• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018, 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 <tmmintrin.h>
13 
14 #include "config/aom_config.h"
15 #include "config/aom_dsp_rtcd.h"
16 
17 #include "aom_dsp/blend.h"
18 #include "aom/aom_integer.h"
19 #include "aom_dsp/x86/synonyms.h"
20 #include "aom_dsp/x86/masked_sad_intrin_ssse3.h"
21 
masked_sad32xh_avx2(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)22 static INLINE unsigned int masked_sad32xh_avx2(
23     const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
24     const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
25     int width, int height) {
26   int x, y;
27   __m256i res = _mm256_setzero_si256();
28   const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
29   const __m256i round_scale =
30       _mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
31   for (y = 0; y < height; y++) {
32     for (x = 0; x < width; x += 32) {
33       const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
34       const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
35       const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
36       const __m256i m = _mm256_lddqu_si256((const __m256i *)&m_ptr[x]);
37       const __m256i m_inv = _mm256_sub_epi8(mask_max, m);
38 
39       // Calculate 16 predicted pixels.
40       // Note that the maximum value of any entry of 'pred_l' or 'pred_r'
41       // is 64 * 255, so we have plenty of space to add rounding constants.
42       const __m256i data_l = _mm256_unpacklo_epi8(a, b);
43       const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv);
44       __m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l);
45       pred_l = _mm256_mulhrs_epi16(pred_l, round_scale);
46 
47       const __m256i data_r = _mm256_unpackhi_epi8(a, b);
48       const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv);
49       __m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r);
50       pred_r = _mm256_mulhrs_epi16(pred_r, round_scale);
51 
52       const __m256i pred = _mm256_packus_epi16(pred_l, pred_r);
53       res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src));
54     }
55 
56     src_ptr += src_stride;
57     a_ptr += a_stride;
58     b_ptr += b_stride;
59     m_ptr += m_stride;
60   }
61   // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
62   res = _mm256_shuffle_epi32(res, 0xd8);
63   res = _mm256_permute4x64_epi64(res, 0xd8);
64   res = _mm256_hadd_epi32(res, res);
65   res = _mm256_hadd_epi32(res, res);
66   int32_t sad = _mm256_extract_epi32(res, 0);
67   return sad;
68 }
69 
xx_loadu2_m128i(const void * hi,const void * lo)70 static INLINE __m256i xx_loadu2_m128i(const void *hi, const void *lo) {
71   __m128i a0 = _mm_lddqu_si128((const __m128i *)(lo));
72   __m128i a1 = _mm_lddqu_si128((const __m128i *)(hi));
73   __m256i a = _mm256_castsi128_si256(a0);
74   return _mm256_inserti128_si256(a, a1, 1);
75 }
76 
masked_sad16xh_avx2(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)77 static INLINE unsigned int masked_sad16xh_avx2(
78     const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
79     const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
80     int height) {
81   int y;
82   __m256i res = _mm256_setzero_si256();
83   const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
84   const __m256i round_scale =
85       _mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
86   for (y = 0; y < height; y += 2) {
87     const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr);
88     const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr);
89     const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr);
90     const __m256i m = xx_loadu2_m128i(m_ptr + m_stride, m_ptr);
91     const __m256i m_inv = _mm256_sub_epi8(mask_max, m);
92 
93     // Calculate 16 predicted pixels.
94     // Note that the maximum value of any entry of 'pred_l' or 'pred_r'
95     // is 64 * 255, so we have plenty of space to add rounding constants.
