/work/libde265/libde265/x86/sse-motion.cc
Line | Count | Source |
1 | | /* |
2 | | * H.265 video codec. |
3 | | * Copyright (c) 2013 openHEVC contributors |
4 | | * Copyright (c) 2013-2014 struktur AG, Dirk Farin <farin@struktur.de> |
5 | | * |
6 | | * This file is part of libde265. |
7 | | * |
8 | | * libde265 is free software: you can redistribute it and/or modify |
9 | | * it under the terms of the GNU Lesser General Public License as |
10 | | * published by the Free Software Foundation, either version 3 of |
11 | | * the License, or (at your option) any later version. |
12 | | * |
13 | | * libde265 is distributed in the hope that it will be useful, |
14 | | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 | | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
16 | | * GNU Lesser General Public License for more details. |
17 | | * |
18 | | * You should have received a copy of the GNU Lesser General Public License |
19 | | * along with libde265. If not, see <http://www.gnu.org/licenses/>. |
20 | | */ |
21 | | |
22 | | #ifdef HAVE_CONFIG_H |
23 | | #include "config.h" |
24 | | #endif |
25 | | |
26 | | #include <stdio.h> |
27 | | #include <emmintrin.h> |
28 | | #include <tmmintrin.h> // SSSE3 |
29 | | #if HAVE_SSE4_1 |
30 | | #include <smmintrin.h> |
31 | | #endif |
32 | | |
33 | | #include "sse-motion.h" |
34 | | #include "libde265/util.h" |
35 | | |
36 | | |
37 | | ALIGNED_16(const int8_t) epel_filters[7][16] = { |
38 | | { -2, 58, 10, -2,-2, 58, 10, -2,-2, 58, 10, -2,-2, 58, 10, -2 }, |
39 | | { -4, 54, 16, -2,-4, 54, 16, -2,-4, 54, 16, -2,-4, 54, 16, -2 }, |
40 | | { -6, 46, 28, -4,-6, 46, 28, -4,-6, 46, 28, -4,-6, 46, 28, -4 }, |
41 | | { -4, 36, 36, -4,-4, 36, 36, -4,-4, 36, 36, -4,-4, 36, 36, -4 }, |
42 | | { -4, 28, 46, -6,-4, 28, 46, -6,-4, 28, 46, -6,-4, 28, 46, -6 }, |
43 | | { -2, 16, 54, -4,-2, 16, 54, -4,-2, 16, 54, -4,-2, 16, 54, -4 }, |
44 | | { -2, 10, 58, -2,-2, 10, 58, -2,-2, 10, 58, -2,-2, 10, 58, -2 }, |
45 | | }; |
46 | | |
47 | | static const uint8_t qpel_extra_before[4] = { 0, 3, 3, 2 }; |
48 | | //static const uint8_t qpel_extra_after[4] = { 0, 3, 4, 4 }; |
49 | | static const uint8_t qpel_extra[4] = { 0, 6, 7, 6 }; |
50 | | |
51 | | static const int epel_extra_before = 1; |
52 | | //static const int epel_extra_after = 2; |
53 | | static const int epel_extra = 3; |
54 | | |
55 | 203M | #define MAX_PB_SIZE 64 |
56 | | |
57 | | #define MASKMOVE 0 |
58 | | |
59 | | void print128(const char* prefix, __m128i r) |
60 | 0 | { |
61 | 0 | unsigned char buf[16]; |
62 | |
|
63 | 0 | *(__m128i*)buf = r; |
64 | |
|
65 | 0 | printf("%s ",prefix); |
66 | 0 | for (int i=0;i<16;i++) |
67 | 0 | { |
68 | 0 | if (i>0) { printf(":"); } |
69 | 0 | printf("%02x", buf[i]); |
70 | 0 | } |
71 | |
|
72 | 0 | printf("\n"); |
73 | 0 | } |
74 | | |
75 | | |
76 | | void printm32(const char* prefix, unsigned char* p) |
77 | 0 | { |
78 | 0 | printf("%s ",prefix); |
79 | |
|
80 | 0 | for (int i=0;i<4;i++) |
81 | 0 | { |
82 | 0 | if (i>0) { printf(":"); } |
83 | 0 | printf("%02x", p[i]); |
84 | 0 | } |
85 | |
|
86 | 0 | printf("\n"); |
87 | 0 | } |
88 | | |
89 | | |
90 | 4.97M | #define BIT_DEPTH 8 |
91 | | |
92 | | void ff_hevc_put_unweighted_pred_8_sse(uint8_t *_dst, ptrdiff_t dststride, |
93 | | const int16_t *src, ptrdiff_t srcstride, |
94 | 12.4M | int width, int height) { |
95 | 12.4M | int x, y; |
96 | 12.4M | uint8_t *dst = (uint8_t*) _dst; |
97 | 12.4M | __m128i r0, r1, f0; |
98 | | |
99 | 12.4M | f0 = _mm_set1_epi16(32); |
100 | | |
101 | | |
102 | 12.4M | if(!(width & 15)) |
103 | 568k | { |
104 | 9.64M | for (y = 0; y < height; y++) { |
105 | 20.1M | for (x = 0; x < width; x += 16) { |
106 | 11.1M | r0 = _mm_load_si128((__m128i *) (src+x)); |
107 | | |
108 | 11.1M | r1 = _mm_load_si128((__m128i *) (src+x + 8)); |
109 | 11.1M | r0 = _mm_adds_epi16(r0, f0); |
110 | | |
111 | 11.1M | r1 = _mm_adds_epi16(r1, f0); |
112 | 11.1M | r0 = _mm_srai_epi16(r0, 6); |
113 | 11.1M | r1 = _mm_srai_epi16(r1, 6); |
114 | 11.1M | r0 = _mm_packus_epi16(r0, r1); |
115 | | |
116 | 11.1M | _mm_storeu_si128((__m128i *) (dst+x), r0); |
117 | 11.1M | } |
118 | 9.07M | dst += dststride; |
119 | 9.07M | src += srcstride; |
120 | 9.07M | } |
121 | 11.9M | }else if(!(width & 7)) |
122 | 3.01M | { |
123 | 26.0M | for (y = 0; y < height; y++) { |
124 | 46.3M | for (x = 0; x < width; x += 8) { |
125 | 23.3M | r0 = _mm_load_si128((__m128i *) (src+x)); |
126 | | |
127 | 23.3M | r0 = _mm_adds_epi16(r0, f0); |
128 | | |
129 | 23.3M | r0 = _mm_srai_epi16(r0, 6); |
130 | 23.3M | r0 = _mm_packus_epi16(r0, r0); |
131 | | |
132 | 23.3M | _mm_storel_epi64((__m128i *) (dst+x), r0); |
133 | 23.3M | } |
134 | 23.0M | dst += dststride; |
135 | 23.0M | src += srcstride; |
136 | 23.0M | } |
137 | 8.90M | }else if(!(width & 3)){ |
138 | 39.9M | for (y = 0; y < height; y++) { |
139 | 66.2M | for(x = 0;x < width; x+=4){ |
140 | 33.5M | r0 = _mm_loadl_epi64((__m128i *) (src+x)); |
141 | 33.5M | r0 = _mm_adds_epi16(r0, f0); |
142 | | |
143 | 33.5M | r0 = _mm_srai_epi16(r0, 6); |
144 | 33.5M | r0 = _mm_packus_epi16(r0, r0); |
145 | | #if MASKMOVE |
146 | | _mm_maskmoveu_si128(r0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
147 | | #else |
148 | | //r0 = _mm_shuffle_epi32 (r0, 0x00); |
149 | 33.5M | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
150 | 33.5M | #endif |
151 | 33.5M | } |
152 | 32.7M | dst += dststride; |
153 | 32.7M | src += srcstride; |
154 | 32.7M | } |
155 | 7.24M | }else{ |
156 | 10.7M | for (y = 0; y < height; y++) { |
157 | 18.7M | for(x = 0;x < width; x+=2){ |
158 | 9.63M | r0 = _mm_loadl_epi64((__m128i *) (src+x)); |
159 | 9.63M | r0 = _mm_adds_epi16(r0, f0); |
160 | | |
161 | 9.63M | r0 = _mm_srai_epi16(r0, 6); |
162 | 9.63M | r0 = _mm_packus_epi16(r0, r0); |
163 | | #if MASKMOVE |
164 | | _mm_maskmoveu_si128(r0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,-1),(char *) (dst+x)); |
165 | | #else |
166 | 9.63M | *((uint16_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
167 | 9.63M | #endif |
168 | 9.63M | } |
169 | 9.11M | dst += dststride; |
170 | 9.11M | src += srcstride; |
171 | 9.11M | } |
172 | 1.65M | } |
173 | | |
174 | 12.4M | } |
175 | | |
176 | | void ff_hevc_put_unweighted_pred_sse(uint8_t *_dst, ptrdiff_t _dststride, |
177 | | const int16_t *src, ptrdiff_t srcstride, |
178 | 0 | int width, int height) { |
179 | 0 | int x, y; |
180 | 0 | uint8_t *dst = (uint8_t*) _dst; |
181 | 0 | ptrdiff_t dststride = _dststride / sizeof(uint8_t); |
182 | 0 | __m128i r0, r1, f0; |
183 | 0 | int shift = 14 - BIT_DEPTH; |
184 | 0 | #if BIT_DEPTH < 14 |
185 | 0 | int16_t offset = 1 << (shift - 1); |
186 | | #else |
187 | | int16_t offset = 0; |
188 | | |
189 | | #endif |
190 | 0 | f0 = _mm_set1_epi16(offset); |
191 | |
|
192 | 0 | for (y = 0; y < height; y++) { |
193 | 0 | for (x = 0; x < width; x += 16) { |
194 | 0 | r0 = _mm_load_si128((__m128i *) &src[x]); |
195 | |
|
196 | 0 | r1 = _mm_load_si128((__m128i *) &src[x + 8]); |
197 | 0 | r0 = _mm_adds_epi16(r0, f0); |
198 | |
|
199 | 0 | r1 = _mm_adds_epi16(r1, f0); |
200 | 0 | r0 = _mm_srai_epi16(r0, shift); |
201 | 0 | r1 = _mm_srai_epi16(r1, shift); |
202 | 0 | r0 = _mm_packus_epi16(r0, r1); |
203 | |
|
204 | 0 | _mm_storeu_si128((__m128i *) &dst[x], r0); |
205 | 0 | } |
206 | 0 | dst += dststride; |
207 | 0 | src += srcstride; |
208 | 0 | } |
209 | 0 | } |
210 | | |
211 | | void ff_hevc_put_weighted_pred_avg_8_sse(uint8_t *_dst, ptrdiff_t dststride, |
212 | | const int16_t *src1, const int16_t *src2, |
213 | | ptrdiff_t srcstride, int width, |
214 | 2.20M | int height) { |
215 | 2.20M | int x, y; |
216 | 2.20M | uint8_t *dst = (uint8_t*) _dst; |
217 | 2.20M | __m128i r0, r1, f0, r2, r3; |
218 | | |
219 | 2.20M | f0 = _mm_set1_epi16(64); |
220 | 2.20M | if(!(width & 15)){ |
221 | 4.80M | for (y = 0; y < height; y++) { |
222 | | |
223 | 10.3M | for (x = 0; x < width; x += 16) { |
224 | 5.84M | r0 = _mm_load_si128((__m128i *) &src1[x]); |
225 | 5.84M | r1 = _mm_load_si128((__m128i *) &src1[x + 8]); |
226 | 5.84M | r2 = _mm_load_si128((__m128i *) &src2[x]); |
227 | 5.84M | r3 = _mm_load_si128((__m128i *) &src2[x + 8]); |
228 | | |
229 | 5.84M | r0 = _mm_adds_epi16(r0, f0); |
230 | 5.84M | r1 = _mm_adds_epi16(r1, f0); |
231 | 5.84M | r0 = _mm_adds_epi16(r0, r2); |
232 | 5.84M | r1 = _mm_adds_epi16(r1, r3); |
233 | 5.84M | r0 = _mm_srai_epi16(r0, 7); |
234 | 5.84M | r1 = _mm_srai_epi16(r1, 7); |
235 | 5.84M | r0 = _mm_packus_epi16(r0, r1); |
236 | | |
237 | 5.84M | _mm_storeu_si128((__m128i *) (dst + x), r0); |
238 | 5.84M | } |
239 | 4.52M | dst += dststride; |
240 | 4.52M | src1 += srcstride; |
241 | 4.52M | src2 += srcstride; |
242 | 4.52M | } |
243 | 1.92M | }else if(!(width & 7)){ |
244 | 8.36M | for (y = 0; y < height; y++) { |
245 | 15.1M | for(x=0;x<width;x+=8){ |
246 | 7.65M | r0 = _mm_load_si128((__m128i *) (src1+x)); |
247 | 7.65M | r2 = _mm_load_si128((__m128i *) (src2+x)); |
248 | | |
249 | 7.65M | r0 = _mm_adds_epi16(r0, f0); |
250 | 7.65M | r0 = _mm_adds_epi16(r0, r2); |
251 | 7.65M | r0 = _mm_srai_epi16(r0, 7); |
252 | 7.65M | r0 = _mm_packus_epi16(r0, r0); |
253 | | |
254 | 7.65M | _mm_storel_epi64((__m128i *) (dst+x), r0); |
255 | 7.65M | } |
256 | 7.53M | dst += dststride; |
257 | 7.53M | src1 += srcstride; |
258 | 7.53M | src2 += srcstride; |
259 | 7.53M | } |
260 | 1.08M | }else if(!(width & 3)){ |
261 | | #if MASKMOVE |
262 | | r1= _mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1); |
263 | | #endif |
264 | 7.10M | for (y = 0; y < height; y++) { |
265 | | |
266 | 12.2M | for(x=0;x<width;x+=4) |
267 | 6.23M | { |
268 | 6.23M | r0 = _mm_loadl_epi64((__m128i *) (src1+x)); |
269 | 6.23M | r2 = _mm_loadl_epi64((__m128i *) (src2+x)); |
270 | | |
271 | 6.23M | r0 = _mm_adds_epi16(r0, f0); |
272 | 6.23M | r0 = _mm_adds_epi16(r0, r2); |
273 | 6.23M | r0 = _mm_srai_epi16(r0, 7); |
274 | 6.23M | r0 = _mm_packus_epi16(r0, r0); |
275 | | |
276 | | #if MASKMOVE |
277 | | _mm_maskmoveu_si128(r0,r1,(char *) (dst+x)); |
278 | | #else |
279 | 6.23M | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
280 | 6.23M | #endif |
281 | 6.23M | } |
282 | 6.02M | dst += dststride; |
283 | 6.02M | src1 += srcstride; |
284 | 6.02M | src2 += srcstride; |
285 | 6.02M | } |
286 | 1.07M | }else{ |
287 | | #if MASKMOVE |
288 | | r1= _mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,-1); |
289 | | #endif |
290 | 130k | for (y = 0; y < height; y++) { |
291 | 353k | for(x=0;x<width;x+=2) |
292 | 235k | { |
293 | 235k | r0 = _mm_loadl_epi64((__m128i *) (src1+x)); |
294 | 235k | r2 = _mm_loadl_epi64((__m128i *) (src2+x)); |
295 | | |
296 | 235k | r0 = _mm_adds_epi16(r0, f0); |
297 | 235k | r0 = _mm_adds_epi16(r0, r2); |
298 | 235k | r0 = _mm_srai_epi16(r0, 7); |
299 | 235k | r0 = _mm_packus_epi16(r0, r0); |
300 | | |
301 | | #if MASKMOVE |
302 | | _mm_maskmoveu_si128(r0,r1,(char *) (dst+x)); |
303 | | #else |
304 | 235k | *((uint16_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
305 | 235k | #endif |
306 | 235k | } |
307 | 117k | dst += dststride; |
308 | 117k | src1 += srcstride; |
309 | 117k | src2 += srcstride; |
310 | 117k | } |
311 | 12.2k | } |
312 | | |
313 | | |
314 | 2.20M | } |
315 | | |
316 | | void ff_hevc_put_weighted_pred_avg_sse(uint8_t *_dst, ptrdiff_t _dststride, |
317 | | const int16_t *src1, const int16_t *src2, |
318 | | ptrdiff_t srcstride, int width, |
319 | 0 | int height) { |
320 | 0 | int x, y; |
321 | 0 | uint8_t *dst = (uint8_t*) _dst; |
322 | 0 | ptrdiff_t dststride = _dststride / sizeof(uint8_t); |
323 | 0 | __m128i r0, r1, f0, r2, r3; |
324 | 0 | int shift = 14 + 1 - BIT_DEPTH; |
325 | 0 | #if BIT_DEPTH < 14 |
326 | 0 | int offset = 1 << (shift - 1); |
327 | | #else |
328 | | int offset = 0; |
329 | | #endif |
330 | 0 | f0 = _mm_set1_epi16(offset); |
331 | 0 | for (y = 0; y < height; y++) { |
332 | |
|
333 | 0 | for (x = 0; x < width; x += 16) { |
334 | 0 | r0 = _mm_load_si128((__m128i *) &src1[x]); |
335 | 0 | r1 = _mm_load_si128((__m128i *) &src1[x + 8]); |
336 | 0 | r2 = _mm_load_si128((__m128i *) &src2[x]); |
337 | 0 | r3 = _mm_load_si128((__m128i *) &src2[x + 8]); |
338 | |
|
339 | 0 | r0 = _mm_adds_epi16(r0, f0); |
340 | 0 | r1 = _mm_adds_epi16(r1, f0); |
341 | 0 | r0 = _mm_adds_epi16(r0, r2); |
342 | 0 | r1 = _mm_adds_epi16(r1, r3); |
343 | 0 | r0 = _mm_srai_epi16(r0, shift); |
344 | 0 | r1 = _mm_srai_epi16(r1, shift); |
345 | 0 | r0 = _mm_packus_epi16(r0, r1); |
346 | |
|
347 | 0 | _mm_storeu_si128((__m128i *) (dst + x), r0); |
348 | 0 | } |
349 | 0 | dst += dststride; |
350 | 0 | src1 += srcstride; |
351 | 0 | src2 += srcstride; |
352 | 0 | } |
353 | 0 | } |
354 | | |
355 | | #if 0 |
356 | | void ff_hevc_weighted_pred_8_sse4(uint8_t denom, int16_t wlxFlag, int16_t olxFlag, |
357 | | uint8_t *_dst, ptrdiff_t _dststride, |
358 | | const int16_t *src, ptrdiff_t srcstride, |
359 | | int width, int height) { |
360 | | |
361 | | int log2Wd; |
362 | | int x, y; |
363 | | |
364 | | uint8_t *dst = (uint8_t*) _dst; |
365 | | ptrdiff_t dststride = _dststride / sizeof(uint8_t); |
366 | | __m128i x0, x1, x2, x3, c0, add, add2; |
367 | | |
368 | | log2Wd = denom + 14 - BIT_DEPTH; |
369 | | |
370 | | add = _mm_set1_epi32(olxFlag * (1 << (BIT_DEPTH - 8))); |
371 | | add2 = _mm_set1_epi32(1 << (log2Wd - 1)); |
372 | | c0 = _mm_set1_epi16(wlxFlag); |
373 | | if (log2Wd >= 1){ |
374 | | if(!(width & 15)){ |
375 | | for (y = 0; y < height; y++) { |
376 | | for (x = 0; x < width; x += 16) { |
377 | | x0 = _mm_load_si128((__m128i *) &src[x]); |
378 | | x2 = _mm_load_si128((__m128i *) &src[x + 8]); |
379 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
380 | | _mm_mulhi_epi16(x0, c0)); |
381 | | x3 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c0), |
382 | | _mm_mulhi_epi16(x2, c0)); |
383 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
384 | | _mm_mulhi_epi16(x0, c0)); |
385 | | x2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c0), |
386 | | _mm_mulhi_epi16(x2, c0)); |
387 | | x0 = _mm_add_epi32(x0, add2); |
388 | | x1 = _mm_add_epi32(x1, add2); |
389 | | x2 = _mm_add_epi32(x2, add2); |
390 | | x3 = _mm_add_epi32(x3, add2); |
391 | | x0 = _mm_srai_epi32(x0, log2Wd); |
392 | | x1 = _mm_srai_epi32(x1, log2Wd); |
393 | | x2 = _mm_srai_epi32(x2, log2Wd); |
394 | | x3 = _mm_srai_epi32(x3, log2Wd); |
395 | | x0 = _mm_add_epi32(x0, add); |
396 | | x1 = _mm_add_epi32(x1, add); |
397 | | x2 = _mm_add_epi32(x2, add); |
398 | | x3 = _mm_add_epi32(x3, add); |
399 | | x0 = _mm_packus_epi32(x0, x1); |
400 | | x2 = _mm_packus_epi32(x2, x3); |
401 | | x0 = _mm_packus_epi16(x0, x2); |
402 | | |
403 | | _mm_storeu_si128((__m128i *) (dst + x), x0); |
404 | | |
405 | | } |
406 | | dst += dststride; |
407 | | src += srcstride; |
408 | | } |
409 | | }else if(!(width & 7)){ |
410 | | for (y = 0; y < height; y++) { |
411 | | for(x=0;x<width;x+=8){ |
412 | | x0 = _mm_load_si128((__m128i *) (src+x)); |
413 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
414 | | _mm_mulhi_epi16(x0, c0)); |
415 | | |
416 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
417 | | _mm_mulhi_epi16(x0, c0)); |
418 | | |
419 | | x0 = _mm_add_epi32(x0, add2); |
420 | | x1 = _mm_add_epi32(x1, add2); |
421 | | |
422 | | x0 = _mm_srai_epi32(x0, log2Wd); |
423 | | x1 = _mm_srai_epi32(x1, log2Wd); |
424 | | |
425 | | x0 = _mm_add_epi32(x0, add); |
426 | | x1 = _mm_add_epi32(x1, add); |
427 | | |
428 | | x0 = _mm_packus_epi32(x0, x1); |
429 | | x0 = _mm_packus_epi16(x0, x0); |
430 | | |
431 | | _mm_storel_epi64((__m128i *) (dst+x), x0); |
432 | | |
433 | | } |
434 | | dst += dststride; |
435 | | src += srcstride; |
436 | | } |
437 | | }else if(!(width & 3)){ |
438 | | for (y = 0; y < height; y++) { |
439 | | for(x=0;x<width;x+=4){ |
440 | | x0 = _mm_loadl_epi64((__m128i *)(src+x)); |
441 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
442 | | _mm_mulhi_epi16(x0, c0)); |
443 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
444 | | _mm_mulhi_epi16(x0, c0)); |
445 | | |
446 | | x0 = _mm_add_epi32(x0, add2); |
447 | | x1 = _mm_add_epi32(x1, add2); |
448 | | x0 = _mm_srai_epi32(x0, log2Wd); |
449 | | x1 = _mm_srai_epi32(x1, log2Wd); |
450 | | x0 = _mm_add_epi32(x0, add); |
451 | | x1 = _mm_add_epi32(x1, add); |
452 | | x0 = _mm_packus_epi32(x0, x1); |
453 | | x0 = _mm_packus_epi16(x0, x0); |
454 | | |
455 | | _mm_maskmoveu_si128(x0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
456 | | // _mm_storeu_si128((__m128i *) (dst + x), x0); |
457 | | } |
458 | | dst += dststride; |
459 | | src += srcstride; |
460 | | } |
461 | | }else{ |
462 | | for (y = 0; y < height; y++) { |
463 | | for(x=0;x<width;x+=2){ |
464 | | x0 = _mm_loadl_epi64((__m128i *)(src+x)); |
465 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
466 | | _mm_mulhi_epi16(x0, c0)); |
467 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
468 | | _mm_mulhi_epi16(x0, c0)); |
469 | | |
470 | | x0 = _mm_add_epi32(x0, add2); |
471 | | x1 = _mm_add_epi32(x1, add2); |
472 | | x0 = _mm_srai_epi32(x0, log2Wd); |
473 | | x1 = _mm_srai_epi32(x1, log2Wd); |
474 | | x0 = _mm_add_epi32(x0, add); |
475 | | x1 = _mm_add_epi32(x1, add); |
476 | | x0 = _mm_packus_epi32(x0, x1); |
477 | | x0 = _mm_packus_epi16(x0, x0); |
478 | | |
479 | | _mm_maskmoveu_si128(x0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,-1),(char *) (dst+x)); |
480 | | // _mm_storeu_si128((__m128i *) (dst + x), x0); |
481 | | } |
482 | | dst += dststride; |
483 | | src += srcstride; |
484 | | } |
485 | | } |
486 | | }else{ |
487 | | if(!(width & 15)){ |
488 | | for (y = 0; y < height; y++) { |
489 | | for (x = 0; x < width; x += 16) { |
490 | | |
491 | | x0 = _mm_load_si128((__m128i *) &src[x]); |
492 | | x2 = _mm_load_si128((__m128i *) &src[x + 8]); |
493 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
494 | | _mm_mulhi_epi16(x0, c0)); |
495 | | x3 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c0), |
496 | | _mm_mulhi_epi16(x2, c0)); |
497 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
498 | | _mm_mulhi_epi16(x0, c0)); |
499 | | x2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c0), |
500 | | _mm_mulhi_epi16(x2, c0)); |
501 | | |
502 | | x0 = _mm_add_epi32(x0, add2); |
503 | | x1 = _mm_add_epi32(x1, add2); |
504 | | x2 = _mm_add_epi32(x2, add2); |
505 | | x3 = _mm_add_epi32(x3, add2); |
506 | | |
507 | | x0 = _mm_packus_epi32(x0, x1); |
508 | | x2 = _mm_packus_epi32(x2, x3); |
509 | | x0 = _mm_packus_epi16(x0, x2); |
510 | | |
511 | | _mm_storeu_si128((__m128i *) (dst + x), x0); |
512 | | |
513 | | } |
514 | | dst += dststride; |
515 | | src += srcstride; |
516 | | } |
517 | | }else if(!(width & 7)){ |
518 | | for (y = 0; y < height; y++) { |
519 | | for(x=0;x<width;x+=8){ |
520 | | x0 = _mm_load_si128((__m128i *) (src+x)); |
521 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
522 | | _mm_mulhi_epi16(x0, c0)); |
523 | | |
524 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
525 | | _mm_mulhi_epi16(x0, c0)); |
526 | | |
527 | | |
528 | | x0 = _mm_add_epi32(x0, add2); |
529 | | x1 = _mm_add_epi32(x1, add2); |
530 | | |
531 | | x0 = _mm_packus_epi32(x0, x1); |
532 | | x0 = _mm_packus_epi16(x0, x0); |
533 | | |
534 | | _mm_storeu_si128((__m128i *) (dst+x), x0); |
535 | | } |
536 | | |
537 | | dst += dststride; |
538 | | src += srcstride; |
539 | | } |
540 | | }else if(!(width & 3)){ |
541 | | for (y = 0; y < height; y++) { |
542 | | for(x=0;x<width;x+=4){ |
543 | | x0 = _mm_loadl_epi64((__m128i *) (src+x)); |
544 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
545 | | _mm_mulhi_epi16(x0, c0)); |
546 | | |
547 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
548 | | _mm_mulhi_epi16(x0, c0)); |
549 | | |
550 | | |
551 | | x0 = _mm_add_epi32(x0, add2); |
552 | | x1 = _mm_add_epi32(x1, add2); |
553 | | |
554 | | |
555 | | x0 = _mm_packus_epi32(x0, x1); |
556 | | x0 = _mm_packus_epi16(x0, x0); |
557 | | |
558 | | |
559 | | _mm_maskmoveu_si128(x0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
560 | | } |
561 | | dst += dststride; |
562 | | src += srcstride; |
563 | | } |
564 | | }else{ |
565 | | for (y = 0; y < height; y++) { |
566 | | for(x=0;x<width;x+=2){ |
567 | | x0 = _mm_loadl_epi64((__m128i *) (src+x)); |
568 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
569 | | _mm_mulhi_epi16(x0, c0)); |
570 | | |
571 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
572 | | _mm_mulhi_epi16(x0, c0)); |
573 | | |
574 | | |
575 | | x0 = _mm_add_epi32(x0, add2); |
576 | | x1 = _mm_add_epi32(x1, add2); |
577 | | |
578 | | |
579 | | x0 = _mm_packus_epi32(x0, x1); |
580 | | x0 = _mm_packus_epi16(x0, x0); |
581 | | |
582 | | |
583 | | _mm_maskmoveu_si128(x0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,-1),(char *) (dst+x)); |
584 | | } |
585 | | dst += dststride; |
586 | | src += srcstride; |
587 | | } |
588 | | |
589 | | } |
590 | | |
591 | | } |
592 | | |
593 | | } |
594 | | #endif |
595 | | |
596 | | |
597 | | #if 0 |
598 | | void ff_hevc_weighted_pred_sse(uint8_t denom, int16_t wlxFlag, int16_t olxFlag, |
599 | | uint8_t *_dst, ptrdiff_t _dststride, |
600 | | const int16_t *src, ptrdiff_t srcstride, |
601 | | int width, int height) { |
602 | | |
603 | | int log2Wd; |
604 | | int x, y; |
605 | | |
606 | | uint8_t *dst = (uint8_t*) _dst; |
607 | | ptrdiff_t dststride = _dststride / sizeof(uint8_t); |
608 | | __m128i x0, x1, x2, x3, c0, add, add2; |
609 | | |
610 | | log2Wd = denom + 14 - BIT_DEPTH; |
611 | | |
612 | | add = _mm_set1_epi32(olxFlag * (1 << (BIT_DEPTH - 8))); |
613 | | add2 = _mm_set1_epi32(1 << (log2Wd - 1)); |
614 | | c0 = _mm_set1_epi16(wlxFlag); |
615 | | if (log2Wd >= 1) |
616 | | for (y = 0; y < height; y++) { |
617 | | for (x = 0; x < width; x += 16) { |
618 | | x0 = _mm_load_si128((__m128i *) &src[x]); |
619 | | x2 = _mm_load_si128((__m128i *) &src[x + 8]); |
620 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
621 | | _mm_mulhi_epi16(x0, c0)); |
622 | | x3 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c0), |
623 | | _mm_mulhi_epi16(x2, c0)); |
624 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
625 | | _mm_mulhi_epi16(x0, c0)); |
626 | | x2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c0), |
627 | | _mm_mulhi_epi16(x2, c0)); |
628 | | x0 = _mm_add_epi32(x0, add2); |
629 | | x1 = _mm_add_epi32(x1, add2); |
630 | | x2 = _mm_add_epi32(x2, add2); |
631 | | x3 = _mm_add_epi32(x3, add2); |
632 | | x0 = _mm_srai_epi32(x0, log2Wd); |
633 | | x1 = _mm_srai_epi32(x1, log2Wd); |
634 | | x2 = _mm_srai_epi32(x2, log2Wd); |
635 | | x3 = _mm_srai_epi32(x3, log2Wd); |
636 | | x0 = _mm_add_epi32(x0, add); |
637 | | x1 = _mm_add_epi32(x1, add); |
638 | | x2 = _mm_add_epi32(x2, add); |
639 | | x3 = _mm_add_epi32(x3, add); |
640 | | x0 = _mm_packus_epi32(x0, x1); |
641 | | x2 = _mm_packus_epi32(x2, x3); |
642 | | x0 = _mm_packus_epi16(x0, x2); |
643 | | |
644 | | _mm_storeu_si128((__m128i *) (dst + x), x0); |
645 | | |
646 | | } |
647 | | dst += dststride; |
648 | | src += srcstride; |
649 | | } |
650 | | else |
651 | | for (y = 0; y < height; y++) { |
652 | | for (x = 0; x < width; x += 16) { |
653 | | |
654 | | x0 = _mm_load_si128((__m128i *) &src[x]); |
655 | | x2 = _mm_load_si128((__m128i *) &src[x + 8]); |
656 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
657 | | _mm_mulhi_epi16(x0, c0)); |
658 | | x3 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c0), |
659 | | _mm_mulhi_epi16(x2, c0)); |
660 | | x0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
661 | | _mm_mulhi_epi16(x0, c0)); |
662 | | x2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c0), |
663 | | _mm_mulhi_epi16(x2, c0)); |
664 | | |
665 | | x0 = _mm_add_epi32(x0, add2); |
666 | | x1 = _mm_add_epi32(x1, add2); |
667 | | x2 = _mm_add_epi32(x2, add2); |
668 | | x3 = _mm_add_epi32(x3, add2); |
669 | | |
670 | | x0 = _mm_packus_epi32(x0, x1); |
671 | | x2 = _mm_packus_epi32(x2, x3); |
672 | | x0 = _mm_packus_epi16(x0, x2); |
673 | | |
674 | | _mm_storeu_si128((__m128i *) (dst + x), x0); |
675 | | |
676 | | } |
677 | | dst += dststride; |
678 | | src += srcstride; |
679 | | } |
680 | | } |
681 | | #endif |
682 | | |
683 | | #if HAVE_SSE4_1 |
684 | | void ff_hevc_weighted_pred_avg_8_sse4(uint8_t denom, int16_t wl0Flag, |
685 | | int16_t wl1Flag, int16_t ol0Flag, int16_t ol1Flag, |
686 | | uint8_t *_dst, ptrdiff_t _dststride, |
687 | | const int16_t *src1, const int16_t *src2, ptrdiff_t srcstride, |
688 | 0 | int width, int height) { |
689 | 0 | int shift, shift2; |
690 | 0 | int log2Wd; |
691 | 0 | int o0; |
692 | 0 | int o1; |
693 | 0 | int x, y; |
694 | 0 | uint8_t *dst = (uint8_t*) _dst; |
695 | 0 | ptrdiff_t dststride = _dststride / sizeof(uint8_t); |
696 | 0 | __m128i x0, x1, x2, x3, r0, r1, r2, r3, c0, c1, c2; |
697 | 0 | shift = 14 - BIT_DEPTH; |
698 | 0 | log2Wd = denom + shift; |
699 | |
|
700 | 0 | o0 = (ol0Flag) * (1 << (BIT_DEPTH - 8)); |
701 | 0 | o1 = (ol1Flag) * (1 << (BIT_DEPTH - 8)); |
702 | 0 | shift2 = (log2Wd + 1); |
703 | 0 | c0 = _mm_set1_epi16(wl0Flag); |
704 | 0 | c1 = _mm_set1_epi16(wl1Flag); |
705 | 0 | c2 = _mm_set1_epi32((o0 + o1 + 1) << log2Wd); |
706 | |
|
707 | 0 | if(!(width & 15)){ |
708 | 0 | for (y = 0; y < height; y++) { |
709 | 0 | for (x = 0; x < width; x += 16) { |
710 | 0 | x0 = _mm_load_si128((__m128i *) &src1[x]); |
711 | 0 | x1 = _mm_load_si128((__m128i *) &src1[x + 8]); |
