• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017, Alliance for Open Media. All rights reserved
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include <stdio.h>
13 #include <tmmintrin.h>
14 
15 #include "config/aom_config.h"
16 #include "config/aom_dsp_rtcd.h"
17 
18 #include "aom_dsp/blend.h"
19 #include "aom/aom_integer.h"
20 #include "aom_dsp/x86/synonyms.h"
21 
22 #include "aom_dsp/x86/masked_sad_intrin_ssse3.h"
23 
24 // For width a multiple of 16
25 static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr,
26                                             int src_stride,
27                                             const uint8_t *a_ptr, int a_stride,
28                                             const uint8_t *b_ptr, int b_stride,
29                                             const uint8_t *m_ptr, int m_stride,
30                                             int width, int height);
31 
32 #define MASKSADMXN_SSSE3(m, n)                                                \
33   unsigned int aom_masked_sad##m##x##n##_ssse3(                               \
34       const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
35       const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
36       int invert_mask) {                                                      \
37     if (!invert_mask)                                                         \
38       return masked_sad_ssse3(src, src_stride, ref, ref_stride, second_pred,  \
39                               m, msk, msk_stride, m, n);                      \
40     else                                                                      \
41       return masked_sad_ssse3(src, src_stride, second_pred, m, ref,           \
42                               ref_stride, msk, msk_stride, m, n);             \
43   }
44 
45 #define MASKSAD8XN_SSSE3(n)                                                   \
46   unsigned int aom_masked_sad8x##n##_ssse3(                                   \
47       const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
48       const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
49       int invert_mask) {                                                      \
50     if (!invert_mask)                                                         \
51       return aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,        \
52                                      second_pred, 8, msk, msk_stride, n);     \
53     else                                                                      \
54       return aom_masked_sad8xh_ssse3(src, src_stride, second_pred, 8, ref,    \
55                                      ref_stride, msk, msk_stride, n);         \
56   }
57 
58 #define MASKSAD4XN_SSSE3(n)                                                   \
59   unsigned int aom_masked_sad4x##n##_ssse3(                                   \
60       const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
61       const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
62       int invert_mask) {                                                      \
63     if (!invert_mask)                                                         \
64       return aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,        \
65                                      second_pred, 4, msk, msk_stride, n);     \
66     else                                                                      \
67       return aom_masked_sad4xh_ssse3(src, src_stride, second_pred, 4, ref,    \
68                                      ref_stride, msk, msk_stride, n);         \
69   }
70 
71 MASKSADMXN_SSSE3(128, 128)
72 MASKSADMXN_SSSE3(128, 64)
73 MASKSADMXN_SSSE3(64, 128)
74 MASKSADMXN_SSSE3(64, 64)
75 MASKSADMXN_SSSE3(64, 32)
76 MASKSADMXN_SSSE3(32, 64)
77 MASKSADMXN_SSSE3(32, 32)
78 MASKSADMXN_SSSE3(32, 16)
79 MASKSADMXN_SSSE3(16, 32)
80 MASKSADMXN_SSSE3(16, 16)
81 MASKSADMXN_SSSE3(16, 8)
82 MASKSAD8XN_SSSE3(16)
83 MASKSAD8XN_SSSE3(8)
84 MASKSAD8XN_SSSE3(4)
85 MASKSAD4XN_SSSE3(8)
86 MASKSAD4XN_SSSE3(4)
87 MASKSAD4XN_SSSE3(16)
88 MASKSADMXN_SSSE3(16, 4)
89 MASKSAD8XN_SSSE3(32)
90 MASKSADMXN_SSSE3(32, 8)
91 MASKSADMXN_SSSE3(16, 64)
92 MASKSADMXN_SSSE3(64, 16)
93 
masked_sad_ssse3(const uint8_t * src_ptr,int src_stride,const uint8_t * a_ptr,int a_stride,const uint8_t * b_ptr,int b_stride,const uint8_t * m_ptr,int m_stride,int width,int height)94 static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr,
95                                             int src_stride,
96                                             const uint8_t *a_ptr, int a_stride,
97                                             const uint8_t *b_ptr, int b_stride,
98                                             const uint8_t *m_ptr, int m_stride,
99                                             int width, int height) {
100   int x, y;
101   __m128i res = _mm_setzero_si128();
102   const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
103 
104   for (y = 0; y < height; y++) {
105     for (x = 0; x < width; x += 16) {
106       const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
107       const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]);
108       const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
109       const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]);
110       const __m128i m_inv = _mm_sub_epi8(mask_max, m);
111 
112       // Calculate 16 predicted pixels.
