1 /*
2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "./vpx_config.h"
12 #include "./vpx_dsp_rtcd.h"
13 #include "vpx_ports/mem.h"
14
15 typedef uint32_t (*high_variance_fn_t)(const uint16_t *src, int src_stride,
16 const uint16_t *ref, int ref_stride,
17 uint32_t *sse, int *sum);
18
19 uint32_t vpx_highbd_calc8x8var_sse2(const uint16_t *src, int src_stride,
20 const uint16_t *ref, int ref_stride,
21 uint32_t *sse, int *sum);
22
23 uint32_t vpx_highbd_calc16x16var_sse2(const uint16_t *src, int src_stride,
24 const uint16_t *ref, int ref_stride,
25 uint32_t *sse, int *sum);
26
highbd_8_variance_sse2(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int w,int h,uint32_t * sse,int * sum,high_variance_fn_t var_fn,int block_size)27 static void highbd_8_variance_sse2(const uint16_t *src, int src_stride,
28 const uint16_t *ref, int ref_stride, int w,
29 int h, uint32_t *sse, int *sum,
30 high_variance_fn_t var_fn, int block_size) {
31 int i, j;
32
33 *sse = 0;
34 *sum = 0;
35
36 for (i = 0; i < h; i += block_size) {
37 for (j = 0; j < w; j += block_size) {
38 unsigned int sse0;
39 int sum0;
40 var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
41 ref_stride, &sse0, &sum0);
42 *sse += sse0;
43 *sum += sum0;
44 }
45 }
46 }
47
highbd_10_variance_sse2(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int w,int h,uint32_t * sse,int * sum,high_variance_fn_t var_fn,int block_size)48 static void highbd_10_variance_sse2(const uint16_t *src, int src_stride,
49 const uint16_t *ref, int ref_stride, int w,
50 int h, uint32_t *sse, int *sum,
51 high_variance_fn_t var_fn, int block_size) {
52 int i, j;
53 uint64_t sse_long = 0;
54 int32_t sum_long = 0;
55
56 for (i = 0; i < h; i += block_size) {
57 for (j = 0; j < w; j += block_size) {
58 unsigned int sse0;
59 int sum0;
60 var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
61 ref_stride, &sse0, &sum0);
62 sse_long += sse0;
63 sum_long += sum0;
64 }
65 }
66 *sum = ROUND_POWER_OF_TWO(sum_long, 2);
67 *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
68 }
69
highbd_12_variance_sse2(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int w,int h,uint32_t * sse,int * sum,high_variance_fn_t var_fn,int block_size)70 static void highbd_12_variance_sse2(const uint16_t *src, int src_stride,
71 const uint16_t *ref, int ref_stride, int w,
72 int h, uint32_t *sse, int *sum,
73 high_variance_fn_t var_fn, int block_size) {
74 int i, j;
75 uint64_t sse_long = 0;
76 int32_t sum_long = 0;
77
78 for (i = 0; i < h; i += block_size) {
79 for (j = 0; j < w; j += block_size) {
80 unsigned int sse0;
81 int sum0;
82 var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
83 ref_stride, &sse0, &sum0);
84 sse_long += sse0;
85 sum_long += sum0;
86 }
87 }
88 *sum = ROUND_POWER_OF_TWO(sum_long, 4);
89 *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);
90 }
91
92 #define HIGH_GET_VAR(S) \
93 void vpx_highbd_8_get##S##x##S##var_sse2( \
94 const uint8_t *src8, int src_stride, const uint8_t *ref8, \
95 int ref_stride, uint32_t *sse, int *sum) { \
96 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
97 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
98 vpx_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \
99 sum); \
100 } \
101 \
102 void vpx_highbd_10_get##S##x##S##var_sse2( \
103 const uint8_t *src8, int src_stride, const uint8_t *ref8, \
104 int ref_stride, uint32_t *sse, int *sum) { \
105 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
106 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
107 vpx_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \
108 sum); \
109 *sum = ROUND_POWER_OF_TWO(*sum, 2); \
110 *sse = ROUND_POWER_OF_TWO(*sse, 4); \
111 } \
112 \
113 void vpx_highbd_12_get##S##x##S##var_sse2( \
114 const uint8_t *src8, int src_stride, const uint8_t *ref8, \
115 int ref_stride, uint32_t *sse, int *sum) { \
116 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
117 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
118 vpx_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \
119 sum); \
120 *sum = ROUND_POWER_OF_TWO(*sum, 4); \
121 *sse = ROUND_POWER_OF_TWO(*sse, 8); \
122 }
123
124 HIGH_GET_VAR(16);
125 HIGH_GET_VAR(8);
126
127 #undef HIGH_GET_VAR
128
129 #define VAR_FN(w, h, block_size, shift) \
130 uint32_t vpx_highbd_8_variance##w##x##h##_sse2( \
131 const uint8_t *src8, int src_stride, const uint8_t *ref8, \
132 int ref_stride, uint32_t *sse) { \
133 int sum; \
134 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
135 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
136 highbd_8_variance_sse2( \
137 src, src_stride, ref, ref_stride, w, h, sse, &sum, \