712 | 0 | x2 = _mm_load_si128((__m128i *) &src2[x]); |
713 | 0 | x3 = _mm_load_si128((__m128i *) &src2[x + 8]); |
714 | |
|
715 | 0 | r0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
716 | 0 | _mm_mulhi_epi16(x0, c0)); |
717 | 0 | r1 = _mm_unpacklo_epi16(_mm_mullo_epi16(x1, c0), |
718 | 0 | _mm_mulhi_epi16(x1, c0)); |
719 | 0 | r2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c1), |
720 | 0 | _mm_mulhi_epi16(x2, c1)); |
721 | 0 | r3 = _mm_unpacklo_epi16(_mm_mullo_epi16(x3, c1), |
722 | 0 | _mm_mulhi_epi16(x3, c1)); |
723 | 0 | x0 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
724 | 0 | _mm_mulhi_epi16(x0, c0)); |
725 | 0 | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x1, c0), |
726 | 0 | _mm_mulhi_epi16(x1, c0)); |
727 | 0 | x2 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c1), |
728 | 0 | _mm_mulhi_epi16(x2, c1)); |
729 | 0 | x3 = _mm_unpackhi_epi16(_mm_mullo_epi16(x3, c1), |
730 | 0 | _mm_mulhi_epi16(x3, c1)); |
731 | 0 | r0 = _mm_add_epi32(r0, r2); |
732 | 0 | r1 = _mm_add_epi32(r1, r3); |
733 | 0 | r2 = _mm_add_epi32(x0, x2); |
734 | 0 | r3 = _mm_add_epi32(x1, x3); |
735 | |
|
736 | 0 | r0 = _mm_add_epi32(r0, c2); |
737 | 0 | r1 = _mm_add_epi32(r1, c2); |
738 | 0 | r2 = _mm_add_epi32(r2, c2); |
739 | 0 | r3 = _mm_add_epi32(r3, c2); |
740 | |
|
741 | 0 | r0 = _mm_srai_epi32(r0, shift2); |
742 | 0 | r1 = _mm_srai_epi32(r1, shift2); |
743 | 0 | r2 = _mm_srai_epi32(r2, shift2); |
744 | 0 | r3 = _mm_srai_epi32(r3, shift2); |
745 | |
|
746 | 0 | r0 = _mm_packus_epi32(r0, r2); |
747 | 0 | r1 = _mm_packus_epi32(r1, r3); |
748 | 0 | r0 = _mm_packus_epi16(r0, r1); |
749 | |
|
750 | 0 | _mm_storeu_si128((__m128i *) (dst + x), r0); |
751 | |
|
752 | 0 | } |
753 | 0 | dst += dststride; |
754 | 0 | src1 += srcstride; |
755 | 0 | src2 += srcstride; |
756 | 0 | } |
757 | 0 | }else if(!(width & 7)){ |
758 | 0 | for (y = 0; y < height; y++) { |
759 | 0 | for(x=0;x<width;x+=8){ |
760 | 0 | x0 = _mm_load_si128((__m128i *) (src1+x)); |
761 | 0 | x2 = _mm_load_si128((__m128i *) (src2+x)); |
762 | |
|
763 | 0 | r0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
764 | 0 | _mm_mulhi_epi16(x0, c0)); |
765 | |
|
766 | 0 | r2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c1), |
767 | 0 | _mm_mulhi_epi16(x2, c1)); |
768 | |
|
769 | 0 | x0 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
770 | 0 | _mm_mulhi_epi16(x0, c0)); |
771 | |
|
772 | 0 | x2 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c1), |
773 | 0 | _mm_mulhi_epi16(x2, c1)); |
774 | |
|
775 | 0 | r0 = _mm_add_epi32(r0, r2); |
776 | 0 | r2 = _mm_add_epi32(x0, x2); |
777 | | |
778 | |
|
779 | 0 | r0 = _mm_add_epi32(r0, c2); |
780 | 0 | r2 = _mm_add_epi32(r2, c2); |
781 | |
|
782 | 0 | r0 = _mm_srai_epi32(r0, shift2); |
783 | 0 | r2 = _mm_srai_epi32(r2, shift2); |
784 | |
|
785 | 0 | r0 = _mm_packus_epi32(r0, r2); |
786 | 0 | r0 = _mm_packus_epi16(r0, r0); |
787 | |
|
788 | 0 | _mm_storel_epi64((__m128i *) (dst+x), r0); |
789 | 0 | } |
790 | |
|
791 | 0 | dst += dststride; |
792 | 0 | src1 += srcstride; |
793 | 0 | src2 += srcstride; |
794 | 0 | } |
795 | 0 | }else if(!(width & 3)){ |
796 | 0 | for (y = 0; y < height; y++) { |
797 | 0 | for(x=0;x<width;x+=4){ |
798 | 0 | x0 = _mm_loadl_epi64((__m128i *) (src1+x)); |
799 | 0 | x2 = _mm_loadl_epi64((__m128i *) (src2+x)); |
800 | |
|
801 | 0 | r0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
802 | 0 | _mm_mulhi_epi16(x0, c0)); |
803 | |
|
804 | 0 | r2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c1), |
805 | 0 | _mm_mulhi_epi16(x2, c1)); |
806 | |
|
807 | 0 | x0 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
808 | 0 | _mm_mulhi_epi16(x0, c0)); |
809 | |
|
810 | 0 | x2 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c1), |
811 | 0 | _mm_mulhi_epi16(x2, c1)); |
812 | |
|
813 | 0 | r0 = _mm_add_epi32(r0, r2); |
814 | 0 | r2 = _mm_add_epi32(x0, x2); |
815 | |
|
816 | 0 | r0 = _mm_add_epi32(r0, c2); |
817 | 0 | r2 = _mm_add_epi32(r2, c2); |
818 | |
|
819 | 0 | r0 = _mm_srai_epi32(r0, shift2); |
820 | 0 | r2 = _mm_srai_epi32(r2, shift2); |
821 | |
|
822 | 0 | r0 = _mm_packus_epi32(r0, r2); |
823 | 0 | r0 = _mm_packus_epi16(r0, r0); |
824 | |
|
825 | | #if MASKMOVE |
826 | | _mm_maskmoveu_si128(r0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
827 | | #else |
828 | 0 | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
829 | 0 | #endif |
830 | 0 | } |
831 | 0 | dst += dststride; |
832 | 0 | src1 += srcstride; |
833 | 0 | src2 += srcstride; |
834 | 0 | } |
835 | 0 | }else{ |
836 | 0 | for (y = 0; y < height; y++) { |
837 | 0 | for(x=0;x<width;x+=2){ |
838 | 0 | x0 = _mm_loadl_epi64((__m128i *) (src1+x)); |
839 | 0 | x2 = _mm_loadl_epi64((__m128i *) (src2+x)); |
840 | |
|
841 | 0 | r0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
842 | 0 | _mm_mulhi_epi16(x0, c0)); |
843 | |
|
844 | 0 | r2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c1), |
845 | 0 | _mm_mulhi_epi16(x2, c1)); |
846 | |
|
847 | 0 | x0 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
848 | 0 | _mm_mulhi_epi16(x0, c0)); |
849 | |
|
850 | 0 | x2 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c1), |
851 | 0 | _mm_mulhi_epi16(x2, c1)); |
852 | |
|
853 | 0 | r0 = _mm_add_epi32(r0, r2); |
854 | 0 | r2 = _mm_add_epi32(x0, x2); |
855 | |
|
856 | 0 | r0 = _mm_add_epi32(r0, c2); |
857 | 0 | r2 = _mm_add_epi32(r2, c2); |
858 | |
|
859 | 0 | r0 = _mm_srai_epi32(r0, shift2); |
860 | 0 | r2 = _mm_srai_epi32(r2, shift2); |
861 | |
|
862 | 0 | r0 = _mm_packus_epi32(r0, r2); |
863 | 0 | r0 = _mm_packus_epi16(r0, r0); |
864 | |
|
865 | | #if MASKMOVE |
866 | | _mm_maskmoveu_si128(r0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,-1),(char *) (dst+x)); |
867 | | #else |
868 | 0 | *((uint16_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
869 | 0 | #endif |
870 | 0 | } |
871 | 0 | dst += dststride; |
872 | 0 | src1 += srcstride; |
873 | 0 | src2 += srcstride; |
874 | 0 | } |
875 | 0 | } |
876 | 0 | } |
877 | | #endif |
878 | | |
879 | | |
880 | | #if 0 |
881 | | void ff_hevc_weighted_pred_avg_sse(uint8_t denom, int16_t wl0Flag, |
882 | | int16_t wl1Flag, int16_t ol0Flag, int16_t ol1Flag, uint8_t *_dst, |
883 | | ptrdiff_t _dststride, const int16_t *src1, const int16_t *src2, ptrdiff_t srcstride, |
884 | | int width, int height) { |
885 | | int shift, shift2; |
886 | | int log2Wd; |
887 | | int o0; |
888 | | int o1; |
889 | | int x, y; |
890 | | uint8_t *dst = (uint8_t*) _dst; |
891 | | ptrdiff_t dststride = _dststride / sizeof(uint8_t); |
892 | | __m128i x0, x1, x2, x3, r0, r1, r2, r3, c0, c1, c2; |
893 | | shift = 14 - BIT_DEPTH; |
894 | | log2Wd = denom + shift; |
895 | | |
896 | | o0 = (ol0Flag) * (1 << (BIT_DEPTH - 8)); |
897 | | o1 = (ol1Flag) * (1 << (BIT_DEPTH - 8)); |
898 | | shift2 = (log2Wd + 1); |
899 | | c0 = _mm_set1_epi16(wl0Flag); |
900 | | c1 = _mm_set1_epi16(wl1Flag); |
901 | | c2 = _mm_set1_epi32((o0 + o1 + 1) << log2Wd); |
902 | | |
903 | | for (y = 0; y < height; y++) { |
904 | | for (x = 0; x < width; x += 16) { |
905 | | x0 = _mm_load_si128((__m128i *) &src1[x]); |
906 | | x1 = _mm_load_si128((__m128i *) &src1[x + 8]); |
907 | | x2 = _mm_load_si128((__m128i *) &src2[x]); |
908 | | x3 = _mm_load_si128((__m128i *) &src2[x + 8]); |
909 | | |
910 | | r0 = _mm_unpacklo_epi16(_mm_mullo_epi16(x0, c0), |
911 | | _mm_mulhi_epi16(x0, c0)); |
912 | | r1 = _mm_unpacklo_epi16(_mm_mullo_epi16(x1, c0), |
913 | | _mm_mulhi_epi16(x1, c0)); |
914 | | r2 = _mm_unpacklo_epi16(_mm_mullo_epi16(x2, c1), |
915 | | _mm_mulhi_epi16(x2, c1)); |
916 | | r3 = _mm_unpacklo_epi16(_mm_mullo_epi16(x3, c1), |
917 | | _mm_mulhi_epi16(x3, c1)); |
918 | | x0 = _mm_unpackhi_epi16(_mm_mullo_epi16(x0, c0), |
919 | | _mm_mulhi_epi16(x0, c0)); |
920 | | x1 = _mm_unpackhi_epi16(_mm_mullo_epi16(x1, c0), |
921 | | _mm_mulhi_epi16(x1, c0)); |
922 | | x2 = _mm_unpackhi_epi16(_mm_mullo_epi16(x2, c1), |
923 | | _mm_mulhi_epi16(x2, c1)); |
924 | | x3 = _mm_unpackhi_epi16(_mm_mullo_epi16(x3, c1), |
925 | | _mm_mulhi_epi16(x3, c1)); |
926 | | r0 = _mm_add_epi32(r0, r2); |
927 | | r1 = _mm_add_epi32(r1, r3); |
928 | | r2 = _mm_add_epi32(x0, x2); |
929 | | r3 = _mm_add_epi32(x1, x3); |
930 | | |
931 | | r0 = _mm_add_epi32(r0, c2); |
932 | | r1 = _mm_add_epi32(r1, c2); |
933 | | r2 = _mm_add_epi32(r2, c2); |
934 | | r3 = _mm_add_epi32(r3, c2); |
935 | | |
936 | | r0 = _mm_srai_epi32(r0, shift2); |
937 | | r1 = _mm_srai_epi32(r1, shift2); |
938 | | r2 = _mm_srai_epi32(r2, shift2); |
939 | | r3 = _mm_srai_epi32(r3, shift2); |
940 | | |
941 | | r0 = _mm_packus_epi32(r0, r2); |
942 | | r1 = _mm_packus_epi32(r1, r3); |
943 | | r0 = _mm_packus_epi16(r0, r1); |
944 | | |
945 | | _mm_storeu_si128((__m128i *) (dst + x), r0); |
946 | | |
947 | | } |
948 | | dst += dststride; |
949 | | src1 += srcstride; |
950 | | src2 += srcstride; |
951 | | } |
952 | | } |
953 | | #endif |
954 | | |
955 | | |
956 | | void ff_hevc_put_hevc_epel_pixels_8_sse(int16_t *dst, ptrdiff_t dststride, |
957 | | const uint8_t *_src, ptrdiff_t srcstride, |
958 | | int width, int height, int mx, |
959 | 6.64M | int my, int16_t* mcbuffer) { |
960 | 6.64M | int x, y; |
961 | 6.64M | __m128i x1, x2,x3; |
962 | 6.64M | uint8_t *src = (uint8_t*) _src; |
963 | 6.64M | if(!(width & 15)){ |
964 | 263k | x3= _mm_setzero_si128(); |
965 | 4.36M | for (y = 0; y < height; y++) { |
966 | 8.97M | for (x = 0; x < width; x += 16) { |
967 | | |
968 | 4.86M | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
969 | 4.86M | x2 = _mm_unpacklo_epi8(x1, x3); |
970 | | |
971 | 4.86M | x1 = _mm_unpackhi_epi8(x1, x3); |
972 | | |
973 | 4.86M | x2 = _mm_slli_epi16(x2, 6); |
974 | 4.86M | x1 = _mm_slli_epi16(x1, 6); |
975 | 4.86M | _mm_store_si128((__m128i *) &dst[x], x2); |
976 | 4.86M | _mm_store_si128((__m128i *) &dst[x + 8], x1); |
977 | | |
978 | 4.86M | } |
979 | 4.10M | src += srcstride; |
980 | 4.10M | dst += dststride; |
981 | 4.10M | } |
982 | 6.38M | }else if(!(width & 7)){ |
983 | 876k | x1= _mm_setzero_si128(); |
984 | 8.67M | for (y = 0; y < height; y++) { |
985 | 15.7M | for (x = 0; x < width; x += 8) { |
986 | | |
987 | 7.99M | x2 = _mm_loadl_epi64((__m128i *) &src[x]); |
988 | 7.99M | x2 = _mm_unpacklo_epi8(x2, x1); |
989 | 7.99M | x2 = _mm_slli_epi16(x2, 6); |
990 | 7.99M | _mm_store_si128((__m128i *) &dst[x], x2); |
991 | | |
992 | 7.99M | } |
993 | 7.79M | src += srcstride; |
994 | 7.79M | dst += dststride; |
995 | 7.79M | } |
996 | 5.50M | }else if(!(width & 3)){ |
997 | 4.89M | x1= _mm_setzero_si128(); |
998 | 26.1M | for (y = 0; y < height; y++) { |
999 | 42.9M | for (x = 0; x < width; x += 4) { |
1000 | | |
1001 | 21.6M | x2 = _mm_loadl_epi64((__m128i *) &src[x]); |
1002 | 21.6M | x2 = _mm_unpacklo_epi8(x2,x1); |
1003 | | |
1004 | 21.6M | x2 = _mm_slli_epi16(x2, 6); |
1005 | | |
1006 | 21.6M | _mm_storel_epi64((__m128i *) &dst[x], x2); |
1007 | | |
1008 | 21.6M | } |
1009 | 21.3M | src += srcstride; |
1010 | 21.3M | dst += dststride; |
1011 | 21.3M | } |
1012 | 4.89M | }else{ |
1013 | 614k | x1= _mm_setzero_si128(); |
1014 | 3.95M | for (y = 0; y < height; y++) { |
1015 | 7.16M | for (x = 0; x < width; x += 2) { |
1016 | | |
1017 | 3.82M | x2 = _mm_loadl_epi64((__m128i *) &src[x]); |
1018 | 3.82M | x2 = _mm_unpacklo_epi8(x2, x1); |
1019 | 3.82M | x2 = _mm_slli_epi16(x2, 6); |
1020 | | #if MASKMOVE |
1021 | | _mm_maskmoveu_si128(x2,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
1022 | | #else |
1023 | 3.82M | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(x2); |
1024 | 3.82M | #endif |
1025 | 3.82M | } |
1026 | 3.34M | src += srcstride; |
1027 | 3.34M | dst += dststride; |
1028 | 3.34M | } |
1029 | 614k | } |
1030 | | |
1031 | 6.64M | } |
1032 | | |
1033 | | #ifndef __native_client__ |
1034 | | void ff_hevc_put_hevc_epel_pixels_10_sse(int16_t *dst, ptrdiff_t dststride, |
1035 | | const uint8_t *_src, ptrdiff_t _srcstride, |
1036 | | int width, int height, int mx, |
1037 | 0 | int my, int16_t* mcbuffer) { |
1038 | 0 | int x, y; |
1039 | 0 | __m128i x2; |
1040 | 0 | uint16_t *src = (uint16_t*) _src; |
1041 | 0 | ptrdiff_t srcstride = _srcstride>>1; |
1042 | 0 | if(!(width & 7)){ |
1043 | | //x1= _mm_setzero_si128(); |
1044 | 0 | for (y = 0; y < height; y++) { |
1045 | 0 | for (x = 0; x < width; x += 8) { |
1046 | |
|
1047 | 0 | x2 = _mm_loadu_si128((__m128i *) &src[x]); |
1048 | 0 | x2 = _mm_slli_epi16(x2, 4); //shift 14 - BIT LENGTH |
1049 | 0 | _mm_store_si128((__m128i *) &dst[x], x2); |
1050 | |
|
1051 | 0 | } |
1052 | 0 | src += srcstride; |
1053 | 0 | dst += dststride; |
1054 | 0 | } |
1055 | 0 | }else if(!(width & 3)){ |
1056 | | //x1= _mm_setzero_si128(); |
1057 | 0 | for (y = 0; y < height; y++) { |
1058 | 0 | for (x = 0; x < width; x += 4) { |
1059 | |
|
1060 | 0 | x2 = _mm_loadl_epi64((__m128i *) &src[x]); |
1061 | 0 | x2 = _mm_slli_epi16(x2, 4); //shift 14 - BIT LENGTH |
1062 | |
|
1063 | 0 | _mm_storel_epi64((__m128i *) &dst[x], x2); |
1064 | |
|
1065 | 0 | } |
1066 | 0 | src += srcstride; |
1067 | 0 | dst += dststride; |
1068 | 0 | } |
1069 | 0 | }else{ |
1070 | | //x1= _mm_setzero_si128(); |
1071 | 0 | for (y = 0; y < height; y++) { |
1072 | 0 | for (x = 0; x < width; x += 2) { |
1073 | |
|
1074 | 0 | x2 = _mm_loadl_epi64((__m128i *) &src[x]); |
1075 | 0 | x2 = _mm_slli_epi16(x2, 4); //shift 14 - BIT LENGTH |
1076 | 0 | _mm_maskmoveu_si128(x2,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
1077 | 0 | } |
1078 | 0 | src += srcstride; |
1079 | 0 | dst += dststride; |
1080 | 0 | } |
1081 | 0 | } |
1082 | |
|
1083 | 0 | } |
1084 | | #endif |
1085 | | |
1086 | | void ff_hevc_put_hevc_epel_h_8_sse(int16_t *dst, ptrdiff_t dststride, |
1087 | | const uint8_t *_src, ptrdiff_t _srcstride, |
1088 | | int width, int height, int mx, |
1089 | 1.44M | int my, int16_t* mcbuffer, int bit_depth) { |
1090 | 1.44M | int x, y; |
1091 | 1.44M | const uint8_t *src = (const uint8_t*) _src; |
1092 | 1.44M | ptrdiff_t srcstride = _srcstride; |
1093 | 1.44M | const int8_t *filter = epel_filters[mx - 1]; |
1094 | 1.44M | __m128i r0, bshuffle1, bshuffle2, x1, x2, x3; |
1095 | 1.44M | int8_t filter_0 = filter[0]; |
1096 | 1.44M | int8_t filter_1 = filter[1]; |
1097 | 1.44M | int8_t filter_2 = filter[2]; |
1098 | 1.44M | int8_t filter_3 = filter[3]; |
1099 | 1.44M | r0 = _mm_set_epi8(filter_3, filter_2, filter_1, filter_0, filter_3, |
1100 | 1.44M | filter_2, filter_1, filter_0, filter_3, filter_2, filter_1, |
1101 | 1.44M | filter_0, filter_3, filter_2, filter_1, filter_0); |
1102 | 1.44M | bshuffle1 = _mm_set_epi8(6, 5, 4, 3, 5, 4, 3, 2, 4, 3, 2, 1, 3, 2, 1, 0); |
1103 | | |
1104 | | |
1105 | | /* |
1106 | | printf("---IN---SSE\n"); |
1107 | | |
1108 | | int extra_top = 1; |
1109 | | int extra_left = 1; |
1110 | | int extra_right = 2; |
1111 | | int extra_bottom = 2; |
1112 | | |
1113 | | for (int y=-extra_top;y<height+extra_bottom;y++) { |
1114 | | uint8_t* p = &_src[y*_srcstride -extra_left]; |
1115 | | |
1116 | | for (int x=-extra_left;x<width+extra_right;x++) { |
1117 | | printf("%05d ",*p << 6); |
1118 | | p++; |
1119 | | } |
1120 | | printf("\n"); |
1121 | | } |
1122 | | */ |
1123 | | |
1124 | 1.44M | if(!(width & 7)){ |
1125 | 249k | bshuffle2 = _mm_set_epi8(10, 9, 8, 7, 9, 8, 7, 6, 8, 7, 6, 5, 7, 6, 5, |
1126 | 249k | 4); |
1127 | 2.75M | for (y = 0; y < height; y++) { |
1128 | 6.13M | for (x = 0; x < width; x += 8) { |
1129 | | |
1130 | 3.62M | x1 = _mm_loadu_si128((__m128i *) &src[x - 1]); |
1131 | 3.62M | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1132 | 3.62M | x3 = _mm_shuffle_epi8(x1, bshuffle2); |
1133 | | |
1134 | | /* PMADDUBSW then PMADDW */ |
1135 | 3.62M | x2 = _mm_maddubs_epi16(x2, r0); |
1136 | 3.62M | x3 = _mm_maddubs_epi16(x3, r0); |
1137 | 3.62M | x2 = _mm_hadd_epi16(x2, x3); |
1138 | 3.62M | _mm_store_si128((__m128i *) &dst[x], x2); |
1139 | 3.62M | } |
1140 | 2.50M | src += srcstride; |
1141 | 2.50M | dst += dststride; |
1142 | 2.50M | } |
1143 | 1.19M | }else if(!(width & 3)){ |
1144 | | |
1145 | 5.39M | for (y = 0; y < height; y++) { |
1146 | 8.90M | for (x = 0; x < width; x += 4) { |
1147 | | /* load data in register */ |
1148 | 4.48M | x1 = _mm_loadu_si128((__m128i *) &src[x-1]); |
1149 | 4.48M | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1150 | | |
1151 | | /* PMADDUBSW then PMADDW */ |
1152 | 4.48M | x2 = _mm_maddubs_epi16(x2, r0); |
1153 | 4.48M | x2 = _mm_hadd_epi16(x2, _mm_setzero_si128()); |
1154 | | /* give results back */ |
1155 | 4.48M | _mm_storel_epi64((__m128i *) &dst[x], x2); |
1156 | 4.48M | } |
1157 | 4.42M | src += srcstride; |
1158 | 4.42M | dst += dststride; |
1159 | 4.42M | } |
1160 | 965k | }else{ |
1161 | 1.57M | for (y = 0; y < height; y++) { |
1162 | 2.80M | for (x = 0; x < width; x += 2) { |
1163 | | /* load data in register */ |
1164 | 1.46M | x1 = _mm_loadu_si128((__m128i *) &src[x-1]); |
1165 | 1.46M | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1166 | | |
1167 | | /* PMADDUBSW then PMADDW */ |
1168 | 1.46M | x2 = _mm_maddubs_epi16(x2, r0); |
1169 | 1.46M | x2 = _mm_hadd_epi16(x2, _mm_setzero_si128()); |
1170 | | /* give results back */ |
1171 | | #if MASKMOVE |
1172 | | _mm_maskmoveu_si128(x2,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
1173 | | #else |
1174 | 1.46M | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(x2); |
1175 | 1.46M | #endif |
1176 | 1.46M | } |
1177 | 1.33M | src += srcstride; |
1178 | 1.33M | dst += dststride; |
1179 | 1.33M | } |
1180 | 234k | } |
1181 | 1.44M | } |
1182 | | |
1183 | | #ifndef __native_client__ |
1184 | | void ff_hevc_put_hevc_epel_h_10_sse(int16_t *dst, ptrdiff_t dststride, |
1185 | | const uint8_t *_src, ptrdiff_t _srcstride, |
1186 | | int width, int height, int mx, |
1187 | 0 | int my, int16_t* mcbuffer) { |
1188 | 0 | int x, y; |
1189 | 0 | uint16_t *src = (uint16_t*) _src; |
1190 | 0 | ptrdiff_t srcstride = _srcstride>>1; |
1191 | 0 | const int8_t *filter = epel_filters[mx - 1]; |
1192 | 0 | __m128i r0, bshuffle1, bshuffle2, x1, x2, x3, r1; |
1193 | 0 | int8_t filter_0 = filter[0]; |
1194 | 0 | int8_t filter_1 = filter[1]; |
1195 | 0 | int8_t filter_2 = filter[2]; |
1196 | 0 | int8_t filter_3 = filter[3]; |
1197 | 0 | r0 = _mm_set_epi16(filter_3, filter_2, filter_1, |
1198 | 0 | filter_0, filter_3, filter_2, filter_1, filter_0); |
1199 | 0 | bshuffle1 = _mm_set_epi8(9,8,7,6,5,4, 3, 2,7,6,5,4, 3, 2, 1, 0); |
1200 | |
|
1201 | 0 | if(!(width & 3)){ |
1202 | 0 | bshuffle2 = _mm_set_epi8(13,12,11,10,9,8,7,6,11,10, 9,8,7,6,5, 4); |
1203 | 0 | for (y = 0; y < height; y++) { |
1204 | 0 | for (x = 0; x < width; x += 4) { |
1205 | |
|
1206 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x-1]); |
1207 | 0 | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1208 | 0 | x3 = _mm_shuffle_epi8(x1, bshuffle2); |
1209 | | |
1210 | |
|
1211 | 0 | x2 = _mm_madd_epi16(x2, r0); |
1212 | 0 | x3 = _mm_madd_epi16(x3, r0); |
1213 | 0 | x2 = _mm_hadd_epi32(x2, x3); |
1214 | 0 | x2= _mm_srai_epi32(x2,2); //>> (BIT_DEPTH - 8) |
1215 | |
|
1216 | 0 | x2 = _mm_packs_epi32(x2,r0); |
1217 | | //give results back |
1218 | 0 | _mm_storel_epi64((__m128i *) &dst[x], x2); |
1219 | 0 | } |
1220 | 0 | src += srcstride; |
1221 | 0 | dst += dststride; |
1222 | 0 | } |
1223 | 0 | }else{ |
1224 | 0 | r1= _mm_setzero_si128(); |
1225 | 0 | for (y = 0; y < height; y++) { |
1226 | 0 | for (x = 0; x < width; x += 2) { |
1227 | | /* load data in register */ |
1228 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x-1]); |
1229 | 0 | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1230 | | |
1231 | | /* PMADDUBSW then PMADDW */ |
1232 | 0 | x2 = _mm_madd_epi16(x2, r0); |
1233 | 0 | x2 = _mm_hadd_epi32(x2, r1); |
1234 | 0 | x2= _mm_srai_epi32(x2,2); //>> (BIT_DEPTH - 8) |
1235 | 0 | x2 = _mm_packs_epi32(x2, r1); |
1236 | | /* give results back */ |
1237 | 0 | _mm_maskmoveu_si128(x2,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
1238 | 0 | } |
1239 | 0 | src += srcstride; |
1240 | 0 | dst += dststride; |
1241 | 0 | } |
1242 | 0 | } |
1243 | 0 | } |
1244 | | #endif |
1245 | | |
1246 | | |
1247 | | void ff_hevc_put_hevc_epel_v_8_sse(int16_t *dst, ptrdiff_t dststride, |
1248 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, int mx, |
1249 | 1.56M | int my, int16_t* mcbuffer, int bit_depth) { |
1250 | 1.56M | int x, y; |
1251 | 1.56M | __m128i x0, x1, x2, x3, t0, t1, t2, t3, r0, f0, f1, f2, f3, r1; |
1252 | 1.56M | uint8_t *src = (uint8_t*) _src; |
1253 | 1.56M | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
1254 | 1.56M | const int8_t *filter = epel_filters[my - 1]; |
1255 | 1.56M | int8_t filter_0 = filter[0]; |
1256 | 1.56M | int8_t filter_1 = filter[1]; |
1257 | 1.56M | int8_t filter_2 = filter[2]; |
1258 | 1.56M | int8_t filter_3 = filter[3]; |
1259 | 1.56M | f0 = _mm_set1_epi16(filter_0); |
1260 | 1.56M | f1 = _mm_set1_epi16(filter_1); |
1261 | 1.56M | f2 = _mm_set1_epi16(filter_2); |
1262 | 1.56M | f3 = _mm_set1_epi16(filter_3); |
1263 | | |
1264 | 1.56M | if(!(width & 15)){ |
1265 | 1.04M | for (y = 0; y < height; y++) { |
1266 | 2.05M | for (x = 0; x < width; x += 16) { |
1267 | | /* check if memory needs to be reloaded */ |
1268 | | |
1269 | 1.07M | x0 = _mm_loadu_si128((__m128i *) &src[x - srcstride]); |
1270 | 1.07M | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
1271 | 1.07M | x2 = _mm_loadu_si128((__m128i *) &src[x + srcstride]); |
1272 | 1.07M | x3 = _mm_loadu_si128((__m128i *) &src[x + 2 * srcstride]); |
1273 | | |
1274 | 1.07M | t0 = _mm_unpacklo_epi8(x0, _mm_setzero_si128()); |
1275 | 1.07M | t1 = _mm_unpacklo_epi8(x1, _mm_setzero_si128()); |
1276 | 1.07M | t2 = _mm_unpacklo_epi8(x2, _mm_setzero_si128()); |
1277 | 1.07M | t3 = _mm_unpacklo_epi8(x3, _mm_setzero_si128()); |
1278 | | |
1279 | 1.07M | x0 = _mm_unpackhi_epi8(x0, _mm_setzero_si128()); |
1280 | 1.07M | x1 = _mm_unpackhi_epi8(x1, _mm_setzero_si128()); |
1281 | 1.07M | x2 = _mm_unpackhi_epi8(x2, _mm_setzero_si128()); |
1282 | 1.07M | x3 = _mm_unpackhi_epi8(x3, _mm_setzero_si128()); |
1283 | | |
1284 | | /* multiply by correct value : */ |
1285 | 1.07M | r0 = _mm_mullo_epi16(t0, f0); |
1286 | 1.07M | r1 = _mm_mullo_epi16(x0, f0); |
1287 | 1.07M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t1, f1)); |
1288 | 1.07M | r1 = _mm_adds_epi16(r1, _mm_mullo_epi16(x1, f1)); |
1289 | 1.07M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t2, f2)); |
1290 | 1.07M | r1 = _mm_adds_epi16(r1, _mm_mullo_epi16(x2, f2)); |
1291 | 1.07M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t3, f3)); |
1292 | 1.07M | r1 = _mm_adds_epi16(r1, _mm_mullo_epi16(x3, f3)); |
1293 | | /* give results back */ |
1294 | 1.07M | _mm_store_si128((__m128i *) &dst[x], r0); |
1295 | 1.07M | _mm_storeu_si128((__m128i *) &dst[x + 8], r1); |
1296 | 1.07M | } |
1297 | 975k | src += srcstride; |
1298 | 975k | dst += dststride; |
1299 | 975k | } |
1300 | 1.49M | }else if(!(width & 7)){ |
1301 | 185k | r1= _mm_setzero_si128(); |
1302 | 1.72M | for (y = 0; y < height; y++) { |
1303 | 3.12M | for(x=0;x<width;x+=8){ |
1304 | 1.58M | x0 = _mm_loadl_epi64((__m128i *) &src[x - srcstride]); |
1305 | 1.58M | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
1306 | 1.58M | x2 = _mm_loadl_epi64((__m128i *) &src[x + srcstride]); |
1307 | 1.58M | x3 = _mm_loadl_epi64((__m128i *) &src[x + 2 * srcstride]); |
1308 | | |
1309 | 1.58M | t0 = _mm_unpacklo_epi8(x0, r1); |
1310 | 1.58M | t1 = _mm_unpacklo_epi8(x1, r1); |
1311 | 1.58M | t2 = _mm_unpacklo_epi8(x2, r1); |
1312 | 1.58M | t3 = _mm_unpacklo_epi8(x3, r1); |
1313 | | |
1314 | | |
1315 | | /* multiply by correct value : */ |
1316 | 1.58M | r0 = _mm_mullo_epi16(t0, f0); |
1317 | 1.58M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t1, f1)); |
1318 | 1.58M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t2, f2)); |
1319 | 1.58M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t3, f3)); |
1320 | | /* give results back */ |
1321 | 1.58M | _mm_storeu_si128((__m128i *) &dst[x], r0); |
1322 | 1.58M | } |
1323 | 1.54M | src += srcstride; |
1324 | 1.54M | dst += dststride; |
1325 | 1.54M | } |
1326 | 1.30M | }else if(!(width & 3)){ |
1327 | 1.14M | r1= _mm_setzero_si128(); |
1328 | 6.00M | for (y = 0; y < height; y++) { |
1329 | 9.79M | for(x=0;x<width;x+=4){ |
1330 | 4.93M | x0 = _mm_loadl_epi64((__m128i *) &src[x - srcstride]); |
1331 | 4.93M | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
1332 | 4.93M | x2 = _mm_loadl_epi64((__m128i *) &src[x + srcstride]); |
1333 | 4.93M | x3 = _mm_loadl_epi64((__m128i *) &src[x + 2 * srcstride]); |
1334 | | |
1335 | 4.93M | t0 = _mm_unpacklo_epi8(x0, r1); |
1336 | 4.93M | t1 = _mm_unpacklo_epi8(x1, r1); |
1337 | 4.93M | t2 = _mm_unpacklo_epi8(x2, r1); |
1338 | 4.93M | t3 = _mm_unpacklo_epi8(x3, r1); |
1339 | | |
1340 | | |
1341 | | /* multiply by correct value : */ |
1342 | 4.93M | r0 = _mm_mullo_epi16(t0, f0); |
1343 | 4.93M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t1, f1)); |
1344 | 4.93M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t2, f2)); |
1345 | 4.93M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t3, f3)); |
1346 | | /* give results back */ |
1347 | 4.93M | _mm_storel_epi64((__m128i *) &dst[x], r0); |
1348 | 4.93M | } |
1349 | 4.85M | src += srcstride; |
1350 | 4.85M | dst += dststride; |
1351 | 4.85M | } |
1352 | 1.14M | }else{ |
1353 | 162k | r1= _mm_setzero_si128(); |
1354 | 1.09M | for (y = 0; y < height; y++) { |
