1 /*
2 * Copyright (c) 2010 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 "vp9/common/vp9_loopfilter.h"
14 #include "vp9/common/vp9_onyxc_int.h"
15 #include "vp9/common/vp9_reconinter.h"
16 #include "vpx_dsp/vpx_dsp_common.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vpx_ports/mem.h"
19
20 #include "vp9/common/vp9_seg_common.h"
21
22 // 64 bit masks for left transform size. Each 1 represents a position where
23 // we should apply a loop filter across the left border of an 8x8 block
24 // boundary.
25 //
26 // In the case of TX_16X16-> ( in low order byte first we end up with
27 // a mask that looks like this
28 //
29 // 10101010
30 // 10101010
31 // 10101010
32 // 10101010
33 // 10101010
34 // 10101010
35 // 10101010
36 // 10101010
37 //
38 // A loopfilter should be applied to every other 8x8 horizontally.
39 static const uint64_t left_64x64_txform_mask[TX_SIZES]= {
40 0xffffffffffffffffULL, // TX_4X4
41 0xffffffffffffffffULL, // TX_8x8
42 0x5555555555555555ULL, // TX_16x16
43 0x1111111111111111ULL, // TX_32x32
44 };
45
46 // 64 bit masks for above transform size. Each 1 represents a position where
47 // we should apply a loop filter across the top border of an 8x8 block
48 // boundary.
49 //
50 // In the case of TX_32x32 -> ( in low order byte first we end up with
51 // a mask that looks like this
52 //
53 // 11111111
54 // 00000000
55 // 00000000
56 // 00000000
57 // 11111111
58 // 00000000
59 // 00000000
60 // 00000000
61 //
62 // A loopfilter should be applied to every other 4 the row vertically.
63 static const uint64_t above_64x64_txform_mask[TX_SIZES]= {
64 0xffffffffffffffffULL, // TX_4X4
65 0xffffffffffffffffULL, // TX_8x8
66 0x00ff00ff00ff00ffULL, // TX_16x16
67 0x000000ff000000ffULL, // TX_32x32
68 };
69
70 // 64 bit masks for prediction sizes (left). Each 1 represents a position
71 // where left border of an 8x8 block. These are aligned to the right most
72 // appropriate bit, and then shifted into place.
73 //
74 // In the case of TX_16x32 -> ( low order byte first ) we end up with
75 // a mask that looks like this :
76 //
77 // 10000000
78 // 10000000
79 // 10000000
80 // 10000000
81 // 00000000
82 // 00000000
83 // 00000000
84 // 00000000
85 static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
86 0x0000000000000001ULL, // BLOCK_4X4,
87 0x0000000000000001ULL, // BLOCK_4X8,
88 0x0000000000000001ULL, // BLOCK_8X4,
89 0x0000000000000001ULL, // BLOCK_8X8,
90 0x0000000000000101ULL, // BLOCK_8X16,
91 0x0000000000000001ULL, // BLOCK_16X8,
92 0x0000000000000101ULL, // BLOCK_16X16,
93 0x0000000001010101ULL, // BLOCK_16X32,
94 0x0000000000000101ULL, // BLOCK_32X16,
95 0x0000000001010101ULL, // BLOCK_32X32,
96 0x0101010101010101ULL, // BLOCK_32X64,
97 0x0000000001010101ULL, // BLOCK_64X32,
98 0x0101010101010101ULL, // BLOCK_64X64
99 };
100
101 // 64 bit mask to shift and set for each prediction size.
102 static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
103 0x0000000000000001ULL, // BLOCK_4X4
104 0x0000000000000001ULL, // BLOCK_4X8
105 0x0000000000000001ULL, // BLOCK_8X4
106 0x0000000000000001ULL, // BLOCK_8X8
107 0x0000000000000001ULL, // BLOCK_8X16,
108 0x0000000000000003ULL, // BLOCK_16X8
109 0x0000000000000003ULL, // BLOCK_16X16
110 0x0000000000000003ULL, // BLOCK_16X32,
111 0x000000000000000fULL, // BLOCK_32X16,
112 0x000000000000000fULL, // BLOCK_32X32,
113 0x000000000000000fULL, // BLOCK_32X64,
114 0x00000000000000ffULL, // BLOCK_64X32,
115 0x00000000000000ffULL, // BLOCK_64X64
116 };
117 // 64 bit mask to shift and set for each prediction size. A bit is set for
118 // each 8x8 block that would be in the left most block of the given block
119 // size in the 64x64 block.
120 static const uint64_t size_mask[BLOCK_SIZES] = {
121 0x0000000000000001ULL, // BLOCK_4X4
122 0x0000000000000001ULL, // BLOCK_4X8
123 0x0000000000000001ULL, // BLOCK_8X4
124 0x0000000000000001ULL, // BLOCK_8X8
125 0x0000000000000101ULL, // BLOCK_8X16,
126 0x0000000000000003ULL, // BLOCK_16X8
127 0x0000000000000303ULL, // BLOCK_16X16
128 0x0000000003030303ULL, // BLOCK_16X32,
129 0x0000000000000f0fULL, // BLOCK_32X16,
130 0x000000000f0f0f0fULL, // BLOCK_32X32,
131 0x0f0f0f0f0f0f0f0fULL, // BLOCK_32X64,
132 0x00000000ffffffffULL, // BLOCK_64X32,
133 0xffffffffffffffffULL, // BLOCK_64X64
134 };
135
136 // These are used for masking the left and above borders.
137 static const uint64_t left_border = 0x1111111111111111ULL;
138 static const uint64_t above_border = 0x000000ff000000ffULL;
139
140 // 16 bit masks for uv transform sizes.
141 static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= {
142 0xffff, // TX_4X4
143 0xffff, // TX_8x8
144 0x5555, // TX_16x16
145 0x1111, // TX_32x32
146 };
147
148 static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= {
149 0xffff, // TX_4X4
150 0xffff, // TX_8x8
151 0x0f0f, // TX_16x16
152 0x000f, // TX_32x32
153 };
154
155 // 16 bit left mask to shift and set for each uv prediction size.
156 static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
157 0x0001, // BLOCK_4X4,
158 0x0001, // BLOCK_4X8,
159 0x0001, // BLOCK_8X4,
160 0x0001, // BLOCK_8X8,
161 0x0001, // BLOCK_8X16,
162 0x0001, // BLOCK_16X8,
163 0x0001, // BLOCK_16X16,
164 0x0011, // BLOCK_16X32,
165 0x0001, // BLOCK_32X16,
166 0x0011, // BLOCK_32X32,
167 0x1111, // BLOCK_32X64
168 0x0011, // BLOCK_64X32,
169 0x1111, // BLOCK_64X64
170 };
171 // 16 bit above mask to shift and set for uv each prediction size.
172 static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
173 0x0001, // BLOCK_4X4
174 0x0001, // BLOCK_4X8
175 0x0001, // BLOCK_8X4
176 0x0001, // BLOCK_8X8
177 0x0001, // BLOCK_8X16,
178 0x0001, // BLOCK_16X8
179 0x0001, // BLOCK_16X16
180 0x0001, // BLOCK_16X32,
181 0x0003, // BLOCK_32X16,
182 0x0003, // BLOCK_32X32,
183 0x0003, // BLOCK_32X64,
184 0x000f, // BLOCK_64X32,
185 0x000f, // BLOCK_64X64
186 };
187
188 // 64 bit mask to shift and set for each uv prediction size
189 static const uint16_t size_mask_uv[BLOCK_SIZES] = {
190 0x0001, // BLOCK_4X4
191 0x0001, // BLOCK_4X8
192 0x0001, // BLOCK_8X4
193 0x0001, // BLOCK_8X8
194 0x0001, // BLOCK_8X16,
195 0x0001, // BLOCK_16X8
196 0x0001, // BLOCK_16X16
197 0x0011, // BLOCK_16X32,
198 0x0003, // BLOCK_32X16,
199 0x0033, // BLOCK_32X32,
200 0x3333, // BLOCK_32X64,
201 0x00ff, // BLOCK_64X32,
202 0xffff, // BLOCK_64X64
203 };
204 static const uint16_t left_border_uv = 0x1111;
205 static const uint16_t above_border_uv = 0x000f;
206
207 static const int mode_lf_lut[MB_MODE_COUNT] = {
208 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // INTRA_MODES
209 1, 1, 0, 1 // INTER_MODES (ZEROMV == 0)
210 };
211
update_sharpness(loop_filter_info_n * lfi,int sharpness_lvl)212 static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
213 int lvl;
214
215 // For each possible value for the loop filter fill out limits
216 for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
217 // Set loop filter parameters that control sharpness.
