2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
11 #include "./vpx_config.h"
12 #include "./vpx_dsp_rtcd.h"
13 #include "vp10/common/loopfilter.h"
14 #include "vp10/common/onyxc_int.h"
15 #include "vp10/common/reconinter.h"
16 #include "vpx_dsp/vpx_dsp_common.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vpx_ports/mem.h"
20 #include "vp10/common/seg_common.h"
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
26 // In the case of TX_16X16-> ( in low order byte first we end up with
27 // a mask that looks like this
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
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
50 // In the case of TX_32x32 -> ( in low order byte first we end up with
51 // a mask that looks like this
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
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.
74 // In the case of TX_16x32 -> ( low order byte first ) we end up with
75 // a mask that looks like this :
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
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
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
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;
140 // 16 bit masks for uv transform sizes.
141 static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= {
148 static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= {
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
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] = {
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
188 // 64 bit mask to shift and set for each uv prediction size
189 static const uint16_t size_mask_uv[BLOCK_SIZES] = {
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
204 static const uint16_t left_border_uv = 0x1111;
205 static const uint16_t above_border_uv = 0x000f;
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)
212 static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
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));
220 if (sharpness_lvl > 0) {
221 if (block_inside_limit > (9 - sharpness_lvl))
222 block_inside_limit = (9 - sharpness_lvl);
225 if (block_inside_limit < 1)
226 block_inside_limit = 1;
228 memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
229 memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
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]];
240 void vp10_loop_filter_init(VP10_COMMON *cm) {
241 loop_filter_info_n *lfi = &cm->lf_info;
242 struct loopfilter *lf = &cm->lf;
245 // init limits for given sharpness
246 update_sharpness(lfi, lf->sharpness_level);
247 lf->last_sharpness_level = lf->sharpness_level;
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);
254 void vp10_loop_filter_frame_init(VP10_COMMON *cm, int default_filt_lvl) {
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;
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;
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,
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]));
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);
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);
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;
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;
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;
324 const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
325 const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
327 // TODO(yunqingwang): count in loopfilter functions should be removed.
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,
333 } else if (mask_16x16_0 & 1) {
334 vpx_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
337 vpx_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
338 lfi1->lim, lfi1->hev_thr);
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,
347 } else if (mask_8x8_0 & 1) {
348 vpx_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
351 vpx_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
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,
361 } else if (mask_4x4_0 & 1) {
362 vpx_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
365 vpx_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
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,
375 } else if (mask_4x4_int_0 & 1) {
376 vpx_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
379 vpx_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
390 mask_4x4_int_0 >>= 1;
394 mask_4x4_int_1 >>= 1;
398 #if CONFIG_VP9_HIGHBITDEPTH
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;
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;
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;
424 const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
425 const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
427 // TODO(yunqingwang): count in loopfilter functions should be removed.
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,
433 } else if (mask_16x16_0 & 1) {
434 vpx_highbd_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
437 vpx_highbd_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
438 lfi1->lim, lfi1->hev_thr, bd);
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,
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);
451 vpx_highbd_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim,
452 lfi1->lim, lfi1->hev_thr, 1, bd);
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,
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);
465 vpx_highbd_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim,
466 lfi1->lim, lfi1->hev_thr, 1, bd);
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,
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);
479 vpx_highbd_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim,
480 lfi1->lim, lfi1->hev_thr, 1, bd);
490 mask_4x4_int_0 >>= 1;
494 mask_4x4_int_1 >>= 1;
497 #endif // CONFIG_VP9_HIGHBITDEPTH
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) {
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;
515 if (mask_16x16 & 1) {
516 if ((mask_16x16 & 3) == 3) {
517 vpx_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
521 vpx_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
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);
529 vpx_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
530 lfi->hev_thr, lfin->mblim, lfin->lim,
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);
538 if (mask_4x4_int & 1)
539 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
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);
547 vpx_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
549 if (mask_4x4_int & 1)
550 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
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);
558 vpx_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
559 lfi->hev_thr, lfin->mblim, lfin->lim,
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);
566 if (mask_4x4_int & 1)
567 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
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);
575 vpx_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
577 if (mask_4x4_int & 1)
578 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
581 } else if (mask_4x4_int & 1) {
582 vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
588 mask_16x16 >>= count;
591 mask_4x4_int >>= count;
595 #if CONFIG_VP9_HIGHBITDEPTH
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) {
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;
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);
618 vpx_highbd_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
619 lfi->hev_thr, 1, bd);
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);
626 vpx_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
627 lfi->hev_thr, lfin->mblim, lfin->lim,
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,
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);
646 vpx_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
647 lfi->hev_thr, 1, bd);
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);
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);
659 vpx_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
660 lfi->hev_thr, lfin->mblim, lfin->lim,
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,
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);
678 vpx_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
679 lfi->hev_thr, 1, bd);
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);
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);
693 mask_16x16 >>= count;
696 mask_4x4_int >>= count;
699 #endif // CONFIG_VP9_HIGHBITDEPTH
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
707 // TODO(JBB) Need another function for different resolution color..
708 static void build_masks(const loop_filter_info_n *const lfi_n,
709 const MODE_INFO *mi, const int shift_y,
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;
725 // If filter level is 0 we don't loop filter.
729 const int w = num_8x8_blocks_wide_lookup[block_size];
730 const int h = num_8x8_blocks_high_lookup[block_size];
732 for (i = 0; i < h; i++) {
733 memset(&lfm->lfl_y[index], filter_level, w);
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:
745 // NOTE : In this example the low bit is left most ( 1000 ) is stored as
748 // U and V set things on a 16 bit scale.
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;
755 // If the block has no coefficients and is not intra we skip applying
756 // the loop filter on block edges.