1355 | 1.93M | for(x=0;x<width;x+=2){ |
1356 | 1.00M | x0 = _mm_loadl_epi64((__m128i *) &src[x - srcstride]); |
1357 | 1.00M | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
1358 | 1.00M | x2 = _mm_loadl_epi64((__m128i *) &src[x + srcstride]); |
1359 | 1.00M | x3 = _mm_loadl_epi64((__m128i *) &src[x + 2 * srcstride]); |
1360 | | |
1361 | 1.00M | t0 = _mm_unpacklo_epi8(x0, r1); |
1362 | 1.00M | t1 = _mm_unpacklo_epi8(x1, r1); |
1363 | 1.00M | t2 = _mm_unpacklo_epi8(x2, r1); |
1364 | 1.00M | t3 = _mm_unpacklo_epi8(x3, r1); |
1365 | | |
1366 | | |
1367 | | /* multiply by correct value : */ |
1368 | 1.00M | r0 = _mm_mullo_epi16(t0, f0); |
1369 | 1.00M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t1, f1)); |
1370 | 1.00M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t2, f2)); |
1371 | 1.00M | r0 = _mm_adds_epi16(r0, _mm_mullo_epi16(t3, f3)); |
1372 | | /* give results back */ |
1373 | | #if MASKMOVE |
1374 | | _mm_maskmoveu_si128(r0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
1375 | | #else |
1376 | 1.00M | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
1377 | 1.00M | #endif |
1378 | 1.00M | } |
1379 | 932k | src += srcstride; |
1380 | 932k | dst += dststride; |
1381 | 932k | } |
1382 | 162k | } |
1383 | 1.56M | } |
1384 | | |
1385 | | #ifndef __native_client__ |
1386 | | void ff_hevc_put_hevc_epel_v_10_sse(int16_t *dst, ptrdiff_t dststride, |
1387 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, int mx, |
1388 | 0 | int my, int16_t* mcbuffer) { |
1389 | 0 | int x, y; |
1390 | 0 | __m128i x0, x1, x2, x3, t0, t1, t2, t3, r0, f0, f1, f2, f3, r1, r2, r3; |
1391 | 0 | uint16_t *src = (uint16_t*) _src; |
1392 | 0 | ptrdiff_t srcstride = _srcstride >>1; |
1393 | 0 | const int8_t *filter = epel_filters[my - 1]; |
1394 | 0 | int8_t filter_0 = filter[0]; |
1395 | 0 | int8_t filter_1 = filter[1]; |
1396 | 0 | int8_t filter_2 = filter[2]; |
1397 | 0 | int8_t filter_3 = filter[3]; |
1398 | 0 | f0 = _mm_set1_epi16(filter_0); |
1399 | 0 | f1 = _mm_set1_epi16(filter_1); |
1400 | 0 | f2 = _mm_set1_epi16(filter_2); |
1401 | 0 | f3 = _mm_set1_epi16(filter_3); |
1402 | |
|
1403 | 0 | if(!(width & 7)){ |
1404 | 0 | r1= _mm_setzero_si128(); |
1405 | 0 | for (y = 0; y < height; y++) { |
1406 | 0 | for(x=0;x<width;x+=8){ |
1407 | 0 | x0 = _mm_loadu_si128((__m128i *) &src[x - srcstride]); |
1408 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
1409 | 0 | x2 = _mm_loadu_si128((__m128i *) &src[x + srcstride]); |
1410 | 0 | x3 = _mm_loadu_si128((__m128i *) &src[x + 2 * srcstride]); |
1411 | | |
1412 | | // multiply by correct value : |
1413 | 0 | r0 = _mm_mullo_epi16(x0, f0); |
1414 | 0 | t0 = _mm_mulhi_epi16(x0, f0); |
1415 | |
|
1416 | 0 | x0= _mm_unpacklo_epi16(r0,t0); |
1417 | 0 | t0= _mm_unpackhi_epi16(r0,t0); |
1418 | |
|
1419 | 0 | r1 = _mm_mullo_epi16(x1, f1); |
1420 | 0 | t1 = _mm_mulhi_epi16(x1, f1); |
1421 | |
|
1422 | 0 | x1= _mm_unpacklo_epi16(r1,t1); |
1423 | 0 | t1= _mm_unpackhi_epi16(r1,t1); |
1424 | | |
1425 | |
|
1426 | 0 | r2 = _mm_mullo_epi16(x2, f2); |
1427 | 0 | t2 = _mm_mulhi_epi16(x2, f2); |
1428 | |
|
1429 | 0 | x2= _mm_unpacklo_epi16(r2,t2); |
1430 | 0 | t2= _mm_unpackhi_epi16(r2,t2); |
1431 | | |
1432 | |
|
1433 | 0 | r3 = _mm_mullo_epi16(x3, f3); |
1434 | 0 | t3 = _mm_mulhi_epi16(x3, f3); |
1435 | |
|
1436 | 0 | x3= _mm_unpacklo_epi16(r3,t3); |
1437 | 0 | t3= _mm_unpackhi_epi16(r3,t3); |
1438 | | |
1439 | |
|
1440 | 0 | r0= _mm_add_epi32(x0,x1); |
1441 | 0 | r1= _mm_add_epi32(x2,x3); |
1442 | |
|
1443 | 0 | t0= _mm_add_epi32(t0,t1); |
1444 | 0 | t1= _mm_add_epi32(t2,t3); |
1445 | |
|
1446 | 0 | r0= _mm_add_epi32(r0,r1); |
1447 | 0 | t0= _mm_add_epi32(t0,t1); |
1448 | |
|
1449 | 0 | r0= _mm_srai_epi32(r0,2);//>> (BIT_DEPTH - 8) |
1450 | 0 | t0= _mm_srai_epi32(t0,2);//>> (BIT_DEPTH - 8) |
1451 | |
|
1452 | 0 | r0= _mm_packs_epi32(r0, t0); |
1453 | | // give results back |
1454 | 0 | _mm_storeu_si128((__m128i *) &dst[x], r0); |
1455 | 0 | } |
1456 | 0 | src += srcstride; |
1457 | 0 | dst += dststride; |
1458 | 0 | } |
1459 | 0 | }else if(!(width & 3)){ |
1460 | 0 | r1= _mm_setzero_si128(); |
1461 | 0 | for (y = 0; y < height; y++) { |
1462 | 0 | for(x=0;x<width;x+=4){ |
1463 | 0 | x0 = _mm_loadl_epi64((__m128i *) &src[x - srcstride]); |
1464 | 0 | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
1465 | 0 | x2 = _mm_loadl_epi64((__m128i *) &src[x + srcstride]); |
1466 | 0 | x3 = _mm_loadl_epi64((__m128i *) &src[x + 2 * srcstride]); |
1467 | | |
1468 | | /* multiply by correct value : */ |
1469 | 0 | r0 = _mm_mullo_epi16(x0, f0); |
1470 | 0 | t0 = _mm_mulhi_epi16(x0, f0); |
1471 | |
|
1472 | 0 | x0= _mm_unpacklo_epi16(r0,t0); |
1473 | |
|
1474 | 0 | r1 = _mm_mullo_epi16(x1, f1); |
1475 | 0 | t1 = _mm_mulhi_epi16(x1, f1); |
1476 | |
|
1477 | 0 | x1= _mm_unpacklo_epi16(r1,t1); |
1478 | | |
1479 | |
|
1480 | 0 | r2 = _mm_mullo_epi16(x2, f2); |
1481 | 0 | t2 = _mm_mulhi_epi16(x2, f2); |
1482 | |
|
1483 | 0 | x2= _mm_unpacklo_epi16(r2,t2); |
1484 | | |
1485 | |
|
1486 | 0 | r3 = _mm_mullo_epi16(x3, f3); |
1487 | 0 | t3 = _mm_mulhi_epi16(x3, f3); |
1488 | |
|
1489 | 0 | x3= _mm_unpacklo_epi16(r3,t3); |
1490 | | |
1491 | |
|
1492 | 0 | r0= _mm_add_epi32(x0,x1); |
1493 | 0 | r1= _mm_add_epi32(x2,x3); |
1494 | 0 | r0= _mm_add_epi32(r0,r1); |
1495 | 0 | r0= _mm_srai_epi32(r0,2);//>> (BIT_DEPTH - 8) |
1496 | |
|
1497 | 0 | r0= _mm_packs_epi32(r0, r0); |
1498 | | |
1499 | | // give results back |
1500 | 0 | _mm_storel_epi64((__m128i *) &dst[x], r0); |
1501 | 0 | } |
1502 | 0 | src += srcstride; |
1503 | 0 | dst += dststride; |
1504 | 0 | } |
1505 | 0 | }else{ |
1506 | 0 | r1= _mm_setzero_si128(); |
1507 | 0 | for (y = 0; y < height; y++) { |
1508 | 0 | for(x=0;x<width;x+=2){ |
1509 | 0 | x0 = _mm_loadl_epi64((__m128i *) &src[x - srcstride]); |
1510 | 0 | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
1511 | 0 | x2 = _mm_loadl_epi64((__m128i *) &src[x + srcstride]); |
1512 | 0 | x3 = _mm_loadl_epi64((__m128i *) &src[x + 2 * srcstride]); |
1513 | | |
1514 | | /* multiply by correct value : */ |
1515 | 0 | r0 = _mm_mullo_epi16(x0, f0); |
1516 | 0 | t0 = _mm_mulhi_epi16(x0, f0); |
1517 | |
|
1518 | 0 | x0= _mm_unpacklo_epi16(r0,t0); |
1519 | |
|
1520 | 0 | r1 = _mm_mullo_epi16(x1, f1); |
1521 | 0 | t1 = _mm_mulhi_epi16(x1, f1); |
1522 | |
|
1523 | 0 | x1= _mm_unpacklo_epi16(r1,t1); |
1524 | |
|
1525 | 0 | r2 = _mm_mullo_epi16(x2, f2); |
1526 | 0 | t2 = _mm_mulhi_epi16(x2, f2); |
1527 | |
|
1528 | 0 | x2= _mm_unpacklo_epi16(r2,t2); |
1529 | |
|
1530 | 0 | r3 = _mm_mullo_epi16(x3, f3); |
1531 | 0 | t3 = _mm_mulhi_epi16(x3, f3); |
1532 | |
|
1533 | 0 | x3= _mm_unpacklo_epi16(r3,t3); |
1534 | |
|
1535 | 0 | r0= _mm_add_epi32(x0,x1); |
1536 | 0 | r1= _mm_add_epi32(x2,x3); |
1537 | 0 | r0= _mm_add_epi32(r0,r1); |
1538 | 0 | r0= _mm_srai_epi32(r0,2);//>> (BIT_DEPTH - 8) |
1539 | |
|
1540 | 0 | r0= _mm_packs_epi32(r0, r0); |
1541 | | |
1542 | | /* give results back */ |
1543 | 0 | _mm_maskmoveu_si128(r0,_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1),(char *) (dst+x)); |
1544 | |
|
1545 | 0 | } |
1546 | 0 | src += srcstride; |
1547 | 0 | dst += dststride; |
1548 | 0 | } |
1549 | 0 | } |
1550 | 0 | } |
1551 | | #endif |
1552 | | |
1553 | | void ff_hevc_put_hevc_epel_hv_8_sse(int16_t *dst, ptrdiff_t dststride, |
1554 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, int mx, |
1555 | 4.69M | int my, int16_t* mcbuffer, int bit_depth) { |
1556 | 4.69M | int x, y; |
1557 | 4.69M | uint8_t *src = (uint8_t*) _src; |
1558 | 4.69M | ptrdiff_t srcstride = _srcstride; |
1559 | 4.69M | const int8_t *filter_h = epel_filters[mx - 1]; |
1560 | 4.69M | const int8_t *filter_v = epel_filters[my - 1]; |
1561 | 4.69M | __m128i r0, bshuffle1, bshuffle2, x0, x1, x2, x3, t0, t1, t2, t3, f0, f1, |
1562 | 4.69M | f2, f3, r1, r2; |
1563 | 4.69M | int8_t filter_0 = filter_h[0]; |
1564 | 4.69M | int8_t filter_1 = filter_h[1]; |
1565 | 4.69M | int8_t filter_2 = filter_h[2]; |
1566 | 4.69M | int8_t filter_3 = filter_h[3]; |
1567 | 4.69M | int16_t *tmp = mcbuffer; |
1568 | 4.69M | r0 = _mm_set_epi8(filter_3, filter_2, filter_1, filter_0, filter_3, |
1569 | 4.69M | filter_2, filter_1, filter_0, filter_3, filter_2, filter_1, |
1570 | 4.69M | filter_0, filter_3, filter_2, filter_1, filter_0); |
1571 | 4.69M | bshuffle1 = _mm_set_epi8(6, 5, 4, 3, 5, 4, 3, 2, 4, 3, 2, 1, 3, 2, 1, 0); |
1572 | | |
1573 | 4.69M | src -= epel_extra_before * srcstride; |
1574 | | |
1575 | 4.69M | f3 = _mm_set1_epi16(filter_v[3]); |
1576 | 4.69M | f1 = _mm_set1_epi16(filter_v[1]); |
1577 | 4.69M | f2 = _mm_set1_epi16(filter_v[2]); |
1578 | 4.69M | f0 = _mm_set1_epi16(filter_v[0]); |
1579 | | |
1580 | | /* horizontal treatment */ |
1581 | 4.69M | if(!(width & 7)){ |
1582 | 712k | bshuffle2 = _mm_set_epi8(10, 9, 8, 7, 9, 8, 7, 6, 8, 7, 6, 5, 7, 6, 5, |
1583 | 712k | 4); |
1584 | 9.71M | for (y = 0; y < height + epel_extra; y++) { |
1585 | 21.7M | for (x = 0; x < width; x += 8) { |
1586 | | |
1587 | 12.7M | x1 = _mm_loadu_si128((__m128i *) &src[x - 1]); |
1588 | 12.7M | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1589 | 12.7M | x3 = _mm_shuffle_epi8(x1, bshuffle2); |
1590 | | |
1591 | | /* PMADDUBSW then PMADDW */ |
1592 | 12.7M | x2 = _mm_maddubs_epi16(x2, r0); |
1593 | 12.7M | x3 = _mm_maddubs_epi16(x3, r0); |
1594 | 12.7M | x2 = _mm_hadd_epi16(x2, x3); |
1595 | 12.7M | _mm_store_si128((__m128i *) &tmp[x], x2); |
1596 | 12.7M | } |
1597 | 9.00M | src += srcstride; |
1598 | 9.00M | tmp += MAX_PB_SIZE; |
1599 | 9.00M | } |
1600 | 712k | tmp = mcbuffer + epel_extra_before * MAX_PB_SIZE; |
1601 | | |
1602 | | /* vertical treatment */ |
1603 | | |
1604 | 7.57M | for (y = 0; y < height; y++) { |
1605 | 16.8M | for (x = 0; x < width; x += 8) { |
1606 | | /* check if memory needs to be reloaded */ |
1607 | 10.0M | x0 = _mm_load_si128((__m128i *) &tmp[x - MAX_PB_SIZE]); |
1608 | 10.0M | x1 = _mm_load_si128((__m128i *) &tmp[x]); |
1609 | 10.0M | x2 = _mm_load_si128((__m128i *) &tmp[x + MAX_PB_SIZE]); |
1610 | 10.0M | x3 = _mm_load_si128((__m128i *) &tmp[x + 2 * MAX_PB_SIZE]); |
1611 | | |
1612 | 10.0M | r0 = _mm_mullo_epi16(x0, f0); |
1613 | 10.0M | r1 = _mm_mulhi_epi16(x0, f0); |
1614 | 10.0M | r2 = _mm_mullo_epi16(x1, f1); |
1615 | 10.0M | t0 = _mm_unpacklo_epi16(r0, r1); |
1616 | 10.0M | x0 = _mm_unpackhi_epi16(r0, r1); |
1617 | 10.0M | r0 = _mm_mulhi_epi16(x1, f1); |
1618 | 10.0M | r1 = _mm_mullo_epi16(x2, f2); |
1619 | 10.0M | t1 = _mm_unpacklo_epi16(r2, r0); |
1620 | 10.0M | x1 = _mm_unpackhi_epi16(r2, r0); |
1621 | 10.0M | r2 = _mm_mulhi_epi16(x2, f2); |
1622 | 10.0M | r0 = _mm_mullo_epi16(x3, f3); |
1623 | 10.0M | t2 = _mm_unpacklo_epi16(r1, r2); |
1624 | 10.0M | x2 = _mm_unpackhi_epi16(r1, r2); |
1625 | 10.0M | r1 = _mm_mulhi_epi16(x3, f3); |
1626 | 10.0M | t3 = _mm_unpacklo_epi16(r0, r1); |
1627 | 10.0M | x3 = _mm_unpackhi_epi16(r0, r1); |
1628 | | |
1629 | | /* multiply by correct value : */ |
1630 | 10.0M | r0 = _mm_add_epi32(t0, t1); |
1631 | 10.0M | r1 = _mm_add_epi32(x0, x1); |
1632 | 10.0M | r0 = _mm_add_epi32(r0, t2); |
1633 | 10.0M | r1 = _mm_add_epi32(r1, x2); |
1634 | 10.0M | r0 = _mm_add_epi32(r0, t3); |
1635 | 10.0M | r1 = _mm_add_epi32(r1, x3); |
1636 | 10.0M | r0 = _mm_srai_epi32(r0, 6); |
1637 | 10.0M | r1 = _mm_srai_epi32(r1, 6); |
1638 | | |
1639 | | /* give results back */ |
1640 | 10.0M | r0 = _mm_packs_epi32(r0, r1); |
1641 | 10.0M | _mm_store_si128((__m128i *) &dst[x], r0); |
1642 | 10.0M | } |
1643 | 6.86M | tmp += MAX_PB_SIZE; |
1644 | 6.86M | dst += dststride; |
1645 | 6.86M | } |
1646 | 3.98M | }else if(!(width & 3)){ |
1647 | 26.4M | for (y = 0; y < height + epel_extra; y ++) { |
1648 | 47.0M | for(x=0;x<width;x+=4){ |
1649 | | /* load data in register */ |
1650 | 23.7M | x1 = _mm_loadl_epi64((__m128i *) &src[x-1]); |
1651 | | |
1652 | 23.7M | x1 = _mm_shuffle_epi8(x1, bshuffle1); |
1653 | | |
1654 | | /* PMADDUBSW then PMADDW */ |
1655 | 23.7M | x1 = _mm_maddubs_epi16(x1, r0); |
1656 | 23.7M | x1 = _mm_hadd_epi16(x1, _mm_setzero_si128()); |
1657 | | |
1658 | | /* give results back */ |
1659 | 23.7M | _mm_storel_epi64((__m128i *) &tmp[x], x1); |
1660 | | |
1661 | 23.7M | } |
1662 | 23.3M | src += srcstride; |
1663 | 23.3M | tmp += MAX_PB_SIZE; |
1664 | 23.3M | } |
1665 | 3.10M | tmp = mcbuffer + epel_extra_before * MAX_PB_SIZE; |
1666 | | |
1667 | | /* vertical treatment */ |
1668 | | |
1669 | | |
1670 | 17.1M | for (y = 0; y < height; y++) { |
1671 | 28.3M | for (x = 0; x < width; x += 4) { |
1672 | | /* check if memory needs to be reloaded */ |
1673 | 14.3M | x0 = _mm_loadl_epi64((__m128i *) &tmp[x - MAX_PB_SIZE]); |
1674 | 14.3M | x1 = _mm_loadl_epi64((__m128i *) &tmp[x]); |
1675 | 14.3M | x2 = _mm_loadl_epi64((__m128i *) &tmp[x + MAX_PB_SIZE]); |
1676 | 14.3M | x3 = _mm_loadl_epi64((__m128i *) &tmp[x + 2 * MAX_PB_SIZE]); |
1677 | | |
1678 | 14.3M | r0 = _mm_mullo_epi16(x0, f0); |
1679 | 14.3M | r1 = _mm_mulhi_epi16(x0, f0); |
1680 | 14.3M | r2 = _mm_mullo_epi16(x1, f1); |
1681 | 14.3M | t0 = _mm_unpacklo_epi16(r0, r1); |
1682 | | |
1683 | 14.3M | r0 = _mm_mulhi_epi16(x1, f1); |
1684 | 14.3M | r1 = _mm_mullo_epi16(x2, f2); |
1685 | 14.3M | t1 = _mm_unpacklo_epi16(r2, r0); |
1686 | | |
1687 | 14.3M | r2 = _mm_mulhi_epi16(x2, f2); |
1688 | 14.3M | r0 = _mm_mullo_epi16(x3, f3); |
1689 | 14.3M | t2 = _mm_unpacklo_epi16(r1, r2); |
1690 | | |
1691 | 14.3M | r1 = _mm_mulhi_epi16(x3, f3); |
1692 | 14.3M | t3 = _mm_unpacklo_epi16(r0, r1); |
1693 | | |
1694 | | |
1695 | | /* multiply by correct value : */ |
1696 | 14.3M | r0 = _mm_add_epi32(t0, t1); |
1697 | 14.3M | r0 = _mm_add_epi32(r0, t2); |
1698 | 14.3M | r0 = _mm_add_epi32(r0, t3); |
1699 | 14.3M | r0 = _mm_srai_epi32(r0, 6); |
1700 | | |
1701 | | /* give results back */ |
1702 | 14.3M | r0 = _mm_packs_epi32(r0, r0); |
1703 | 14.3M | _mm_storel_epi64((__m128i *) &dst[x], r0); |
1704 | 14.3M | } |
1705 | 14.0M | tmp += MAX_PB_SIZE; |
1706 | 14.0M | dst += dststride; |
1707 | 14.0M | } |
1708 | 3.10M | }else{ |
1709 | | #if MASKMOVE |
1710 | | bshuffle2=_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1); |
1711 | | #endif |
1712 | 8.18M | for (y = 0; y < height + epel_extra; y ++) { |
1713 | 15.0M | for(x=0;x<width;x+=2){ |
1714 | | /* load data in register */ |
1715 | 7.73M | x1 = _mm_loadl_epi64((__m128i *) &src[x-1]); |
1716 | 7.73M | x1 = _mm_shuffle_epi8(x1, bshuffle1); |
1717 | | |
1718 | | /* PMADDUBSW then PMADDW */ |
1719 | 7.73M | x1 = _mm_maddubs_epi16(x1, r0); |
1720 | 7.73M | x1 = _mm_hadd_epi16(x1, _mm_setzero_si128()); |
1721 | | |
1722 | | /* give results back */ |
1723 | | #if MASKMOVE |
1724 | | _mm_maskmoveu_si128(x1,bshuffle2,(char *) (tmp+x)); |
1725 | | #else |
1726 | 7.73M | *((uint32_t*)(tmp+x)) = _mm_cvtsi128_si32(x1); |
1727 | 7.73M | #endif |
1728 | 7.73M | } |
1729 | 7.31M | src += srcstride; |
1730 | 7.31M | tmp += MAX_PB_SIZE; |
1731 | 7.31M | } |
1732 | | |
1733 | 871k | tmp = mcbuffer + epel_extra_before * MAX_PB_SIZE; |
1734 | | |
1735 | | /* vertical treatment */ |
1736 | | |
1737 | 5.56M | for (y = 0; y < height; y++) { |
1738 | 9.71M | for (x = 0; x < width; x += 2) { |
1739 | | /* check if memory needs to be reloaded */ |
1740 | 5.01M | x0 = _mm_loadl_epi64((__m128i *) &tmp[x - MAX_PB_SIZE]); |
1741 | 5.01M | x1 = _mm_loadl_epi64((__m128i *) &tmp[x]); |
1742 | 5.01M | x2 = _mm_loadl_epi64((__m128i *) &tmp[x + MAX_PB_SIZE]); |
1743 | 5.01M | x3 = _mm_loadl_epi64((__m128i *) &tmp[x + 2 * MAX_PB_SIZE]); |
1744 | | |
1745 | 5.01M | r0 = _mm_mullo_epi16(x0, f0); |
1746 | 5.01M | r1 = _mm_mulhi_epi16(x0, f0); |
1747 | 5.01M | r2 = _mm_mullo_epi16(x1, f1); |
1748 | 5.01M | t0 = _mm_unpacklo_epi16(r0, r1); |
1749 | 5.01M | r0 = _mm_mulhi_epi16(x1, f1); |
1750 | 5.01M | r1 = _mm_mullo_epi16(x2, f2); |
1751 | 5.01M | t1 = _mm_unpacklo_epi16(r2, r0); |
1752 | 5.01M | r2 = _mm_mulhi_epi16(x2, f2); |
1753 | 5.01M | r0 = _mm_mullo_epi16(x3, f3); |
1754 | 5.01M | t2 = _mm_unpacklo_epi16(r1, r2); |
1755 | 5.01M | r1 = _mm_mulhi_epi16(x3, f3); |
1756 | 5.01M | t3 = _mm_unpacklo_epi16(r0, r1); |
1757 | | |
1758 | | /* multiply by correct value : */ |
1759 | 5.01M | r0 = _mm_add_epi32(t0, t1); |
1760 | 5.01M | r0 = _mm_add_epi32(r0, t2); |
1761 | 5.01M | r0 = _mm_add_epi32(r0, t3); |
1762 | 5.01M | r0 = _mm_srai_epi32(r0, 6); |
1763 | | /* give results back */ |
1764 | 5.01M | r0 = _mm_packs_epi32(r0, r0); |
1765 | | #if MASKMOVE |
1766 | | _mm_maskmoveu_si128(r0,bshuffle2,(char *) (dst+x)); |
1767 | | #else |
1768 | 5.01M | *((uint32_t*)(dst+x)) = _mm_cvtsi128_si32(r0); |
1769 | 5.01M | #endif |
1770 | 5.01M | } |
1771 | 4.69M | tmp += MAX_PB_SIZE; |
1772 | 4.69M | dst += dststride; |
1773 | 4.69M | } |
1774 | 871k | } |
1775 | | |
1776 | 4.69M | } |
1777 | | |
1778 | | |
1779 | | #ifndef __native_client__ |
1780 | | void ff_hevc_put_hevc_epel_hv_10_sse(int16_t *dst, ptrdiff_t dststride, |
1781 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, int mx, |
1782 | 0 | int my, int16_t* mcbuffer) { |
1783 | 0 | int x, y; |
1784 | 0 | uint16_t *src = (uint16_t*) _src; |
1785 | 0 | ptrdiff_t srcstride = _srcstride>>1; |
1786 | 0 | const int8_t *filter_h = epel_filters[mx - 1]; |
1787 | 0 | const int8_t *filter_v = epel_filters[my - 1]; |
1788 | 0 | __m128i r0, bshuffle1, bshuffle2, x0, x1, x2, x3, t0, t1, t2, t3, f0, f1, |
1789 | 0 | f2, f3, r1, r2, r3; |
1790 | 0 | int8_t filter_0 = filter_h[0]; |
1791 | 0 | int8_t filter_1 = filter_h[1]; |
1792 | 0 | int8_t filter_2 = filter_h[2]; |
1793 | 0 | int8_t filter_3 = filter_h[3]; |
1794 | 0 | int16_t *tmp = mcbuffer; |
1795 | |
|
1796 | 0 | r0 = _mm_set_epi16(filter_3, filter_2, filter_1, |
1797 | 0 | filter_0, filter_3, filter_2, filter_1, filter_0); |
1798 | 0 | bshuffle1 = _mm_set_epi8(9,8,7,6,5,4, 3, 2,7,6,5,4, 3, 2, 1, 0); |
1799 | |
|
1800 | 0 | src -= epel_extra_before * srcstride; |
1801 | |
|
1802 | 0 | f0 = _mm_set1_epi16(filter_v[0]); |
1803 | 0 | f1 = _mm_set1_epi16(filter_v[1]); |
1804 | 0 | f2 = _mm_set1_epi16(filter_v[2]); |
1805 | 0 | f3 = _mm_set1_epi16(filter_v[3]); |
1806 | | |
1807 | | |
1808 | | /* horizontal treatment */ |
1809 | 0 | if(!(width & 3)){ |
1810 | 0 | bshuffle2 = _mm_set_epi8(13,12,11,10,9,8,7,6,11,10, 9,8,7,6,5, 4); |
1811 | 0 | for (y = 0; y < height + epel_extra; y ++) { |
1812 | 0 | for(x=0;x<width;x+=4){ |
1813 | |
|
1814 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x-1]); |
1815 | 0 | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1816 | 0 | x3 = _mm_shuffle_epi8(x1, bshuffle2); |
1817 | | |
1818 | |
|
1819 | 0 | x2 = _mm_madd_epi16(x2, r0); |
1820 | 0 | x3 = _mm_madd_epi16(x3, r0); |
1821 | 0 | x2 = _mm_hadd_epi32(x2, x3); |
1822 | 0 | x2= _mm_srai_epi32(x2,2); //>> (BIT_DEPTH - 8) |
1823 | |
|
1824 | 0 | x2 = _mm_packs_epi32(x2,r0); |
1825 | | //give results back |
1826 | 0 | _mm_storel_epi64((__m128i *) &tmp[x], x2); |
1827 | |
|
1828 | 0 | } |
1829 | 0 | src += srcstride; |
1830 | 0 | tmp += MAX_PB_SIZE; |
1831 | 0 | } |
1832 | 0 | tmp = mcbuffer + epel_extra_before * MAX_PB_SIZE; |
1833 | | |
1834 | | // vertical treatment |
1835 | | |
1836 | |
|
1837 | 0 | for (y = 0; y < height; y++) { |
1838 | 0 | for (x = 0; x < width; x += 4) { |
1839 | 0 | x0 = _mm_loadl_epi64((__m128i *) &tmp[x - MAX_PB_SIZE]); |
1840 | 0 | x1 = _mm_loadl_epi64((__m128i *) &tmp[x]); |
1841 | 0 | x2 = _mm_loadl_epi64((__m128i *) &tmp[x + MAX_PB_SIZE]); |
1842 | 0 | x3 = _mm_loadl_epi64((__m128i *) &tmp[x + 2 * MAX_PB_SIZE]); |
1843 | |
|
1844 | 0 | r0 = _mm_mullo_epi16(x0, f0); |
1845 | 0 | r1 = _mm_mulhi_epi16(x0, f0); |
1846 | 0 | r2 = _mm_mullo_epi16(x1, f1); |
1847 | 0 | t0 = _mm_unpacklo_epi16(r0, r1); |
1848 | |
|
1849 | 0 | r0 = _mm_mulhi_epi16(x1, f1); |
1850 | 0 | r1 = _mm_mullo_epi16(x2, f2); |
1851 | 0 | t1 = _mm_unpacklo_epi16(r2, r0); |
1852 | |
|
1853 | 0 | r2 = _mm_mulhi_epi16(x2, f2); |
1854 | 0 | r0 = _mm_mullo_epi16(x3, f3); |
1855 | 0 | t2 = _mm_unpacklo_epi16(r1, r2); |
1856 | |
|
1857 | 0 | r1 = _mm_mulhi_epi16(x3, f3); |
1858 | 0 | t3 = _mm_unpacklo_epi16(r0, r1); |
1859 | | |
1860 | | |
1861 | |
|
1862 | 0 | r0 = _mm_add_epi32(t0, t1); |
1863 | 0 | r0 = _mm_add_epi32(r0, t2); |
1864 | 0 | r0 = _mm_add_epi32(r0, t3); |
1865 | 0 | r0 = _mm_srai_epi32(r0, 6); |
1866 | | |
1867 | | // give results back |
1868 | 0 | r0 = _mm_packs_epi32(r0, r0); |
1869 | 0 | _mm_storel_epi64((__m128i *) &dst[x], r0); |
1870 | 0 | } |
1871 | 0 | tmp += MAX_PB_SIZE; |
1872 | 0 | dst += dststride; |
1873 | 0 | } |
1874 | 0 | }else{ |
1875 | 0 | bshuffle2=_mm_set_epi8(0,0,0,0,0,0,0,0,0,0,0,0,-1,-1,-1,-1); |
1876 | 0 | r1= _mm_setzero_si128(); |
1877 | 0 | for (y = 0; y < height + epel_extra; y ++) { |
1878 | 0 | for(x=0;x<width;x+=2){ |
1879 | | /* load data in register */ |
1880 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x-1]); |
1881 | 0 | x2 = _mm_shuffle_epi8(x1, bshuffle1); |
1882 | | |
1883 | | /* PMADDUBSW then PMADDW */ |
1884 | 0 | x2 = _mm_madd_epi16(x2, r0); |
1885 | 0 | x2 = _mm_hadd_epi32(x2, r1); |
1886 | 0 | x2= _mm_srai_epi32(x2,2); //>> (BIT_DEPTH - 8) |
1887 | 0 | x2 = _mm_packs_epi32(x2, r1); |
1888 | | /* give results back */ |
1889 | 0 | _mm_maskmoveu_si128(x2,bshuffle2,(char *) (tmp+x)); |
1890 | 0 | } |
1891 | 0 | src += srcstride; |
1892 | 0 | tmp += MAX_PB_SIZE; |
1893 | 0 | } |
1894 | |
|
1895 | 0 | tmp = mcbuffer + epel_extra_before * MAX_PB_SIZE; |
1896 | | |
1897 | | /* vertical treatment */ |
1898 | |
|
1899 | 0 | for (y = 0; y < height; y++) { |
1900 | 0 | for (x = 0; x < width; x += 2) { |
1901 | | /* check if memory needs to be reloaded */ |
1902 | 0 | x0 = _mm_loadl_epi64((__m128i *) &tmp[x - MAX_PB_SIZE]); |
1903 | 0 | x1 = _mm_loadl_epi64((__m128i *) &tmp[x]); |
1904 | 0 | x2 = _mm_loadl_epi64((__m128i *) &tmp[x + MAX_PB_SIZE]); |
1905 | 0 | x3 = _mm_loadl_epi64((__m128i *) &tmp[x + 2 * MAX_PB_SIZE]); |
1906 | |
|
1907 | 0 | r0 = _mm_mullo_epi16(x0, f0); |
1908 | 0 | t0 = _mm_mulhi_epi16(x0, f0); |
1909 | |
|
1910 | 0 | x0= _mm_unpacklo_epi16(r0,t0); |
1911 | |
|
1912 | 0 | r1 = _mm_mullo_epi16(x1, f1); |
1913 | 0 | t1 = _mm_mulhi_epi16(x1, f1); |
1914 | |
|
1915 | 0 | x1= _mm_unpacklo_epi16(r1,t1); |
1916 | |
|
1917 | 0 | r2 = _mm_mullo_epi16(x2, f2); |
1918 | 0 | t2 = _mm_mulhi_epi16(x2, f2); |
1919 | |
|
1920 | 0 | x2= _mm_unpacklo_epi16(r2,t2); |
1921 | |
|
1922 | 0 | r3 = _mm_mullo_epi16(x3, f3); |
1923 | 0 | t3 = _mm_mulhi_epi16(x3, f3); |
1924 | |
|
1925 | 0 | x3= _mm_unpacklo_epi16(r3,t3); |
1926 | |
|
1927 | 0 | r0= _mm_add_epi32(x0,x1); |
1928 | 0 | r1= _mm_add_epi32(x2,x3); |
1929 | 0 | r0= _mm_add_epi32(r0,r1); |
1930 | 0 | r0 = _mm_srai_epi32(r0, 6); |
1931 | | /* give results back */ |
1932 | 0 | r0 = _mm_packs_epi32(r0, r0); |
1933 | 0 | _mm_maskmoveu_si128(r0,bshuffle2,(char *) (dst+x)); |
1934 | 0 | } |
1935 | 0 | tmp += MAX_PB_SIZE; |
1936 | 0 | dst += dststride; |
1937 | 0 | } |
1938 | 0 | } |
1939 | 0 | } |
1940 | | #endif |
1941 | | |
1942 | | void ff_hevc_put_hevc_qpel_pixels_8_sse(int16_t *dst, ptrdiff_t dststride, |
1943 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
1944 | 3.45M | int16_t* mcbuffer) { |
1945 | 3.45M | int x, y; |
1946 | 3.45M | __m128i x1, x2, x3, x0; |
1947 | 3.45M | uint8_t *src = (uint8_t*) _src; |
1948 | 3.45M | ptrdiff_t srcstride = _srcstride; |
1949 | 3.45M | x0= _mm_setzero_si128(); |
1950 | 3.45M | if(!(width & 15)){ |
1951 | 11.3M | for (y = 0; y < height; y++) { |
1952 | 26.7M | for (x = 0; x < width; x += 16) { |
1953 | | |
1954 | 15.9M | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
1955 | 15.9M | x2 = _mm_unpacklo_epi8(x1, x0); |
1956 | | |
1957 | 15.9M | x3 = _mm_unpackhi_epi8(x1, x0); |
1958 | | |
1959 | 15.9M | x2 = _mm_slli_epi16(x2, 6); |
1960 | 15.9M | x3 = _mm_slli_epi16(x3, 6); |
1961 | 15.9M | _mm_storeu_si128((__m128i *) &dst[x], x2); |
1962 | 15.9M | _mm_storeu_si128((__m128i *) &dst[x + 8], x3); |
1963 | | |
1964 | 15.9M | } |
1965 | 10.7M | src += srcstride; |
1966 | 10.7M | dst += dststride; |
1967 | 10.7M | } |
1968 | 2.90M | }else if(!(width & 7)){ |
1969 | 22.6M | for (y = 0; y < height; y++) { |
1970 | 40.6M | for (x = 0; x < width; x += 8) { |
1971 | | |
1972 | 20.4M | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
1973 | 20.4M | x2 = _mm_unpacklo_epi8(x1, x0); |
1974 | 20.4M | x2 = _mm_slli_epi16(x2, 6); |
1975 | 20.4M | _mm_storeu_si128((__m128i *) &dst[x], x2); |
1976 | | |
1977 | 20.4M | } |
1978 | 20.1M | src += srcstride; |
1979 | 20.1M | dst += dststride; |
1980 | 20.1M | } |
1981 | 2.52M | }else if(!(width & 3)){ |
1982 | 3.62M | for (y = 0; y < height; y++) { |
1983 | 6.90M | for(x=0;x<width;x+=4){ |
1984 | 3.66M | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
1985 | 3.66M | x2 = _mm_unpacklo_epi8(x1, x0); |
1986 | 3.66M | x2 = _mm_slli_epi16(x2, 6); |
1987 | 3.66M | _mm_storel_epi64((__m128i *) &dst[x], x2); |
1988 | 3.66M | } |
1989 | 3.24M | src += srcstride; |
1990 | 3.24M | dst += dststride; |
1991 | 3.24M | } |
1992 | 379k | }else{ |
1993 | | #if MASKMOVE |
1994 | | x4= _mm_set_epi32(0,0,0,-1); //mask to store |
1995 | | #endif |
1996 | 0 | for (y = 0; y < height; y++) { |
1997 | 0 | for(x=0;x<width;x+=2){ |
1998 | 0 | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
1999 | 0 | x2 = _mm_unpacklo_epi8(x1, x0); |
2000 | 0 | x2 = _mm_slli_epi16(x2, 6); |
2001 | | #if MASKMOVE |
2002 | | _mm_maskmoveu_si128(x2,x4,(char *) (dst+x)); |