96     const __m256i data_l = _mm256_unpacklo_epi8(a, b);
97     const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv);
98     __m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l);
99     pred_l = _mm256_mulhrs_epi16(pred_l, round_scale);
100 
101     const __m256i data_r = _mm256_unpackhi_epi8(a, b);
102     const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv);
103     __m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r);
104     pred_r = _mm256_mulhrs_epi16(pred_r, round_scale);
105 
106     const __m256i pred = _mm256_packus_epi16(pred_l, pred_r);
107     res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src));
108 
109     src_ptr += src_stride << 1;
110     a_ptr += a_stride << 1;
111     b_ptr += b_stride << 1;
112     m_ptr += m_stride << 1;
113   }
114   // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
115   res = _mm256_shuffle_epi32(res, 0xd8);
116   res = _mm256_permute4x64_epi64(res, 0xd8);
117   res = _mm256_hadd_epi32(res, res);
118   res = _mm256_hadd_epi32(res, res);
119   int32_t sad = _mm256_extract_epi32(res, 0);
120   return sad;
121 }
122 
aom_masked_sad_avx2(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,const uint8_t * second_pred,const uint8_t * msk,int msk_stride,int invert_mask,int m,int n)123 static INLINE unsigned int aom_masked_sad_avx2(
124     const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
125     const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
126     int invert_mask, int m, int n) {
127   unsigned int sad;
128   if (!invert_mask) {
129     switch (m) {
130       case 4:
131         sad = aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
132                                       second_pred, m, msk, msk_stride, n);
133         break;
134       case 8:
135         sad = aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,
136                                       second_pred, m, msk, msk_stride, n);
137         break;
138       case 16:
139         sad = masked_sad16xh_avx2(src, src_stride, ref, ref_stride, second_pred,
140                                   m, msk, msk_stride, n);
141         break;
142       default:
143         sad = masked_sad32xh_avx2(src, src_stride, ref, ref_stride, second_pred,
144                                   m, msk, msk_stride, m, n);
145         break;
146     }
147   } else {
148     switch (m) {
149       case 4:
150         sad = aom_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
151                                       ref_stride, msk, msk_stride, n);
152         break;
153       case 8:
154         sad = aom_masked_sad8xh_ssse3(src, src_stride, second_pred, m, ref,
155                                       ref_stride, msk, msk_stride, n);
156         break;
157       case 16:
158         sad = masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
159                                   ref_stride, msk, msk_stride, n);
160         break;
161       default:
162         sad = masked_sad32xh_avx2(src, src_stride, second_pred, m, ref,
163                                   ref_stride, msk, msk_stride, m, n);
164         break;
165     }
166   }
167   return sad;
168 }
169 
170 #define MASKSADMXN_AVX2(m, n)                                                 \
171   unsigned int aom_masked_sad##m##x##n##_avx2(                                \
172       const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
173       const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
174       int invert_mask) {                                                      \
175     return aom_masked_sad_avx2(src, src_stride, ref, ref_stride, second_pred, \
176                                msk, msk_stride, invert_mask, m, n);           \
177   }
178 
179 MASKSADMXN_AVX2(4, 4)
180 MASKSADMXN_AVX2(4, 8)
181 MASKSADMXN_AVX2(8, 4)
182 MASKSADMXN_AVX2(8, 8)
183 MASKSADMXN_AVX2(8, 16)
184 MASKSADMXN_AVX2(16, 8)
185 MASKSADMXN_AVX2(16, 16)
186 MASKSADMXN_AVX2(16, 32)
187 MASKSADMXN_AVX2(32, 16)
188 MASKSADMXN_AVX2(32, 32)
189 MASKSADMXN_AVX2(32, 64)
190 MASKSADMXN_AVX2(64, 32)
191 MASKSADMXN_AVX2(64, 64)
192 MASKSADMXN_AVX2(64, 128)
193 MASKSADMXN_AVX2(128, 64)
194 MASKSADMXN_AVX2(128, 128)
195 MASKSADMXN_AVX2(4, 16)
196 MASKSADMXN_AVX2(16, 4)
197 MASKSADMXN_AVX2(8, 32)
198 MASKSADMXN_AVX2(32, 8)
199 MASKSADMXN_AVX2(16, 64)
200 MASKSADMXN_AVX2(64, 16)
201 
highbd_masked_sad8xh_avx2(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)202 static INLINE unsigned int highbd_masked_sad8xh_avx2(
203     const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
204     const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
205     int height) {
206   const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
207   const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
208   const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
209   int y;
210   __m256i res = _mm256_setzero_si256();
211   const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
212   const __m256i round_const =
213       _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
214   const __m256i one = _mm256_set1_epi16(1);
215 
216   for (y = 0; y < height; y += 2) {
217     const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr);
218     const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr);
219     const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr);
220     // Zero-extend mask to 16 bits
221     const __m256i m = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(
222         _mm_loadl_epi64((const __m128i *)(m_ptr)),
223         _mm_loadl_epi64((const __m128i *)(m_ptr + m_stride))));
224     const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
225 
226     const __m256i data_l = _mm256_unpacklo_epi16(a, b);
227     const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
228     __m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
229     pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
230                                AOM_BLEND_A64_ROUND_BITS);
231 
232     const __m256i data_r = _mm256_unpackhi_epi16(a, b);
233     const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
234     __m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
235     pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
236                                AOM_BLEND_A64_ROUND_BITS);
237 
238     // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
239     // so it is safe to do signed saturation here.
240     const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
241     // There is no 16-bit SAD instruction, so we have to synthesize
242     // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
243     // and accumulating them at the end
244     const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
245     res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
246 
247     src_ptr += src_stride << 1;
248     a_ptr += a_stride << 1;
249     b_ptr += b_stride << 1;
250     m_ptr += m_stride << 1;
251   }
252   // At this point, we have four 32-bit partial SADs stored in 'res'.