218 int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
219
220 if (sharpness_lvl > 0) {
221 if (block_inside_limit > (9 - sharpness_lvl))
222 block_inside_limit = (9 - sharpness_lvl);
223 }
224
225 if (block_inside_limit < 1)
226 block_inside_limit = 1;
227
228 memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
229 memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
230 SIMD_WIDTH);
231 }
232 }
233
get_filter_level(const loop_filter_info_n * lfi_n,const MB_MODE_INFO * mbmi)234 static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
235 const MB_MODE_INFO *mbmi) {
236 return lfi_n->lvl[mbmi->segment_id][mbmi->ref_frame[0]]
237 [mode_lf_lut[mbmi->mode]];
238 }
239
vp9_loop_filter_init(VP9_COMMON * cm)240 void vp9_loop_filter_init(VP9_COMMON *cm) {
241 loop_filter_info_n *lfi = &cm->lf_info;
242 struct loopfilter *lf = &cm->lf;
243 int lvl;
244
245 // init limits for given sharpness
246 update_sharpness(lfi, lf->sharpness_level);
247 lf->last_sharpness_level = lf->sharpness_level;
248
249 // init hev threshold const vectors
250 for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
251 memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
252 }
253
vp9_loop_filter_frame_init(VP9_COMMON * cm,int default_filt_lvl)254 void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
255 int seg_id;
256 // n_shift is the multiplier for lf_deltas
257 // the multiplier is 1 for when filter_lvl is between 0 and 31;
258 // 2 when filter_lvl is between 32 and 63
259 const int scale = 1 << (default_filt_lvl >> 5);
260 loop_filter_info_n *const lfi = &cm->lf_info;
261 struct loopfilter *const lf = &cm->lf;
262 const struct segmentation *const seg = &cm->seg;
263
264 // update limits if sharpness has changed
265 if (lf->last_sharpness_level != lf->sharpness_level) {
266 update_sharpness(lfi, lf->sharpness_level);
267 lf->last_sharpness_level = lf->sharpness_level;
268 }
269
270 for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
271 int lvl_seg = default_filt_lvl;
272 if (segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
273 const int data = get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
274 lvl_seg = clamp(seg->abs_delta == SEGMENT_ABSDATA ?
275 data : default_filt_lvl + data,
276 0, MAX_LOOP_FILTER);
277 }
278
279 if (!lf->mode_ref_delta_enabled) {
280 // we could get rid of this if we assume that deltas are set to
281 // zero when not in use; encoder always uses deltas
282 memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
283 } else {
284 int ref, mode;
285 const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
286 lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
287
288 for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref) {
289 for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
290 const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale
291 + lf->mode_deltas[mode] * scale;
292 lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
293 }
294 }
295 }
296 }
297 }
298
filter_selectively_vert_row2(int subsampling_factor,uint8_t * s,int pitch,unsigned int mask_16x16_l,unsigned int mask_8x8_l,unsigned int mask_4x4_l,unsigned int mask_4x4_int_l,const loop_filter_info_n * lfi_n,const uint8_t * lfl)299 static void filter_selectively_vert_row2(int subsampling_factor,
300 uint8_t *s, int pitch,
301 unsigned int mask_16x16_l,
302 unsigned int mask_8x8_l,
303 unsigned int mask_4x4_l,
304 unsigned int mask_4x4_int_l,
305 const loop_filter_info_n *lfi_n,
306 const uint8_t *lfl) {
307 const int mask_shift = subsampling_factor ? 4 : 8;
308 const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
309 const int lfl_forward = subsampling_factor ? 4 : 8;
310
311 unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
312 unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
313 unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
314 unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
315 unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
316 unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
317 unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
318 unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
319 unsigned int mask;
320
321 for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
322 mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
323 mask; mask >>= 1) {
324 const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
325 const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
326
327 // TODO(yunqingwang): count in loopfilter functions should be removed.
328 if (mask & 1) {
329 if ((mask_16x16_0 | mask_16x16_1) & 1) {
330 if ((mask_16x16_0 & mask_16x16_1) & 1) {
331 vpx_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
332 lfi0->hev_thr);
333 } else if (mask_16x16_0 & 1) {
334 vpx_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
335 lfi0->hev_thr);
336 } else {
337 vpx_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
338 lfi1->lim, lfi1->hev_thr);
339 }
340 }
341
342 if ((mask_8x8_0 | mask_8x8_1) & 1) {
343 if ((mask_8x8_0 & mask_8x8_1) & 1) {
344 vpx_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
345 lfi0->hev_thr, lfi1->mblim, lfi1->lim,
346 lfi1->hev_thr);
347 } else if (mask_8x8_0 & 1) {
348 vpx_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
349 1);
350 } else {
351 vpx_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
352 lfi1->hev_thr, 1);
353 }
354 }
355
356 if ((mask_4x4_0 | mask_4x4_1) & 1) {
357 if ((mask_4x4_0 & mask_4x4_1) & 1) {
358 vpx_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
359 lfi0->hev_thr, lfi1->mblim, lfi1->lim,
360 lfi1->hev_thr);
361 } else if (mask_4x4_0 & 1) {
362 vpx_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
363 1);
364 } else {
365 vpx_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
366 lfi1->hev_thr, 1);
367 }
368 }
369
370 if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
371 if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
372 vpx_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
373 lfi0->hev_thr, lfi1->mblim, lfi1->lim,
374 lfi1->hev_thr);
375 } else if (mask_4x4_int_0 & 1) {
376 vpx_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
377 lfi0->hev_thr, 1);
378 } else {
379 vpx_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
380 lfi1->hev_thr, 1);
381 }
382 }
383 }
384
385 s += 8;
386 lfl += 1;
387 mask_16x16_0 >>= 1;
388 mask_8x8_0 >>= 1;
389 mask_4x4_0 >>= 1;
390 mask_4x4_int_0 >>= 1;
391 mask_16x16_1 >>= 1;
392 mask_8x8_1 >>= 1;
393 mask_4x4_1 >>= 1;
394 mask_4x4_int_1 >>= 1;
395 }
396 }
397
398 #if CONFIG_VP9_HIGHBITDEPTH
highbd_filter_selectively_vert_row2(int subsampling_factor,uint16_t * s,int pitch,unsigned int mask_16x16_l,unsigned int mask_8x8_l,unsigned int mask_4x4_l,unsigned int mask_4x4_int_l,const loop_filter_info_n * lfi_n,const uint8_t * lfl,int bd)399 static void highbd_filter_selectively_vert_row2(int subsampling_factor,
400 uint16_t *s, int pitch,
401 unsigned int mask_16x16_l,
402 unsigned int mask_8x8_l,
403 unsigned int mask_4x4_l,
404 unsigned int mask_4x4_int_l,
405 const loop_filter_info_n *lfi_n,
406 const uint8_t *lfl, int bd) {
407 const int mask_shift = subsampling_factor ? 4 : 8;
408 const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
409 const int lfl_forward = subsampling_factor ? 4 : 8;
410
411 unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
412 unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
413 unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
414 unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
415 unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
416 unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
417 unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
418 unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
419 unsigned int mask;
420
421 for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
422 mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
423 mask; mask >>= 1) {
424 const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
425 const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
426
427 // TODO(yunqingwang): count in loopfilter functions should be removed.