2003 | | #else |
2004 | 0 | *((uint16_t*)(dst+x)) = _mm_cvtsi128_si32(x2); |
2005 | 0 | #endif |
2006 | 0 | } |
2007 | 0 | src += srcstride; |
2008 | 0 | dst += dststride; |
2009 | 0 | } |
2010 | 0 | } |
2011 | | |
2012 | | |
2013 | 3.45M | } |
2014 | | |
2015 | | #ifndef __native_client__ |
2016 | | void ff_hevc_put_hevc_qpel_pixels_10_sse(int16_t *dst, ptrdiff_t dststride, |
2017 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2018 | 0 | int16_t* mcbuffer) { |
2019 | 0 | int x, y; |
2020 | 0 | __m128i x1, x2, x4; |
2021 | 0 | uint16_t *src = (uint16_t*) _src; |
2022 | 0 | ptrdiff_t srcstride = _srcstride>>1; |
2023 | 0 | if(!(width & 7)){ |
2024 | 0 | for (y = 0; y < height; y++) { |
2025 | 0 | for (x = 0; x < width; x += 8) { |
2026 | |
|
2027 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x]); |
2028 | 0 | x2 = _mm_slli_epi16(x1, 4); //14-BIT DEPTH |
2029 | 0 | _mm_storeu_si128((__m128i *) &dst[x], x2); |
2030 | |
|
2031 | 0 | } |
2032 | 0 | src += srcstride; |
2033 | 0 | dst += dststride; |
2034 | 0 | } |
2035 | 0 | }else if(!(width & 3)){ |
2036 | 0 | for (y = 0; y < height; y++) { |
2037 | 0 | for(x=0;x<width;x+=4){ |
2038 | 0 | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
2039 | 0 | x2 = _mm_slli_epi16(x1, 4);//14-BIT DEPTH |
2040 | 0 | _mm_storel_epi64((__m128i *) &dst[x], x2); |
2041 | 0 | } |
2042 | 0 | src += srcstride; |
2043 | 0 | dst += dststride; |
2044 | 0 | } |
2045 | 0 | }else{ |
2046 | 0 | x4= _mm_set_epi32(0,0,0,-1); //mask to store |
2047 | 0 | for (y = 0; y < height; y++) { |
2048 | 0 | for(x=0;x<width;x+=2){ |
2049 | 0 | x1 = _mm_loadl_epi64((__m128i *) &src[x]); |
2050 | 0 | x2 = _mm_slli_epi16(x1, 4);//14-BIT DEPTH |
2051 | 0 | _mm_maskmoveu_si128(x2,x4,(char *) (dst+x)); |
2052 | 0 | } |
2053 | 0 | src += srcstride; |
2054 | 0 | dst += dststride; |
2055 | 0 | } |
2056 | 0 | } |
2057 | | |
2058 | |
|
2059 | 0 | } |
2060 | | #endif |
2061 | | |
2062 | | |
2063 | | void ff_hevc_put_hevc_qpel_h_1_8_sse(int16_t *dst, ptrdiff_t dststride, |
2064 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2065 | 253k | int16_t* mcbuffer) { |
2066 | 253k | int x, y; |
2067 | 253k | const uint8_t *src = _src; |
2068 | 253k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2069 | 253k | __m128i x1, r0, x2, x3, x4, x5; |
2070 | | |
2071 | 253k | r0 = _mm_set_epi8(0, 1, -5, 17, 58, -10, 4, -1, 0, 1, -5, 17, 58, -10, 4, |
2072 | 253k | -1); |
2073 | | |
2074 | 253k | if(!(width & 7)){ |
2075 | 2.40M | for (y = 0; y < height; y++) { |
2076 | 5.70M | for (x = 0; x < width; x += 8) { |
2077 | | /* load data in register */ |
2078 | 3.51M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
2079 | 3.51M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2080 | 3.51M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2081 | 3.51M | _mm_srli_si128(x1, 3)); |
2082 | 3.51M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
2083 | 3.51M | _mm_srli_si128(x1, 5)); |
2084 | 3.51M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
2085 | 3.51M | _mm_srli_si128(x1, 7)); |
2086 | | |
2087 | | /* PMADDUBSW then PMADDW */ |
2088 | 3.51M | x2 = _mm_maddubs_epi16(x2, r0); |
2089 | 3.51M | x3 = _mm_maddubs_epi16(x3, r0); |
2090 | 3.51M | x4 = _mm_maddubs_epi16(x4, r0); |
2091 | 3.51M | x5 = _mm_maddubs_epi16(x5, r0); |
2092 | 3.51M | x2 = _mm_hadd_epi16(x2, x3); |
2093 | 3.51M | x4 = _mm_hadd_epi16(x4, x5); |
2094 | 3.51M | x2 = _mm_hadd_epi16(x2, x4); |
2095 | | /* give results back */ |
2096 | 3.51M | _mm_store_si128((__m128i *) &dst[x],x2); |
2097 | | |
2098 | 3.51M | } |
2099 | 2.18M | src += srcstride; |
2100 | 2.18M | dst += dststride; |
2101 | 2.18M | } |
2102 | 217k | }else if(!(width &3)){ |
2103 | | |
2104 | 351k | for (y = 0; y < height; y ++) { |
2105 | 677k | for(x=0;x<width;x+=4){ |
2106 | | /* load data in register */ |
2107 | 362k | x1 = _mm_loadu_si128((__m128i *) &src[x-3]); |
2108 | 362k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2109 | 362k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2110 | 362k | _mm_srli_si128(x1, 3)); |
2111 | | |
2112 | | /* PMADDUBSW then PMADDW */ |
2113 | 362k | x2 = _mm_maddubs_epi16(x2, r0); |
2114 | 362k | x3 = _mm_maddubs_epi16(x3, r0); |
2115 | 362k | x2 = _mm_hadd_epi16(x2, x3); |
2116 | 362k | x2 = _mm_hadd_epi16(x2, x2); |
2117 | | |
2118 | | /* give results back */ |
2119 | 362k | _mm_storel_epi64((__m128i *) &dst[x], x2); |
2120 | 362k | } |
2121 | | |
2122 | 314k | src += srcstride; |
2123 | 314k | dst += dststride; |
2124 | 314k | } |
2125 | 36.4k | }else{ |
2126 | 0 | x5= _mm_setzero_si128(); |
2127 | | #if MASKMOVE |
2128 | | x3= _mm_set_epi32(0,0,0,-1); |
2129 | | #endif |
2130 | 0 | for (y = 0; y < height; y ++) { |
2131 | 0 | for(x=0;x<width;x+=4){ |
2132 | | /* load data in register */ |
2133 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x-3]); |
2134 | 0 | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2135 | | |
2136 | | |
2137 | | |
2138 | | /* PMADDUBSW then PMADDW */ |
2139 | 0 | x2 = _mm_maddubs_epi16(x2, r0); |
2140 | 0 | x2 = _mm_hadd_epi16(x2,x5 ); |
2141 | 0 | x2 = _mm_hadd_epi16(x2,x5 ); |
2142 | | |
2143 | | /* give results back */ |
2144 | | //_mm_storel_epi64((__m128i *) &dst[x], x2); |
2145 | | #if MASKMOVE |
2146 | | _mm_maskmoveu_si128(x2,x3,(char *) (dst+x)); |
2147 | | #else |
2148 | 0 | *((uint16_t*)(dst+x)) = _mm_cvtsi128_si32(x2); |
2149 | 0 | #endif |
2150 | 0 | } |
2151 | |
|
2152 | 0 | src += srcstride; |
2153 | 0 | dst += dststride; |
2154 | 0 | } |
2155 | 0 | } |
2156 | | |
2157 | 253k | } |
2158 | | #ifndef __native_client__ |
2159 | | /* |
2160 | | * @TODO : Valgrind to see if it's useful to use SSE or wait for AVX2 implementation |
2161 | | */ |
2162 | | void ff_hevc_put_hevc_qpel_h_1_10_sse(int16_t *dst, ptrdiff_t dststride, |
2163 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2164 | 0 | int16_t* mcbuffer) { |
2165 | 0 | int x, y; |
2166 | 0 | uint16_t *src = (uint16_t*)_src; |
2167 | 0 | ptrdiff_t srcstride = _srcstride>>1; |
2168 | 0 | __m128i x0, x1, x2, x3, r0; |
2169 | |
|
2170 | 0 | r0 = _mm_set_epi16(0, 1, -5, 17, 58, -10, 4, -1); |
2171 | 0 | x0= _mm_setzero_si128(); |
2172 | 0 | x3= _mm_set_epi32(0,0,0,-1); |
2173 | 0 | for (y = 0; y < height; y ++) { |
2174 | 0 | for(x=0;x<width;x+=2){ |
2175 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x-3]); |
2176 | 0 | x2 = _mm_srli_si128(x1,2); //last 16bit not used so 1 load can be used for 2 dst |
2177 | |
|
2178 | 0 | x1 = _mm_madd_epi16(x1,r0); |
2179 | 0 | x2 = _mm_madd_epi16(x2,r0); |
2180 | |
|
2181 | 0 | x1 = _mm_hadd_epi32(x1,x2); |
2182 | 0 | x1 = _mm_hadd_epi32(x1,x0); |
2183 | 0 | x1= _mm_srai_epi32(x1,2); //>>BIT_DEPTH-8 |
2184 | 0 | x1= _mm_packs_epi32(x1,x0); |
2185 | | // dst[x]= _mm_extract_epi16(x1,0); |
2186 | 0 | _mm_maskmoveu_si128(x1,x3,(char *) (dst+x)); |
2187 | 0 | } |
2188 | 0 | src += srcstride; |
2189 | 0 | dst += dststride; |
2190 | 0 | } |
2191 | |
|
2192 | 0 | } |
2193 | | #endif |
2194 | | |
2195 | | |
2196 | | void ff_hevc_put_hevc_qpel_h_2_8_sse(int16_t *dst, ptrdiff_t dststride, |
2197 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2198 | 155k | int16_t* mcbuffer) { |
2199 | 155k | int x, y; |
2200 | 155k | const uint8_t *src = _src; |
2201 | 155k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2202 | 155k | __m128i x1, r0, x2, x3, x4, x5; |
2203 | | |
2204 | 155k | r0 = _mm_set_epi8(-1, 4, -11, 40, 40, -11, 4, -1, -1, 4, -11, 40, 40, -11, |
2205 | 155k | 4, -1); |
2206 | | |
2207 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
2208 | 155k | if(!(width - 15)){ |
2209 | 0 | for (y = 0; y < height; y++) { |
2210 | 0 | for (x = 0; x < width; x += 8) { |
2211 | | /* load data in register */ |
2212 | 0 | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
2213 | 0 | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2214 | 0 | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2215 | 0 | _mm_srli_si128(x1, 3)); |
2216 | 0 | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
2217 | 0 | _mm_srli_si128(x1, 5)); |
2218 | 0 | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
2219 | 0 | _mm_srli_si128(x1, 7)); |
2220 | | |
2221 | | /* PMADDUBSW then PMADDW */ |
2222 | 0 | x2 = _mm_maddubs_epi16(x2, r0); |
2223 | 0 | x3 = _mm_maddubs_epi16(x3, r0); |
2224 | 0 | x4 = _mm_maddubs_epi16(x4, r0); |
2225 | 0 | x5 = _mm_maddubs_epi16(x5, r0); |
2226 | 0 | x2 = _mm_hadd_epi16(x2, x3); |
2227 | 0 | x4 = _mm_hadd_epi16(x4, x5); |
2228 | 0 | x2 = _mm_hadd_epi16(x2, x4); |
2229 | | /* give results back */ |
2230 | 0 | _mm_store_si128((__m128i *) &dst[x],x2); |
2231 | 0 | } |
2232 | 0 | src += srcstride; |
2233 | 0 | dst += dststride; |
2234 | 0 | } |
2235 | |
|
2236 | 155k | }else{ |
2237 | | |
2238 | 1.61M | for (y = 0; y < height; y ++) { |
2239 | 5.75M | for(x=0;x<width;x+=4){ |
2240 | | /* load data in register */ |
2241 | 4.30M | x1 = _mm_loadu_si128((__m128i *) &src[x-3]); |
2242 | | |
2243 | 4.30M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2244 | 4.30M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2245 | 4.30M | _mm_srli_si128(x1, 3)); |
2246 | | |
2247 | | |
2248 | | /* PMADDUBSW then PMADDW */ |
2249 | 4.30M | x2 = _mm_maddubs_epi16(x2, r0); |
2250 | 4.30M | x3 = _mm_maddubs_epi16(x3, r0); |
2251 | 4.30M | x2 = _mm_hadd_epi16(x2, x3); |
2252 | 4.30M | x2 = _mm_hadd_epi16(x2, _mm_setzero_si128()); |
2253 | | |
2254 | | /* give results back */ |
2255 | 4.30M | _mm_storel_epi64((__m128i *) &dst[x], x2); |
2256 | | |
2257 | 4.30M | } |
2258 | 1.45M | src += srcstride; |
2259 | 1.45M | dst += dststride; |
2260 | 1.45M | } |
2261 | 155k | } |
2262 | | |
2263 | 155k | } |
2264 | | |
2265 | | #if 0 |
2266 | | static void ff_hevc_put_hevc_qpel_h_2_sse(int16_t *dst, ptrdiff_t dststride, |
2267 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2268 | | int16_t* mcbuffer) { |
2269 | | int x, y; |
2270 | | uint8_t *src = _src; |
2271 | | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2272 | | __m128i x1, r0, x2, x3, x4, x5; |
2273 | | |
2274 | | r0 = _mm_set_epi8(-1, 4, -11, 40, 40, -11, 4, -1, -1, 4, -11, 40, 40, -11, |
2275 | | 4, -1); |
2276 | | |
2277 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
2278 | | if(!(width & 7)){ |
2279 | | for (y = 0; y < height; y++) { |
2280 | | for (x = 0; x < width; x += 8) { |
2281 | | /* load data in register */ |
2282 | | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
2283 | | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2284 | | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2285 | | _mm_srli_si128(x1, 3)); |
2286 | | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
2287 | | _mm_srli_si128(x1, 5)); |
2288 | | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
2289 | | _mm_srli_si128(x1, 7)); |
2290 | | |
2291 | | /* PMADDUBSW then PMADDW */ |
2292 | | x2 = _mm_maddubs_epi16(x2, r0); |
2293 | | x3 = _mm_maddubs_epi16(x3, r0); |
2294 | | x4 = _mm_maddubs_epi16(x4, r0); |
2295 | | x5 = _mm_maddubs_epi16(x5, r0); |
2296 | | x2 = _mm_hadd_epi16(x2, x3); |
2297 | | x4 = _mm_hadd_epi16(x4, x5); |
2298 | | x2 = _mm_hadd_epi16(x2, x4); |
2299 | | /* give results back */ |
2300 | | _mm_store_si128((__m128i *) &dst[x],x2); |
2301 | | } |
2302 | | src += srcstride; |
2303 | | dst += dststride; |
2304 | | } |
2305 | | |
2306 | | }else{ |
2307 | | |
2308 | | for (y = 0; y < height; y ++) { |
2309 | | for(x=0;x<width;x+=4){ |
2310 | | /* load data in register */ |
2311 | | x1 = _mm_loadu_si128((__m128i *) &src[x-3]); |
2312 | | |
2313 | | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2314 | | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2315 | | _mm_srli_si128(x1, 3)); |
2316 | | |
2317 | | |
2318 | | /* PMADDUBSW then PMADDW */ |
2319 | | x2 = _mm_maddubs_epi16(x2, r0); |
2320 | | x3 = _mm_maddubs_epi16(x3, r0); |
2321 | | x2 = _mm_hadd_epi16(x2, x3); |
2322 | | x2 = _mm_hadd_epi16(x2, _mm_setzero_si128()); |
2323 | | |
2324 | | /* give results back */ |
2325 | | _mm_storel_epi64((__m128i *) &dst[x], x2); |
2326 | | |
2327 | | } |
2328 | | src += srcstride; |
2329 | | dst += dststride; |
2330 | | } |
2331 | | } |
2332 | | |
2333 | | } |
2334 | | static void ff_hevc_put_hevc_qpel_h_3_sse(int16_t *dst, ptrdiff_t dststride, |
2335 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2336 | | int16_t* mcbuffer) { |
2337 | | int x, y; |
2338 | | uint8_t *src = _src; |
2339 | | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2340 | | __m128i x1, r0, x2, x3, x4, x5; |
2341 | | |
2342 | | r0 = _mm_set_epi8(-1, 4, -10, 58, 17, -5, 1, 0, -1, 4, -10, 58, 17, -5, 1, |
2343 | | 0); |
2344 | | |
2345 | | if(!(width & 7)){ |
2346 | | for (y = 0; y < height; y++) { |
2347 | | for (x = 0; x < width; x += 8) { |
2348 | | /* load data in register */ |
2349 | | x1 = _mm_loadu_si128((__m128i *) &src[x - 2]); |
2350 | | x1 = _mm_slli_si128(x1, 1); |
2351 | | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2352 | | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2353 | | _mm_srli_si128(x1, 3)); |
2354 | | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
2355 | | _mm_srli_si128(x1, 5)); |
2356 | | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
2357 | | _mm_srli_si128(x1, 7)); |
2358 | | |
2359 | | /* PMADDUBSW then PMADDW */ |
2360 | | x2 = _mm_maddubs_epi16(x2, r0); |
2361 | | x3 = _mm_maddubs_epi16(x3, r0); |
2362 | | x4 = _mm_maddubs_epi16(x4, r0); |
2363 | | x5 = _mm_maddubs_epi16(x5, r0); |
2364 | | x2 = _mm_hadd_epi16(x2, x3); |
2365 | | x4 = _mm_hadd_epi16(x4, x5); |
2366 | | x2 = _mm_hadd_epi16(x2, x4); |
2367 | | /* give results back */ |
2368 | | _mm_store_si128((__m128i *) &dst[x], |
2369 | | _mm_srli_si128(x2, BIT_DEPTH - 8)); |
2370 | | } |
2371 | | src += srcstride; |
2372 | | dst += dststride; |
2373 | | } |
2374 | | }else{ |
2375 | | for (y = 0; y < height; y ++) { |
2376 | | for(x=0;x<width;x+=4){ |
2377 | | /* load data in register */ |
2378 | | x1 = _mm_loadu_si128((__m128i *) &src[x-2]); |
2379 | | x1 = _mm_slli_si128(x1, 1); |
2380 | | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2381 | | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2382 | | _mm_srli_si128(x1, 3)); |
2383 | | |
2384 | | /* PMADDUBSW then PMADDW */ |
2385 | | x2 = _mm_maddubs_epi16(x2, r0); |
2386 | | x3 = _mm_maddubs_epi16(x3, r0); |
2387 | | x2 = _mm_hadd_epi16(x2, x3); |
2388 | | x2 = _mm_hadd_epi16(x2, _mm_setzero_si128()); |
2389 | | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
2390 | | /* give results back */ |
2391 | | _mm_storel_epi64((__m128i *) &dst[x], x2); |
2392 | | |
2393 | | } |
2394 | | src += srcstride; |
2395 | | dst += dststride; |
2396 | | } |
2397 | | } |
2398 | | } |
2399 | | #endif |
2400 | | |
2401 | | void ff_hevc_put_hevc_qpel_h_3_8_sse(int16_t *dst, ptrdiff_t dststride, |
2402 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2403 | 219k | int16_t* mcbuffer) { |
2404 | 219k | int x, y; |
2405 | 219k | const uint8_t *src = _src; |
2406 | 219k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2407 | 219k | __m128i x1, r0, x2, x3, x4, x5; |
2408 | | |
2409 | 219k | r0 = _mm_set_epi8(-1, 4, -10, 58, 17, -5, 1, 0, -1, 4, -10, 58, 17, -5, 1, |
2410 | 219k | 0); |
2411 | | |
2412 | 219k | if(!(width & 7)){ |
2413 | 2.04M | for (y = 0; y < height; y++) { |
2414 | 5.06M | for (x = 0; x < width; x += 8) { |
2415 | | /* load data in register */ |
2416 | 3.19M | x1 = _mm_loadu_si128((__m128i *) &src[x - 2]); |
2417 | 3.19M | x1 = _mm_slli_si128(x1, 1); |
2418 | 3.19M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2419 | 3.19M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2420 | 3.19M | _mm_srli_si128(x1, 3)); |
2421 | 3.19M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
2422 | 3.19M | _mm_srli_si128(x1, 5)); |
2423 | 3.19M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
2424 | 3.19M | _mm_srli_si128(x1, 7)); |
2425 | | |
2426 | | /* PMADDUBSW then PMADDW */ |
2427 | 3.19M | x2 = _mm_maddubs_epi16(x2, r0); |
2428 | 3.19M | x3 = _mm_maddubs_epi16(x3, r0); |
2429 | 3.19M | x4 = _mm_maddubs_epi16(x4, r0); |
2430 | 3.19M | x5 = _mm_maddubs_epi16(x5, r0); |
2431 | 3.19M | x2 = _mm_hadd_epi16(x2, x3); |
2432 | 3.19M | x4 = _mm_hadd_epi16(x4, x5); |
2433 | 3.19M | x2 = _mm_hadd_epi16(x2, x4); |
2434 | | /* give results back */ |
2435 | 3.19M | _mm_store_si128((__m128i *) &dst[x],x2); |
2436 | 3.19M | } |
2437 | 1.86M | src += srcstride; |
2438 | 1.86M | dst += dststride; |
2439 | 1.86M | } |
2440 | 182k | }else{ |
2441 | 355k | for (y = 0; y < height; y ++) { |
2442 | 695k | for(x=0;x<width;x+=4){ |
2443 | | /* load data in register */ |
2444 | 376k | x1 = _mm_loadu_si128((__m128i *) &src[x-2]); |
2445 | 376k | x1 = _mm_slli_si128(x1, 1); |
2446 | 376k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
2447 | 376k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
2448 | 376k | _mm_srli_si128(x1, 3)); |
2449 | | |
2450 | | /* PMADDUBSW then PMADDW */ |
2451 | 376k | x2 = _mm_maddubs_epi16(x2, r0); |
2452 | 376k | x3 = _mm_maddubs_epi16(x3, r0); |
2453 | 376k | x2 = _mm_hadd_epi16(x2, x3); |
2454 | 376k | x2 = _mm_hadd_epi16(x2, _mm_setzero_si128()); |
2455 | | /* give results back */ |
2456 | 376k | _mm_storel_epi64((__m128i *) &dst[x], x2); |
2457 | | |
2458 | 376k | } |
2459 | 319k | src += srcstride; |
2460 | 319k | dst += dststride; |
2461 | 319k | } |
2462 | 36.7k | } |
2463 | 219k | } |
2464 | | /** |
2465 | | for column MC treatment, we will calculate 8 pixels at the same time by multiplying the values |
2466 | | of each row. |
2467 | | |
2468 | | */ |
2469 | | void ff_hevc_put_hevc_qpel_v_1_8_sse(int16_t *dst, ptrdiff_t dststride, |
2470 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2471 | 217k | int16_t* mcbuffer) { |
2472 | 217k | int x, y; |
2473 | 217k | uint8_t *src = (uint8_t*) _src; |
2474 | 217k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2475 | 217k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, r0, r1, r2; |
2476 | 217k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
2477 | 217k | r1 = _mm_set_epi16(0, 1, -5, 17, 58, -10, 4, -1); |
2478 | | |
2479 | 217k | if(!(width & 15)){ |
2480 | 41.3k | x8 = _mm_setzero_si128(); |
2481 | 761k | for (y = 0; y < height; y++) { |
2482 | 1.71M | for (x = 0; x < width; x += 16) { |
2483 | | /* check if memory needs to be reloaded */ |
2484 | 995k | x1 = _mm_loadu_si128((__m128i *) &src[x - 3 * srcstride]); |
2485 | 995k | x2 = _mm_loadu_si128((__m128i *) &src[x - 2 * srcstride]); |
2486 | 995k | x3 = _mm_loadu_si128((__m128i *) &src[x - srcstride]); |
2487 | 995k | x4 = _mm_loadu_si128((__m128i *) &src[x]); |
2488 | 995k | x5 = _mm_loadu_si128((__m128i *) &src[x + srcstride]); |
2489 | 995k | x6 = _mm_loadu_si128((__m128i *) &src[x + 2 * srcstride]); |
2490 | 995k | x7 = _mm_loadu_si128((__m128i *) &src[x + 3 * srcstride]); |
2491 | | |
2492 | 995k | t1 = _mm_unpacklo_epi8(x1,x8); |
2493 | 995k | t2 = _mm_unpacklo_epi8(x2, x8); |
2494 | 995k | t3 = _mm_unpacklo_epi8(x3, x8); |
2495 | 995k | t4 = _mm_unpacklo_epi8(x4, x8); |
2496 | 995k | t5 = _mm_unpacklo_epi8(x5, x8); |
2497 | 995k | t6 = _mm_unpacklo_epi8(x6, x8); |
2498 | 995k | t7 = _mm_unpacklo_epi8(x7, x8); |
2499 | | |
2500 | 995k | x1 = _mm_unpackhi_epi8(x1,x8); |
2501 | 995k | x2 = _mm_unpackhi_epi8(x2, x8); |
2502 | 995k | x3 = _mm_unpackhi_epi8(x3, x8); |
2503 | 995k | x4 = _mm_unpackhi_epi8(x4, x8); |
2504 | 995k | x5 = _mm_unpackhi_epi8(x5, x8); |
2505 | 995k | x6 = _mm_unpackhi_epi8(x6, x8); |
2506 | 995k | x7 = _mm_unpackhi_epi8(x7, x8); |
2507 | | |
2508 | | /* multiply by correct value : */ |
2509 | 995k | r0 = _mm_mullo_epi16(t1, |
2510 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 0))); |
2511 | 995k | r2 = _mm_mullo_epi16(x1, |
2512 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 0))); |
2513 | 995k | r0 = _mm_adds_epi16(r0, |
2514 | 995k | _mm_mullo_epi16(t2, |
2515 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 1)))); |
2516 | 995k | r2 = _mm_adds_epi16(r2, |
2517 | 995k | _mm_mullo_epi16(x2, |
2518 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 1)))); |
2519 | 995k | r0 = _mm_adds_epi16(r0, |
2520 | 995k | _mm_mullo_epi16(t3, |
2521 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
2522 | 995k | r2 = _mm_adds_epi16(r2, |
2523 | 995k | _mm_mullo_epi16(x3, |
2524 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
2525 | | |
2526 | 995k | r0 = _mm_adds_epi16(r0, |
2527 | 995k | _mm_mullo_epi16(t4, |
2528 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
2529 | 995k | r2 = _mm_adds_epi16(r2, |
2530 | 995k | _mm_mullo_epi16(x4, |
2531 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
2532 | | |
2533 | 995k | r0 = _mm_adds_epi16(r0, |
2534 | 995k | _mm_mullo_epi16(t5, |
2535 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
2536 | 995k | r2 = _mm_adds_epi16(r2, |
2537 | 995k | _mm_mullo_epi16(x5, |
2538 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
2539 | | |
2540 | 995k | r0 = _mm_adds_epi16(r0, |
2541 | 995k | _mm_mullo_epi16(t6, |
2542 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
2543 | 995k | r2 = _mm_adds_epi16(r2, |
2544 | 995k | _mm_mullo_epi16(x6, |
2545 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
2546 | | |
2547 | 995k | r0 = _mm_adds_epi16(r0, |
2548 | 995k | _mm_mullo_epi16(t7, |
2549 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
2550 | 995k | r2 = _mm_adds_epi16(r2, |
2551 | 995k | _mm_mullo_epi16(x7, |
2552 | 995k | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
2553 | | |
2554 | | |
2555 | | /* give results back */ |
2556 | 995k | _mm_store_si128((__m128i *) &dst[x],r0); |
2557 | 995k | _mm_store_si128((__m128i *) &dst[x + 8],r2); |
2558 | 995k | } |
2559 | 720k | src += srcstride; |
2560 | 720k | dst += dststride; |
2561 | 720k | } |
2562 | | |
2563 | 176k | }else{ |
2564 | 176k | x = 0; |
2565 | 176k | x8 = _mm_setzero_si128(); |
2566 | 176k | t8 = _mm_setzero_si128(); |
2567 | 1.63M | for (y = 0; y < height; y ++) { |
2568 | 4.18M | for(x=0;x<width;x+=4){ |
2569 | | /* load data in register */ |
2570 | 2.72M | x1 = _mm_loadl_epi64((__m128i *) &src[x-(3 * srcstride)]); |
2571 | 2.72M | x2 = _mm_loadl_epi64((__m128i *) &src[x-(2 * srcstride)]); |
2572 | 2.72M | x3 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
2573 | 2.72M | x4 = _mm_loadl_epi64((__m128i *) &src[x]); |
2574 | 2.72M | x5 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
2575 | 2.72M | x6 = _mm_loadl_epi64((__m128i *) &src[x+(2 * srcstride)]); |
2576 | 2.72M | x7 = _mm_loadl_epi64((__m128i *) &src[x+(3 * srcstride)]); |
2577 | | |
2578 | | |
2579 | | |
2580 | 2.72M | x1 = _mm_unpacklo_epi8(x1, t8); |
2581 | 2.72M | x2 = _mm_unpacklo_epi8(x2, t8); |
2582 | 2.72M | x3 = _mm_unpacklo_epi8(x3, t8); |
2583 | 2.72M | x4 = _mm_unpacklo_epi8(x4, t8); |
2584 | 2.72M | x5 = _mm_unpacklo_epi8(x5, t8); |
2585 | 2.72M | x6 = _mm_unpacklo_epi8(x6, t8); |
2586 | 2.72M | x7 = _mm_unpacklo_epi8(x7, t8); |
2587 | | |
2588 | | |
2589 | 2.72M | r0 = _mm_mullo_epi16(x1, _mm_set1_epi16(_mm_extract_epi16(r1, 0))); |
2590 | | |
2591 | 2.72M | r0 = _mm_adds_epi16(r0, |
2592 | 2.72M | _mm_mullo_epi16(x2, |
2593 | 2.72M | _mm_set1_epi16(_mm_extract_epi16(r1, 1)))); |
2594 | | |
2595 | | |
2596 | 2.72M | r0 = _mm_adds_epi16(r0, |
2597 | 2.72M | _mm_mullo_epi16(x3, |
2598 | 2.72M | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
2599 | | |
2600 | 2.72M | r0 = _mm_adds_epi16(r0, |
2601 | 2.72M | _mm_mullo_epi16(x4, |
2602 | 2.72M | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
2603 | | |
2604 | 2.72M | r0 = _mm_adds_epi16(r0, |
2605 | 2.72M | _mm_mullo_epi16(x5, |
2606 | 2.72M | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
2607 | | |
2608 | | |
2609 | 2.72M | r0 = _mm_adds_epi16(r0, |
2610 | 2.72M | _mm_mullo_epi16(x6, |
2611 | 2.72M | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
2612 | | |
2613 | | |
2614 | 2.72M | r0 = _mm_adds_epi16(r0, |
2615 | 2.72M | _mm_mullo_epi16(x7, |
2616 | 2.72M | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
2617 | | |
2618 | | /* give results back */ |
2619 | 2.72M | _mm_storel_epi64((__m128i *) &dst[x], r0); |
2620 | 2.72M | } |
2621 | 1.45M | src += srcstride; |
2622 | 1.45M | dst += dststride; |
2623 | 1.45M | } |
2624 | 176k | } |
2625 | 217k | } |
2626 | | |
2627 | | #if 0 |
2628 | | void ff_hevc_put_hevc_qpel_v_1_10_sse4(int16_t *dst, ptrdiff_t dststride, |
2629 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2630 | | int16_t* mcbuffer) { |
2631 | | int x, y; |
2632 | | uint16_t *src = (uint16_t*) _src; |
2633 | | ptrdiff_t srcstride = _srcstride >> 1; |
2634 | | __m128i x1, x2, x3, x4, x5, x6, x7, r1; |
2635 | | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
2636 | | |
2637 | | t7= _mm_set1_epi32(1); |
2638 | | t6= _mm_set1_epi32(-5); |
2639 | | t5= _mm_set1_epi32(17); |
2640 | | t4= _mm_set1_epi32(58); |
2641 | | t3= _mm_set1_epi32(-10); |
2642 | | t2= _mm_set1_epi32(4); |
2643 | | t1= _mm_set1_epi32(-1); |
2644 | | t8= _mm_setzero_si128(); |
2645 | | |
2646 | | for (y = 0; y < height; y ++) { |
2647 | | for(x=0;x<width;x+=4){ |
2648 | | /* load data in register */ |
2649 | | x1 = _mm_loadl_epi64((__m128i *) &src[x-(3 * srcstride)]); |
2650 | | x2 = _mm_loadl_epi64((__m128i *) &src[x-(2 * srcstride)]); |
2651 | | x3 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
2652 | | x4 = _mm_loadl_epi64((__m128i *) &src[x]); |