113       // Note that the maximum value of any entry of 'pred_l' or 'pred_r'
114       // is 64 * 255, so we have plenty of space to add rounding constants.
115       const __m128i data_l = _mm_unpacklo_epi8(a, b);
116       const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv);
117       __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l);
118       pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS);
119 
120       const __m128i data_r = _mm_unpackhi_epi8(a, b);
121       const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv);
122       __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r);
123       pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS);
124 
125       const __m128i pred = _mm_packus_epi16(pred_l, pred_r);
126       res = _mm_add_epi32(res, _mm_sad_epu8(pred, src));
127     }
128 
129     src_ptr += src_stride;
130     a_ptr += a_stride;
131     b_ptr += b_stride;
132     m_ptr += m_stride;
133   }
134   // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
135   unsigned int sad = (unsigned int)(_mm_cvtsi128_si32(res) +
136                                     _mm_cvtsi128_si32(_mm_srli_si128(res, 8)));
137   return sad;
138 }
139 
aom_masked_sad8xh_ssse3(const uint8_t * src_ptr,int src_stride,const uint8_t * a_ptr,int a_stride,const uint8_t * b_ptr,int b_stride,const uint8_t * m_ptr,int m_stride,int height)140 unsigned int aom_masked_sad8xh_ssse3(const uint8_t *src_ptr, int src_stride,
141                                      const uint8_t *a_ptr, int a_stride,
142                                      const uint8_t *b_ptr, int b_stride,
143                                      const uint8_t *m_ptr, int m_stride,
144                                      int height) {
145   int y;
146   __m128i res = _mm_setzero_si128();
147   const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
148 
149   for (y = 0; y < height; y += 2) {
150     const __m128i src = _mm_unpacklo_epi64(
151         _mm_loadl_epi64((const __m128i *)src_ptr),
152         _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride]));
153     const __m128i a0 = _mm_loadl_epi64((const __m128i *)a_ptr);
154     const __m128i a1 = _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride]);
155     const __m128i b0 = _mm_loadl_epi64((const __m128i *)b_ptr);
156     const __m128i b1 = _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride]);
157     const __m128i m =
158         _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr),
159                            _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride]));
160     const __m128i m_inv = _mm_sub_epi8(mask_max, m);
161 
162     const __m128i data_l = _mm_unpacklo_epi8(a0, b0);
163     const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv);
164     __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l);
165     pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS);
166 
167     const __m128i data_r = _mm_unpacklo_epi8(a1, b1);
168     const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv);
169     __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r);
170     pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS);
171 
172     const __m128i pred = _mm_packus_epi16(pred_l, pred_r);
173     res = _mm_add_epi32(res, _mm_sad_epu8(pred, src));
174 
175     src_ptr += src_stride * 2;
176     a_ptr += a_stride * 2;
177     b_ptr += b_stride * 2;
178     m_ptr += m_stride * 2;
179   }
180   unsigned int sad = (unsigned int)(_mm_cvtsi128_si32(res) +
181                                     _mm_cvtsi128_si32(_mm_srli_si128(res, 8)));
182   return sad;
183 }
184 
aom_masked_sad4xh_ssse3(const uint8_t * src_ptr,int src_stride,const uint8_t * a_ptr,int a_stride,const uint8_t * b_ptr,int b_stride,const uint8_t * m_ptr,int m_stride,int height)185 unsigned int aom_masked_sad4xh_ssse3(const uint8_t *src_ptr, int src_stride,
186                                      const uint8_t *a_ptr, int a_stride,
187                                      const uint8_t *b_ptr, int b_stride,
188                                      const uint8_t *m_ptr, int m_stride,
189                                      int height) {
190   int y;
191   __m128i res = _mm_setzero_si128();
192   const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
193 
194   for (y = 0; y < height; y += 2) {
195     // Load two rows at a time, this seems to be a bit faster
196     // than four rows at a time in this case.