428 if (mask & 1) {
429 if ((mask_16x16_0 | mask_16x16_1) & 1) {
430 if ((mask_16x16_0 & mask_16x16_1) & 1) {
431 vpx_highbd_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
432 lfi0->hev_thr, bd);
433 } else if (mask_16x16_0 & 1) {
434 vpx_highbd_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
435 lfi0->hev_thr, bd);
436 } else {
437 vpx_highbd_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
438 lfi1->lim, lfi1->hev_thr, bd);
439 }
440 }
441
442 if ((mask_8x8_0 | mask_8x8_1) & 1) {
443 if ((mask_8x8_0 & mask_8x8_1) & 1) {
444 vpx_highbd_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
445 lfi0->hev_thr, lfi1->mblim, lfi1->lim,
446 lfi1->hev_thr, bd);
447 } else if (mask_8x8_0 & 1) {
448 vpx_highbd_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim,
449 lfi0->hev_thr, 1, bd);
450 } else {
451 vpx_highbd_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim,
452 lfi1->lim, lfi1->hev_thr, 1, bd);
453 }
454 }
455
456 if ((mask_4x4_0 | mask_4x4_1) & 1) {
457 if ((mask_4x4_0 & mask_4x4_1) & 1) {
458 vpx_highbd_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
459 lfi0->hev_thr, lfi1->mblim, lfi1->lim,
460 lfi1->hev_thr, bd);
461 } else if (mask_4x4_0 & 1) {
462 vpx_highbd_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim,
463 lfi0->hev_thr, 1, bd);
464 } else {
465 vpx_highbd_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim,
466 lfi1->lim, lfi1->hev_thr, 1, bd);
467 }
468 }
469
470 if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
471 if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
472 vpx_highbd_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
473 lfi0->hev_thr, lfi1->mblim, lfi1->lim,
474 lfi1->hev_thr, bd);
475 } else if (mask_4x4_int_0 & 1) {
476 vpx_highbd_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
477 lfi0->hev_thr, 1, bd);
478 } else {
479 vpx_highbd_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim,
480 lfi1->lim, lfi1->hev_thr, 1, bd);
481 }
482 }
483 }
484
485 s += 8;
486 lfl += 1;
487 mask_16x16_0 >>= 1;
488 mask_8x8_0 >>= 1;
489 mask_4x4_0 >>= 1;
490 mask_4x4_int_0 >>= 1;
491 mask_16x16_1 >>= 1;
492 mask_8x8_1 >>= 1;
493 mask_4x4_1 >>= 1;
494 mask_4x4_int_1 >>= 1;
495 }
496 }
497 #endif // CONFIG_VP9_HIGHBITDEPTH
498
filter_selectively_horiz(uint8_t * s,int pitch,unsigned int mask_16x16,unsigned int mask_8x8,unsigned int mask_4x4,unsigned int mask_4x4_int,const loop_filter_info_n * lfi_n,const uint8_t * lfl)499 static void filter_selectively_horiz(uint8_t *s, int pitch,
500 unsigned int mask_16x16,
501 unsigned int mask_8x8,
502 unsigned int mask_4x4,
503 unsigned int mask_4x4_int,
504 const loop_filter_info_n *lfi_n,
505 const uint8_t *lfl) {
506 unsigned int mask;
507 int count;
508
509 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
510 mask; mask >>= count) {
511 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
512
513 count = 1;
514 if (mask & 1) {
515 if (mask_16x16 & 1) {
516 if ((mask_16x16 & 3) == 3) {
517 vpx_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
518 lfi->hev_thr, 2);
519 count = 2;
520 } else {
521 vpx_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
522 lfi->hev_thr, 1);
523 }
524 } else if (mask_8x8 & 1) {
525 if ((mask_8x8 & 3) == 3) {
526 // Next block's thresholds.
527 const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
528
529 vpx_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
530 lfi->hev_thr, lfin->mblim, lfin->lim,
531 lfin->hev_thr);
532
533 if ((mask_4x4_int & 3) == 3) {
534 vpx_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
535 lfi->lim, lfi->hev_thr, lfin->mblim,
536 lfin->lim, lfin->hev_thr);
537 } else {
538 if (mask_4x4_int & 1)
539 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
540 lfi->hev_thr, 1);
541 else if (mask_4x4_int & 2)
542 vpx_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
543 lfin->lim, lfin->hev_thr, 1);
544 }
545 count = 2;
546 } else {
547 vpx_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
548
549 if (mask_4x4_int & 1)
550 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
551 lfi->hev_thr, 1);
552 }
553 } else if (mask_4x4 & 1) {
554 if ((mask_4x4 & 3) == 3) {
555 // Next block's thresholds.
556 const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
557
558 vpx_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
559 lfi->hev_thr, lfin->mblim, lfin->lim,
560 lfin->hev_thr);
561 if ((mask_4x4_int & 3) == 3) {
562 vpx_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
563 lfi->lim, lfi->hev_thr, lfin->mblim,
564 lfin->lim, lfin->hev_thr);
565 } else {
566 if (mask_4x4_int & 1)
567 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
568 lfi->hev_thr, 1);
569 else if (mask_4x4_int & 2)
570 vpx_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
571 lfin->lim, lfin->hev_thr, 1);
572 }
573 count = 2;
574 } else {
575 vpx_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
576
577 if (mask_4x4_int & 1)
578 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
579 lfi->hev_thr, 1);
580 }
581 } else if (mask_4x4_int & 1) {
582 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
583 lfi->hev_thr, 1);
584 }
585 }
586 s += 8 * count;
587 lfl += count;
588 mask_16x16 >>= count;
589 mask_8x8 >>= count;
590 mask_4x4 >>= count;
591 mask_4x4_int >>= count;
592 }
593 }
594
595 #if CONFIG_VP9_HIGHBITDEPTH
highbd_filter_selectively_horiz(uint16_t * s,int pitch,unsigned int mask_16x16,unsigned int mask_8x8,unsigned int mask_4x4,unsigned int mask_4x4_int,const loop_filter_info_n * lfi_n,const uint8_t * lfl,int bd)596 static void highbd_filter_selectively_horiz(uint16_t *s, int pitch,
597 unsigned int mask_16x16,
598 unsigned int mask_8x8,
599 unsigned int mask_4x4,
600 unsigned int mask_4x4_int,
601 const loop_filter_info_n *lfi_n,
602 const uint8_t *lfl, int bd) {
603 unsigned int mask;
604 int count;
605
606 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
607 mask; mask >>= count) {
608 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
609
610 count = 1;
611 if (mask & 1) {
612 if (mask_16x16 & 1) {
613 if ((mask_16x16 & 3) == 3) {
614 vpx_highbd_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
615 lfi->hev_thr, 2, bd);
616 count = 2;
617 } else {
618 vpx_highbd_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
619 lfi->hev_thr, 1, bd);
620 }
621 } else if (mask_8x8 & 1) {
622 if ((mask_8x8 & 3) == 3) {
623 // Next block's thresholds.
624 const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
625
626 vpx_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
627 lfi->hev_thr, lfin->mblim, lfin->lim,
628 lfin->hev_thr, bd);
629
630 if ((mask_4x4_int & 3) == 3) {
631 vpx_highbd_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
632 lfi->lim, lfi->hev_thr,
633 lfin->mblim, lfin->lim,
634 lfin->hev_thr, bd);
635 } else {
636 if (mask_4x4_int & 1) {
637 vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
638 lfi->lim, lfi->hev_thr, 1, bd);
639 } else if (mask_4x4_int & 2) {
640 vpx_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
641 lfin->lim, lfin->hev_thr, 1, bd);
642 }
643 }
644 count = 2;
645 } else {
646 vpx_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
647 lfi->hev_thr, 1, bd);
648
649 if (mask_4x4_int & 1) {
650 vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
651 lfi->lim, lfi->hev_thr, 1, bd);
652 }
653 }
654 } else if (mask_4x4 & 1) {
655 if ((mask_4x4 & 3) == 3) {
656 // Next block's thresholds.
657 const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
658
659 vpx_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
660 lfi->hev_thr, lfin->mblim, lfin->lim,
661 lfin->hev_thr, bd);
662 if ((mask_4x4_int & 3) == 3) {
663 vpx_highbd_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
664 lfi->lim, lfi->hev_thr,
665 lfin->mblim, lfin->lim,
666 lfin->hev_thr, bd);
667 } else {
668 if (mask_4x4_int & 1) {
669 vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
670 lfi->lim, lfi->hev_thr, 1, bd);
671 } else if (mask_4x4_int & 2) {
672 vpx_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
673 lfin->lim, lfin->hev_thr, 1, bd);
674 }
675 }
676 count = 2;
677 } else {
678 vpx_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
679 lfi->hev_thr, 1, bd);
680
681 if (mask_4x4_int & 1) {
682 vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
683 lfi->lim, lfi->hev_thr, 1, bd);
684 }
685 }
686 } else if (mask_4x4_int & 1) {
687 vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
688 lfi->hev_thr, 1, bd);
689 }
690 }
691 s += 8 * count;
692 lfl += count;
693 mask_16x16 >>= count;
694 mask_8x8 >>= count;
695 mask_4x4 >>= count;
696 mask_4x4_int >>= count;
697 }
698 }
699 #endif // CONFIG_VP9_HIGHBITDEPTH
700
701 // This function ors into the current lfm structure, where to do loop
702 // filters for the specific mi we are looking at. It uses information
703 // including the block_size_type (32x16, 32x32, etc.), the transform size,
704 // whether there were any coefficients encoded, and the loop filter strength
705 // block we are currently looking at. Shift is used to position the
706 // 1's we produce.