2653 | | x5 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
2654 | | x6 = _mm_loadl_epi64((__m128i *) &src[x+(2 * srcstride)]); |
2655 | | x7 = _mm_loadl_epi64((__m128i *) &src[x+(3 * srcstride)]); |
2656 | | |
2657 | | |
2658 | | x1 = _mm_unpacklo_epi16(x1, t8); |
2659 | | x2 = _mm_unpacklo_epi16(x2, t8); |
2660 | | x3 = _mm_unpacklo_epi16(x3, t8); |
2661 | | x4 = _mm_unpacklo_epi16(x4, t8); |
2662 | | x5 = _mm_unpacklo_epi16(x5, t8); |
2663 | | x6 = _mm_unpacklo_epi16(x6, t8); |
2664 | | x7 = _mm_unpacklo_epi16(x7, t8); |
2665 | | |
2666 | | |
2667 | | r1 = _mm_mullo_epi32(x1,t1); |
2668 | | |
2669 | | r1 = _mm_add_epi32(r1, |
2670 | | _mm_mullo_epi32(x2,t2)); |
2671 | | |
2672 | | |
2673 | | r1 = _mm_add_epi32(r1, |
2674 | | _mm_mullo_epi32(x3,t3)); |
2675 | | |
2676 | | r1 = _mm_add_epi32(r1, |
2677 | | _mm_mullo_epi32(x4,t4)); |
2678 | | |
2679 | | r1 = _mm_add_epi32(r1, |
2680 | | _mm_mullo_epi32(x5,t5)); |
2681 | | |
2682 | | |
2683 | | r1 = _mm_add_epi32(r1, |
2684 | | _mm_mullo_epi32(x6,t6)); |
2685 | | |
2686 | | |
2687 | | r1 = _mm_add_epi32(r1, _mm_mullo_epi32(x7,t7)); |
2688 | | r1 = _mm_srai_epi32(r1,2); //bit depth - 8 |
2689 | | |
2690 | | |
2691 | | r1 = _mm_packs_epi32(r1,t8); |
2692 | | |
2693 | | // give results back |
2694 | | _mm_storel_epi64((__m128i *) (dst + x), r1); |
2695 | | } |
2696 | | src += srcstride; |
2697 | | dst += dststride; |
2698 | | } |
2699 | | |
2700 | | } |
2701 | | #endif |
2702 | | |
2703 | | |
2704 | | |
2705 | | void ff_hevc_put_hevc_qpel_v_2_8_sse(int16_t *dst, ptrdiff_t dststride, |
2706 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2707 | 138k | int16_t* mcbuffer) { |
2708 | 138k | int x, y; |
2709 | 138k | uint8_t *src = (uint8_t*) _src; |
2710 | 138k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2711 | 138k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, r0, r1, r2; |
2712 | 138k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
2713 | 138k | r1 = _mm_set_epi16(-1, 4, -11, 40, 40, -11, 4, -1); |
2714 | | |
2715 | 138k | if(!(width & 15)){ |
2716 | 427k | for (y = 0; y < height; y++) { |
2717 | 998k | for (x = 0; x < width; x += 16) { |
2718 | 593k | r0 = _mm_setzero_si128(); |
2719 | | /* check if memory needs to be reloaded */ |
2720 | 593k | x1 = _mm_loadu_si128((__m128i *) &src[x - 3 * srcstride]); |
2721 | 593k | x2 = _mm_loadu_si128((__m128i *) &src[x - 2 * srcstride]); |
2722 | 593k | x3 = _mm_loadu_si128((__m128i *) &src[x - srcstride]); |
2723 | 593k | x4 = _mm_loadu_si128((__m128i *) &src[x]); |
2724 | 593k | x5 = _mm_loadu_si128((__m128i *) &src[x + srcstride]); |
2725 | 593k | x6 = _mm_loadu_si128((__m128i *) &src[x + 2 * srcstride]); |
2726 | 593k | x7 = _mm_loadu_si128((__m128i *) &src[x + 3 * srcstride]); |
2727 | 593k | x8 = _mm_loadu_si128((__m128i *) &src[x + 4 * srcstride]); |
2728 | | |
2729 | 593k | t1 = _mm_unpacklo_epi8(x1, r0); |
2730 | 593k | t2 = _mm_unpacklo_epi8(x2, r0); |
2731 | 593k | t3 = _mm_unpacklo_epi8(x3, r0); |
2732 | 593k | t4 = _mm_unpacklo_epi8(x4, r0); |
2733 | 593k | t5 = _mm_unpacklo_epi8(x5, r0); |
2734 | 593k | t6 = _mm_unpacklo_epi8(x6, r0); |
2735 | 593k | t7 = _mm_unpacklo_epi8(x7, r0); |
2736 | 593k | t8 = _mm_unpacklo_epi8(x8, r0); |
2737 | | |
2738 | 593k | x1 = _mm_unpackhi_epi8(x1, r0); |
2739 | 593k | x2 = _mm_unpackhi_epi8(x2, r0); |
2740 | 593k | x3 = _mm_unpackhi_epi8(x3, r0); |
2741 | 593k | x4 = _mm_unpackhi_epi8(x4, r0); |
2742 | 593k | x5 = _mm_unpackhi_epi8(x5, r0); |
2743 | 593k | x6 = _mm_unpackhi_epi8(x6, r0); |
2744 | 593k | x7 = _mm_unpackhi_epi8(x7, r0); |
2745 | 593k | x8 = _mm_unpackhi_epi8(x8, r0); |
2746 | | |
2747 | | /* multiply by correct value : */ |
2748 | 593k | r0 = _mm_mullo_epi16(t1, |
2749 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 0))); |
2750 | 593k | r2 = _mm_mullo_epi16(x1, |
2751 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 0))); |
2752 | 593k | r0 = _mm_adds_epi16(r0, |
2753 | 593k | _mm_mullo_epi16(t2, |
2754 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 1)))); |
2755 | 593k | r2 = _mm_adds_epi16(r2, |
2756 | 593k | _mm_mullo_epi16(x2, |
2757 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 1)))); |
2758 | 593k | r0 = _mm_adds_epi16(r0, |
2759 | 593k | _mm_mullo_epi16(t3, |
2760 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
2761 | 593k | r2 = _mm_adds_epi16(r2, |
2762 | 593k | _mm_mullo_epi16(x3, |
2763 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
2764 | | |
2765 | 593k | r0 = _mm_adds_epi16(r0, |
2766 | 593k | _mm_mullo_epi16(t4, |
2767 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
2768 | 593k | r2 = _mm_adds_epi16(r2, |
2769 | 593k | _mm_mullo_epi16(x4, |
2770 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
2771 | | |
2772 | 593k | r0 = _mm_adds_epi16(r0, |
2773 | 593k | _mm_mullo_epi16(t5, |
2774 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
2775 | 593k | r2 = _mm_adds_epi16(r2, |
2776 | 593k | _mm_mullo_epi16(x5, |
2777 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
2778 | | |
2779 | 593k | r0 = _mm_adds_epi16(r0, |
2780 | 593k | _mm_mullo_epi16(t6, |
2781 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
2782 | 593k | r2 = _mm_adds_epi16(r2, |
2783 | 593k | _mm_mullo_epi16(x6, |
2784 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
2785 | | |
2786 | 593k | r0 = _mm_adds_epi16(r0, |
2787 | 593k | _mm_mullo_epi16(t7, |
2788 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
2789 | 593k | r2 = _mm_adds_epi16(r2, |
2790 | 593k | _mm_mullo_epi16(x7, |
2791 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
2792 | | |
2793 | 593k | r0 = _mm_adds_epi16(r0, |
2794 | 593k | _mm_mullo_epi16(t8, |
2795 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
2796 | 593k | r2 = _mm_adds_epi16(r2, |
2797 | 593k | _mm_mullo_epi16(x8, |
2798 | 593k | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
2799 | | |
2800 | | /* give results back */ |
2801 | 593k | _mm_store_si128((__m128i *) &dst[x],r0); |
2802 | 593k | _mm_store_si128((__m128i *) &dst[x + 8],r2); |
2803 | 593k | } |
2804 | 405k | src += srcstride; |
2805 | 405k | dst += dststride; |
2806 | 405k | } |
2807 | 115k | }else{ |
2808 | 115k | x = 0; |
2809 | 1.01M | for (y = 0; y < height; y ++) { |
2810 | 2.56M | for(x=0;x<width;x+=4){ |
2811 | 1.66M | r0 = _mm_setzero_si128(); |
2812 | | /* load data in register */ |
2813 | 1.66M | x1 = _mm_loadl_epi64((__m128i *) &src[x - 3 * srcstride]); |
2814 | 1.66M | x2 = _mm_loadl_epi64((__m128i *) &src[x-2 * srcstride]); |
2815 | 1.66M | x3 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
2816 | 1.66M | x4 = _mm_loadl_epi64((__m128i *) &src[x]); |
2817 | 1.66M | x5 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
2818 | 1.66M | x6 = _mm_loadl_epi64((__m128i *) &src[x+2 * srcstride]); |
2819 | 1.66M | x7 = _mm_loadl_epi64((__m128i *) &src[x+3 * srcstride]); |
2820 | 1.66M | x8 = _mm_loadl_epi64((__m128i *) &src[x + 4 * srcstride]); |
2821 | | |
2822 | 1.66M | x1 = _mm_unpacklo_epi8(x1,r0); |
2823 | 1.66M | x2 = _mm_unpacklo_epi8(x2, r0); |
2824 | 1.66M | x3 = _mm_unpacklo_epi8(x3, r0); |
2825 | 1.66M | x4 = _mm_unpacklo_epi8(x4, r0); |
2826 | 1.66M | x5 = _mm_unpacklo_epi8(x5, r0); |
2827 | 1.66M | x6 = _mm_unpacklo_epi8(x6, r0); |
2828 | 1.66M | x7 = _mm_unpacklo_epi8(x7, r0); |
2829 | 1.66M | x8 = _mm_unpacklo_epi8(x8, r0); |
2830 | | |
2831 | | |
2832 | 1.66M | r0 = _mm_mullo_epi16(x1, _mm_set1_epi16(_mm_extract_epi16(r1, 0))); |
2833 | | |
2834 | 1.66M | r0 = _mm_adds_epi16(r0, |
2835 | 1.66M | _mm_mullo_epi16(x2, |
2836 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 1)))); |
2837 | | |
2838 | | |
2839 | 1.66M | r0 = _mm_adds_epi16(r0, |
2840 | 1.66M | _mm_mullo_epi16(x3, |
2841 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
2842 | | |
2843 | | |
2844 | 1.66M | r0 = _mm_adds_epi16(r0, |
2845 | 1.66M | _mm_mullo_epi16(x4, |
2846 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
2847 | | |
2848 | | |
2849 | 1.66M | r0 = _mm_adds_epi16(r0, |
2850 | 1.66M | _mm_mullo_epi16(x5, |
2851 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
2852 | | |
2853 | | |
2854 | 1.66M | r0 = _mm_adds_epi16(r0, |
2855 | 1.66M | _mm_mullo_epi16(x6, |
2856 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
2857 | | |
2858 | | |
2859 | 1.66M | r0 = _mm_adds_epi16(r0, |
2860 | 1.66M | _mm_mullo_epi16(x7, |
2861 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
2862 | | |
2863 | | |
2864 | 1.66M | r0 = _mm_adds_epi16(r0, |
2865 | 1.66M | _mm_mullo_epi16(x8, |
2866 | 1.66M | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
2867 | | |
2868 | | |
2869 | | /* give results back */ |
2870 | 1.66M | _mm_storel_epi64((__m128i *) &dst[x], r0); |
2871 | | |
2872 | 1.66M | } |
2873 | 894k | src += srcstride; |
2874 | 894k | dst += dststride; |
2875 | 894k | } |
2876 | 115k | } |
2877 | 138k | } |
2878 | | |
2879 | | #if 0 |
2880 | | void ff_hevc_put_hevc_qpel_v_2_10_sse(int16_t *dst, ptrdiff_t dststride, |
2881 | | cosnt uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2882 | | int16_t* mcbuffer) { |
2883 | | int x, y; |
2884 | | uint16_t *src = (uint16_t*) _src; |
2885 | | ptrdiff_t srcstride = _srcstride >> 1; |
2886 | | __m128i x1, x2, x3, x4, x5, x6, x7, x8, r0, r1, r2; |
2887 | | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
2888 | | r1 = _mm_set_epi16(-1, 4, -11, 40, 40, -11, 4, -1); |
2889 | | |
2890 | | t1= _mm_set1_epi32(-1); |
2891 | | t2= _mm_set1_epi32(4); |
2892 | | t3= _mm_set1_epi32(-11); |
2893 | | t4= _mm_set1_epi32(40); |
2894 | | t5= _mm_set1_epi32(40); |
2895 | | t6= _mm_set1_epi32(-11); |
2896 | | t7= _mm_set1_epi32(4); |
2897 | | t8= _mm_set1_epi32(-1); |
2898 | | |
2899 | | { |
2900 | | x = 0; |
2901 | | r0 = _mm_setzero_si128(); |
2902 | | for (y = 0; y < height; y ++) { |
2903 | | for(x=0;x<width;x+=4){ |
2904 | | |
2905 | | /* load data in register */ |
2906 | | x1 = _mm_loadl_epi64((__m128i *) &src[x - 3 * srcstride]); |
2907 | | x2 = _mm_loadl_epi64((__m128i *) &src[x-2 * srcstride]); |
2908 | | x3 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
2909 | | x4 = _mm_loadl_epi64((__m128i *) &src[x]); |
2910 | | x5 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
2911 | | x6 = _mm_loadl_epi64((__m128i *) &src[x+2 * srcstride]); |
2912 | | x7 = _mm_loadl_epi64((__m128i *) &src[x+3 * srcstride]); |
2913 | | x8 = _mm_loadl_epi64((__m128i *) &src[x + 4 * srcstride]); |
2914 | | |
2915 | | x1 = _mm_unpacklo_epi16(x1, r0); |
2916 | | x2 = _mm_unpacklo_epi16(x2, r0); |
2917 | | x3 = _mm_unpacklo_epi16(x3, r0); |
2918 | | x4 = _mm_unpacklo_epi16(x4, r0); |
2919 | | x5 = _mm_unpacklo_epi16(x5, r0); |
2920 | | x6 = _mm_unpacklo_epi16(x6, r0); |
2921 | | x7 = _mm_unpacklo_epi16(x7, r0); |
2922 | | x8 = _mm_unpacklo_epi16(x8, r0); |
2923 | | |
2924 | | |
2925 | | r1 = _mm_mullo_epi32(x1, t1); |
2926 | | |
2927 | | r1 = _mm_add_epi32(r1, |
2928 | | _mm_mullo_epi32(x2,t2)); |
2929 | | |
2930 | | |
2931 | | r1 = _mm_add_epi32(r1, |
2932 | | _mm_mullo_epi32(x3,t3)); |
2933 | | |
2934 | | |
2935 | | r1 = _mm_add_epi32(r1, |
2936 | | _mm_mullo_epi32(x4,t4)); |
2937 | | |
2938 | | |
2939 | | r1 = _mm_add_epi32(r1, |
2940 | | _mm_mullo_epi32(x5,t5)); |
2941 | | |
2942 | | |
2943 | | r1 = _mm_add_epi32(r1, |
2944 | | _mm_mullo_epi32(x6,t6)); |
2945 | | |
2946 | | |
2947 | | r1 = _mm_add_epi32(r1, |
2948 | | _mm_mullo_epi32(x7,t7)); |
2949 | | |
2950 | | |
2951 | | r1 = _mm_add_epi32(r1, |
2952 | | _mm_mullo_epi32(x8,t8)); |
2953 | | |
2954 | | |
2955 | | r1= _mm_srai_epi32(r1,2); //bit depth - 8 |
2956 | | |
2957 | | r1= _mm_packs_epi32(r1,t8); |
2958 | | |
2959 | | /* give results back */ |
2960 | | _mm_storel_epi64((__m128i *) (dst+x), r1); |
2961 | | |
2962 | | } |
2963 | | src += srcstride; |
2964 | | dst += dststride; |
2965 | | } |
2966 | | } |
2967 | | } |
2968 | | #endif |
2969 | | |
2970 | | #if 0 |
2971 | | static void ff_hevc_put_hevc_qpel_v_3_sse(int16_t *dst, ptrdiff_t dststride, |
2972 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
2973 | | int16_t* mcbuffer) { |
2974 | | int x, y; |
2975 | | uint8_t *src = (uint8_t*) _src; |
2976 | | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
2977 | | __m128i x1, x2, x3, x4, x5, x6, x7, x8, r0, r1, r2; |
2978 | | __m128i t2, t3, t4, t5, t6, t7, t8; |
2979 | | r1 = _mm_set_epi16(-1, 4, -10, 58, 17, -5, 1, 0); |
2980 | | |
2981 | | if(!(width & 15)){ |
2982 | | for (y = 0; y < height; y++) { |
2983 | | for (x = 0; x < width; x += 16) { |
2984 | | /* check if memory needs to be reloaded */ |
2985 | | x1 = _mm_setzero_si128(); |
2986 | | x2 = _mm_loadu_si128((__m128i *) &src[x - 2 * srcstride]); |
2987 | | x3 = _mm_loadu_si128((__m128i *) &src[x - srcstride]); |
2988 | | x4 = _mm_loadu_si128((__m128i *) &src[x]); |
2989 | | x5 = _mm_loadu_si128((__m128i *) &src[x + srcstride]); |
2990 | | x6 = _mm_loadu_si128((__m128i *) &src[x + 2 * srcstride]); |
2991 | | x7 = _mm_loadu_si128((__m128i *) &src[x + 3 * srcstride]); |
2992 | | x8 = _mm_loadu_si128((__m128i *) &src[x + 4 * srcstride]); |
2993 | | |
2994 | | t2 = _mm_unpacklo_epi8(x2, x1); |
2995 | | t3 = _mm_unpacklo_epi8(x3, x1); |
2996 | | t4 = _mm_unpacklo_epi8(x4, x1); |
2997 | | t5 = _mm_unpacklo_epi8(x5, x1); |
2998 | | t6 = _mm_unpacklo_epi8(x6, x1); |
2999 | | t7 = _mm_unpacklo_epi8(x7, x1); |
3000 | | t8 = _mm_unpacklo_epi8(x8, x1); |
3001 | | |
3002 | | x2 = _mm_unpackhi_epi8(x2, x1); |
3003 | | x3 = _mm_unpackhi_epi8(x3, x1); |
3004 | | x4 = _mm_unpackhi_epi8(x4, x1); |
3005 | | x5 = _mm_unpackhi_epi8(x5, x1); |
3006 | | x6 = _mm_unpackhi_epi8(x6, x1); |
3007 | | x7 = _mm_unpackhi_epi8(x7, x1); |
3008 | | x8 = _mm_unpackhi_epi8(x8, x1); |
3009 | | |
3010 | | /* multiply by correct value : */ |
3011 | | r0 = _mm_mullo_epi16(t2, |
3012 | | _mm_set1_epi16(_mm_extract_epi16(r1, 1))); |
3013 | | r2 = _mm_mullo_epi16(x2, |
3014 | | _mm_set1_epi16(_mm_extract_epi16(r1, 1))); |
3015 | | |
3016 | | r0 = _mm_adds_epi16(r0, |
3017 | | _mm_mullo_epi16(t3, |
3018 | | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
3019 | | r2 = _mm_adds_epi16(r2, |
3020 | | _mm_mullo_epi16(x3, |
3021 | | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
3022 | | |
3023 | | r0 = _mm_adds_epi16(r0, |
3024 | | _mm_mullo_epi16(t4, |
3025 | | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
3026 | | r2 = _mm_adds_epi16(r2, |
3027 | | _mm_mullo_epi16(x4, |
3028 | | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
3029 | | |
3030 | | r0 = _mm_adds_epi16(r0, |
3031 | | _mm_mullo_epi16(t5, |
3032 | | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
3033 | | r2 = _mm_adds_epi16(r2, |
3034 | | _mm_mullo_epi16(x5, |
3035 | | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
3036 | | |
3037 | | r0 = _mm_adds_epi16(r0, |
3038 | | _mm_mullo_epi16(t6, |
3039 | | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
3040 | | r2 = _mm_adds_epi16(r2, |
3041 | | _mm_mullo_epi16(x6, |
3042 | | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
3043 | | |
3044 | | r0 = _mm_adds_epi16(r0, |
3045 | | _mm_mullo_epi16(t7, |
3046 | | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
3047 | | r2 = _mm_adds_epi16(r2, |
3048 | | _mm_mullo_epi16(x7, |
3049 | | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
3050 | | |
3051 | | r0 = _mm_adds_epi16(r0, |
3052 | | _mm_mullo_epi16(t8, |
3053 | | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
3054 | | r2 = _mm_adds_epi16(r2, |
3055 | | _mm_mullo_epi16(x8, |
3056 | | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
3057 | | |
3058 | | /* give results back */ |
3059 | | _mm_store_si128((__m128i *) &dst[x], |
3060 | | _mm_srli_epi16(r0, BIT_DEPTH - 8)); |
3061 | | _mm_store_si128((__m128i *) &dst[x + 8], |
3062 | | _mm_srli_epi16(r2, BIT_DEPTH - 8)); |
3063 | | } |
3064 | | src += srcstride; |
3065 | | dst += dststride; |
3066 | | } |
3067 | | }else{ |
3068 | | x = 0; |
3069 | | for (y = 0; y < height; y ++) { |
3070 | | for(x=0;x<width;x+=4){ |
3071 | | r0 = _mm_set1_epi16(0); |
3072 | | /* load data in register */ |
3073 | | //x1 = _mm_setzero_si128(); |
3074 | | x2 = _mm_loadl_epi64((__m128i *) &src[x-2 * srcstride]); |
3075 | | x3 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
3076 | | x4 = _mm_loadl_epi64((__m128i *) &src[x]); |
3077 | | x5 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
3078 | | x6 = _mm_loadl_epi64((__m128i *) &src[x+2 * srcstride]); |
3079 | | x7 = _mm_loadl_epi64((__m128i *) &src[x+3 * srcstride]); |
3080 | | x8 = _mm_loadl_epi64((__m128i *) &src[x + 4 * srcstride]); |
3081 | | |
3082 | | x1 = _mm_unpacklo_epi8(x1,r0); |
3083 | | x2 = _mm_unpacklo_epi8(x2, r0); |
3084 | | x3 = _mm_unpacklo_epi8(x3, r0); |
3085 | | x4 = _mm_unpacklo_epi8(x4, r0); |
3086 | | x5 = _mm_unpacklo_epi8(x5, r0); |
3087 | | x6 = _mm_unpacklo_epi8(x6, r0); |
3088 | | x7 = _mm_unpacklo_epi8(x7, r0); |
3089 | | x8 = _mm_unpacklo_epi8(x8, r0); |
3090 | | |
3091 | | |
3092 | | r0 = _mm_mullo_epi16(x2, _mm_set1_epi16(_mm_extract_epi16(r1, 1))); |
3093 | | |
3094 | | |
3095 | | r0 = _mm_adds_epi16(r0, |
3096 | | _mm_mullo_epi16(x3, |
3097 | | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
3098 | | |
3099 | | |
3100 | | r0 = _mm_adds_epi16(r0, |
3101 | | _mm_mullo_epi16(x4, |
3102 | | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
3103 | | |
3104 | | |
3105 | | r0 = _mm_adds_epi16(r0, |
3106 | | _mm_mullo_epi16(x5, |
3107 | | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
3108 | | |
3109 | | |
3110 | | r0 = _mm_adds_epi16(r0, |
3111 | | _mm_mullo_epi16(x6, |
3112 | | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
3113 | | |
3114 | | |
3115 | | r0 = _mm_adds_epi16(r0, |
3116 | | _mm_mullo_epi16(x7, |
3117 | | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
3118 | | |
3119 | | |
3120 | | r0 = _mm_adds_epi16(r0, |
3121 | | _mm_mullo_epi16(x8, |
3122 | | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
3123 | | |
3124 | | |
3125 | | r0 = _mm_srli_epi16(r0, BIT_DEPTH - 8); |
3126 | | /* give results back */ |
3127 | | _mm_storel_epi64((__m128i *) &dst[x], r0); |
3128 | | |
3129 | | } |
3130 | | src += srcstride; |
3131 | | dst += dststride; |
3132 | | } |
3133 | | } |
3134 | | |
3135 | | } |
3136 | | #endif |
3137 | | |
3138 | | void ff_hevc_put_hevc_qpel_v_3_8_sse(int16_t *dst, ptrdiff_t dststride, |
3139 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
3140 | 251k | int16_t* mcbuffer) { |
3141 | 251k | int x, y; |
3142 | 251k | uint8_t *src = (uint8_t*) _src; |
3143 | 251k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
3144 | 251k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, r0, r1, r2; |
3145 | 251k | __m128i t2, t3, t4, t5, t6, t7, t8; |
3146 | 251k | r1 = _mm_set_epi16(-1, 4, -10, 58, 17, -5, 1, 0); |
3147 | | |
3148 | 251k | if(!(width & 15)){ |
3149 | 703k | for (y = 0; y < height; y++) { |
3150 | 1.69M | for (x = 0; x < width; x += 16) { |
3151 | | /* check if memory needs to be reloaded */ |
3152 | 1.02M | x1 = _mm_setzero_si128(); |
3153 | 1.02M | x2 = _mm_loadu_si128((__m128i *) &src[x - 2 * srcstride]); |
3154 | 1.02M | x3 = _mm_loadu_si128((__m128i *) &src[x - srcstride]); |
3155 | 1.02M | x4 = _mm_loadu_si128((__m128i *) &src[x]); |
3156 | 1.02M | x5 = _mm_loadu_si128((__m128i *) &src[x + srcstride]); |
3157 | 1.02M | x6 = _mm_loadu_si128((__m128i *) &src[x + 2 * srcstride]); |
3158 | 1.02M | x7 = _mm_loadu_si128((__m128i *) &src[x + 3 * srcstride]); |
3159 | 1.02M | x8 = _mm_loadu_si128((__m128i *) &src[x + 4 * srcstride]); |
3160 | | |
3161 | 1.02M | t2 = _mm_unpacklo_epi8(x2, x1); |
3162 | 1.02M | t3 = _mm_unpacklo_epi8(x3, x1); |
3163 | 1.02M | t4 = _mm_unpacklo_epi8(x4, x1); |
3164 | 1.02M | t5 = _mm_unpacklo_epi8(x5, x1); |
3165 | 1.02M | t6 = _mm_unpacklo_epi8(x6, x1); |
3166 | 1.02M | t7 = _mm_unpacklo_epi8(x7, x1); |
3167 | 1.02M | t8 = _mm_unpacklo_epi8(x8, x1); |
3168 | | |
3169 | 1.02M | x2 = _mm_unpackhi_epi8(x2, x1); |
3170 | 1.02M | x3 = _mm_unpackhi_epi8(x3, x1); |
3171 | 1.02M | x4 = _mm_unpackhi_epi8(x4, x1); |
3172 | 1.02M | x5 = _mm_unpackhi_epi8(x5, x1); |
3173 | 1.02M | x6 = _mm_unpackhi_epi8(x6, x1); |
3174 | 1.02M | x7 = _mm_unpackhi_epi8(x7, x1); |
3175 | 1.02M | x8 = _mm_unpackhi_epi8(x8, x1); |
3176 | | |
3177 | | /* multiply by correct value : */ |
3178 | 1.02M | r0 = _mm_mullo_epi16(t2, |
3179 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 1))); |
3180 | 1.02M | r2 = _mm_mullo_epi16(x2, |
3181 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 1))); |
3182 | | |
3183 | 1.02M | r0 = _mm_adds_epi16(r0, |
3184 | 1.02M | _mm_mullo_epi16(t3, |
3185 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
3186 | 1.02M | r2 = _mm_adds_epi16(r2, |
3187 | 1.02M | _mm_mullo_epi16(x3, |
3188 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
3189 | | |
3190 | 1.02M | r0 = _mm_adds_epi16(r0, |
3191 | 1.02M | _mm_mullo_epi16(t4, |
3192 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
3193 | 1.02M | r2 = _mm_adds_epi16(r2, |
3194 | 1.02M | _mm_mullo_epi16(x4, |
3195 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
3196 | | |
3197 | 1.02M | r0 = _mm_adds_epi16(r0, |
3198 | 1.02M | _mm_mullo_epi16(t5, |
3199 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
3200 | 1.02M | r2 = _mm_adds_epi16(r2, |
3201 | 1.02M | _mm_mullo_epi16(x5, |
3202 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
3203 | | |
3204 | 1.02M | r0 = _mm_adds_epi16(r0, |
3205 | 1.02M | _mm_mullo_epi16(t6, |
3206 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
3207 | 1.02M | r2 = _mm_adds_epi16(r2, |
3208 | 1.02M | _mm_mullo_epi16(x6, |
3209 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
3210 | | |
3211 | 1.02M | r0 = _mm_adds_epi16(r0, |
3212 | 1.02M | _mm_mullo_epi16(t7, |
3213 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
3214 | 1.02M | r2 = _mm_adds_epi16(r2, |
3215 | 1.02M | _mm_mullo_epi16(x7, |
3216 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
3217 | | |
3218 | 1.02M | r0 = _mm_adds_epi16(r0, |
3219 | 1.02M | _mm_mullo_epi16(t8, |
3220 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
3221 | 1.02M | r2 = _mm_adds_epi16(r2, |
3222 | 1.02M | _mm_mullo_epi16(x8, |
3223 | 1.02M | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
3224 | | |
3225 | | /* give results back */ |
3226 | 1.02M | _mm_store_si128((__m128i *) &dst[x],r0); |
3227 | 1.02M | _mm_store_si128((__m128i *) &dst[x + 8],r2); |
3228 | 1.02M | } |
3229 | 666k | src += srcstride; |
3230 | 666k | dst += dststride; |
3231 | 666k | } |
3232 | 213k | }else{ |
3233 | 213k | x = 0; |
3234 | 1.92M | for (y = 0; y < height; y ++) { |
3235 | 4.87M | for(x=0;x<width;x+=4){ |
3236 | 3.16M | r0 = _mm_set1_epi16(0); |
3237 | | /* load data in register */ |
3238 | 3.16M | x2 = _mm_loadl_epi64((__m128i *) &src[x-2 * srcstride]); |
3239 | 3.16M | x3 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
3240 | 3.16M | x4 = _mm_loadl_epi64((__m128i *) &src[x]); |
3241 | 3.16M | x5 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
3242 | 3.16M | x6 = _mm_loadl_epi64((__m128i *) &src[x+2 * srcstride]); |
3243 | 3.16M | x7 = _mm_loadl_epi64((__m128i *) &src[x+3 * srcstride]); |
3244 | 3.16M | x8 = _mm_loadl_epi64((__m128i *) &src[x + 4 * srcstride]); |
3245 | | |
3246 | 3.16M | x2 = _mm_unpacklo_epi8(x2, r0); |
3247 | 3.16M | x3 = _mm_unpacklo_epi8(x3, r0); |
3248 | 3.16M | x4 = _mm_unpacklo_epi8(x4, r0); |
3249 | 3.16M | x5 = _mm_unpacklo_epi8(x5, r0); |
3250 | 3.16M | x6 = _mm_unpacklo_epi8(x6, r0); |
3251 | 3.16M | x7 = _mm_unpacklo_epi8(x7, r0); |
3252 | 3.16M | x8 = _mm_unpacklo_epi8(x8, r0); |
3253 | | |
3254 | 3.16M | r0 = _mm_mullo_epi16(x2, _mm_set1_epi16(_mm_extract_epi16(r1, 1))); |
3255 | | |
3256 | 3.16M | r0 = _mm_adds_epi16(r0, |
3257 | 3.16M | _mm_mullo_epi16(x3, |
3258 | 3.16M | _mm_set1_epi16(_mm_extract_epi16(r1, 2)))); |
3259 | | |
3260 | 3.16M | r0 = _mm_adds_epi16(r0, |
3261 | 3.16M | _mm_mullo_epi16(x4, |
3262 | 3.16M | _mm_set1_epi16(_mm_extract_epi16(r1, 3)))); |
3263 | | |
3264 | 3.16M | r0 = _mm_adds_epi16(r0, |
3265 | 3.16M | _mm_mullo_epi16(x5, |
3266 | 3.16M | _mm_set1_epi16(_mm_extract_epi16(r1, 4)))); |
3267 | | |
3268 | 3.16M | r0 = _mm_adds_epi16(r0, |
3269 | 3.16M | _mm_mullo_epi16(x6, |
3270 | 3.16M | _mm_set1_epi16(_mm_extract_epi16(r1, 5)))); |
3271 | | |
3272 | 3.16M | r0 = _mm_adds_epi16(r0, |
3273 | 3.16M | _mm_mullo_epi16(x7, |
3274 | 3.16M | _mm_set1_epi16(_mm_extract_epi16(r1, 6)))); |
3275 | | |
3276 | 3.16M | r0 = _mm_adds_epi16(r0, |
3277 | 3.16M | _mm_mullo_epi16(x8, |
3278 | 3.16M | _mm_set1_epi16(_mm_extract_epi16(r1, 7)))); |
3279 | | |
3280 | | /* give results back */ |
3281 | 3.16M | _mm_storel_epi64((__m128i *) &dst[x], r0); |
3282 | | |
3283 | 3.16M | } |
3284 | 1.71M | src += srcstride; |
3285 | 1.71M | dst += dststride; |
3286 | 1.71M | } |
3287 | 213k | } |
3288 | | |
3289 | 251k | } |
3290 | | |
3291 | | |
3292 | | #if 0 |
3293 | | void ff_hevc_put_hevc_qpel_v_3_10_sse(int16_t *dst, ptrdiff_t dststride, |
3294 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
3295 | | int16_t* mcbuffer) { |
3296 | | int x, y; |
3297 | | uint16_t *src = (uint16_t*) _src; |
3298 | | ptrdiff_t srcstride = _srcstride >> 1; |
3299 | | __m128i x1, x2, x3, x4, x5, x6, x7, r0; |
3300 | | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
3301 | | |
3302 | | t7 = _mm_set1_epi32(-1); |
3303 | | t6 = _mm_set1_epi32(4); |
