TAPS - 2 + 7) & ~7;
830      // When dst_w is 1 or 2, we need extra padding to avoid heap read overflow
831      const int extra_padding = 16;
832      uint8_t *const temp_buffer =
833          (uint8_t *)malloc(buffer_stride * buffer_height + extra_padding);
834      if (temp_buffer) {
835        scale_plane_4_to_1_general(
836            src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
837            dst_h, vp9_filter_kernels[filter_type][phase_scaler], temp_buffer);
838        scale_plane_4_to_1_general(
839            src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
840            dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
841            temp_buffer);
842        scale_plane_4_to_1_general(
843            src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
844            dst_uv_w, dst_uv_h, vp9_filter_kernels[filter_type][phase_scaler],
845            temp_buffer);
846        free(temp_buffer);
847      } else {
848        scaled = 0;
849      }
850    }
851  } else if (4 * dst_w == 3 * src_w && 4 * dst_h == 3 * src_h) {
852    // 4 to 3
853    const int buffer_stride_hor = (dst_w + 5) - ((dst_w + 5) % 6) + 2;
854    const int buffer_stride_ver = (dst_w + 7) & ~7;
855    const int buffer_height = (4 * dst_h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
856    // When the vertical filter reads more pixels than the horizontal filter
857    // generated in each row, we need extra padding to avoid heap read overflow.
858    // For example, the horizontal filter generates 18 pixels but the vertical
859    // filter reads 24 pixels in a row. The difference is multiplied by 2 since
860    // two rows are interlaced together in the optimization.
861    const int extra_padding = (buffer_stride_ver > buffer_stride_hor)
862                                  ? 2 * (buffer_stride_ver - buffer_stride_hor)
863                                  : 0;
864    const int buffer_size = buffer_stride_hor * buffer_height + extra_padding;
865    uint8_t *const temp_buffer = (uint8_t *)malloc(buffer_size);
866    if (temp_buffer) {
867      scaled = 1;
868      scale_plane_4_to_3_general(
869          src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
870          dst_h, vp9_filter_kernels[filter_type], phase_scaler, temp_buffer);
871      scale_plane_4_to_3_general(src->u_buffer, src->uv_stride, dst->u_buffer,
872                                 dst->uv_stride, dst_uv_w, dst_uv_h,
873                                 vp9_filter_kernels[filter_type], phase_scaler,
874                                 temp_buffer);
875      scale_plane_4_to_3_general(src->v_buffer, src->uv_stride, dst->v_buffer,
876                                 dst->uv_stride, dst_uv_w, dst_uv_h,
877                                 vp9_filter_kernels[filter_type], phase_scaler,
878                                 temp_buffer);
879      free(temp_buffer);
880    }
881  } else if (dst_w == src_w * 2 && dst_h == src_h * 2 && phase_scaler == 0) {
882    // 1 to 2
883    uint8_t *const temp_buffer = (uint8_t *)malloc(8 * ((src_w + 7) & ~7));
884    if (temp_buffer) {
885      scaled = 1;
886      scale_plane_1_to_2_phase_0(
887          src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, src_w,
888          src_h, vp9_filter_kernels[filter_type][8], temp_buffer);
889      scale_plane_1_to_2_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
890                                 dst->uv_stride, src_w / 2, src_h / 2,
891                                 vp9_filter_kernels[filter_type][8],
892                                 temp_buffer);
893      scale_plane_1_to_2_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
894                                 dst->uv_stride, src_w / 2, src_h / 2,
895                                 vp9_filter_kernels[filter_type][8],
896                                 temp_buffer);
897      free(temp_buffer);
898    }
899  }
900
901  if (scaled) {
902    vpx_extend_frame_borders(dst);
903  } else {
904    // Call c version for all other scaling ratios.
905    vp9_scale_and_extend_frame_c(src, dst, filter_type, phase_scaler);
906  }
907}
908