707 // TODO(JBB) Need another function for different resolution color..
build_masks(const loop_filter_info_n * const lfi_n,const MODE_INFO * mi,const int shift_y,const int shift_uv,LOOP_FILTER_MASK * lfm)708 static void build_masks(const loop_filter_info_n *const lfi_n,
709 const MODE_INFO *mi, const int shift_y,
710 const int shift_uv,
711 LOOP_FILTER_MASK *lfm) {
712 const MB_MODE_INFO *mbmi = &mi->mbmi;
713 const BLOCK_SIZE block_size = mbmi->sb_type;
714 const TX_SIZE tx_size_y = mbmi->tx_size;
715 const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
716 const int filter_level = get_filter_level(lfi_n, mbmi);
717 uint64_t *const left_y = &lfm->left_y[tx_size_y];
718 uint64_t *const above_y = &lfm->above_y[tx_size_y];
719 uint64_t *const int_4x4_y = &lfm->int_4x4_y;
720 uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
721 uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
722 uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
723 int i;
724
725 // If filter level is 0 we don't loop filter.
726 if (!filter_level) {
727 return;
728 } else {
729 const int w = num_8x8_blocks_wide_lookup[block_size];
730 const int h = num_8x8_blocks_high_lookup[block_size];
731 int index = shift_y;
732 for (i = 0; i < h; i++) {
733 memset(&lfm->lfl_y[index], filter_level, w);
734 index += 8;
735 }
736 }
737
738 // These set 1 in the current block size for the block size edges.
739 // For instance if the block size is 32x16, we'll set:
740 // above = 1111
741 // 0000
742 // and
743 // left = 1000
744 // = 1000
745 // NOTE : In this example the low bit is left most ( 1000 ) is stored as
746 // 1, not 8...
747 //
748 // U and V set things on a 16 bit scale.
749 //
750 *above_y |= above_prediction_mask[block_size] << shift_y;
751 *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
752 *left_y |= left_prediction_mask[block_size] << shift_y;
753 *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
754
755 // If the block has no coefficients and is not intra we skip applying
756 // the loop filter on block edges.
757 if (mbmi->skip && is_inter_block(mbmi))
758 return;
759
760 // Here we are adding a mask for the transform size. The transform
761 // size mask is set to be correct for a 64x64 prediction block size. We
762 // mask to match the size of the block we are working on and then shift it
763 // into place..
764 *above_y |= (size_mask[block_size] &
765 above_64x64_txform_mask[tx_size_y]) << shift_y;
766 *above_uv |= (size_mask_uv[block_size] &
767 above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
768
769 *left_y |= (size_mask[block_size] &
770 left_64x64_txform_mask[tx_size_y]) << shift_y;
771 *left_uv |= (size_mask_uv[block_size] &
772 left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
773
774 // Here we are trying to determine what to do with the internal 4x4 block
775 // boundaries. These differ from the 4x4 boundaries on the outside edge of
776 // an 8x8 in that the internal ones can be skipped and don't depend on
777 // the prediction block size.
778 if (tx_size_y == TX_4X4)
779 *int_4x4_y |= size_mask[block_size] << shift_y;
780
781 if (tx_size_uv == TX_4X4)
782 *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
783 }
784
785 // This function does the same thing as the one above with the exception that
786 // it only affects the y masks. It exists because for blocks < 16x16 in size,
787 // we only update u and v masks on the first block.
build_y_mask(const loop_filter_info_n * const lfi_n,const MODE_INFO * mi,const int shift_y,LOOP_FILTER_MASK * lfm)788 static void build_y_mask(const loop_filter_info_n *const lfi_n,
789 const MODE_INFO *mi, const int shift_y,
790 LOOP_FILTER_MASK *lfm) {
791 const MB_MODE_INFO *mbmi = &mi->mbmi;
792 const BLOCK_SIZE block_size = mbmi->sb_type;
793 const TX_SIZE tx_size_y = mbmi->tx_size;
794 const int filter_level = get_filter_level(lfi_n, mbmi);
795 uint64_t *const left_y = &lfm->left_y[tx_size_y];
796 uint64_t *const above_y = &lfm->above_y[tx_size_y];
797 uint64_t *const int_4x4_y = &lfm->int_4x4_y;
798 int i;
799
800 if (!filter_level) {
801 return;
802 } else {
803 const int w = num_8x8_blocks_wide_lookup[block_size];
804 const int h = num_8x8_blocks_high_lookup[block_size];
805 int index = shift_y;
806 for (i = 0; i < h; i++) {
807 memset(&lfm->lfl_y[index], filter_level, w);
808 index += 8;
809 }
810 }
811
812 *above_y |= above_prediction_mask[block_size] << shift_y;
813 *left_y |= left_prediction_mask[block_size] << shift_y;
814
815 if (mbmi->skip && is_inter_block(mbmi))
816 return;
817
818 *above_y |= (size_mask[block_size] &
819 above_64x64_txform_mask[tx_size_y]) << shift_y;
820
821 *left_y |= (size_mask[block_size] &
822 left_64x64_txform_mask[tx_size_y]) << shift_y;
823
824 if (tx_size_y == TX_4X4)
825 *int_4x4_y |= size_mask[block_size] << shift_y;
826 }
827
vp9_adjust_mask(VP9_COMMON * const cm,const int mi_row,const int mi_col,LOOP_FILTER_MASK * lfm)828 void vp9_adjust_mask(VP9_COMMON *const cm, const int mi_row,
829 const int mi_col, LOOP_FILTER_MASK *lfm) {
830 int i;
831
832 // The largest loopfilter we have is 16x16 so we use the 16x16 mask
833 // for 32x32 transforms also.
834 lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
835 lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
836 lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
837 lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
838
839 // We do at least 8 tap filter on every 32x32 even if the transform size
840 // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
841 // remove it from the 4x4.
842 lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
843 lfm->left_y[TX_4X4] &= ~left_border;
844 lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
845 lfm->above_y[TX_4X4] &= ~above_border;
846 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
847 lfm->left_uv[TX_4X4] &= ~left_border_uv;
848 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
849 lfm->above_uv[TX_4X4] &= ~above_border_uv;
850
851 // We do some special edge handling.
852 if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
853 const uint64_t rows = cm->mi_rows - mi_row;
854
855 // Each pixel inside the border gets a 1,
856 const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1);
857 const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1);
858
859 // Remove values completely outside our border.
860 for (i = 0; i < TX_32X32; i++) {
861 lfm->left_y[i] &= mask_y;
862 lfm->above_y[i] &= mask_y;
863 lfm->left_uv[i] &= mask_uv;
864 lfm->above_uv[i] &= mask_uv;
865 }
866 lfm->int_4x4_y &= mask_y;
867 lfm->int_4x4_uv &= mask_uv;
868
869 // We don't apply a wide loop filter on the last uv block row. If set
870 // apply the shorter one instead.
871 if (rows == 1) {
872 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
873 lfm->above_uv[TX_16X16] = 0;
874 }
875 if (rows == 5) {
876 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
877 lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
878 }
879 }
880
881 if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
882 const uint64_t columns = cm->mi_cols - mi_col;
883
884 // Each pixel inside the border gets a 1, the multiply copies the border
885 // to where we need it.
886 const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL;
887 const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
888
889 // Internal edges are not applied on the last column of the image so
890 // we mask 1 more for the internal edges
891 const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
892
893 // Remove the bits outside the image edge.
894 for (i = 0; i < TX_32X32; i++) {
895 lfm->left_y[i] &= mask_y;
896 lfm->above_y[i] &= mask_y;
897 lfm->left_uv[i] &= mask_uv;
898 lfm->above_uv[i] &= mask_uv;
899 }
900 lfm->int_4x4_y &= mask_y;
901 lfm->int_4x4_uv &= mask_uv_int;
902
903 // We don't apply a wide loop filter on the last uv column. If set
904 // apply the shorter one instead.