3304 | | t5 = _mm_set1_epi32(-10); |
3305 | | t4 = _mm_set1_epi32(58); |
3306 | | t3 = _mm_set1_epi32(17); |
3307 | | t2 = _mm_set1_epi32(-5); |
3308 | | t1 = _mm_set1_epi32(1); |
3309 | | t8= _mm_setzero_si128(); |
3310 | | { |
3311 | | |
3312 | | for (y = 0; y < height; y ++) { |
3313 | | for(x=0;x<width;x+=4){ |
3314 | | /* load data in register */ |
3315 | | x1 = _mm_loadl_epi64((__m128i *) &src[x-2 * srcstride]); |
3316 | | x2 = _mm_loadl_epi64((__m128i *) &src[x-srcstride]); |
3317 | | x3 = _mm_loadl_epi64((__m128i *) &src[x]); |
3318 | | x4 = _mm_loadl_epi64((__m128i *) &src[x+srcstride]); |
3319 | | x5 = _mm_loadl_epi64((__m128i *) &src[x+2 * srcstride]); |
3320 | | x6 = _mm_loadl_epi64((__m128i *) &src[x+3 * srcstride]); |
3321 | | x7 = _mm_loadl_epi64((__m128i *) &src[x + 4 * srcstride]); |
3322 | | |
3323 | | x1 = _mm_unpacklo_epi16(x1, t8); |
3324 | | x2 = _mm_unpacklo_epi16(x2, t8); |
3325 | | x3 = _mm_unpacklo_epi16(x3, t8); |
3326 | | x4 = _mm_unpacklo_epi16(x4, t8); |
3327 | | x5 = _mm_unpacklo_epi16(x5, t8); |
3328 | | x6 = _mm_unpacklo_epi16(x6, t8); |
3329 | | x7 = _mm_unpacklo_epi16(x7, t8); |
3330 | | |
3331 | | r0 = _mm_mullo_epi32(x1, t1); |
3332 | | |
3333 | | r0 = _mm_add_epi32(r0, |
3334 | | _mm_mullo_epi32(x2,t2)); |
3335 | | |
3336 | | r0 = _mm_add_epi32(r0, |
3337 | | _mm_mullo_epi32(x3,t3)); |
3338 | | |
3339 | | r0 = _mm_add_epi32(r0, |
3340 | | _mm_mullo_epi32(x4,t4)); |
3341 | | |
3342 | | r0 = _mm_add_epi32(r0, |
3343 | | _mm_mullo_epi32(x5,t5)); |
3344 | | |
3345 | | r0 = _mm_add_epi32(r0, |
3346 | | _mm_mullo_epi32(x6,t6)); |
3347 | | |
3348 | | r0 = _mm_add_epi32(r0, |
3349 | | _mm_mullo_epi32(x7,t7)); |
3350 | | |
3351 | | r0= _mm_srai_epi32(r0,2); |
3352 | | |
3353 | | r0= _mm_packs_epi32(r0,t8); |
3354 | | |
3355 | | /* give results back */ |
3356 | | _mm_storel_epi64((__m128i *) &dst[x], r0); |
3357 | | |
3358 | | } |
3359 | | src += srcstride; |
3360 | | dst += dststride; |
3361 | | } |
3362 | | } |
3363 | | |
3364 | | } |
3365 | | #endif |
3366 | | |
3367 | | |
3368 | | |
3369 | | void ff_hevc_put_hevc_qpel_h_1_v_1_sse(int16_t *dst, ptrdiff_t dststride, |
3370 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
3371 | 257k | int16_t* mcbuffer) { |
3372 | 257k | int x, y; |
3373 | 257k | uint8_t* src = (uint8_t*) _src; |
3374 | 257k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
3375 | 257k | int16_t *tmp = mcbuffer; |
3376 | 257k | __m128i x1, x2, x3, x4, x5, x6, x7, rBuffer, rTemp, r0, r1; |
3377 | 257k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
3378 | | |
3379 | 257k | src -= qpel_extra_before[1] * srcstride; |
3380 | 257k | r0 = _mm_set_epi8(0, 1, -5, 17, 58, -10, 4, -1, 0, 1, -5, 17, 58, -10, 4, |
3381 | 257k | -1); |
3382 | | |
3383 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
3384 | 257k | if (width == 4) { |
3385 | | |
3386 | 515k | for (y = 0; y < height + qpel_extra[1]; y += 2) { |
3387 | | /* load data in register */ |
3388 | 451k | x1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3389 | 451k | src += srcstride; |
3390 | 451k | t1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3391 | 451k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3392 | 451k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
3393 | 451k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3394 | 451k | _mm_srli_si128(x1, 3)); |
3395 | 451k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
3396 | 451k | _mm_srli_si128(t1, 3)); |
3397 | | |
3398 | | /* PMADDUBSW then PMADDW */ |
3399 | 451k | x2 = _mm_maddubs_epi16(x2, r0); |
3400 | 451k | t2 = _mm_maddubs_epi16(t2, r0); |
3401 | 451k | x3 = _mm_maddubs_epi16(x3, r0); |
3402 | 451k | t3 = _mm_maddubs_epi16(t3, r0); |
3403 | 451k | x2 = _mm_hadd_epi16(x2, x3); |
3404 | 451k | t2 = _mm_hadd_epi16(t2, t3); |
3405 | 451k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
3406 | 451k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
3407 | 451k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
3408 | 451k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
3409 | | /* give results back */ |
3410 | 451k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
3411 | | |
3412 | 451k | tmp += MAX_PB_SIZE; |
3413 | 451k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
3414 | | |
3415 | 451k | src += srcstride; |
3416 | 451k | tmp += MAX_PB_SIZE; |
3417 | 451k | } |
3418 | 63.8k | } else |
3419 | 3.20M | for (y = 0; y < height + qpel_extra[1]; y++) { |
3420 | 7.76M | for (x = 0; x < width; x += 8) { |
3421 | | /* load data in register */ |
3422 | 4.75M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
3423 | 4.75M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3424 | 4.75M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3425 | 4.75M | _mm_srli_si128(x1, 3)); |
3426 | 4.75M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
3427 | 4.75M | _mm_srli_si128(x1, 5)); |
3428 | 4.75M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
3429 | 4.75M | _mm_srli_si128(x1, 7)); |
3430 | | |
3431 | | /* PMADDUBSW then PMADDW */ |
3432 | 4.75M | x2 = _mm_maddubs_epi16(x2, r0); |
3433 | 4.75M | x3 = _mm_maddubs_epi16(x3, r0); |
3434 | 4.75M | x4 = _mm_maddubs_epi16(x4, r0); |
3435 | 4.75M | x5 = _mm_maddubs_epi16(x5, r0); |
3436 | 4.75M | x2 = _mm_hadd_epi16(x2, x3); |
3437 | 4.75M | x4 = _mm_hadd_epi16(x4, x5); |
3438 | 4.75M | x2 = _mm_hadd_epi16(x2, x4); |
3439 | 4.75M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
3440 | | |
3441 | | /* give results back */ |
3442 | 4.75M | _mm_store_si128((__m128i *) &tmp[x], x2); |
3443 | | |
3444 | 4.75M | } |
3445 | 3.01M | src += srcstride; |
3446 | 3.01M | tmp += MAX_PB_SIZE; |
3447 | 3.01M | } |
3448 | | |
3449 | 257k | tmp = mcbuffer + qpel_extra_before[1] * MAX_PB_SIZE; |
3450 | 257k | srcstride = MAX_PB_SIZE; |
3451 | | |
3452 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
3453 | | for register calculations */ |
3454 | 257k | rTemp = _mm_set_epi16(0, 1, -5, 17, 58, -10, 4, -1); |
3455 | 2.62M | for (y = 0; y < height; y++) { |
3456 | 6.13M | for (x = 0; x < width; x += 8) { |
3457 | | |
3458 | 3.76M | x1 = _mm_load_si128((__m128i *) &tmp[x - 3 * srcstride]); |
3459 | 3.76M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
3460 | 3.76M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
3461 | 3.76M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
3462 | 3.76M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
3463 | 3.76M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
3464 | 3.76M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
3465 | | |
3466 | 3.76M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 0)); |
3467 | 3.76M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
3468 | 3.76M | t8 = _mm_mullo_epi16(x1, r0); |
3469 | 3.76M | rBuffer = _mm_mulhi_epi16(x1, r0); |
3470 | 3.76M | t7 = _mm_mullo_epi16(x2, r1); |
3471 | 3.76M | t1 = _mm_unpacklo_epi16(t8, rBuffer); |
3472 | 3.76M | x1 = _mm_unpackhi_epi16(t8, rBuffer); |
3473 | | |
3474 | 3.76M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
3475 | 3.76M | rBuffer = _mm_mulhi_epi16(x2, r1); |
3476 | 3.76M | t8 = _mm_mullo_epi16(x3, r0); |
3477 | 3.76M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
3478 | 3.76M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
3479 | | |
3480 | 3.76M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
3481 | 3.76M | rBuffer = _mm_mulhi_epi16(x3, r0); |
3482 | 3.76M | t7 = _mm_mullo_epi16(x4, r1); |
3483 | 3.76M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
3484 | 3.76M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
3485 | | |
3486 | 3.76M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
3487 | 3.76M | rBuffer = _mm_mulhi_epi16(x4, r1); |
3488 | 3.76M | t8 = _mm_mullo_epi16(x5, r0); |
3489 | 3.76M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
3490 | 3.76M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
3491 | | |
3492 | 3.76M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
3493 | 3.76M | rBuffer = _mm_mulhi_epi16(x5, r0); |
3494 | 3.76M | t7 = _mm_mullo_epi16(x6, r1); |
3495 | 3.76M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
3496 | 3.76M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
3497 | | |
3498 | 3.76M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
3499 | 3.76M | rBuffer = _mm_mulhi_epi16(x6, r1); |
3500 | 3.76M | t8 = _mm_mullo_epi16(x7, r0); |
3501 | 3.76M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
3502 | 3.76M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
3503 | | |
3504 | 3.76M | rBuffer = _mm_mulhi_epi16(x7, r0); |
3505 | 3.76M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
3506 | 3.76M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
3507 | | |
3508 | | |
3509 | | |
3510 | | /* add calculus by correct value : */ |
3511 | | |
3512 | 3.76M | r1 = _mm_add_epi32(x1, x2); |
3513 | 3.76M | x3 = _mm_add_epi32(x3, x4); |
3514 | 3.76M | x5 = _mm_add_epi32(x5, x6); |
3515 | 3.76M | r1 = _mm_add_epi32(r1, x3); |
3516 | | |
3517 | 3.76M | r1 = _mm_add_epi32(r1, x5); |
3518 | | |
3519 | 3.76M | r0 = _mm_add_epi32(t1, t2); |
3520 | 3.76M | t3 = _mm_add_epi32(t3, t4); |
3521 | 3.76M | t5 = _mm_add_epi32(t5, t6); |
3522 | 3.76M | r0 = _mm_add_epi32(r0, t3); |
3523 | 3.76M | r0 = _mm_add_epi32(r0, t5); |
3524 | 3.76M | r1 = _mm_add_epi32(r1, x7); |
3525 | 3.76M | r0 = _mm_add_epi32(r0, t7); |
3526 | 3.76M | r1 = _mm_srli_epi32(r1, 6); |
3527 | 3.76M | r0 = _mm_srli_epi32(r0, 6); |
3528 | | |
3529 | 3.76M | r1 = _mm_and_si128(r1, |
3530 | 3.76M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
3531 | 3.76M | r0 = _mm_and_si128(r0, |
3532 | 3.76M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
3533 | 3.76M | r0 = _mm_hadd_epi16(r0, r1); |
3534 | 3.76M | _mm_store_si128((__m128i *) &dst[x], r0); |
3535 | | |
3536 | 3.76M | } |
3537 | 2.37M | tmp += MAX_PB_SIZE; |
3538 | 2.37M | dst += dststride; |
3539 | 2.37M | } |
3540 | 257k | } |
3541 | | void ff_hevc_put_hevc_qpel_h_1_v_2_sse(int16_t *dst, ptrdiff_t dststride, |
3542 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
3543 | 137k | int16_t* mcbuffer) { |
3544 | 137k | int x, y; |
3545 | 137k | uint8_t *src = (uint8_t*) _src; |
3546 | 137k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
3547 | 137k | int16_t *tmp = mcbuffer; |
3548 | 137k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, rBuffer, rTemp, r0, r1; |
3549 | 137k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
3550 | | |
3551 | 137k | src -= qpel_extra_before[2] * srcstride; |
3552 | 137k | r0 = _mm_set_epi8(0, 1, -5, 17, 58, -10, 4, -1, 0, 1, -5, 17, 58, -10, 4, |
3553 | 137k | -1); |
3554 | | |
3555 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
3556 | 137k | if (width == 4) { |
3557 | | |
3558 | 286k | for (y = 0; y < height + qpel_extra[2]; y += 2) { |
3559 | | /* load data in register */ |
3560 | 254k | x1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3561 | 254k | src += srcstride; |
3562 | 254k | t1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3563 | 254k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3564 | 254k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
3565 | 254k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3566 | 254k | _mm_srli_si128(x1, 3)); |
3567 | 254k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
3568 | 254k | _mm_srli_si128(t1, 3)); |
3569 | | |
3570 | | /* PMADDUBSW then PMADDW */ |
3571 | 254k | x2 = _mm_maddubs_epi16(x2, r0); |
3572 | 254k | t2 = _mm_maddubs_epi16(t2, r0); |
3573 | 254k | x3 = _mm_maddubs_epi16(x3, r0); |
3574 | 254k | t3 = _mm_maddubs_epi16(t3, r0); |
3575 | 254k | x2 = _mm_hadd_epi16(x2, x3); |
3576 | 254k | t2 = _mm_hadd_epi16(t2, t3); |
3577 | 254k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
3578 | 254k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
3579 | 254k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
3580 | 254k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
3581 | | /* give results back */ |
3582 | 254k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
3583 | | |
3584 | 254k | tmp += MAX_PB_SIZE; |
3585 | 254k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
3586 | | |
3587 | 254k | src += srcstride; |
3588 | 254k | tmp += MAX_PB_SIZE; |
3589 | 254k | } |
3590 | 31.2k | } else |
3591 | 1.94M | for (y = 0; y < height + qpel_extra[2]; y++) { |
3592 | 4.68M | for (x = 0; x < width; x += 8) { |
3593 | | /* load data in register */ |
3594 | 2.85M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
3595 | 2.85M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3596 | 2.85M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3597 | 2.85M | _mm_srli_si128(x1, 3)); |
3598 | 2.85M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
3599 | 2.85M | _mm_srli_si128(x1, 5)); |
3600 | 2.85M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
3601 | 2.85M | _mm_srli_si128(x1, 7)); |
3602 | | |
3603 | | /* PMADDUBSW then PMADDW */ |
3604 | 2.85M | x2 = _mm_maddubs_epi16(x2, r0); |
3605 | 2.85M | x3 = _mm_maddubs_epi16(x3, r0); |
3606 | 2.85M | x4 = _mm_maddubs_epi16(x4, r0); |
3607 | 2.85M | x5 = _mm_maddubs_epi16(x5, r0); |
3608 | 2.85M | x2 = _mm_hadd_epi16(x2, x3); |
3609 | 2.85M | x4 = _mm_hadd_epi16(x4, x5); |
3610 | 2.85M | x2 = _mm_hadd_epi16(x2, x4); |
3611 | 2.85M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
3612 | | |
3613 | | /* give results back */ |
3614 | 2.85M | _mm_store_si128((__m128i *) &tmp[x], x2); |
3615 | | |
3616 | 2.85M | } |
3617 | 1.83M | src += srcstride; |
3618 | 1.83M | tmp += MAX_PB_SIZE; |
3619 | 1.83M | } |
3620 | | |
3621 | 137k | tmp = mcbuffer + qpel_extra_before[2] * MAX_PB_SIZE; |
3622 | 137k | srcstride = MAX_PB_SIZE; |
3623 | | |
3624 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
3625 | | for register calculations */ |
3626 | 137k | rTemp = _mm_set_epi16(-1, 4, -11, 40, 40, -11, 4, -1); |
3627 | 1.49M | for (y = 0; y < height; y++) { |
3628 | 3.47M | for (x = 0; x < width; x += 8) { |
3629 | | |
3630 | 2.12M | x1 = _mm_load_si128((__m128i *) &tmp[x - 3 * srcstride]); |
3631 | 2.12M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
3632 | 2.12M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
3633 | 2.12M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
3634 | 2.12M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
3635 | 2.12M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
3636 | 2.12M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
3637 | 2.12M | x8 = _mm_loadu_si128((__m128i *) &tmp[x + 4 * srcstride]); |
3638 | | |
3639 | 2.12M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 0)); |
3640 | 2.12M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
3641 | 2.12M | t8 = _mm_mullo_epi16(x1, r0); |
3642 | 2.12M | rBuffer = _mm_mulhi_epi16(x1, r0); |
3643 | 2.12M | t7 = _mm_mullo_epi16(x2, r1); |
3644 | 2.12M | t1 = _mm_unpacklo_epi16(t8, rBuffer); |
3645 | 2.12M | x1 = _mm_unpackhi_epi16(t8, rBuffer); |
3646 | | |
3647 | 2.12M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
3648 | 2.12M | rBuffer = _mm_mulhi_epi16(x2, r1); |
3649 | 2.12M | t8 = _mm_mullo_epi16(x3, r0); |
3650 | 2.12M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
3651 | 2.12M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
3652 | | |
3653 | 2.12M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
3654 | 2.12M | rBuffer = _mm_mulhi_epi16(x3, r0); |
3655 | 2.12M | t7 = _mm_mullo_epi16(x4, r1); |
3656 | 2.12M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
3657 | 2.12M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
3658 | | |
3659 | 2.12M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
3660 | 2.12M | rBuffer = _mm_mulhi_epi16(x4, r1); |
3661 | 2.12M | t8 = _mm_mullo_epi16(x5, r0); |
3662 | 2.12M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
3663 | 2.12M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
3664 | | |
3665 | 2.12M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
3666 | 2.12M | rBuffer = _mm_mulhi_epi16(x5, r0); |
3667 | 2.12M | t7 = _mm_mullo_epi16(x6, r1); |
3668 | 2.12M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
3669 | 2.12M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
3670 | | |
3671 | 2.12M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
3672 | 2.12M | rBuffer = _mm_mulhi_epi16(x6, r1); |
3673 | 2.12M | t8 = _mm_mullo_epi16(x7, r0); |
3674 | 2.12M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
3675 | 2.12M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
3676 | | |
3677 | 2.12M | rBuffer = _mm_mulhi_epi16(x7, r0); |
3678 | 2.12M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
3679 | 2.12M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
3680 | | |
3681 | 2.12M | t8 = _mm_unpacklo_epi16( |
3682 | 2.12M | _mm_mullo_epi16(x8, |
3683 | 2.12M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
3684 | 2.12M | _mm_mulhi_epi16(x8, |
3685 | 2.12M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
3686 | 2.12M | x8 = _mm_unpackhi_epi16( |
3687 | 2.12M | _mm_mullo_epi16(x8, |
3688 | 2.12M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
3689 | 2.12M | _mm_mulhi_epi16(x8, |
3690 | 2.12M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
3691 | | |
3692 | | /* add calculus by correct value : */ |
3693 | | |
3694 | 2.12M | r1 = _mm_add_epi32(x1, x2); |
3695 | 2.12M | x3 = _mm_add_epi32(x3, x4); |
3696 | 2.12M | x5 = _mm_add_epi32(x5, x6); |
3697 | 2.12M | r1 = _mm_add_epi32(r1, x3); |
3698 | 2.12M | x7 = _mm_add_epi32(x7, x8); |
3699 | 2.12M | r1 = _mm_add_epi32(r1, x5); |
3700 | | |
3701 | 2.12M | r0 = _mm_add_epi32(t1, t2); |
3702 | 2.12M | t3 = _mm_add_epi32(t3, t4); |
3703 | 2.12M | t5 = _mm_add_epi32(t5, t6); |
3704 | 2.12M | r0 = _mm_add_epi32(r0, t3); |
3705 | 2.12M | t7 = _mm_add_epi32(t7, t8); |
3706 | 2.12M | r0 = _mm_add_epi32(r0, t5); |
3707 | 2.12M | r1 = _mm_add_epi32(r1, x7); |
3708 | 2.12M | r0 = _mm_add_epi32(r0, t7); |
3709 | 2.12M | r1 = _mm_srli_epi32(r1, 6); |
3710 | 2.12M | r0 = _mm_srli_epi32(r0, 6); |
3711 | | |
3712 | 2.12M | r1 = _mm_and_si128(r1, |
3713 | 2.12M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
3714 | 2.12M | r0 = _mm_and_si128(r0, |
3715 | 2.12M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
3716 | 2.12M | r0 = _mm_hadd_epi16(r0, r1); |
3717 | 2.12M | _mm_store_si128((__m128i *) &dst[x], r0); |
3718 | | |
3719 | 2.12M | } |
3720 | 1.35M | tmp += MAX_PB_SIZE; |
3721 | 1.35M | dst += dststride; |
3722 | 1.35M | } |
3723 | 137k | } |
3724 | | void ff_hevc_put_hevc_qpel_h_1_v_3_sse(int16_t *dst, ptrdiff_t dststride, |
3725 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
3726 | 145k | int16_t* mcbuffer) { |
3727 | 145k | int x, y; |
3728 | 145k | uint8_t *src = (uint8_t*) _src; |
3729 | 145k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
3730 | 145k | int16_t *tmp = mcbuffer; |
3731 | 145k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, rBuffer, rTemp, r0, r1; |
3732 | 145k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
3733 | | |
3734 | 145k | src -= qpel_extra_before[3] * srcstride; |
3735 | 145k | r0 = _mm_set_epi8(0, 1, -5, 17, 58, -10, 4, -1, 0, 1, -5, 17, 58, -10, 4, |
3736 | 145k | -1); |
3737 | | |
3738 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
3739 | 145k | if (width == 4) { |
3740 | | |
3741 | 190k | for (y = 0; y < height + qpel_extra[3]; y += 2) { |
3742 | | /* load data in register */ |
3743 | 167k | x1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3744 | 167k | src += srcstride; |
3745 | 167k | t1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3746 | 167k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3747 | 167k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
3748 | 167k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3749 | 167k | _mm_srli_si128(x1, 3)); |
3750 | 167k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
3751 | 167k | _mm_srli_si128(t1, 3)); |
3752 | | |
3753 | | /* PMADDUBSW then PMADDW */ |
3754 | 167k | x2 = _mm_maddubs_epi16(x2, r0); |
3755 | 167k | t2 = _mm_maddubs_epi16(t2, r0); |
3756 | 167k | x3 = _mm_maddubs_epi16(x3, r0); |
3757 | 167k | t3 = _mm_maddubs_epi16(t3, r0); |
3758 | 167k | x2 = _mm_hadd_epi16(x2, x3); |
3759 | 167k | t2 = _mm_hadd_epi16(t2, t3); |
3760 | 167k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
3761 | 167k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
3762 | 167k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
3763 | 167k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
3764 | | /* give results back */ |
3765 | 167k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
3766 | | |
3767 | 167k | tmp += MAX_PB_SIZE; |
3768 | 167k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
3769 | | |
3770 | 167k | src += srcstride; |
3771 | 167k | tmp += MAX_PB_SIZE; |
3772 | 167k | } |
3773 | 23.6k | } else |
3774 | 2.05M | for (y = 0; y < height + qpel_extra[3]; y++) { |
3775 | 4.96M | for (x = 0; x < width; x += 8) { |
3776 | | /* load data in register */ |
3777 | 3.03M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
3778 | 3.03M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3779 | 3.03M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3780 | 3.03M | _mm_srli_si128(x1, 3)); |
3781 | 3.03M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
3782 | 3.03M | _mm_srli_si128(x1, 5)); |
3783 | 3.03M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
3784 | 3.03M | _mm_srli_si128(x1, 7)); |
3785 | | |
3786 | | /* PMADDUBSW then PMADDW */ |
3787 | 3.03M | x2 = _mm_maddubs_epi16(x2, r0); |
3788 | 3.03M | x3 = _mm_maddubs_epi16(x3, r0); |
3789 | 3.03M | x4 = _mm_maddubs_epi16(x4, r0); |
3790 | 3.03M | x5 = _mm_maddubs_epi16(x5, r0); |
3791 | 3.03M | x2 = _mm_hadd_epi16(x2, x3); |
3792 | 3.03M | x4 = _mm_hadd_epi16(x4, x5); |
3793 | 3.03M | x2 = _mm_hadd_epi16(x2, x4); |
3794 | 3.03M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
3795 | | |
3796 | | /* give results back */ |
3797 | 3.03M | _mm_store_si128((__m128i *) &tmp[x], x2); |
3798 | | |
3799 | 3.03M | } |
3800 | 1.93M | src += srcstride; |
3801 | 1.93M | tmp += MAX_PB_SIZE; |
3802 | 1.93M | } |
3803 | | |
3804 | 145k | tmp = mcbuffer + qpel_extra_before[3] * MAX_PB_SIZE; |
3805 | 145k | srcstride = MAX_PB_SIZE; |
3806 | | |
3807 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
3808 | | for register calculations */ |
3809 | 145k | rTemp = _mm_set_epi16(-1, 4, -10, 58, 17, -5, 1, 0); |
3810 | 1.53M | for (y = 0; y < height; y++) { |
3811 | 3.64M | for (x = 0; x < width; x += 8) { |
3812 | | |
3813 | 2.25M | x1 = _mm_setzero_si128(); |
3814 | 2.25M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
3815 | 2.25M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
3816 | 2.25M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
3817 | 2.25M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
3818 | 2.25M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
3819 | 2.25M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
3820 | 2.25M | x8 = _mm_load_si128((__m128i *) &tmp[x + 4 * srcstride]); |
3821 | | |
3822 | | |
3823 | 2.25M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
3824 | | |
3825 | 2.25M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
3826 | 2.25M | t7 = _mm_mullo_epi16(x2, r1); |
3827 | 2.25M | rBuffer = _mm_mulhi_epi16(x2, r1); |
3828 | 2.25M | t8 = _mm_mullo_epi16(x3, r0); |
3829 | 2.25M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
3830 | 2.25M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
3831 | | |
3832 | 2.25M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
3833 | 2.25M | rBuffer = _mm_mulhi_epi16(x3, r0); |
3834 | 2.25M | t7 = _mm_mullo_epi16(x4, r1); |
3835 | 2.25M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
3836 | 2.25M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
3837 | | |
3838 | 2.25M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
3839 | 2.25M | rBuffer = _mm_mulhi_epi16(x4, r1); |
3840 | 2.25M | t8 = _mm_mullo_epi16(x5, r0); |
3841 | 2.25M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
3842 | 2.25M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
3843 | | |
3844 | 2.25M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
3845 | 2.25M | rBuffer = _mm_mulhi_epi16(x5, r0); |
3846 | 2.25M | t7 = _mm_mullo_epi16(x6, r1); |
3847 | 2.25M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
3848 | 2.25M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
3849 | | |
3850 | 2.25M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
3851 | 2.25M | rBuffer = _mm_mulhi_epi16(x6, r1); |
3852 | 2.25M | t8 = _mm_mullo_epi16(x7, r0); |
3853 | 2.25M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
3854 | 2.25M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
3855 | | |
3856 | 2.25M | rBuffer = _mm_mulhi_epi16(x7, r0); |
3857 | 2.25M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
3858 | 2.25M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
3859 | | |
3860 | 2.25M | t8 = _mm_unpacklo_epi16( |
3861 | 2.25M | _mm_mullo_epi16(x8, |
3862 | 2.25M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
3863 | 2.25M | _mm_mulhi_epi16(x8, |
3864 | 2.25M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
3865 | 2.25M | x8 = _mm_unpackhi_epi16( |
3866 | 2.25M | _mm_mullo_epi16(x8, |
3867 | 2.25M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