197     const __m128i src =
198         _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)src_ptr),
199                            _mm_cvtsi32_si128(*(int *)&src_ptr[src_stride]));
200     const __m128i a =
201         _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)a_ptr),
202                            _mm_cvtsi32_si128(*(int *)&a_ptr[a_stride]));
203     const __m128i b =
204         _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)b_ptr),
205                            _mm_cvtsi32_si128(*(int *)&b_ptr[b_stride]));
206     const __m128i m =
207         _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)m_ptr),
208                            _mm_cvtsi32_si128(*(int *)&m_ptr[m_stride]));
209     const __m128i m_inv = _mm_sub_epi8(mask_max, m);
210 
211     const __m128i data = _mm_unpacklo_epi8(a, b);
212     const __m128i mask = _mm_unpacklo_epi8(m, m_inv);
213     __m128i pred_16bit = _mm_maddubs_epi16(data, mask);
214     pred_16bit = xx_roundn_epu16(pred_16bit, AOM_BLEND_A64_ROUND_BITS);
215 
216     const __m128i pred = _mm_packus_epi16(pred_16bit, _mm_setzero_si128());
217     res = _mm_add_epi32(res, _mm_sad_epu8(pred, src));
218 
219     src_ptr += src_stride * 2;
220     a_ptr += a_stride * 2;
221     b_ptr += b_stride * 2;
222     m_ptr += m_stride * 2;
223   }
224   // At this point, the SAD is stored in lane 0 of 'res'
225   return (unsigned int)_mm_cvtsi128_si32(res);
226 }
227 
228 // For width a multiple of 8
229 static INLINE unsigned int highbd_masked_sad_ssse3(
230     const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
231     const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
232     int width, int height);
233 
234 #define HIGHBD_MASKSADMXN_SSSE3(m, n)                                         \
235   unsigned int aom_highbd_masked_sad##m##x##n##_ssse3(                        \
236       const uint8_t *src8, int src_stride, const uint8_t *ref8,               \
237       int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,        \
238       int msk_stride, int invert_mask) {                                      \
239     if (!invert_mask)                                                         \
240       return highbd_masked_sad_ssse3(src8, src_stride, ref8, ref_stride,      \
241                                      second_pred8, m, msk, msk_stride, m, n); \
242     else                                                                      \
243       return highbd_masked_sad_ssse3(src8, src_stride, second_pred8, m, ref8, \
244                                      ref_stride, msk, msk_stride, m, n);      \
245   }
246 
247 #define HIGHBD_MASKSAD4XN_SSSE3(n)                                             \
248   unsigned int aom_highbd_masked_sad4x##n##_ssse3(                             \
249       const uint8_t *src8, int src_stride, const uint8_t *ref8,                \
250       int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,         \
251       int msk_stride, int invert_mask) {                                       \
252     if (!invert_mask)                                                          \
253       return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, ref8,            \
254                                             ref_stride, second_pred8, 4, msk,  \
255                                             msk_stride, n);                    \
256     else                                                                       \
257       return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, second_pred8, 4, \
258                                             ref8, ref_stride, msk, msk_stride, \
259                                             n);                                \
260   }
261 
262 HIGHBD_MASKSADMXN_SSSE3(128, 128)
263 HIGHBD_MASKSADMXN_SSSE3(128, 64)
264 HIGHBD_MASKSADMXN_SSSE3(64, 128)
265 HIGHBD_MASKSADMXN_SSSE3(64, 64)
266 HIGHBD_MASKSADMXN_SSSE3(64, 32)
267 HIGHBD_MASKSADMXN_SSSE3(32, 64)
268 HIGHBD_MASKSADMXN_SSSE3(32, 32)
269 HIGHBD_MASKSADMXN_SSSE3(32, 16)
270 HIGHBD_MASKSADMXN_SSSE3(16, 32)
271 HIGHBD_MASKSADMXN_SSSE3(16, 16)
272 HIGHBD_MASKSADMXN_SSSE3(16, 8)
273 HIGHBD_MASKSADMXN_SSSE3(8, 16)
274 HIGHBD_MASKSADMXN_SSSE3(8, 8)
275 HIGHBD_MASKSADMXN_SSSE3(8, 4)
276 HIGHBD_MASKSAD4XN_SSSE3(8)
277 HIGHBD_MASKSAD4XN_SSSE3(4)
278 HIGHBD_MASKSAD4XN_SSSE3(16)
279 HIGHBD_MASKSADMXN_SSSE3(16, 4)
280 HIGHBD_MASKSADMXN_SSSE3(8, 32)
281 HIGHBD_MASKSADMXN_SSSE3(32, 8)
282 HIGHBD_MASKSADMXN_SSSE3(16, 64)
283 HIGHBD_MASKSADMXN_SSSE3(64, 16)
284 
highbd_masked_sad_ssse3(const uint8_t * src8,int src_stride,const uint8_t * a8,int a_stride,const uint8_t * b8,int b_stride,const uint8_t * m_ptr,int m_stride,int width,int height)285 static INLINE unsigned int highbd_masked_sad_ssse3(
286     const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
287     const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
288     int width, int height) {
289   const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
290   const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
291   const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
292   int x, y;
293   __m128i res = _mm_setzero_si128();
294   const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
295   const __m128i round_const =
296       _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
297   const __m128i one = _mm_set1_epi16(1);
298 
299   for (y = 0; y < height; y++) {
300     for (x = 0; x < width; x += 8) {
301       const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
302       const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]);
303       const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
304       // Zero-extend mask to 16 bits
305       const __m128i m = _mm_unpacklo_epi8(
306           _mm_loadl_epi64((const __m128i *)&m_ptr[x]), _mm_setzero_si128());
307       const __m128i m_inv = _mm_sub_epi16(mask_max, m);
308 
309       const __m128i data_l = _mm_unpacklo_epi16(a, b);
310       const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv);
311       __m128i pred_l = _mm_madd_epi16(data_l, mask_l);
312       pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const),
313                               AOM_BLEND_A64_ROUND_BITS);
314 
315       const __m128i data_r = _mm_unpackhi_epi16(a, b);
316       const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv);
317       __m128i pred_r = _mm_madd_epi16(data_r, mask_r);
318       pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const),
319                               AOM_BLEND_A64_ROUND_BITS);
320 
321       // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
322       // so it is safe to do signed saturation here.