253   res = _mm256_hadd_epi32(res, res);
254   res = _mm256_hadd_epi32(res, res);
255   int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
256   return sad;
257 }
258 
highbd_masked_sad16xh_avx2(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)259 static INLINE unsigned int highbd_masked_sad16xh_avx2(
260     const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
261     const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
262     int width, int height) {
263   const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
264   const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
265   const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
266   int x, y;
267   __m256i res = _mm256_setzero_si256();
268   const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
269   const __m256i round_const =
270       _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
271   const __m256i one = _mm256_set1_epi16(1);
272 
273   for (y = 0; y < height; y++) {
274     for (x = 0; x < width; x += 16) {
275       const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
276       const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
277       const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
278       // Zero-extend mask to 16 bits
279       const __m256i m =
280           _mm256_cvtepu8_epi16(_mm_lddqu_si128((const __m128i *)&m_ptr[x]));
281       const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
282 
283       const __m256i data_l = _mm256_unpacklo_epi16(a, b);
284       const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
285       __m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
286       pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
287                                  AOM_BLEND_A64_ROUND_BITS);
288 
289       const __m256i data_r = _mm256_unpackhi_epi16(a, b);
290       const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
291       __m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
292       pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
293                                  AOM_BLEND_A64_ROUND_BITS);
294 
295       // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
296       // so it is safe to do signed saturation here.
297       const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
298       // There is no 16-bit SAD instruction, so we have to synthesize
299       // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
300       // and accumulating them at the end
301       const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
302       res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
303     }
304 
305     src_ptr += src_stride;
306     a_ptr += a_stride;
307     b_ptr += b_stride;
308     m_ptr += m_stride;
309   }
310   // At this point, we have four 32-bit partial SADs stored in 'res'.
311   res = _mm256_hadd_epi32(res, res);
312   res = _mm256_hadd_epi32(res, res);
313   int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
314   return sad;
315 }
316 
aom_highbd_masked_sad_avx2(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,const uint8_t * second_pred,const uint8_t * msk,int msk_stride,int invert_mask,int m,int n)317 static INLINE unsigned int aom_highbd_masked_sad_avx2(
318     const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
319     const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
320     int invert_mask, int m, int n) {
321   unsigned int sad;
322   if (!invert_mask) {
323     switch (m) {
324       case 4:
325         sad =
326             aom_highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
327                                            second_pred, m, msk, msk_stride, n);
328         break;
329       case 8:
330         sad = highbd_masked_sad8xh_avx2(src, src_stride, ref, ref_stride,
331                                         second_pred, m, msk, msk_stride, n);
332         break;
333       default:
334         sad = highbd_masked_sad16xh_avx2(src, src_stride, ref, ref_stride,
335                                          second_pred, m, msk, msk_stride, m, n);
336         break;
337     }
338   } else {
339     switch (m) {
340       case 4:
341         sad =
342             aom_highbd_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
343                                            ref_stride, msk, msk_stride, n);
344         break;
345       case 8:
346         sad = highbd_masked_sad8xh_avx2(src, src_stride, second_pred, m, ref,
347                                         ref_stride, msk, msk_stride, n);
348         break;
349       default:
350         sad = highbd_masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
351                                          ref_stride, msk, msk_stride, m, n);
352         break;
353     }
354   }
355   return sad;
356 }
357 
358 #define HIGHBD_MASKSADMXN_AVX2(m, n)                                      \
359   unsigned int aom_highbd_masked_sad##m##x##n##_avx2(                     \
360       const uint8_t *src8, int src_stride, const uint8_t *ref8,           \
361       int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,    \
362       int msk_stride, int invert_mask) {                                  \
363     return aom_highbd_masked_sad_avx2(src8, src_stride, ref8, ref_stride, \
364                                       second_pred8, msk, msk_stride,      \
365                                       invert_mask, m, n);                 \
366   }
367 
368 HIGHBD_MASKSADMXN_AVX2(4, 4)
369 HIGHBD_MASKSADMXN_AVX2(4, 8)
370 HIGHBD_MASKSADMXN_AVX2(8, 4)
371 HIGHBD_MASKSADMXN_AVX2(8, 8)
372 HIGHBD_MASKSADMXN_AVX2(8, 16)
373 HIGHBD_MASKSADMXN_AVX2(16, 8)
374 HIGHBD_MASKSADMXN_AVX2(16, 16)
375 HIGHBD_MASKSADMXN_AVX2(16, 32)
376 HIGHBD_MASKSADMXN_AVX2(32, 16)
377 HIGHBD_MASKSADMXN_AVX2(32, 32)
378 HIGHBD_MASKSADMXN_AVX2(32, 64)
379 HIGHBD_MASKSADMXN_AVX2(64, 32)
380 HIGHBD_MASKSADMXN_AVX2(64, 64)
381 HIGHBD_MASKSADMXN_AVX2(64, 128)
382 HIGHBD_MASKSADMXN_AVX2(128, 64)
383 HIGHBD_MASKSADMXN_AVX2(128, 128)
384 HIGHBD_MASKSADMXN_AVX2(4, 16)
385 HIGHBD_MASKSADMXN_AVX2(16, 4)
386 HIGHBD_MASKSADMXN_AVX2(8, 32)
387 HIGHBD_MASKSADMXN_AVX2(32, 8)
388 HIGHBD_MASKSADMXN_AVX2(16, 64)
389 HIGHBD_MASKSADMXN_AVX2(64, 16)
390