3868 | 2.25M | _mm_mulhi_epi16(x8, |
3869 | 2.25M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
3870 | | |
3871 | | /* add calculus by correct value : */ |
3872 | | |
3873 | 2.25M | x3 = _mm_add_epi32(x3, x4); |
3874 | 2.25M | x5 = _mm_add_epi32(x5, x6); |
3875 | 2.25M | r1 = _mm_add_epi32(x2, x3); |
3876 | 2.25M | x7 = _mm_add_epi32(x7, x8); |
3877 | 2.25M | r1 = _mm_add_epi32(r1, x5); |
3878 | | |
3879 | 2.25M | t3 = _mm_add_epi32(t3, t4); |
3880 | 2.25M | t5 = _mm_add_epi32(t5, t6); |
3881 | 2.25M | r0 = _mm_add_epi32(t2, t3); |
3882 | 2.25M | t7 = _mm_add_epi32(t7, t8); |
3883 | 2.25M | r0 = _mm_add_epi32(r0, t5); |
3884 | 2.25M | r1 = _mm_add_epi32(r1, x7); |
3885 | 2.25M | r0 = _mm_add_epi32(r0, t7); |
3886 | 2.25M | r1 = _mm_srli_epi32(r1, 6); |
3887 | 2.25M | r0 = _mm_srli_epi32(r0, 6); |
3888 | | |
3889 | 2.25M | r1 = _mm_and_si128(r1, |
3890 | 2.25M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
3891 | 2.25M | r0 = _mm_and_si128(r0, |
3892 | 2.25M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
3893 | 2.25M | r0 = _mm_hadd_epi16(r0, r1); |
3894 | 2.25M | _mm_store_si128((__m128i *) &dst[x], r0); |
3895 | | |
3896 | 2.25M | } |
3897 | 1.39M | tmp += MAX_PB_SIZE; |
3898 | 1.39M | dst += dststride; |
3899 | 1.39M | } |
3900 | 145k | } |
3901 | | void ff_hevc_put_hevc_qpel_h_2_v_1_sse(int16_t *dst, ptrdiff_t dststride, |
3902 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
3903 | 190k | int16_t* mcbuffer) { |
3904 | 190k | int x, y; |
3905 | 190k | uint8_t *src = (uint8_t*) _src; |
3906 | 190k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
3907 | 190k | int16_t *tmp = mcbuffer; |
3908 | 190k | __m128i x1, x2, x3, x4, x5, x6, x7, rBuffer, rTemp, r0, r1; |
3909 | 190k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
3910 | | |
3911 | 190k | src -= qpel_extra_before[1] * srcstride; |
3912 | 190k | r0 = _mm_set_epi8(-1, 4, -11, 40, 40, -11, 4, -1, -1, 4, -11, 40, 40, -11, |
3913 | 190k | 4, -1); |
3914 | | |
3915 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
3916 | 190k | if (width == 4) { |
3917 | | |
3918 | 260k | for (y = 0; y < height + qpel_extra[1]; y += 2) { |
3919 | | /* load data in register */ |
3920 | 228k | x1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3921 | 228k | src += srcstride; |
3922 | 228k | t1 = _mm_loadu_si128((__m128i *) &src[-3]); |
3923 | 228k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3924 | 228k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
3925 | 228k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3926 | 228k | _mm_srli_si128(x1, 3)); |
3927 | 228k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
3928 | 228k | _mm_srli_si128(t1, 3)); |
3929 | | |
3930 | | /* PMADDUBSW then PMADDW */ |
3931 | 228k | x2 = _mm_maddubs_epi16(x2, r0); |
3932 | 228k | t2 = _mm_maddubs_epi16(t2, r0); |
3933 | 228k | x3 = _mm_maddubs_epi16(x3, r0); |
3934 | 228k | t3 = _mm_maddubs_epi16(t3, r0); |
3935 | 228k | x2 = _mm_hadd_epi16(x2, x3); |
3936 | 228k | t2 = _mm_hadd_epi16(t2, t3); |
3937 | 228k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
3938 | 228k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
3939 | 228k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
3940 | 228k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
3941 | | /* give results back */ |
3942 | 228k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
3943 | | |
3944 | 228k | tmp += MAX_PB_SIZE; |
3945 | 228k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
3946 | | |
3947 | 228k | src += srcstride; |
3948 | 228k | tmp += MAX_PB_SIZE; |
3949 | 228k | } |
3950 | 32.3k | } else |
3951 | 2.66M | for (y = 0; y < height + qpel_extra[1]; y++) { |
3952 | 6.28M | for (x = 0; x < width; x += 8) { |
3953 | | /* load data in register */ |
3954 | 3.78M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
3955 | 3.78M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
3956 | 3.78M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
3957 | 3.78M | _mm_srli_si128(x1, 3)); |
3958 | 3.78M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
3959 | 3.78M | _mm_srli_si128(x1, 5)); |
3960 | 3.78M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
3961 | 3.78M | _mm_srli_si128(x1, 7)); |
3962 | | |
3963 | | /* PMADDUBSW then PMADDW */ |
3964 | 3.78M | x2 = _mm_maddubs_epi16(x2, r0); |
3965 | 3.78M | x3 = _mm_maddubs_epi16(x3, r0); |
3966 | 3.78M | x4 = _mm_maddubs_epi16(x4, r0); |
3967 | 3.78M | x5 = _mm_maddubs_epi16(x5, r0); |
3968 | 3.78M | x2 = _mm_hadd_epi16(x2, x3); |
3969 | 3.78M | x4 = _mm_hadd_epi16(x4, x5); |
3970 | 3.78M | x2 = _mm_hadd_epi16(x2, x4); |
3971 | 3.78M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
3972 | | |
3973 | | /* give results back */ |
3974 | 3.78M | _mm_store_si128((__m128i *) &tmp[x], x2); |
3975 | | |
3976 | 3.78M | } |
3977 | 2.50M | src += srcstride; |
3978 | 2.50M | tmp += MAX_PB_SIZE; |
3979 | 2.50M | } |
3980 | | |
3981 | 190k | tmp = mcbuffer + qpel_extra_before[1] * MAX_PB_SIZE; |
3982 | 190k | srcstride = MAX_PB_SIZE; |
3983 | | |
3984 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
3985 | | for register calculations */ |
3986 | 190k | rTemp = _mm_set_epi16(0, 1, -5, 17, 58, -10, 4, -1); |
3987 | 2.00M | for (y = 0; y < height; y++) { |
3988 | 4.67M | for (x = 0; x < width; x += 8) { |
3989 | | |
3990 | 2.85M | x1 = _mm_load_si128((__m128i *) &tmp[x - 3 * srcstride]); |
3991 | 2.85M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
3992 | 2.85M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
3993 | 2.85M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
3994 | 2.85M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
3995 | 2.85M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
3996 | 2.85M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
3997 | | |
3998 | 2.85M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 0)); |
3999 | 2.85M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
4000 | 2.85M | t8 = _mm_mullo_epi16(x1, r0); |
4001 | 2.85M | rBuffer = _mm_mulhi_epi16(x1, r0); |
4002 | 2.85M | t7 = _mm_mullo_epi16(x2, r1); |
4003 | 2.85M | t1 = _mm_unpacklo_epi16(t8, rBuffer); |
4004 | 2.85M | x1 = _mm_unpackhi_epi16(t8, rBuffer); |
4005 | | |
4006 | 2.85M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
4007 | 2.85M | rBuffer = _mm_mulhi_epi16(x2, r1); |
4008 | 2.85M | t8 = _mm_mullo_epi16(x3, r0); |
4009 | 2.85M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
4010 | 2.85M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
4011 | | |
4012 | 2.85M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
4013 | 2.85M | rBuffer = _mm_mulhi_epi16(x3, r0); |
4014 | 2.85M | t7 = _mm_mullo_epi16(x4, r1); |
4015 | 2.85M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
4016 | 2.85M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
4017 | | |
4018 | 2.85M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
4019 | 2.85M | rBuffer = _mm_mulhi_epi16(x4, r1); |
4020 | 2.85M | t8 = _mm_mullo_epi16(x5, r0); |
4021 | 2.85M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
4022 | 2.85M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
4023 | | |
4024 | 2.85M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
4025 | 2.85M | rBuffer = _mm_mulhi_epi16(x5, r0); |
4026 | 2.85M | t7 = _mm_mullo_epi16(x6, r1); |
4027 | 2.85M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
4028 | 2.85M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
4029 | | |
4030 | 2.85M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
4031 | 2.85M | rBuffer = _mm_mulhi_epi16(x6, r1); |
4032 | 2.85M | t8 = _mm_mullo_epi16(x7, r0); |
4033 | 2.85M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
4034 | 2.85M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
4035 | | |
4036 | 2.85M | rBuffer = _mm_mulhi_epi16(x7, r0); |
4037 | 2.85M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
4038 | 2.85M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
4039 | | |
4040 | | |
4041 | | |
4042 | | /* add calculus by correct value : */ |
4043 | | |
4044 | 2.85M | r1 = _mm_add_epi32(x1, x2); |
4045 | 2.85M | x3 = _mm_add_epi32(x3, x4); |
4046 | 2.85M | x5 = _mm_add_epi32(x5, x6); |
4047 | 2.85M | r1 = _mm_add_epi32(r1, x3); |
4048 | 2.85M | r1 = _mm_add_epi32(r1, x5); |
4049 | | |
4050 | 2.85M | r0 = _mm_add_epi32(t1, t2); |
4051 | 2.85M | t3 = _mm_add_epi32(t3, t4); |
4052 | 2.85M | t5 = _mm_add_epi32(t5, t6); |
4053 | 2.85M | r0 = _mm_add_epi32(r0, t3); |
4054 | 2.85M | r0 = _mm_add_epi32(r0, t5); |
4055 | 2.85M | r1 = _mm_add_epi32(r1, x7); |
4056 | 2.85M | r0 = _mm_add_epi32(r0, t7); |
4057 | 2.85M | r1 = _mm_srli_epi32(r1, 6); |
4058 | 2.85M | r0 = _mm_srli_epi32(r0, 6); |
4059 | | |
4060 | 2.85M | r1 = _mm_and_si128(r1, |
4061 | 2.85M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4062 | 2.85M | r0 = _mm_and_si128(r0, |
4063 | 2.85M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4064 | 2.85M | r0 = _mm_hadd_epi16(r0, r1); |
4065 | 2.85M | _mm_store_si128((__m128i *) &dst[x], r0); |
4066 | | |
4067 | 2.85M | } |
4068 | 1.81M | tmp += MAX_PB_SIZE; |
4069 | 1.81M | dst += dststride; |
4070 | 1.81M | } |
4071 | 190k | } |
4072 | | void ff_hevc_put_hevc_qpel_h_2_v_2_sse(int16_t *dst, ptrdiff_t dststride, |
4073 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
4074 | 295k | int16_t* mcbuffer) { |
4075 | 295k | int x, y; |
4076 | 295k | uint8_t *src = (uint8_t*) _src; |
4077 | 295k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
4078 | 295k | int16_t *tmp = mcbuffer; |
4079 | 295k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, rBuffer, rTemp, r0, r1; |
4080 | 295k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
4081 | | |
4082 | 295k | src -= qpel_extra_before[2] * srcstride; |
4083 | 295k | r0 = _mm_set_epi8(-1, 4, -11, 40, 40, -11, 4, -1, -1, 4, -11, 40, 40, -11, |
4084 | 295k | 4, -1); |
4085 | | |
4086 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
4087 | 295k | if (width == 4) { |
4088 | | |
4089 | 546k | for (y = 0; y < height + qpel_extra[2]; y += 2) { |
4090 | | /* load data in register */ |
4091 | 486k | x1 = _mm_loadu_si128((__m128i *) &src[-3]); |
4092 | 486k | src += srcstride; |
4093 | 486k | t1 = _mm_loadu_si128((__m128i *) &src[-3]); |
4094 | 486k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4095 | 486k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
4096 | 486k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4097 | 486k | _mm_srli_si128(x1, 3)); |
4098 | 486k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
4099 | 486k | _mm_srli_si128(t1, 3)); |
4100 | | |
4101 | | /* PMADDUBSW then PMADDW */ |
4102 | 486k | x2 = _mm_maddubs_epi16(x2, r0); |
4103 | 486k | t2 = _mm_maddubs_epi16(t2, r0); |
4104 | 486k | x3 = _mm_maddubs_epi16(x3, r0); |
4105 | 486k | t3 = _mm_maddubs_epi16(t3, r0); |
4106 | 486k | x2 = _mm_hadd_epi16(x2, x3); |
4107 | 486k | t2 = _mm_hadd_epi16(t2, t3); |
4108 | 486k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
4109 | 486k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
4110 | 486k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
4111 | 486k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
4112 | | /* give results back */ |
4113 | 486k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
4114 | | |
4115 | 486k | tmp += MAX_PB_SIZE; |
4116 | 486k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
4117 | | |
4118 | 486k | src += srcstride; |
4119 | 486k | tmp += MAX_PB_SIZE; |
4120 | 486k | } |
4121 | 60.3k | } else |
4122 | 3.92M | for (y = 0; y < height + qpel_extra[2]; y++) { |
4123 | 8.75M | for (x = 0; x < width; x += 8) { |
4124 | | /* load data in register */ |
4125 | 5.05M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
4126 | 5.05M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4127 | 5.05M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4128 | 5.05M | _mm_srli_si128(x1, 3)); |
4129 | 5.05M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
4130 | 5.05M | _mm_srli_si128(x1, 5)); |
4131 | 5.05M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
4132 | 5.05M | _mm_srli_si128(x1, 7)); |
4133 | | |
4134 | | /* PMADDUBSW then PMADDW */ |
4135 | 5.05M | x2 = _mm_maddubs_epi16(x2, r0); |
4136 | 5.05M | x3 = _mm_maddubs_epi16(x3, r0); |
4137 | 5.05M | x4 = _mm_maddubs_epi16(x4, r0); |
4138 | 5.05M | x5 = _mm_maddubs_epi16(x5, r0); |
4139 | 5.05M | x2 = _mm_hadd_epi16(x2, x3); |
4140 | 5.05M | x4 = _mm_hadd_epi16(x4, x5); |
4141 | 5.05M | x2 = _mm_hadd_epi16(x2, x4); |
4142 | 5.05M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
4143 | | |
4144 | | /* give results back */ |
4145 | 5.05M | _mm_store_si128((__m128i *) &tmp[x], x2); |
4146 | | |
4147 | 5.05M | } |
4148 | 3.69M | src += srcstride; |
4149 | 3.69M | tmp += MAX_PB_SIZE; |
4150 | 3.69M | } |
4151 | | |
4152 | 295k | tmp = mcbuffer + qpel_extra_before[2] * MAX_PB_SIZE; |
4153 | 295k | srcstride = MAX_PB_SIZE; |
4154 | | |
4155 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
4156 | | for register calculations */ |
4157 | 295k | rTemp = _mm_set_epi16(-1, 4, -11, 40, 40, -11, 4, -1); |
4158 | 2.83M | for (y = 0; y < height; y++) { |
4159 | 6.11M | for (x = 0; x < width; x += 8) { |
4160 | | |
4161 | 3.57M | x1 = _mm_load_si128((__m128i *) &tmp[x - 3 * srcstride]); |
4162 | 3.57M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
4163 | 3.57M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
4164 | 3.57M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
4165 | 3.57M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
4166 | 3.57M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
4167 | 3.57M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
4168 | 3.57M | x8 = _mm_load_si128((__m128i *) &tmp[x + 4 * srcstride]); |
4169 | | |
4170 | 3.57M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 0)); |
4171 | 3.57M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
4172 | 3.57M | t8 = _mm_mullo_epi16(x1, r0); |
4173 | 3.57M | rBuffer = _mm_mulhi_epi16(x1, r0); |
4174 | 3.57M | t7 = _mm_mullo_epi16(x2, r1); |
4175 | 3.57M | t1 = _mm_unpacklo_epi16(t8, rBuffer); |
4176 | 3.57M | x1 = _mm_unpackhi_epi16(t8, rBuffer); |
4177 | | |
4178 | 3.57M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
4179 | 3.57M | rBuffer = _mm_mulhi_epi16(x2, r1); |
4180 | 3.57M | t8 = _mm_mullo_epi16(x3, r0); |
4181 | 3.57M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
4182 | 3.57M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
4183 | | |
4184 | 3.57M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
4185 | 3.57M | rBuffer = _mm_mulhi_epi16(x3, r0); |
4186 | 3.57M | t7 = _mm_mullo_epi16(x4, r1); |
4187 | 3.57M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
4188 | 3.57M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
4189 | | |
4190 | 3.57M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
4191 | 3.57M | rBuffer = _mm_mulhi_epi16(x4, r1); |
4192 | 3.57M | t8 = _mm_mullo_epi16(x5, r0); |
4193 | 3.57M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
4194 | 3.57M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
4195 | | |
4196 | 3.57M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
4197 | 3.57M | rBuffer = _mm_mulhi_epi16(x5, r0); |
4198 | 3.57M | t7 = _mm_mullo_epi16(x6, r1); |
4199 | 3.57M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
4200 | 3.57M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
4201 | | |
4202 | 3.57M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
4203 | 3.57M | rBuffer = _mm_mulhi_epi16(x6, r1); |
4204 | 3.57M | t8 = _mm_mullo_epi16(x7, r0); |
4205 | 3.57M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
4206 | 3.57M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
4207 | | |
4208 | 3.57M | rBuffer = _mm_mulhi_epi16(x7, r0); |
4209 | 3.57M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
4210 | 3.57M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
4211 | | |
4212 | 3.57M | t8 = _mm_unpacklo_epi16( |
4213 | 3.57M | _mm_mullo_epi16(x8, |
4214 | 3.57M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4215 | 3.57M | _mm_mulhi_epi16(x8, |
4216 | 3.57M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4217 | 3.57M | x8 = _mm_unpackhi_epi16( |
4218 | 3.57M | _mm_mullo_epi16(x8, |
4219 | 3.57M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4220 | 3.57M | _mm_mulhi_epi16(x8, |
4221 | 3.57M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4222 | | |
4223 | | /* add calculus by correct value : */ |
4224 | | |
4225 | 3.57M | r1 = _mm_add_epi32(x1, x2); |
4226 | 3.57M | x3 = _mm_add_epi32(x3, x4); |
4227 | 3.57M | x5 = _mm_add_epi32(x5, x6); |
4228 | 3.57M | r1 = _mm_add_epi32(r1, x3); |
4229 | 3.57M | x7 = _mm_add_epi32(x7, x8); |
4230 | 3.57M | r1 = _mm_add_epi32(r1, x5); |
4231 | | |
4232 | 3.57M | r0 = _mm_add_epi32(t1, t2); |
4233 | 3.57M | t3 = _mm_add_epi32(t3, t4); |
4234 | 3.57M | t5 = _mm_add_epi32(t5, t6); |
4235 | 3.57M | r0 = _mm_add_epi32(r0, t3); |
4236 | 3.57M | t7 = _mm_add_epi32(t7, t8); |
4237 | 3.57M | r0 = _mm_add_epi32(r0, t5); |
4238 | 3.57M | r1 = _mm_add_epi32(r1, x7); |
4239 | 3.57M | r0 = _mm_add_epi32(r0, t7); |
4240 | 3.57M | r1 = _mm_srli_epi32(r1, 6); |
4241 | 3.57M | r0 = _mm_srli_epi32(r0, 6); |
4242 | | |
4243 | 3.57M | r1 = _mm_and_si128(r1, |
4244 | 3.57M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4245 | 3.57M | r0 = _mm_and_si128(r0, |
4246 | 3.57M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4247 | 3.57M | r0 = _mm_hadd_epi16(r0, r1); |
4248 | 3.57M | _mm_store_si128((__m128i *) &dst[x], r0); |
4249 | | |
4250 | 3.57M | } |
4251 | 2.54M | tmp += MAX_PB_SIZE; |
4252 | 2.54M | dst += dststride; |
4253 | 2.54M | } |
4254 | 295k | } |
4255 | | void ff_hevc_put_hevc_qpel_h_2_v_3_sse(int16_t *dst, ptrdiff_t dststride, |
4256 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
4257 | 111k | int16_t* mcbuffer) { |
4258 | 111k | int x, y; |
4259 | 111k | uint8_t *src = (uint8_t*) _src; |
4260 | 111k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
4261 | 111k | int16_t *tmp = mcbuffer; |
4262 | 111k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, rBuffer, rTemp, r0, r1; |
4263 | 111k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
4264 | | |
4265 | 111k | src -= qpel_extra_before[3] * srcstride; |
4266 | 111k | r0 = _mm_set_epi8(-1, 4, -11, 40, 40, -11, 4, -1, -1, 4, -11, 40, 40, -11, |
4267 | 111k | 4, -1); |
4268 | | |
4269 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
4270 | 111k | if (width == 4) { |
4271 | | |
4272 | 189k | for (y = 0; y < height + qpel_extra[3]; y += 2) { |
4273 | | /* load data in register */ |
4274 | 166k | x1 = _mm_loadu_si128((__m128i *) &src[-3]); |
4275 | 166k | src += srcstride; |
4276 | 166k | t1 = _mm_loadu_si128((__m128i *) &src[-3]); |
4277 | 166k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4278 | 166k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
4279 | 166k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4280 | 166k | _mm_srli_si128(x1, 3)); |
4281 | 166k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
4282 | 166k | _mm_srli_si128(t1, 3)); |
4283 | | |
4284 | | /* PMADDUBSW then PMADDW */ |
4285 | 166k | x2 = _mm_maddubs_epi16(x2, r0); |
4286 | 166k | t2 = _mm_maddubs_epi16(t2, r0); |
4287 | 166k | x3 = _mm_maddubs_epi16(x3, r0); |
4288 | 166k | t3 = _mm_maddubs_epi16(t3, r0); |
4289 | 166k | x2 = _mm_hadd_epi16(x2, x3); |
4290 | 166k | t2 = _mm_hadd_epi16(t2, t3); |
4291 | 166k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
4292 | 166k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
4293 | 166k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
4294 | 166k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
4295 | | /* give results back */ |
4296 | 166k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
4297 | | |
4298 | 166k | tmp += MAX_PB_SIZE; |
4299 | 166k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
4300 | | |
4301 | 166k | src += srcstride; |
4302 | 166k | tmp += MAX_PB_SIZE; |
4303 | 166k | } |
4304 | 23.4k | } else |
4305 | 1.51M | for (y = 0; y < height + qpel_extra[3]; y++) { |
4306 | 3.59M | for (x = 0; x < width; x += 8) { |
4307 | | /* load data in register */ |
4308 | 2.17M | x1 = _mm_loadu_si128((__m128i *) &src[x - 3]); |
4309 | 2.17M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4310 | 2.17M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4311 | 2.17M | _mm_srli_si128(x1, 3)); |
4312 | 2.17M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
4313 | 2.17M | _mm_srli_si128(x1, 5)); |
4314 | 2.17M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
4315 | 2.17M | _mm_srli_si128(x1, 7)); |
4316 | | |
4317 | | /* PMADDUBSW then PMADDW */ |
4318 | 2.17M | x2 = _mm_maddubs_epi16(x2, r0); |
4319 | 2.17M | x3 = _mm_maddubs_epi16(x3, r0); |
4320 | 2.17M | x4 = _mm_maddubs_epi16(x4, r0); |
4321 | 2.17M | x5 = _mm_maddubs_epi16(x5, r0); |
4322 | 2.17M | x2 = _mm_hadd_epi16(x2, x3); |
4323 | 2.17M | x4 = _mm_hadd_epi16(x4, x5); |
4324 | 2.17M | x2 = _mm_hadd_epi16(x2, x4); |
4325 | 2.17M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
4326 | | |
4327 | | /* give results back */ |
4328 | 2.17M | _mm_store_si128((__m128i *) &tmp[x], x2); |
4329 | | |
4330 | 2.17M | } |
4331 | 1.42M | src += srcstride; |
4332 | 1.42M | tmp += MAX_PB_SIZE; |
4333 | 1.42M | } |
4334 | | |
4335 | 111k | tmp = mcbuffer + qpel_extra_before[3] * MAX_PB_SIZE; |
4336 | 111k | srcstride = MAX_PB_SIZE; |
4337 | | |
4338 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
4339 | | for register calculations */ |
4340 | 111k | rTemp = _mm_set_epi16(-1, 4, -10, 58, 17, -5, 1, 0); |
4341 | 1.20M | for (y = 0; y < height; y++) { |
4342 | 2.76M | for (x = 0; x < width; x += 8) { |
4343 | | |
4344 | 1.67M | x1 = _mm_setzero_si128(); |
4345 | 1.67M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
4346 | 1.67M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
4347 | 1.67M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
4348 | 1.67M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
4349 | 1.67M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
4350 | 1.67M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
4351 | 1.67M | x8 = _mm_load_si128((__m128i *) &tmp[x + 4 * srcstride]); |
4352 | | |
4353 | 1.67M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
4354 | | |
4355 | 1.67M | t7 = _mm_mullo_epi16(x2, r1); |
4356 | | |
4357 | | |
4358 | 1.67M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
4359 | 1.67M | rBuffer = _mm_mulhi_epi16(x2, r1); |
4360 | 1.67M | t8 = _mm_mullo_epi16(x3, r0); |
4361 | 1.67M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
4362 | 1.67M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
4363 | | |
4364 | 1.67M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
4365 | 1.67M | rBuffer = _mm_mulhi_epi16(x3, r0); |
4366 | 1.67M | t7 = _mm_mullo_epi16(x4, r1); |
4367 | 1.67M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
4368 | 1.67M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
4369 | | |
4370 | 1.67M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
4371 | 1.67M | rBuffer = _mm_mulhi_epi16(x4, r1); |
4372 | 1.67M | t8 = _mm_mullo_epi16(x5, r0); |
4373 | 1.67M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
4374 | 1.67M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
4375 | | |
4376 | 1.67M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
4377 | 1.67M | rBuffer = _mm_mulhi_epi16(x5, r0); |
4378 | 1.67M | t7 = _mm_mullo_epi16(x6, r1); |
4379 | 1.67M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
4380 | 1.67M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
4381 | | |
4382 | 1.67M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
4383 | 1.67M | rBuffer = _mm_mulhi_epi16(x6, r1); |
4384 | 1.67M | t8 = _mm_mullo_epi16(x7, r0); |
4385 | 1.67M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
4386 | 1.67M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
4387 | | |
4388 | 1.67M | rBuffer = _mm_mulhi_epi16(x7, r0); |
4389 | 1.67M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
4390 | 1.67M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
4391 | | |
4392 | 1.67M | t8 = _mm_unpacklo_epi16( |
4393 | 1.67M | _mm_mullo_epi16(x8, |
4394 | 1.67M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4395 | 1.67M | _mm_mulhi_epi16(x8, |
4396 | 1.67M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4397 | 1.67M | x8 = _mm_unpackhi_epi16( |
4398 | 1.67M | _mm_mullo_epi16(x8, |
4399 | 1.67M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4400 | 1.67M | _mm_mulhi_epi16(x8, |
4401 | 1.67M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4402 | | |
4403 | | /* add calculus by correct value : */ |
4404 | | |
4405 | 1.67M | x3 = _mm_add_epi32(x3, x4); |
4406 | 1.67M | x5 = _mm_add_epi32(x5, x6); |
4407 | 1.67M | r1 = _mm_add_epi32(x2, x3); |
4408 | 1.67M | x7 = _mm_add_epi32(x7, x8); |
4409 | 1.67M | r1 = _mm_add_epi32(r1, x5); |
4410 | | |
4411 | 1.67M | t3 = _mm_add_epi32(t3, t4); |
4412 | 1.67M | t5 = _mm_add_epi32(t5, t6); |
4413 | 1.67M | r0 = _mm_add_epi32(t2, t3); |
4414 | 1.67M | t7 = _mm_add_epi32(t7, t8); |
4415 | 1.67M | r0 = _mm_add_epi32(r0, t5); |
4416 | 1.67M | r1 = _mm_add_epi32(r1, x7); |
4417 | 1.67M | r0 = _mm_add_epi32(r0, t7); |
4418 | 1.67M | r1 = _mm_srli_epi32(r1, 6); |
4419 | 1.67M | r0 = _mm_srli_epi32(r0, 6); |