905 if (columns == 1) {
906 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
907 lfm->left_uv[TX_16X16] = 0;
908 }
909 if (columns == 5) {
910 lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
911 lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
912 }
913 }
914 // We don't apply a loop filter on the first column in the image, mask that
915 // out.
916 if (mi_col == 0) {
917 for (i = 0; i < TX_32X32; i++) {
918 lfm->left_y[i] &= 0xfefefefefefefefeULL;
919 lfm->left_uv[i] &= 0xeeee;
920 }
921 }
922
923 // Assert if we try to apply 2 different loop filters at the same position.
924 assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
925 assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
926 assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
927 assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
928 assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8]));
929 assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
930 assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
931 assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
932 assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
933 assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
934 assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
935 assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
936 assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
937 assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
938 assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
939 assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
940 }
941
942 // This function sets up the bit masks for the entire 64x64 region represented
943 // by mi_row, mi_col.
944 // TODO(JBB): This function only works for yv12.
vp9_setup_mask(VP9_COMMON * const cm,const int mi_row,const int mi_col,MODE_INFO ** mi,const int mode_info_stride,LOOP_FILTER_MASK * lfm)945 void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
946 MODE_INFO **mi, const int mode_info_stride,
947 LOOP_FILTER_MASK *lfm) {
948 int idx_32, idx_16, idx_8;
949 const loop_filter_info_n *const lfi_n = &cm->lf_info;
950 MODE_INFO **mip = mi;
951 MODE_INFO **mip2 = mi;
952
953 // These are offsets to the next mi in the 64x64 block. It is what gets
954 // added to the mi ptr as we go through each loop. It helps us to avoid
955 // setting up special row and column counters for each index. The last step
956 // brings us out back to the starting position.
957 const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
958 -(mode_info_stride << 2) - 4};
959 const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2,
960 -(mode_info_stride << 1) - 2};
961 const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};
962
963 // Following variables represent shifts to position the current block
964 // mask over the appropriate block. A shift of 36 to the left will move
965 // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
966 // 4 rows to the appropriate spot.
967 const int shift_32_y[] = {0, 4, 32, 36};
968 const int shift_16_y[] = {0, 2, 16, 18};
969 const int shift_8_y[] = {0, 1, 8, 9};
970 const int shift_32_uv[] = {0, 2, 8, 10};
971 const int shift_16_uv[] = {0, 1, 4, 5};
972 const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ?
973 cm->mi_rows - mi_row : MI_BLOCK_SIZE);
974 const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ?
975 cm->mi_cols - mi_col : MI_BLOCK_SIZE);
976
977 vp9_zero(*lfm);
978 assert(mip[0] != NULL);
979
980 // TODO(jimbankoski): Try moving most of the following code into decode
981 // loop and storing lfm in the mbmi structure so that we don't have to go
982 // through the recursive loop structure multiple times.
983 switch (mip[0]->mbmi.sb_type) {
984 case BLOCK_64X64:
985 build_masks(lfi_n, mip[0] , 0, 0, lfm);
986 break;
987 case BLOCK_64X32:
988 build_masks(lfi_n, mip[0], 0, 0, lfm);
989 mip2 = mip + mode_info_stride * 4;
990 if (4 >= max_rows)
991 break;
992 build_masks(lfi_n, mip2[0], 32, 8, lfm);
993 break;
994 case BLOCK_32X64:
995 build_masks(lfi_n, mip[0], 0, 0, lfm);
996 mip2 = mip + 4;
997 if (4 >= max_cols)
998 break;
999 build_masks(lfi_n, mip2[0], 4, 2, lfm);
1000 break;
1001 default:
1002 for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
1003 const int shift_y = shift_32_y[idx_32];
1004 const int shift_uv = shift_32_uv[idx_32];
1005 const int mi_32_col_offset = ((idx_32 & 1) << 2);
1006 const int mi_32_row_offset = ((idx_32 >> 1) << 2);
1007 if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
1008 continue;
1009 switch (mip[0]->mbmi.sb_type) {
1010 case BLOCK_32X32:
1011 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1012 break;
1013 case BLOCK_32X16:
1014 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1015 if (mi_32_row_offset + 2 >= max_rows)
1016 continue;
1017 mip2 = mip + mode_info_stride * 2;
1018 build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
1019 break;
1020 case BLOCK_16X32:
1021 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1022 if (mi_32_col_offset + 2 >= max_cols)
1023 continue;
1024 mip2 = mip + 2;
1025 build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
1026 break;
1027 default:
1028 for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
1029 const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
1030 const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
1031 const int mi_16_col_offset = mi_32_col_offset +
1032 ((idx_16 & 1) << 1);
1033 const int mi_16_row_offset = mi_32_row_offset +
1034 ((idx_16 >> 1) << 1);
1035
1036 if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
1037 continue;
1038
1039 switch (mip[0]->mbmi.sb_type) {
1040 case BLOCK_16X16:
1041 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1042 break;
1043 case BLOCK_16X8:
1044 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1045 if (mi_16_row_offset + 1 >= max_rows)
1046 continue;
1047 mip2 = mip + mode_info_stride;
1048 build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
1049 break;
1050 case BLOCK_8X16:
1051 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1052 if (mi_16_col_offset +1 >= max_cols)
1053 continue;
1054 mip2 = mip + 1;
1055 build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
1056 break;
1057 default: {
1058 const int shift_y = shift_32_y[idx_32] +
1059 shift_16_y[idx_16] +
1060 shift_8_y[0];
1061 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
1062 mip += offset[0];
1063 for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
1064 const int shift_y = shift_32_y[idx_32] +
1065 shift_16_y[idx_16] +
1066 shift_8_y[idx_8];
1067 const int mi_8_col_offset = mi_16_col_offset +
1068 ((idx_8 & 1));
1069 const int mi_8_row_offset = mi_16_row_offset +
1070 ((idx_8 >> 1));
1071
1072 if (mi_8_col_offset >= max_cols ||
1073 mi_8_row_offset >= max_rows)
1074 continue;
1075 build_y_mask(lfi_n, mip[0], shift_y, lfm);
1076 }
1077 break;
1078 }
1079 }
1080 }
1081 break;
1082 }
1083 }
1084 break;
1085 }
1086
1087 vp9_adjust_mask(cm, mi_row, mi_col, lfm);
1088 }
1089
filter_selectively_vert(uint8_t * s,int pitch,unsigned int mask_16x16,unsigned int mask_8x8,unsigned int mask_4x4,unsigned int mask_4x4_int,const loop_filter_info_n * lfi_n,const uint8_t * lfl)1090 static void filter_selectively_vert(uint8_t *s, int pitch,
1091 unsigned int mask_16x16,
1092 unsigned int mask_8x8,
1093 unsigned int mask_4x4,
1094 unsigned int mask_4x4_int,
1095 const loop_filter_info_n *lfi_n,
1096 const uint8_t *lfl) {
1097 unsigned int mask;
1098
1099 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
1100 mask; mask >>= 1) {
1101 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
1102
1103 if (mask & 1) {
1104 if (mask_16x16 & 1) {
1105 vpx_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
1106 } else if (mask_8x8 & 1) {
1107 vpx_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1108 } else if (mask_4x4 & 1) {
1109 vpx_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1110 }
1111 }
1112 if (mask_4x4_int & 1)
1113 vpx_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1114 s += 8;
1115 lfl += 1;
1116 mask_16x16 >>= 1;
1117 mask_8x8 >>= 1;
1118 mask_4x4 >>= 1;
1119 mask_4x4_int >>= 1;
1120 }
1121 }
1122
1123 #if CONFIG_VP9_HIGHBITDEPTH