138 vpx_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
139 return *sse - (uint32_t)(((int64_t)sum * sum) >> (shift)); \
140 } \
141 \
142 uint32_t vpx_highbd_10_variance##w##x##h##_sse2( \
143 const uint8_t *src8, int src_stride, const uint8_t *ref8, \
144 int ref_stride, uint32_t *sse) { \
145 int sum; \
146 int64_t var; \
147 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
148 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
149 highbd_10_variance_sse2( \
150 src, src_stride, ref, ref_stride, w, h, sse, &sum, \
151 vpx_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
152 var = (int64_t)(*sse) - (((int64_t)sum * sum) >> (shift)); \
153 return (var >= 0) ? (uint32_t)var : 0; \
154 } \
155 \
156 uint32_t vpx_highbd_12_variance##w##x##h##_sse2( \
157 const uint8_t *src8, int src_stride, const uint8_t *ref8, \
158 int ref_stride, uint32_t *sse) { \
159 int sum; \
160 int64_t var; \
161 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
162 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
163 highbd_12_variance_sse2( \
164 src, src_stride, ref, ref_stride, w, h, sse, &sum, \
165 vpx_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
166 var = (int64_t)(*sse) - (((int64_t)sum * sum) >> (shift)); \
167 return (var >= 0) ? (uint32_t)var : 0; \
168 }
169
170 VAR_FN(64, 64, 16, 12);
171 VAR_FN(64, 32, 16, 11);
172 VAR_FN(32, 64, 16, 11);
173 VAR_FN(32, 32, 16, 10);
174 VAR_FN(32, 16, 16, 9);
175 VAR_FN(16, 32, 16, 9);
176 VAR_FN(16, 16, 16, 8);
177 VAR_FN(16, 8, 8, 7);
178 VAR_FN(8, 16, 8, 7);
179 VAR_FN(8, 8, 8, 6);
180
181 #undef VAR_FN
182
vpx_highbd_8_mse16x16_sse2(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,unsigned int * sse)183 unsigned int vpx_highbd_8_mse16x16_sse2(const uint8_t *src8, int src_stride,
184 const uint8_t *ref8, int ref_stride,
185 unsigned int *sse) {
186 int sum;
187 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
188 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
189 highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
190 vpx_highbd_calc16x16var_sse2, 16);
191 return *sse;
192 }
193
vpx_highbd_10_mse16x16_sse2(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,unsigned int * sse)194 unsigned int vpx_highbd_10_mse16x16_sse2(const uint8_t *src8, int src_stride,
195 const uint8_t *ref8, int ref_stride,
196 unsigned int *sse) {
197 int sum;
198 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
199 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
200 highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
201 vpx_highbd_calc16x16var_sse2, 16);
202 return *sse;
203 }
204
vpx_highbd_12_mse16x16_sse2(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,unsigned int * sse)205 unsigned int vpx_highbd_12_mse16x16_sse2(const uint8_t *src8, int src_stride,
206 const uint8_t *ref8, int ref_stride,
207 unsigned int *sse) {
208 int sum;
209 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
210 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
211 highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
212 vpx_highbd_calc16x16var_sse2, 16);
213 return *sse;
214 }
215
vpx_highbd_8_mse8x8_sse2(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,unsigned int * sse)216 unsigned int vpx_highbd_8_mse8x8_sse2(const uint8_t *src8, int src_stride,
217 const uint8_t *ref8, int ref_stride,
218 unsigned int *sse) {
219 int sum;
220 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
221 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
222 highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
223 vpx_highbd_calc8x8var_sse2, 8);
224 return *sse;
225 }
226
vpx_highbd_10_mse8x8_sse2(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,unsigned int * sse)227 unsigned int vpx_highbd_10_mse8x8_sse2(const uint8_t *src8, int src_stride,
228 const uint8_t *ref8, int ref_stride,
229 unsigned int *sse) {
230 int sum;
231 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
232 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
233 highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
234 vpx_highbd_calc8x8var_sse2, 8);
235 return *sse;
236 }
237
vpx_highbd_12_mse8x8_sse2(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,unsigned int * sse)238 unsigned int vpx_highbd_12_mse8x8_sse2(const uint8_t *src8, int src_stride,
239 const uint8_t *ref8, int ref_stride,
240 unsigned int *sse) {
241 int sum;
242 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
243 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
244 highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
245 vpx_highbd_calc8x8var_sse2, 8);
246 return *sse;
247 }
248
249 // The 2 unused parameters are place holders for PIC enabled build.