323       const __m128i pred = _mm_packs_epi32(pred_l, pred_r);
324       // There is no 16-bit SAD instruction, so we have to synthesize
325       // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
326       // and accumulating them at the end
327       const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src));
328       res = _mm_add_epi32(res, _mm_madd_epi16(diff, one));
329     }
330 
331     src_ptr += src_stride;
332     a_ptr += a_stride;
333     b_ptr += b_stride;
334     m_ptr += m_stride;
335   }
336   // At this point, we have four 32-bit partial SADs stored in 'res'.
337   res = _mm_hadd_epi32(res, res);
338   res = _mm_hadd_epi32(res, res);
339   int sad = _mm_cvtsi128_si32(res);
340   return sad;
341 }
342 
aom_highbd_masked_sad4xh_ssse3(const uint8_t * src8,int src_stride,const uint8_t * a8,int a_stride,const uint8_t * b8,int b_stride,const uint8_t * m_ptr,int m_stride,int height)343 unsigned int aom_highbd_masked_sad4xh_ssse3(const uint8_t *src8, int src_stride,
344                                             const uint8_t *a8, int a_stride,
345                                             const uint8_t *b8, int b_stride,
346                                             const uint8_t *m_ptr, int m_stride,
347                                             int height) {
348   const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
349   const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
350   const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
351   int y;
352   __m128i res = _mm_setzero_si128();
353   const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
354   const __m128i round_const =
355       _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
356   const __m128i one = _mm_set1_epi16(1);
357 
358   for (y = 0; y < height; y += 2) {
359     const __m128i src = _mm_unpacklo_epi64(
360         _mm_loadl_epi64((const __m128i *)src_ptr),
361         _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride]));
362     const __m128i a =
363         _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)a_ptr),
364                            _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride]));
365     const __m128i b =
366         _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)b_ptr),
367                            _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride]));
368     // Zero-extend mask to 16 bits
369     const __m128i m = _mm_unpacklo_epi8(
370         _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(const int *)m_ptr),
371                            _mm_cvtsi32_si128(*(const int *)&m_ptr[m_stride])),
372         _mm_setzero_si128());
373     const __m128i m_inv = _mm_sub_epi16(mask_max, m);
374 
375     const __m128i data_l = _mm_unpacklo_epi16(a, b);
376     const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv);
377     __m128i pred_l = _mm_madd_epi16(data_l, mask_l);
378     pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const),
379                             AOM_BLEND_A64_ROUND_BITS);
380 
381     const __m128i data_r = _mm_unpackhi_epi16(a, b);
382     const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv);
383     __m128i pred_r = _mm_madd_epi16(data_r, mask_r);
384     pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const),
385                             AOM_BLEND_A64_ROUND_BITS);
386 
387     const __m128i pred = _mm_packs_epi32(pred_l, pred_r);
388     const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src));
389     res = _mm_add_epi32(res, _mm_madd_epi16(diff, one));
390 
391     src_ptr += src_stride * 2;
392     a_ptr += a_stride * 2;
393     b_ptr += b_stride * 2;
394     m_ptr += m_stride * 2;
395   }
396   res = _mm_hadd_epi32(res, res);
397   res = _mm_hadd_epi32(res, res);
398   int sad = _mm_cvtsi128_si32(res);
399   return sad;
400 }
401