highbd_filter_selectively_vert(uint16_t * s,int pitch,unsigned int mask_16x16,unsigned int mask_8x8,unsigned int mask_4x4,unsigned int mask_4x4_int,const loop_filter_info_n * lfi_n,const uint8_t * lfl,int bd)1124 static void highbd_filter_selectively_vert(uint16_t *s, int pitch,
1125 unsigned int mask_16x16,
1126 unsigned int mask_8x8,
1127 unsigned int mask_4x4,
1128 unsigned int mask_4x4_int,
1129 const loop_filter_info_n *lfi_n,
1130 const uint8_t *lfl, int bd) {
1131 unsigned int mask;
1132
1133 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
1134 mask; mask >>= 1) {
1135 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
1136
1137 if (mask & 1) {
1138 if (mask_16x16 & 1) {
1139 vpx_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim,
1140 lfi->hev_thr, bd);
1141 } else if (mask_8x8 & 1) {
1142 vpx_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim,
1143 lfi->hev_thr, 1, bd);
1144 } else if (mask_4x4 & 1) {
1145 vpx_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim,
1146 lfi->hev_thr, 1, bd);
1147 }
1148 }
1149 if (mask_4x4_int & 1)
1150 vpx_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
1151 lfi->hev_thr, 1, bd);
1152 s += 8;
1153 lfl += 1;
1154 mask_16x16 >>= 1;
1155 mask_8x8 >>= 1;
1156 mask_4x4 >>= 1;
1157 mask_4x4_int >>= 1;
1158 }
1159 }
1160 #endif // CONFIG_VP9_HIGHBITDEPTH
1161
vp9_filter_block_plane_non420(VP9_COMMON * cm,struct macroblockd_plane * plane,MODE_INFO ** mi_8x8,int mi_row,int mi_col)1162 void vp9_filter_block_plane_non420(VP9_COMMON *cm,
1163 struct macroblockd_plane *plane,
1164 MODE_INFO **mi_8x8,
1165 int mi_row, int mi_col) {
1166 const int ss_x = plane->subsampling_x;
1167 const int ss_y = plane->subsampling_y;
1168 const int row_step = 1 << ss_y;
1169 const int col_step = 1 << ss_x;
1170 const int row_step_stride = cm->mi_stride * row_step;
1171 struct buf_2d *const dst = &plane->dst;
1172 uint8_t* const dst0 = dst->buf;
1173 unsigned int mask_16x16[MI_BLOCK_SIZE] = {0};
1174 unsigned int mask_8x8[MI_BLOCK_SIZE] = {0};
1175 unsigned int mask_4x4[MI_BLOCK_SIZE] = {0};
1176 unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
1177 uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
1178 int r, c;
1179
1180 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1181 unsigned int mask_16x16_c = 0;
1182 unsigned int mask_8x8_c = 0;
1183 unsigned int mask_4x4_c = 0;
1184 unsigned int border_mask;
1185
1186 // Determine the vertical edges that need filtering
1187 for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
1188 const MODE_INFO *mi = mi_8x8[c];
1189 const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
1190 const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
1191 // left edge of current unit is block/partition edge -> no skip
1192 const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
1193 !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
1194 const int skip_this_c = skip_this && !block_edge_left;
1195 // top edge of current unit is block/partition edge -> no skip
1196 const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
1197 !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
1198 const int skip_this_r = skip_this && !block_edge_above;
1199 const TX_SIZE tx_size = get_uv_tx_size(&mi[0].mbmi, plane);
1200 const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
1201 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1202
1203 // Filter level can vary per MI
1204 if (!(lfl[(r << 3) + (c >> ss_x)] =
1205 get_filter_level(&cm->lf_info, &mi[0].mbmi)))
1206 continue;
1207
1208 // Build masks based on the transform size of each block
1209 if (tx_size == TX_32X32) {
1210 if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
1211 if (!skip_border_4x4_c)
1212 mask_16x16_c |= 1 << (c >> ss_x);
1213 else
1214 mask_8x8_c |= 1 << (c >> ss_x);
1215 }
1216 if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
1217 if (!skip_border_4x4_r)
1218 mask_16x16[r] |= 1 << (c >> ss_x);
1219 else
1220 mask_8x8[r] |= 1 << (c >> ss_x);
1221 }
1222 } else if (tx_size == TX_16X16) {
1223 if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
1224 if (!skip_border_4x4_c)
1225 mask_16x16_c |= 1 << (c >> ss_x);
1226 else
1227 mask_8x8_c |= 1 << (c >> ss_x);
1228 }
1229 if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
1230 if (!skip_border_4x4_r)
1231 mask_16x16[r] |= 1 << (c >> ss_x);
1232 else
1233 mask_8x8[r] |= 1 << (c >> ss_x);
1234 }
1235 } else {
1236 // force 8x8 filtering on 32x32 boundaries
1237 if (!skip_this_c) {
1238 if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
1239 mask_8x8_c |= 1 << (c >> ss_x);
1240 else
1241 mask_4x4_c |= 1 << (c >> ss_x);
1242 }
1243
1244 if (!skip_this_r) {
1245 if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
1246 mask_8x8[r] |= 1 << (c >> ss_x);
1247 else
1248 mask_4x4[r] |= 1 << (c >> ss_x);
1249 }
1250
1251 if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
1252 mask_4x4_int[r] |= 1 << (c >> ss_x);
1253 }
1254 }
1255
1256 // Disable filtering on the leftmost column
1257 border_mask = ~(mi_col == 0);
1258 #if CONFIG_VP9_HIGHBITDEPTH
1259 if (cm->use_highbitdepth) {
1260 highbd_filter_selectively_vert(CONVERT_TO_SHORTPTR(dst->buf),
1261 dst->stride,
1262 mask_16x16_c & border_mask,
1263 mask_8x8_c & border_mask,
1264 mask_4x4_c & border_mask,
1265 mask_4x4_int[r],
1266 &cm->lf_info, &lfl[r << 3],
1267 (int)cm->bit_depth);
1268 } else {
1269 filter_selectively_vert(dst->buf, dst->stride,
1270 mask_16x16_c & border_mask,
1271 mask_8x8_c & border_mask,
1272 mask_4x4_c & border_mask,
1273 mask_4x4_int[r],
1274 &cm->lf_info, &lfl[r << 3]);
1275 }
1276 #else
1277 filter_selectively_vert(dst->buf, dst->stride,
1278 mask_16x16_c & border_mask,
1279 mask_8x8_c & border_mask,
1280 mask_4x4_c & border_mask,
1281 mask_4x4_int[r],
1282 &cm->lf_info, &lfl[r << 3]);
1283 #endif // CONFIG_VP9_HIGHBITDEPTH
1284 dst->buf += 8 * dst->stride;
1285 mi_8x8 += row_step_stride;
1286 }
1287
1288 // Now do horizontal pass
1289 dst->buf = dst0;
1290 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1291 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1292 const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
1293
1294 unsigned int mask_16x16_r;
1295 unsigned int mask_8x8_r;
1296 unsigned int mask_4x4_r;
1297
1298 if (mi_row + r == 0) {
1299 mask_16x16_r = 0;
1300 mask_8x8_r = 0;
1301 mask_4x4_r = 0;
1302 } else {
1303 mask_16x16_r = mask_16x16[r];
1304 mask_8x8_r = mask_8x8[r];
1305 mask_4x4_r = mask_4x4[r];
1306 }
1307 #if CONFIG_VP9_HIGHBITDEPTH
1308 if (cm->use_highbitdepth) {
1309 highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
1310 dst->stride,
1311 mask_16x16_r,
1312 mask_8x8_r,
1313 mask_4x4_r,
1314 mask_4x4_int_r,
1315 &cm->lf_info, &lfl[r << 3],
1316 (int)cm->bit_depth);
1317 } else {
1318 filter_selectively_horiz(dst->buf, dst->stride,
1319 mask_16x16_r,
1320 mask_8x8_r,
1321 mask_4x4_r,
1322 mask_4x4_int_r,
1323 &cm->lf_info, &lfl[r << 3]);
1324 }
1325 #else
1326 filter_selectively_horiz(dst->buf, dst->stride,
1327 mask_16x16_r,
1328 mask_8x8_r,
1329 mask_4x4_r,
1330 mask_4x4_int_r,
1331 &cm->lf_info, &lfl[r << 3]);
1332 #endif // CONFIG_VP9_HIGHBITDEPTH
1333 dst->buf += 8 * dst->stride;
1334 }
1335 }
1336
vp9_filter_block_plane_ss00(VP9_COMMON * const cm,struct macroblockd_plane * const plane,int mi_row,LOOP_FILTER_MASK * lfm)1337 void vp9_filter_block_plane_ss00(VP9_COMMON *const cm,
1338 struct macroblockd_plane *const plane,
1339 int mi_row,
1340 LOOP_FILTER_MASK *lfm) {
1341 struct buf_2d *const dst = &plane->dst;
1342 uint8_t *const dst0 = dst->buf;
1343 int r;
1344 uint64_t mask_16x16 = lfm->left_y[TX_16X16];
1345 uint64_t mask_8x8 = lfm->left_y[TX_8X8];
1346 uint64_t mask_4x4 = lfm->left_y[TX_4X4];
1347 uint64_t mask_4x4_int = lfm->int_4x4_y;
1348
1349 assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
1350
1351 // Vertical pass: do 2 rows at one time
1352 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
1353 unsigned int mask_16x16_l = mask_16x16 & 0xffff;
1354 unsigned int mask_8x8_l = mask_8x8 & 0xffff;
1355 unsigned int mask_4x4_l = mask_4x4 & 0xffff;
1356 unsigned int mask_4x4_int_l = mask_4x4_int & 0xffff;
1357
1358 // Disable filtering on the leftmost column.