250 // These definitions are for functions defined in
251 // highbd_subpel_variance_impl_sse2.asm
252 #define DECL(w, opt) \
253 int vpx_highbd_sub_pixel_variance##w##xh_##opt( \
254 const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
255 const uint16_t *ref, ptrdiff_t ref_stride, int height, \
256 unsigned int *sse, void *unused0, void *unused);
257 #define DECLS(opt) \
258 DECL(8, opt); \
259 DECL(16, opt)
260
261 DECLS(sse2);
262
263 #undef DECLS
264 #undef DECL
265
266 #define FN(w, h, wf, wlog2, hlog2, opt, cast) \
267 uint32_t vpx_highbd_8_sub_pixel_variance##w##x##h##_##opt( \
268 const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
269 const uint8_t *ref8, int ref_stride, uint32_t *sse_ptr) { \
270 uint32_t sse; \
271 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
272 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
273 int se = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
274 src, src_stride, x_offset, y_offset, ref, ref_stride, h, &sse, NULL, \
275 NULL); \
276 if (w > wf) { \
277 unsigned int sse2; \
278 int se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
279 src + 16, src_stride, x_offset, y_offset, ref + 16, ref_stride, h, \
280 &sse2, NULL, NULL); \
281 se += se2; \
282 sse += sse2; \
283 if (w > wf * 2) { \
284 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
285 src + 32, src_stride, x_offset, y_offset, ref + 32, ref_stride, h, \
286 &sse2, NULL, NULL); \
287 se += se2; \
288 sse += sse2; \
289 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
290 src + 48, src_stride, x_offset, y_offset, ref + 48, ref_stride, h, \
291 &sse2, NULL, NULL); \
292 se += se2; \
293 sse += sse2; \
294 } \
295 } \
296 *sse_ptr = sse; \
297 return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2)); \
298 } \
299 \
300 uint32_t vpx_highbd_10_sub_pixel_variance##w##x##h##_##opt( \
301 const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
302 const uint8_t *ref8, int ref_stride, uint32_t *sse_ptr) { \
303 int64_t var; \
304 uint32_t sse; \
305 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
306 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
307 int se = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
308 src, src_stride, x_offset, y_offset, ref, ref_stride, h, &sse, NULL, \
309 NULL); \
310 if (w > wf) { \
311 uint32_t sse2; \
312 int se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
313 src + 16, src_stride, x_offset, y_offset, ref + 16, ref_stride, h, \
314 &sse2, NULL, NULL); \
315 se += se2; \
316 sse += sse2; \
317 if (w > wf * 2) { \
318 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
319 src + 32, src_stride, x_offset, y_offset, ref + 32, ref_stride, h, \
320 &sse2, NULL, NULL); \
321 se += se2; \
322 sse += sse2; \
323 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
324 src + 48, src_stride, x_offset, y_offset, ref + 48, ref_stride, h, \
325 &sse2, NULL, NULL); \
326 se += se2; \
327 sse += sse2; \
328 } \
329 } \
330 se = ROUND_POWER_OF_TWO(se, 2); \
331 sse = ROUND_POWER_OF_TWO(sse, 4); \
332 *sse_ptr = sse; \
333 var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \
334 return (var >= 0) ? (uint32_t)var : 0; \
335 } \
336 \
337 uint32_t vpx_highbd_12_sub_pixel_variance##w##x##h##_##opt( \
338 const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
339 const uint8_t *ref8, int ref_stride, uint32_t *sse_ptr) { \
340 int start_row; \
341 uint32_t sse; \
342 int se = 0; \
343 int64_t var; \
344 uint64_t long_sse = 0; \
345 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
346 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
347 for (start_row = 0; start_row < h; start_row += 16) { \
348 uint32_t sse2; \
349 int height = h - start_row < 16 ? h - start_row : 16; \
350 int se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
351 src + (start_row * src_stride), src_stride, x_offset, y_offset, \
352 ref + (start_row * ref_stride), ref_stride, height, &sse2, NULL, \
353 NULL); \
354 se += se2; \