4420 | | |
4421 | 1.67M | r1 = _mm_and_si128(r1, |
4422 | 1.67M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4423 | 1.67M | r0 = _mm_and_si128(r0, |
4424 | 1.67M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4425 | 1.67M | r0 = _mm_hadd_epi16(r0, r1); |
4426 | 1.67M | _mm_store_si128((__m128i *) &dst[x], r0); |
4427 | | |
4428 | 1.67M | } |
4429 | 1.09M | tmp += MAX_PB_SIZE; |
4430 | 1.09M | dst += dststride; |
4431 | 1.09M | } |
4432 | 111k | } |
4433 | | void ff_hevc_put_hevc_qpel_h_3_v_1_sse(int16_t *dst, ptrdiff_t dststride, |
4434 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
4435 | 149k | int16_t* mcbuffer) { |
4436 | 149k | int x, y; |
4437 | 149k | uint8_t *src = (uint8_t*) _src; |
4438 | 149k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
4439 | 149k | int16_t *tmp = mcbuffer; |
4440 | 149k | __m128i x1, x2, x3, x4, x5, x6, x7, rBuffer, rTemp, r0, r1; |
4441 | 149k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
4442 | | |
4443 | 149k | src -= qpel_extra_before[1] * srcstride; |
4444 | 149k | r0 = _mm_set_epi8(-1, 4, -10, 58, 17, -5, 1, 0, -1, 4, -10, 58, 17, -5, 1, |
4445 | 149k | 0); |
4446 | | |
4447 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
4448 | 149k | if (width == 4) { |
4449 | | |
4450 | 177k | for (y = 0; y < height + qpel_extra[1]; y += 2) { |
4451 | | /* load data in register */ |
4452 | 155k | x1 = _mm_loadu_si128((__m128i *) &src[-2]); |
4453 | 155k | x1 = _mm_slli_si128(x1, 1); |
4454 | 155k | src += srcstride; |
4455 | 155k | t1 = _mm_loadu_si128((__m128i *) &src[-2]); |
4456 | 155k | t1 = _mm_slli_si128(t1, 1); |
4457 | 155k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4458 | 155k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
4459 | 155k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4460 | 155k | _mm_srli_si128(x1, 3)); |
4461 | 155k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
4462 | 155k | _mm_srli_si128(t1, 3)); |
4463 | | |
4464 | | /* PMADDUBSW then PMADDW */ |
4465 | 155k | x2 = _mm_maddubs_epi16(x2, r0); |
4466 | 155k | t2 = _mm_maddubs_epi16(t2, r0); |
4467 | 155k | x3 = _mm_maddubs_epi16(x3, r0); |
4468 | 155k | t3 = _mm_maddubs_epi16(t3, r0); |
4469 | 155k | x2 = _mm_hadd_epi16(x2, x3); |
4470 | 155k | t2 = _mm_hadd_epi16(t2, t3); |
4471 | 155k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
4472 | 155k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
4473 | 155k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
4474 | 155k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
4475 | | /* give results back */ |
4476 | 155k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
4477 | | |
4478 | 155k | tmp += MAX_PB_SIZE; |
4479 | 155k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
4480 | | |
4481 | 155k | src += srcstride; |
4482 | 155k | tmp += MAX_PB_SIZE; |
4483 | 155k | } |
4484 | 21.8k | } else |
4485 | 2.15M | for (y = 0; y < height + qpel_extra[1]; y++) { |
4486 | 5.16M | for (x = 0; x < width; x += 8) { |
4487 | | /* load data in register */ |
4488 | 3.13M | x1 = _mm_loadu_si128((__m128i *) &src[x - 2]); |
4489 | 3.13M | x1 = _mm_slli_si128(x1, 1); |
4490 | 3.13M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4491 | 3.13M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4492 | 3.13M | _mm_srli_si128(x1, 3)); |
4493 | 3.13M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
4494 | 3.13M | _mm_srli_si128(x1, 5)); |
4495 | 3.13M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
4496 | 3.13M | _mm_srli_si128(x1, 7)); |
4497 | | |
4498 | | /* PMADDUBSW then PMADDW */ |
4499 | 3.13M | x2 = _mm_maddubs_epi16(x2, r0); |
4500 | 3.13M | x3 = _mm_maddubs_epi16(x3, r0); |
4501 | 3.13M | x4 = _mm_maddubs_epi16(x4, r0); |
4502 | 3.13M | x5 = _mm_maddubs_epi16(x5, r0); |
4503 | 3.13M | x2 = _mm_hadd_epi16(x2, x3); |
4504 | 3.13M | x4 = _mm_hadd_epi16(x4, x5); |
4505 | 3.13M | x2 = _mm_hadd_epi16(x2, x4); |
4506 | 3.13M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
4507 | | |
4508 | | /* give results back */ |
4509 | 3.13M | _mm_store_si128((__m128i *) &tmp[x], x2); |
4510 | | |
4511 | 3.13M | } |
4512 | 2.02M | src += srcstride; |
4513 | 2.02M | tmp += MAX_PB_SIZE; |
4514 | 2.02M | } |
4515 | | |
4516 | 149k | tmp = mcbuffer + qpel_extra_before[1] * MAX_PB_SIZE; |
4517 | 149k | srcstride = MAX_PB_SIZE; |
4518 | | |
4519 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
4520 | | for register calculations */ |
4521 | 149k | rTemp = _mm_set_epi16(0, 1, -5, 17, 58, -10, 4, -1); |
4522 | 1.59M | for (y = 0; y < height; y++) { |
4523 | 3.75M | for (x = 0; x < width; x += 8) { |
4524 | | |
4525 | 2.31M | x1 = _mm_load_si128((__m128i *) &tmp[x - 3 * srcstride]); |
4526 | 2.31M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
4527 | 2.31M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
4528 | 2.31M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
4529 | 2.31M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
4530 | 2.31M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
4531 | 2.31M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
4532 | | |
4533 | 2.31M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 0)); |
4534 | 2.31M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
4535 | 2.31M | t8 = _mm_mullo_epi16(x1, r0); |
4536 | 2.31M | rBuffer = _mm_mulhi_epi16(x1, r0); |
4537 | 2.31M | t7 = _mm_mullo_epi16(x2, r1); |
4538 | 2.31M | t1 = _mm_unpacklo_epi16(t8, rBuffer); |
4539 | 2.31M | x1 = _mm_unpackhi_epi16(t8, rBuffer); |
4540 | | |
4541 | 2.31M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
4542 | 2.31M | rBuffer = _mm_mulhi_epi16(x2, r1); |
4543 | 2.31M | t8 = _mm_mullo_epi16(x3, r0); |
4544 | 2.31M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
4545 | 2.31M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
4546 | | |
4547 | 2.31M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
4548 | 2.31M | rBuffer = _mm_mulhi_epi16(x3, r0); |
4549 | 2.31M | t7 = _mm_mullo_epi16(x4, r1); |
4550 | 2.31M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
4551 | 2.31M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
4552 | | |
4553 | 2.31M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
4554 | 2.31M | rBuffer = _mm_mulhi_epi16(x4, r1); |
4555 | 2.31M | t8 = _mm_mullo_epi16(x5, r0); |
4556 | 2.31M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
4557 | 2.31M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
4558 | | |
4559 | 2.31M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
4560 | 2.31M | rBuffer = _mm_mulhi_epi16(x5, r0); |
4561 | 2.31M | t7 = _mm_mullo_epi16(x6, r1); |
4562 | 2.31M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
4563 | 2.31M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
4564 | | |
4565 | 2.31M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
4566 | 2.31M | rBuffer = _mm_mulhi_epi16(x6, r1); |
4567 | 2.31M | t8 = _mm_mullo_epi16(x7, r0); |
4568 | 2.31M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
4569 | 2.31M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
4570 | | |
4571 | 2.31M | rBuffer = _mm_mulhi_epi16(x7, r0); |
4572 | 2.31M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
4573 | 2.31M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
4574 | | |
4575 | | |
4576 | | /* add calculus by correct value : */ |
4577 | | |
4578 | 2.31M | r1 = _mm_add_epi32(x1, x2); |
4579 | 2.31M | x3 = _mm_add_epi32(x3, x4); |
4580 | 2.31M | x5 = _mm_add_epi32(x5, x6); |
4581 | 2.31M | r1 = _mm_add_epi32(r1, x3); |
4582 | 2.31M | r1 = _mm_add_epi32(r1, x5); |
4583 | | |
4584 | 2.31M | r0 = _mm_add_epi32(t1, t2); |
4585 | 2.31M | t3 = _mm_add_epi32(t3, t4); |
4586 | 2.31M | t5 = _mm_add_epi32(t5, t6); |
4587 | 2.31M | r0 = _mm_add_epi32(r0, t3); |
4588 | 2.31M | r0 = _mm_add_epi32(r0, t5); |
4589 | 2.31M | r1 = _mm_add_epi32(r1, x7); |
4590 | 2.31M | r0 = _mm_add_epi32(r0, t7); |
4591 | 2.31M | r1 = _mm_srli_epi32(r1, 6); |
4592 | 2.31M | r0 = _mm_srli_epi32(r0, 6); |
4593 | | |
4594 | 2.31M | r1 = _mm_and_si128(r1, |
4595 | 2.31M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4596 | 2.31M | r0 = _mm_and_si128(r0, |
4597 | 2.31M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4598 | 2.31M | r0 = _mm_hadd_epi16(r0, r1); |
4599 | 2.31M | _mm_store_si128((__m128i *) &dst[x], r0); |
4600 | | |
4601 | 2.31M | } |
4602 | 1.44M | tmp += MAX_PB_SIZE; |
4603 | 1.44M | dst += dststride; |
4604 | 1.44M | } |
4605 | 149k | } |
4606 | | void ff_hevc_put_hevc_qpel_h_3_v_2_sse(int16_t *dst, ptrdiff_t dststride, |
4607 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
4608 | 144k | int16_t* mcbuffer) { |
4609 | 144k | int x, y; |
4610 | 144k | uint8_t *src = (uint8_t*) _src; |
4611 | 144k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
4612 | 144k | int16_t *tmp = mcbuffer; |
4613 | 144k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, rBuffer, rTemp, r0, r1; |
4614 | 144k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
4615 | | |
4616 | 144k | src -= qpel_extra_before[2] * srcstride; |
4617 | 144k | r0 = _mm_set_epi8(-1, 4, -10, 58, 17, -5, 1, 0, -1, 4, -10, 58, 17, -5, 1, |
4618 | 144k | 0); |
4619 | | |
4620 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
4621 | 144k | if (width == 4) { |
4622 | | |
4623 | 259k | for (y = 0; y < height + qpel_extra[2]; y += 2) { |
4624 | | /* load data in register */ |
4625 | 231k | x1 = _mm_loadu_si128((__m128i *) &src[-2]); |
4626 | 231k | x1 = _mm_slli_si128(x1, 1); |
4627 | 231k | src += srcstride; |
4628 | 231k | t1 = _mm_loadu_si128((__m128i *) &src[-2]); |
4629 | 231k | t1 = _mm_slli_si128(t1, 1); |
4630 | 231k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4631 | 231k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
4632 | 231k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4633 | 231k | _mm_srli_si128(x1, 3)); |
4634 | 231k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
4635 | 231k | _mm_srli_si128(t1, 3)); |
4636 | | |
4637 | | /* PMADDUBSW then PMADDW */ |
4638 | 231k | x2 = _mm_maddubs_epi16(x2, r0); |
4639 | 231k | t2 = _mm_maddubs_epi16(t2, r0); |
4640 | 231k | x3 = _mm_maddubs_epi16(x3, r0); |
4641 | 231k | t3 = _mm_maddubs_epi16(t3, r0); |
4642 | 231k | x2 = _mm_hadd_epi16(x2, x3); |
4643 | 231k | t2 = _mm_hadd_epi16(t2, t3); |
4644 | 231k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
4645 | 231k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
4646 | 231k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
4647 | 231k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
4648 | | /* give results back */ |
4649 | 231k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
4650 | | |
4651 | 231k | tmp += MAX_PB_SIZE; |
4652 | 231k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
4653 | | |
4654 | 231k | src += srcstride; |
4655 | 231k | tmp += MAX_PB_SIZE; |
4656 | 231k | } |
4657 | 28.6k | } else |
4658 | 2.09M | for (y = 0; y < height + qpel_extra[2]; y++) { |
4659 | 5.03M | for (x = 0; x < width; x += 8) { |
4660 | | /* load data in register */ |
4661 | 3.06M | x1 = _mm_loadu_si128((__m128i *) &src[x - 2]); |
4662 | 3.06M | x1 = _mm_slli_si128(x1, 1); |
4663 | 3.06M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4664 | 3.06M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4665 | 3.06M | _mm_srli_si128(x1, 3)); |
4666 | 3.06M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
4667 | 3.06M | _mm_srli_si128(x1, 5)); |
4668 | 3.06M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
4669 | 3.06M | _mm_srli_si128(x1, 7)); |
4670 | | |
4671 | | /* PMADDUBSW then PMADDW */ |
4672 | 3.06M | x2 = _mm_maddubs_epi16(x2, r0); |
4673 | 3.06M | x3 = _mm_maddubs_epi16(x3, r0); |
4674 | 3.06M | x4 = _mm_maddubs_epi16(x4, r0); |
4675 | 3.06M | x5 = _mm_maddubs_epi16(x5, r0); |
4676 | 3.06M | x2 = _mm_hadd_epi16(x2, x3); |
4677 | 3.06M | x4 = _mm_hadd_epi16(x4, x5); |
4678 | 3.06M | x2 = _mm_hadd_epi16(x2, x4); |
4679 | 3.06M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
4680 | | |
4681 | | /* give results back */ |
4682 | 3.06M | _mm_store_si128((__m128i *) &tmp[x], x2); |
4683 | | |
4684 | 3.06M | } |
4685 | 1.97M | src += srcstride; |
4686 | 1.97M | tmp += MAX_PB_SIZE; |
4687 | 1.97M | } |
4688 | | |
4689 | 144k | tmp = mcbuffer + qpel_extra_before[2] * MAX_PB_SIZE; |
4690 | 144k | srcstride = MAX_PB_SIZE; |
4691 | | |
4692 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
4693 | | for register calculations */ |
4694 | 144k | rTemp = _mm_set_epi16(-1, 4, -11, 40, 40, -11, 4, -1); |
4695 | 1.53M | for (y = 0; y < height; y++) { |
4696 | 3.60M | for (x = 0; x < width; x += 8) { |
4697 | | |
4698 | 2.21M | x1 = _mm_load_si128((__m128i *) &tmp[x - 3 * srcstride]); |
4699 | 2.21M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
4700 | 2.21M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
4701 | 2.21M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
4702 | 2.21M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
4703 | 2.21M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
4704 | 2.21M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
4705 | 2.21M | x8 = _mm_load_si128((__m128i *) &tmp[x + 4 * srcstride]); |
4706 | | |
4707 | 2.21M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 0)); |
4708 | 2.21M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
4709 | 2.21M | t8 = _mm_mullo_epi16(x1, r0); |
4710 | 2.21M | rBuffer = _mm_mulhi_epi16(x1, r0); |
4711 | 2.21M | t7 = _mm_mullo_epi16(x2, r1); |
4712 | 2.21M | t1 = _mm_unpacklo_epi16(t8, rBuffer); |
4713 | 2.21M | x1 = _mm_unpackhi_epi16(t8, rBuffer); |
4714 | | |
4715 | 2.21M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
4716 | 2.21M | rBuffer = _mm_mulhi_epi16(x2, r1); |
4717 | 2.21M | t8 = _mm_mullo_epi16(x3, r0); |
4718 | 2.21M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
4719 | 2.21M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
4720 | | |
4721 | 2.21M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
4722 | 2.21M | rBuffer = _mm_mulhi_epi16(x3, r0); |
4723 | 2.21M | t7 = _mm_mullo_epi16(x4, r1); |
4724 | 2.21M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
4725 | 2.21M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
4726 | | |
4727 | 2.21M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
4728 | 2.21M | rBuffer = _mm_mulhi_epi16(x4, r1); |
4729 | 2.21M | t8 = _mm_mullo_epi16(x5, r0); |
4730 | 2.21M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
4731 | 2.21M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
4732 | | |
4733 | 2.21M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
4734 | 2.21M | rBuffer = _mm_mulhi_epi16(x5, r0); |
4735 | 2.21M | t7 = _mm_mullo_epi16(x6, r1); |
4736 | 2.21M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
4737 | 2.21M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
4738 | | |
4739 | 2.21M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
4740 | 2.21M | rBuffer = _mm_mulhi_epi16(x6, r1); |
4741 | 2.21M | t8 = _mm_mullo_epi16(x7, r0); |
4742 | 2.21M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
4743 | 2.21M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
4744 | | |
4745 | 2.21M | rBuffer = _mm_mulhi_epi16(x7, r0); |
4746 | 2.21M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
4747 | 2.21M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
4748 | | |
4749 | 2.21M | t8 = _mm_unpacklo_epi16( |
4750 | 2.21M | _mm_mullo_epi16(x8, |
4751 | 2.21M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4752 | 2.21M | _mm_mulhi_epi16(x8, |
4753 | 2.21M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4754 | 2.21M | x8 = _mm_unpackhi_epi16( |
4755 | 2.21M | _mm_mullo_epi16(x8, |
4756 | 2.21M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4757 | 2.21M | _mm_mulhi_epi16(x8, |
4758 | 2.21M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4759 | | |
4760 | | /* add calculus by correct value : */ |
4761 | | |
4762 | 2.21M | r1 = _mm_add_epi32(x1, x2); |
4763 | 2.21M | x3 = _mm_add_epi32(x3, x4); |
4764 | 2.21M | x5 = _mm_add_epi32(x5, x6); |
4765 | 2.21M | r1 = _mm_add_epi32(r1, x3); |
4766 | 2.21M | x7 = _mm_add_epi32(x7, x8); |
4767 | 2.21M | r1 = _mm_add_epi32(r1, x5); |
4768 | | |
4769 | 2.21M | r0 = _mm_add_epi32(t1, t2); |
4770 | 2.21M | t3 = _mm_add_epi32(t3, t4); |
4771 | 2.21M | t5 = _mm_add_epi32(t5, t6); |
4772 | 2.21M | r0 = _mm_add_epi32(r0, t3); |
4773 | 2.21M | t7 = _mm_add_epi32(t7, t8); |
4774 | 2.21M | r0 = _mm_add_epi32(r0, t5); |
4775 | 2.21M | r1 = _mm_add_epi32(r1, x7); |
4776 | 2.21M | r0 = _mm_add_epi32(r0, t7); |
4777 | 2.21M | r1 = _mm_srli_epi32(r1, 6); |
4778 | 2.21M | r0 = _mm_srli_epi32(r0, 6); |
4779 | | |
4780 | 2.21M | r1 = _mm_and_si128(r1, |
4781 | 2.21M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4782 | 2.21M | r0 = _mm_and_si128(r0, |
4783 | 2.21M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4784 | 2.21M | r0 = _mm_hadd_epi16(r0, r1); |
4785 | 2.21M | _mm_store_si128((__m128i *) &dst[x], r0); |
4786 | | |
4787 | 2.21M | } |
4788 | 1.39M | tmp += MAX_PB_SIZE; |
4789 | 1.39M | dst += dststride; |
4790 | 1.39M | } |
4791 | 144k | } |
4792 | | void ff_hevc_put_hevc_qpel_h_3_v_3_sse(int16_t *dst, ptrdiff_t dststride, |
4793 | | const uint8_t *_src, ptrdiff_t _srcstride, int width, int height, |
4794 | 235k | int16_t* mcbuffer) { |
4795 | 235k | int x, y; |
4796 | 235k | uint8_t *src = (uint8_t*) _src; |
4797 | 235k | ptrdiff_t srcstride = _srcstride / sizeof(uint8_t); |
4798 | 235k | int16_t *tmp = mcbuffer; |
4799 | 235k | __m128i x1, x2, x3, x4, x5, x6, x7, x8, rBuffer, rTemp, r0, r1; |
4800 | 235k | __m128i t1, t2, t3, t4, t5, t6, t7, t8; |
4801 | | |
4802 | 235k | src -= qpel_extra_before[3] * srcstride; |
4803 | 235k | r0 = _mm_set_epi8(-1, 4, -10, 58, 17, -5, 1, 0, -1, 4, -10, 58, 17, -5, 1, |
4804 | 235k | 0); |
4805 | | |
4806 | | /* LOAD src from memory to registers to limit memory bandwidth */ |
4807 | 235k | if (width == 4) { |
4808 | | |
4809 | 393k | for (y = 0; y < height + qpel_extra[3]; y += 2) { |
4810 | | /* load data in register */ |
4811 | 344k | x1 = _mm_loadu_si128((__m128i *) &src[-2]); |
4812 | 344k | x1 = _mm_slli_si128(x1, 1); |
4813 | 344k | src += srcstride; |
4814 | 344k | t1 = _mm_loadu_si128((__m128i *) &src[-2]); |
4815 | 344k | t1 = _mm_slli_si128(t1, 1); |
4816 | 344k | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4817 | 344k | t2 = _mm_unpacklo_epi64(t1, _mm_srli_si128(t1, 1)); |
4818 | 344k | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4819 | 344k | _mm_srli_si128(x1, 3)); |
4820 | 344k | t3 = _mm_unpacklo_epi64(_mm_srli_si128(t1, 2), |
4821 | 344k | _mm_srli_si128(t1, 3)); |
4822 | | |
4823 | | /* PMADDUBSW then PMADDW */ |
4824 | 344k | x2 = _mm_maddubs_epi16(x2, r0); |
4825 | 344k | t2 = _mm_maddubs_epi16(t2, r0); |
4826 | 344k | x3 = _mm_maddubs_epi16(x3, r0); |
4827 | 344k | t3 = _mm_maddubs_epi16(t3, r0); |
4828 | 344k | x2 = _mm_hadd_epi16(x2, x3); |
4829 | 344k | t2 = _mm_hadd_epi16(t2, t3); |
4830 | 344k | x2 = _mm_hadd_epi16(x2, _mm_set1_epi16(0)); |
4831 | 344k | t2 = _mm_hadd_epi16(t2, _mm_set1_epi16(0)); |
4832 | 344k | x2 = _mm_srli_epi16(x2, BIT_DEPTH - 8); |
4833 | 344k | t2 = _mm_srli_epi16(t2, BIT_DEPTH - 8); |
4834 | | /* give results back */ |
4835 | 344k | _mm_storel_epi64((__m128i *) &tmp[0], x2); |
4836 | | |
4837 | 344k | tmp += MAX_PB_SIZE; |
4838 | 344k | _mm_storel_epi64((__m128i *) &tmp[0], t2); |
4839 | | |
4840 | 344k | src += srcstride; |
4841 | 344k | tmp += MAX_PB_SIZE; |
4842 | 344k | } |
4843 | 48.2k | } else |
4844 | 3.11M | for (y = 0; y < height + qpel_extra[3]; y++) { |
4845 | 7.50M | for (x = 0; x < width; x += 8) { |
4846 | | /* load data in register */ |
4847 | 4.57M | x1 = _mm_loadu_si128((__m128i *) &src[x - 2]); |
4848 | 4.57M | x1 = _mm_slli_si128(x1, 1); |
4849 | 4.57M | x2 = _mm_unpacklo_epi64(x1, _mm_srli_si128(x1, 1)); |
4850 | 4.57M | x3 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 2), |
4851 | 4.57M | _mm_srli_si128(x1, 3)); |
4852 | 4.57M | x4 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 4), |
4853 | 4.57M | _mm_srli_si128(x1, 5)); |
4854 | 4.57M | x5 = _mm_unpacklo_epi64(_mm_srli_si128(x1, 6), |
4855 | 4.57M | _mm_srli_si128(x1, 7)); |
4856 | | |
4857 | | /* PMADDUBSW then PMADDW */ |
4858 | 4.57M | x2 = _mm_maddubs_epi16(x2, r0); |
4859 | 4.57M | x3 = _mm_maddubs_epi16(x3, r0); |
4860 | 4.57M | x4 = _mm_maddubs_epi16(x4, r0); |
4861 | 4.57M | x5 = _mm_maddubs_epi16(x5, r0); |
4862 | 4.57M | x2 = _mm_hadd_epi16(x2, x3); |
4863 | 4.57M | x4 = _mm_hadd_epi16(x4, x5); |
4864 | 4.57M | x2 = _mm_hadd_epi16(x2, x4); |
4865 | 4.57M | x2 = _mm_srli_si128(x2, BIT_DEPTH - 8); |
4866 | | |
4867 | | /* give results back */ |
4868 | 4.57M | _mm_store_si128((__m128i *) &tmp[x], x2); |
4869 | | |
4870 | 4.57M | } |
4871 | 2.92M | src += srcstride; |
4872 | 2.92M | tmp += MAX_PB_SIZE; |
4873 | 2.92M | } |
4874 | | |
4875 | 235k | tmp = mcbuffer + qpel_extra_before[3] * MAX_PB_SIZE; |
4876 | 235k | srcstride = MAX_PB_SIZE; |
4877 | | |
4878 | | /* vertical treatment on temp table : tmp contains 16 bit values, so need to use 32 bit integers |
4879 | | for register calculations */ |
4880 | 235k | rTemp = _mm_set_epi16(-1, 4, -10, 58, 17, -5, 1, 0); |
4881 | 2.43M | for (y = 0; y < height; y++) { |
4882 | 5.70M | for (x = 0; x < width; x += 8) { |
4883 | | |
4884 | 3.49M | x1 = _mm_setzero_si128(); |
4885 | 3.49M | x2 = _mm_load_si128((__m128i *) &tmp[x - 2 * srcstride]); |
4886 | 3.49M | x3 = _mm_load_si128((__m128i *) &tmp[x - srcstride]); |
4887 | 3.49M | x4 = _mm_load_si128((__m128i *) &tmp[x]); |
4888 | 3.49M | x5 = _mm_load_si128((__m128i *) &tmp[x + srcstride]); |
4889 | 3.49M | x6 = _mm_load_si128((__m128i *) &tmp[x + 2 * srcstride]); |
4890 | 3.49M | x7 = _mm_load_si128((__m128i *) &tmp[x + 3 * srcstride]); |
4891 | 3.49M | x8 = _mm_load_si128((__m128i *) &tmp[x + 4 * srcstride]); |
4892 | | |
4893 | 3.49M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 1)); |
4894 | 3.49M | t7 = _mm_mullo_epi16(x2, r1); |
4895 | | |
4896 | | |
4897 | 3.49M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 2)); |
4898 | 3.49M | rBuffer = _mm_mulhi_epi16(x2, r1); |
4899 | 3.49M | t8 = _mm_mullo_epi16(x3, r0); |
4900 | 3.49M | t2 = _mm_unpacklo_epi16(t7, rBuffer); |
4901 | 3.49M | x2 = _mm_unpackhi_epi16(t7, rBuffer); |
4902 | | |
4903 | 3.49M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 3)); |
4904 | 3.49M | rBuffer = _mm_mulhi_epi16(x3, r0); |
4905 | 3.49M | t7 = _mm_mullo_epi16(x4, r1); |
4906 | 3.49M | t3 = _mm_unpacklo_epi16(t8, rBuffer); |
4907 | 3.49M | x3 = _mm_unpackhi_epi16(t8, rBuffer); |
4908 | | |
4909 | 3.49M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 4)); |
4910 | 3.49M | rBuffer = _mm_mulhi_epi16(x4, r1); |
4911 | 3.49M | t8 = _mm_mullo_epi16(x5, r0); |
4912 | 3.49M | t4 = _mm_unpacklo_epi16(t7, rBuffer); |
4913 | 3.49M | x4 = _mm_unpackhi_epi16(t7, rBuffer); |
4914 | | |
4915 | 3.49M | r1 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 5)); |
4916 | 3.49M | rBuffer = _mm_mulhi_epi16(x5, r0); |
4917 | 3.49M | t7 = _mm_mullo_epi16(x6, r1); |
4918 | 3.49M | t5 = _mm_unpacklo_epi16(t8, rBuffer); |
4919 | 3.49M | x5 = _mm_unpackhi_epi16(t8, rBuffer); |
4920 | | |
4921 | 3.49M | r0 = _mm_set1_epi16(_mm_extract_epi16(rTemp, 6)); |
4922 | 3.49M | rBuffer = _mm_mulhi_epi16(x6, r1); |
4923 | 3.49M | t8 = _mm_mullo_epi16(x7, r0); |
4924 | 3.49M | t6 = _mm_unpacklo_epi16(t7, rBuffer); |
4925 | 3.49M | x6 = _mm_unpackhi_epi16(t7, rBuffer); |
4926 | | |
4927 | 3.49M | rBuffer = _mm_mulhi_epi16(x7, r0); |
4928 | 3.49M | t7 = _mm_unpacklo_epi16(t8, rBuffer); |
4929 | 3.49M | x7 = _mm_unpackhi_epi16(t8, rBuffer); |
4930 | | |
4931 | 3.49M | t8 = _mm_unpacklo_epi16( |
4932 | 3.49M | _mm_mullo_epi16(x8, |
4933 | 3.49M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4934 | 3.49M | _mm_mulhi_epi16(x8, |
4935 | 3.49M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4936 | 3.49M | x8 = _mm_unpackhi_epi16( |
4937 | 3.49M | _mm_mullo_epi16(x8, |
4938 | 3.49M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7))), |
4939 | 3.49M | _mm_mulhi_epi16(x8, |
4940 | 3.49M | _mm_set1_epi16(_mm_extract_epi16(rTemp, 7)))); |
4941 | | |
4942 | | /* add calculus by correct value : */ |
4943 | | |
4944 | 3.49M | x3 = _mm_add_epi32(x3, x4); |
4945 | 3.49M | x5 = _mm_add_epi32(x5, x6); |
4946 | 3.49M | r1 = _mm_add_epi32(x2, x3); |
4947 | 3.49M | x7 = _mm_add_epi32(x7, x8); |
4948 | 3.49M | r1 = _mm_add_epi32(r1, x5); |
4949 | | |
4950 | 3.49M | t3 = _mm_add_epi32(t3, t4); |
4951 | 3.49M | t5 = _mm_add_epi32(t5, t6); |
4952 | 3.49M | r0 = _mm_add_epi32(t2, t3); |
4953 | 3.49M | t7 = _mm_add_epi32(t7, t8); |
4954 | 3.49M | r0 = _mm_add_epi32(r0, t5); |
4955 | 3.49M | r1 = _mm_add_epi32(r1, x7); |
4956 | 3.49M | r0 = _mm_add_epi32(r0, t7); |
4957 | 3.49M | r1 = _mm_srli_epi32(r1, 6); |
4958 | 3.49M | r0 = _mm_srli_epi32(r0, 6); |
4959 | | |
4960 | 3.49M | r1 = _mm_and_si128(r1, |
4961 | 3.49M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4962 | 3.49M | r0 = _mm_and_si128(r0, |
4963 | 3.49M | _mm_set_epi16(0, -1, 0, -1, 0, -1, 0, -1)); |
4964 | 3.49M | r0 = _mm_hadd_epi16(r0, r1); |
4965 | 3.49M | _mm_store_si128((__m128i *) &dst[x], r0); |
4966 | | |
4967 | 3.49M | } |
4968 | 2.20M | tmp += MAX_PB_SIZE; |
4969 | 2.20M | dst += dststride; |
4970 | 2.20M | } |
4971 | 235k | } |