1359 #if CONFIG_VP9_HIGHBITDEPTH
1360 if (cm->use_highbitdepth) {
1361 highbd_filter_selectively_vert_row2(
1362 plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
1363 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1364 &lfm->lfl_y[r << 3], (int)cm->bit_depth);
1365 } else {
1366 filter_selectively_vert_row2(
1367 plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
1368 mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
1369 }
1370 #else
1371 filter_selectively_vert_row2(
1372 plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
1373 mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
1374 #endif // CONFIG_VP9_HIGHBITDEPTH
1375 dst->buf += 16 * dst->stride;
1376 mask_16x16 >>= 16;
1377 mask_8x8 >>= 16;
1378 mask_4x4 >>= 16;
1379 mask_4x4_int >>= 16;
1380 }
1381
1382 // Horizontal pass
1383 dst->buf = dst0;
1384 mask_16x16 = lfm->above_y[TX_16X16];
1385 mask_8x8 = lfm->above_y[TX_8X8];
1386 mask_4x4 = lfm->above_y[TX_4X4];
1387 mask_4x4_int = lfm->int_4x4_y;
1388
1389 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
1390 unsigned int mask_16x16_r;
1391 unsigned int mask_8x8_r;
1392 unsigned int mask_4x4_r;
1393
1394 if (mi_row + r == 0) {
1395 mask_16x16_r = 0;
1396 mask_8x8_r = 0;
1397 mask_4x4_r = 0;
1398 } else {
1399 mask_16x16_r = mask_16x16 & 0xff;
1400 mask_8x8_r = mask_8x8 & 0xff;
1401 mask_4x4_r = mask_4x4 & 0xff;
1402 }
1403
1404 #if CONFIG_VP9_HIGHBITDEPTH
1405 if (cm->use_highbitdepth) {
1406 highbd_filter_selectively_horiz(
1407 CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
1408 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info, &lfm->lfl_y[r << 3],
1409 (int)cm->bit_depth);
1410 } else {
1411 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1412 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
1413 &lfm->lfl_y[r << 3]);
1414 }
1415 #else
1416 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1417 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
1418 &lfm->lfl_y[r << 3]);
1419 #endif // CONFIG_VP9_HIGHBITDEPTH
1420
1421 dst->buf += 8 * dst->stride;
1422 mask_16x16 >>= 8;
1423 mask_8x8 >>= 8;
1424 mask_4x4 >>= 8;
1425 mask_4x4_int >>= 8;
1426 }
1427 }
1428
vp9_filter_block_plane_ss11(VP9_COMMON * const cm,struct macroblockd_plane * const plane,int mi_row,LOOP_FILTER_MASK * lfm)1429 void vp9_filter_block_plane_ss11(VP9_COMMON *const cm,
1430 struct macroblockd_plane *const plane,
1431 int mi_row,
1432 LOOP_FILTER_MASK *lfm) {
1433 struct buf_2d *const dst = &plane->dst;
1434 uint8_t *const dst0 = dst->buf;
1435 int r, c;
1436 uint8_t lfl_uv[16];
1437
1438 uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
1439 uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
1440 uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
1441 uint16_t mask_4x4_int = lfm->int_4x4_uv;
1442
1443 assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
1444
1445 // Vertical pass: do 2 rows at one time
1446 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
1447 for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
1448 lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
1449 lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)];
1450 }
1451
1452 {
1453 unsigned int mask_16x16_l = mask_16x16 & 0xff;
1454 unsigned int mask_8x8_l = mask_8x8 & 0xff;
1455 unsigned int mask_4x4_l = mask_4x4 & 0xff;
1456 unsigned int mask_4x4_int_l = mask_4x4_int & 0xff;
1457
1458 // Disable filtering on the leftmost column.
1459 #if CONFIG_VP9_HIGHBITDEPTH
1460 if (cm->use_highbitdepth) {
1461 highbd_filter_selectively_vert_row2(
1462 plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
1463 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1464 &lfl_uv[r << 1], (int)cm->bit_depth);
1465 } else {
1466 filter_selectively_vert_row2(
1467 plane->subsampling_x, dst->buf, dst->stride,
1468 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1469 &lfl_uv[r << 1]);
1470 }
1471 #else
1472 filter_selectively_vert_row2(
1473 plane->subsampling_x, dst->buf, dst->stride,
1474 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1475 &lfl_uv[r << 1]);
1476 #endif // CONFIG_VP9_HIGHBITDEPTH
1477
1478 dst->buf += 16 * dst->stride;
1479 mask_16x16 >>= 8;
1480 mask_8x8 >>= 8;
1481 mask_4x4 >>= 8;
1482 mask_4x4_int >>= 8;
1483 }
1484 }
1485
1486 // Horizontal pass
1487 dst->buf = dst0;
1488 mask_16x16 = lfm->above_uv[TX_16X16];
1489 mask_8x8 = lfm->above_uv[TX_8X8];
1490 mask_4x4 = lfm->above_uv[TX_4X4];
1491 mask_4x4_int = lfm->int_4x4_uv;
1492
1493 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
1494 const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
1495 const unsigned int mask_4x4_int_r =
1496 skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
1497 unsigned int mask_16x16_r;
1498 unsigned int mask_8x8_r;
1499 unsigned int mask_4x4_r;
1500
1501 if (mi_row + r == 0) {
1502 mask_16x16_r = 0;
1503 mask_8x8_r = 0;
1504 mask_4x4_r = 0;
1505 } else {
1506 mask_16x16_r = mask_16x16 & 0xf;
1507 mask_8x8_r = mask_8x8 & 0xf;
1508 mask_4x4_r = mask_4x4 & 0xf;
1509 }
1510
1511 #if CONFIG_VP9_HIGHBITDEPTH
1512 if (cm->use_highbitdepth) {
1513 highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
1514 dst->stride, mask_16x16_r, mask_8x8_r,
1515 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1516 &lfl_uv[r << 1], (int)cm->bit_depth);
1517 } else {
1518 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1519 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1520 &lfl_uv[r << 1]);
1521 }
1522 #else
1523 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1524 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1525 &lfl_uv[r << 1]);
1526 #endif // CONFIG_VP9_HIGHBITDEPTH
1527
1528 dst->buf += 8 * dst->stride;
1529 mask_16x16 >>= 4;
1530 mask_8x8 >>= 4;
1531 mask_4x4 >>= 4;
1532 mask_4x4_int >>= 4;
1533 }
1534 }
1535
loop_filter_rows(YV12_BUFFER_CONFIG * frame_buffer,VP9_COMMON * cm,struct macroblockd_plane planes[MAX_MB_PLANE],int start,int stop,int y_only)1536 static void loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, VP9_COMMON *cm,
1537 struct macroblockd_plane planes[MAX_MB_PLANE],
1538 int start, int stop, int y_only) {
1539 const int num_planes = y_only ? 1 : MAX_MB_PLANE;
1540 enum lf_path path;
1541 int mi_row, mi_col;
1542
1543 if (y_only)
1544 path = LF_PATH_444;
1545 else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
1546 path = LF_PATH_420;
1547 else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
1548 path = LF_PATH_444;
1549 else
1550 path = LF_PATH_SLOW;
1551
1552 for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
1553 MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
1554 LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0);
1555
1556 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) {
1557 int plane;
1558
1559 vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
1560
1561 // TODO(JBB): Make setup_mask work for non 420.