355 long_sse += sse2; \
356 if (w > wf) { \
357 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
358 src + 16 + (start_row * src_stride), src_stride, x_offset, \
359 y_offset, ref + 16 + (start_row * ref_stride), ref_stride, height, \
360 &sse2, NULL, NULL); \
361 se += se2; \
362 long_sse += sse2; \
363 if (w > wf * 2) { \
364 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
365 src + 32 + (start_row * src_stride), src_stride, x_offset, \
366 y_offset, ref + 32 + (start_row * ref_stride), ref_stride, \
367 height, &sse2, NULL, NULL); \
368 se += se2; \
369 long_sse += sse2; \
370 se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
371 src + 48 + (start_row * src_stride), src_stride, x_offset, \
372 y_offset, ref + 48 + (start_row * ref_stride), ref_stride, \
373 height, &sse2, NULL, NULL); \
374 se += se2; \
375 long_sse += sse2; \
376 } \
377 } \
378 } \
379 se = ROUND_POWER_OF_TWO(se, 4); \
380 sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \
381 *sse_ptr = sse; \
382 var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \
383 return (var >= 0) ? (uint32_t)var : 0; \
384 }
385
386 #define FNS(opt) \
387 FN(64, 64, 16, 6, 6, opt, (int64_t)); \
388 FN(64, 32, 16, 6, 5, opt, (int64_t)); \
389 FN(32, 64, 16, 5, 6, opt, (int64_t)); \
390 FN(32, 32, 16, 5, 5, opt, (int64_t)); \
391 FN(32, 16, 16, 5, 4, opt, (int64_t)); \
392 FN(16, 32, 16, 4, 5, opt, (int64_t)); \
393 FN(16, 16, 16, 4, 4, opt, (int64_t)); \
394 FN(16, 8, 16, 4, 3, opt, (int64_t)); \
395 FN(8, 16, 8, 3, 4, opt, (int64_t)); \
396 FN(8, 8, 8, 3, 3, opt, (int64_t)); \
397 FN(8, 4, 8, 3, 2, opt, (int64_t));
398
399 FNS(sse2);
400
401 #undef FNS
402 #undef FN
403
404 // The 2 unused parameters are place holders for PIC enabled build.
405 #define DECL(w, opt) \
406 int vpx_highbd_sub_pixel_avg_variance##w##xh_##opt( \
407 const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
408 const uint16_t *ref, ptrdiff_t ref_stride, const uint16_t *second, \
409 ptrdiff_t second_stride, int height, unsigned int *sse, void *unused0, \
410 void *unused);
411 #define DECLS(opt1) \
412 DECL(16, opt1) \
413 DECL(8, opt1)
414
415 DECLS(sse2);
416 #undef DECL
417 #undef DECLS
418
419 #define FN(w, h, wf, wlog2, hlog2, opt, cast) \
420 uint32_t vpx_highbd_8_sub_pixel_avg_variance##w##x##h##_##opt( \
421 const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
422 const uint8_t *ref8, int ref_stride, uint32_t *sse_ptr, \
423 const uint8_t *sec8) { \
424 uint32_t sse; \
425 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
426 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
427 uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
428 int se = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
429 src, src_stride, x_offset, y_offset, ref, ref_stride, sec, w, h, &sse, \
430 NULL, NULL); \
431 if (w > wf) { \
432 uint32_t sse2; \
433 int se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
434 src + 16, src_stride, x_offset, y_offset, ref + 16, ref_stride, \
435 sec + 16, w, h, &sse2, NULL, NULL); \
436 se += se2; \
437 sse += sse2; \
438 if (w > wf * 2) { \
439 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
440 src + 32, src_stride, x_offset, y_offset, ref + 32, ref_stride, \
441 sec + 32, w, h, &sse2, NULL, NULL); \
442 se += se2; \
443 sse += sse2; \
444 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
445 src + 48, src_stride, x_offset, y_offset, ref + 48, ref_stride, \
446 sec + 48, w, h, &sse2, NULL, NULL); \
447 se += se2; \
448 sse += sse2; \
449 } \
450 } \
451 *sse_ptr = sse; \
452 return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2)); \
453 } \
454 \
455 uint32_t vpx_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt( \
456 const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
457 const uint8_t *ref8, int ref_stride, uint32_t *sse_ptr, \
458 const uint8_t *sec8) { \
459 int64_t var; \
460 uint32_t sse; \
461 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
462 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
463 uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
464 int se = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
465 src, src_stride, x_offset, y_offset, ref, ref_stride, sec, w, h, &sse, \
466 NULL, NULL); \
467 if (w > wf) { \
468 uint32_t sse2; \
469 int se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
470 src + 16, src_stride, x_offset, y_offset, ref + 16, ref_stride, \
471 sec + 16, w, h, &sse2, NULL, NULL); \
472 se += se2; \
473 sse += sse2; \
474 if (w > wf * 2) { \
475 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
476 src + 32, src_stride, x_offset, y_offset, ref + 32, ref_stride, \
477 sec + 32, w, h, &sse2, NULL, NULL); \
478 se += se2; \
479 sse += sse2; \
480 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
481 src + 48, src_stride, x_offset, y_offset, ref + 48, ref_stride, \
482 sec + 48, w, h, &sse2, NULL, NULL); \
483 se += se2; \
484 sse += sse2; \
485 } \
486 } \
487 se = ROUND_POWER_OF_TWO(se, 2); \
488 sse = ROUND_POWER_OF_TWO(sse, 4); \
489 *sse_ptr = sse; \
490 var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \
491 return (var >= 0) ? (uint32_t)var : 0; \
492 } \
493 \
494 uint32_t vpx_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt( \
495 const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
496 const uint8_t *ref8, int ref_stride, uint32_t *sse_ptr, \
497 const uint8_t *sec8) { \
498 int start_row; \
499 int64_t var; \
500 uint32_t sse; \
501 int se = 0; \
502 uint64_t long_sse = 0; \
503 uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
504 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
505 uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
506 for (start_row = 0; start_row < h; start_row += 16) { \
507 uint32_t sse2; \
508 int height = h - start_row < 16 ? h - start_row : 16; \
509 int se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
510 src + (start_row * src_stride), src_stride, x_offset, y_offset, \
511 ref + (start_row * ref_stride), ref_stride, sec + (start_row * w), \
512 w, height, &sse2, NULL, NULL); \
513 se += se2; \
514 long_sse += sse2; \
515 if (w > wf) { \
516 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
517 src + 16 + (start_row * src_stride), src_stride, x_offset, \
518 y_offset, ref + 16 + (start_row * ref_stride), ref_stride, \
519 sec + 16 + (start_row * w), w, height, &sse2, NULL, NULL); \
520 se += se2; \
521 long_sse += sse2; \
522 if (w > wf * 2) { \
523 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
524 src + 32 + (start_row * src_stride), src_stride, x_offset, \
525 y_offset, ref + 32 + (start_row * ref_stride), ref_stride, \
526 sec + 32 + (start_row * w), w, height, &sse2, NULL, NULL); \
527 se += se2; \
528 long_sse += sse2; \
529 se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
530 src + 48 + (start_row * src_stride), src_stride, x_offset, \
531 y_offset, ref + 48 + (start_row * ref_stride), ref_stride, \
532 sec + 48 + (start_row * w), w, height, &sse2, NULL, NULL); \
533 se += se2; \
534 long_sse += sse2; \
535 } \
536 } \
537 } \
538 se = ROUND_POWER_OF_TWO(se, 4); \
539 sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \
540 *sse_ptr = sse; \
541 var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \
542 return (var >= 0) ? (uint32_t)var : 0; \
543 }
544
545 #define FNS(opt1) \
546 FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
547 FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
548 FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
549 FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
550 FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
551 FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
552 FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
553 FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
554 FN(8, 16, 8, 4, 3, opt1, (int64_t)); \
555 FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
556 FN(8, 4, 8, 3, 2, opt1, (int64_t));
557
558 FNS(sse2);
559
560 #undef FNS
561 #undef FN
562