1562 vp9_adjust_mask(cm, mi_row, mi_col, lfm);
1563
1564 vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm);
1565 for (plane = 1; plane < num_planes; ++plane) {
1566 switch (path) {
1567 case LF_PATH_420:
1568 vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm);
1569 break;
1570 case LF_PATH_444:
1571 vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm);
1572 break;
1573 case LF_PATH_SLOW:
1574 vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
1575 mi_row, mi_col);
1576 break;
1577 }
1578 }
1579 }
1580 }
1581 }
1582
vp9_loop_filter_frame(YV12_BUFFER_CONFIG * frame,VP9_COMMON * cm,MACROBLOCKD * xd,int frame_filter_level,int y_only,int partial_frame)1583 void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
1584 VP9_COMMON *cm, MACROBLOCKD *xd,
1585 int frame_filter_level,
1586 int y_only, int partial_frame) {
1587 int start_mi_row, end_mi_row, mi_rows_to_filter;
1588 if (!frame_filter_level) return;
1589 start_mi_row = 0;
1590 mi_rows_to_filter = cm->mi_rows;
1591 if (partial_frame && cm->mi_rows > 8) {
1592 start_mi_row = cm->mi_rows >> 1;
1593 start_mi_row &= 0xfffffff8;
1594 mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
1595 }
1596 end_mi_row = start_mi_row + mi_rows_to_filter;
1597 loop_filter_rows(frame, cm, xd->plane, start_mi_row, end_mi_row, y_only);
1598 }
1599
1600 // Used by the encoder to build the loopfilter masks.
vp9_build_mask_frame(VP9_COMMON * cm,int frame_filter_level,int partial_frame)1601 void vp9_build_mask_frame(VP9_COMMON *cm, int frame_filter_level,
1602 int partial_frame) {
1603 int start_mi_row, end_mi_row, mi_rows_to_filter;
1604 int mi_col, mi_row;
1605 if (!frame_filter_level) return;
1606 start_mi_row = 0;
1607 mi_rows_to_filter = cm->mi_rows;
1608 if (partial_frame && cm->mi_rows > 8) {
1609 start_mi_row = cm->mi_rows >> 1;
1610 start_mi_row &= 0xfffffff8;
1611 mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
1612 }
1613 end_mi_row = start_mi_row + mi_rows_to_filter;
1614
1615 vp9_loop_filter_frame_init(cm, frame_filter_level);
1616
1617 for (mi_row = start_mi_row; mi_row < end_mi_row; mi_row += MI_BLOCK_SIZE) {
1618 MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
1619 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
1620 // vp9_setup_mask() zeros lfm
1621 vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
1622 get_lfm(&cm->lf, mi_row, mi_col));
1623 }
1624 }
1625 }
1626
1627 // 8x8 blocks in a superblock. A "1" represents the first block in a 16x16
1628 // or greater area.
1629 static const uint8_t first_block_in_16x16[8][8] = {
1630 {1, 0, 1, 0, 1, 0, 1, 0},
1631 {0, 0, 0, 0, 0, 0, 0, 0},
1632 {1, 0, 1, 0, 1, 0, 1, 0},
1633 {0, 0, 0, 0, 0, 0, 0, 0},
1634 {1, 0, 1, 0, 1, 0, 1, 0},
1635 {0, 0, 0, 0, 0, 0, 0, 0},
1636 {1, 0, 1, 0, 1, 0, 1, 0},
1637 {0, 0, 0, 0, 0, 0, 0, 0}
1638 };
1639
1640 // This function sets up the bit masks for a block represented
1641 // by mi_row, mi_col in a 64x64 region.
1642 // TODO(SJL): This function only works for yv12.
vp9_build_mask(VP9_COMMON * cm,const MB_MODE_INFO * mbmi,int mi_row,int mi_col,int bw,int bh)1643 void vp9_build_mask(VP9_COMMON *cm, const MB_MODE_INFO *mbmi, int mi_row,
1644 int mi_col, int bw, int bh) {
1645 const BLOCK_SIZE block_size = mbmi->sb_type;
1646 const TX_SIZE tx_size_y = mbmi->tx_size;
1647 const loop_filter_info_n *const lfi_n = &cm->lf_info;
1648 const int filter_level = get_filter_level(lfi_n, mbmi);
1649 const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
1650 LOOP_FILTER_MASK *const lfm = get_lfm(&cm->lf, mi_row, mi_col);
1651 uint64_t *const left_y = &lfm->left_y[tx_size_y];
1652 uint64_t *const above_y = &lfm->above_y[tx_size_y];
1653 uint64_t *const int_4x4_y = &lfm->int_4x4_y;
1654 uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
1655 uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
1656 uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
1657 const int row_in_sb = (mi_row & 7);
1658 const int col_in_sb = (mi_col & 7);
1659 const int shift_y = col_in_sb + (row_in_sb << 3);
1660 const int shift_uv = (col_in_sb >> 1) + ((row_in_sb >> 1) << 2);
1661 const int build_uv = first_block_in_16x16[row_in_sb][col_in_sb];
1662
1663 if (!filter_level) {
1664 return;
1665 } else {
1666 int index = shift_y;
1667 int i;
1668 for (i = 0; i < bh; i++) {
1669 memset(&lfm->lfl_y[index], filter_level, bw);
1670 index += 8;
1671 }
1672 }
1673
1674 // These set 1 in the current block size for the block size edges.
1675 // For instance if the block size is 32x16, we'll set:
1676 // above = 1111
1677 // 0000
1678 // and
1679 // left = 1000
1680 // = 1000
1681 // NOTE : In this example the low bit is left most ( 1000 ) is stored as
1682 // 1, not 8...
1683 //
1684 // U and V set things on a 16 bit scale.
1685 //
1686 *above_y |= above_prediction_mask[block_size] << shift_y;
1687 *left_y |= left_prediction_mask[block_size] << shift_y;
1688
1689 if (build_uv) {
1690 *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
1691 *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
1692 }
1693
1694 // If the block has no coefficients and is not intra we skip applying
1695 // the loop filter on block edges.
1696 if (mbmi->skip && is_inter_block(mbmi))
1697 return;
1698
1699 // Add a mask for the transform size. The transform size mask is set to
1700 // be correct for a 64x64 prediction block size. Mask to match the size of
1701 // the block we are working on and then shift it into place.
1702 *above_y |= (size_mask[block_size] &
1703 above_64x64_txform_mask[tx_size_y]) << shift_y;
1704 *left_y |= (size_mask[block_size] &
1705 left_64x64_txform_mask[tx_size_y]) << shift_y;
1706
1707 if (build_uv) {
1708 *above_uv |= (size_mask_uv[block_size] &
1709 above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
1710
1711 *left_uv |= (size_mask_uv[block_size] &
1712 left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
1713 }
1714
1715 // Try to determine what to do with the internal 4x4 block boundaries. These
1716 // differ from the 4x4 boundaries on the outside edge of an 8x8 in that the
1717 // internal ones can be skipped and don't depend on the prediction block size.
1718 if (tx_size_y == TX_4X4)
1719 *int_4x4_y |= size_mask[block_size] << shift_y;
1720
1721 if (build_uv && tx_size_uv == TX_4X4)
1722 *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
1723 }
1724
vp9_loop_filter_data_reset(LFWorkerData * lf_data,YV12_BUFFER_CONFIG * frame_buffer,struct VP9Common * cm,const struct macroblockd_plane planes[MAX_MB_PLANE])1725 void vp9_loop_filter_data_reset(
1726 LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
1727 struct VP9Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]) {
1728 lf_data->frame_buffer = frame_buffer;
1729 lf_data->cm = cm;
1730 lf_data->start = 0;
1731 lf_data->stop = 0;
1732 lf_data->y_only = 0;
1733 memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
1734 }
1735
vp9_reset_lfm(VP9_COMMON * const cm)1736 void vp9_reset_lfm(VP9_COMMON *const cm) {
1737 if (cm->lf.filter_level) {
1738 memset(cm->lf.lfm, 0,
1739 ((cm->mi_rows + (MI_BLOCK_SIZE - 1)) >> 3) * cm->lf.lfm_stride *
1740 sizeof(*cm->lf.lfm));
1741 }
1742 }
1743
vp9_loop_filter_worker(LFWorkerData * const lf_data,void * unused)1744 int vp9_loop_filter_worker(LFWorkerData *const lf_data, void *unused) {
1745 (void)unused;
1746 loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
1747 lf_data->start, lf_data->stop, lf_data->y_only);
1748 return 1;
1749 }
1750