757 #if CONFIG_MISC_FIXES
758 if ((mbmi->skip || mbmi->has_no_coeffs) && is_inter_block(mbmi))
761 if (mbmi->skip && is_inter_block(mbmi))
765 // Here we are adding a mask for the transform size. The transform
766 // size mask is set to be correct for a 64x64 prediction block size. We
767 // mask to match the size of the block we are working on and then shift it
769 *above_y |= (size_mask[block_size] &
770 above_64x64_txform_mask[tx_size_y]) << shift_y;
771 *above_uv |= (size_mask_uv[block_size] &
772 above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
774 *left_y |= (size_mask[block_size] &
775 left_64x64_txform_mask[tx_size_y]) << shift_y;
776 *left_uv |= (size_mask_uv[block_size] &
777 left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
779 // Here we are trying to determine what to do with the internal 4x4 block
780 // boundaries. These differ from the 4x4 boundaries on the outside edge of
781 // an 8x8 in that the internal ones can be skipped and don't depend on
782 // the prediction block size.
783 if (tx_size_y == TX_4X4)
784 *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
786 if (tx_size_uv == TX_4X4)
787 *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
790 // This function does the same thing as the one above with the exception that
791 // it only affects the y masks. It exists because for blocks < 16x16 in size,
792 // we only update u and v masks on the first block.
793 static void build_y_mask(const loop_filter_info_n *const lfi_n,
794 const MODE_INFO *mi, const int shift_y,
795 LOOP_FILTER_MASK *lfm) {
796 const MB_MODE_INFO *mbmi = &mi->mbmi;
797 const BLOCK_SIZE block_size = mbmi->sb_type;
798 const TX_SIZE tx_size_y = mbmi->tx_size;
799 const int filter_level = get_filter_level(lfi_n, mbmi);
800 uint64_t *const left_y = &lfm->left_y[tx_size_y];
801 uint64_t *const above_y = &lfm->above_y[tx_size_y];
802 uint64_t *const int_4x4_y = &lfm->int_4x4_y;
808 const int w = num_8x8_blocks_wide_lookup[block_size];
809 const int h = num_8x8_blocks_high_lookup[block_size];
811 for (i = 0; i < h; i++) {
812 memset(&lfm->lfl_y[index], filter_level, w);
817 *above_y |= above_prediction_mask[block_size] << shift_y;
818 *left_y |= left_prediction_mask[block_size] << shift_y;
820 #if CONFIG_MISC_FIXES
821 if ((mbmi->skip || mbmi->has_no_coeffs) && is_inter_block(mbmi))
824 if (mbmi->skip && is_inter_block(mbmi))
828 *above_y |= (size_mask[block_size] &
829 above_64x64_txform_mask[tx_size_y]) << shift_y;
831 *left_y |= (size_mask[block_size] &
832 left_64x64_txform_mask[tx_size_y]) << shift_y;
834 if (tx_size_y == TX_4X4)
835 *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
838 // This function sets up the bit masks for the entire 64x64 region represented
839 // by mi_row, mi_col.
840 // TODO(JBB): This function only works for yv12.
841 void vp10_setup_mask(VP10_COMMON *const cm, const int mi_row, const int mi_col,
842 MODE_INFO **mi, const int mode_info_stride,
843 LOOP_FILTER_MASK *lfm) {
844 int idx_32, idx_16, idx_8;
845 const loop_filter_info_n *const lfi_n = &cm->lf_info;
846 MODE_INFO **mip = mi;
847 MODE_INFO **mip2 = mi;
849 // These are offsets to the next mi in the 64x64 block. It is what gets
850 // added to the mi ptr as we go through each loop. It helps us to avoid
851 // setting up special row and column counters for each index. The last step
852 // brings us out back to the starting position.
853 const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
854 -(mode_info_stride << 2) - 4};
855 const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2,
856 -(mode_info_stride << 1) - 2};
857 const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};
859 // Following variables represent shifts to position the current block
860 // mask over the appropriate block. A shift of 36 to the left will move
861 // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
862 // 4 rows to the appropriate spot.
863 const int shift_32_y[] = {0, 4, 32, 36};
864 const int shift_16_y[] = {0, 2, 16, 18};
865 const int shift_8_y[] = {0, 1, 8, 9};
866 const int shift_32_uv[] = {0, 2, 8, 10};
867 const int shift_16_uv[] = {0, 1, 4, 5};
869 const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ?
870 cm->mi_rows - mi_row : MI_BLOCK_SIZE);
871 const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ?
872 cm->mi_cols - mi_col : MI_BLOCK_SIZE);
875 assert(mip[0] != NULL);
877 // TODO(jimbankoski): Try moving most of the following code into decode
878 // loop and storing lfm in the mbmi structure so that we don't have to go
879 // through the recursive loop structure multiple times.
880 switch (mip[0]->mbmi.sb_type) {
882 build_masks(lfi_n, mip[0] , 0, 0, lfm);
885 build_masks(lfi_n, mip[0], 0, 0, lfm);
886 mip2 = mip + mode_info_stride * 4;
889 build_masks(lfi_n, mip2[0], 32, 8, lfm);
892 build_masks(lfi_n, mip[0], 0, 0, lfm);
896 build_masks(lfi_n, mip2[0], 4, 2, lfm);
899 for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
900 const int shift_y = shift_32_y[idx_32];
901 const int shift_uv = shift_32_uv[idx_32];
902 const int mi_32_col_offset = ((idx_32 & 1) << 2);
903 const int mi_32_row_offset = ((idx_32 >> 1) << 2);
904 if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
906 switch (mip[0]->mbmi.sb_type) {
908 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
911 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
912 if (mi_32_row_offset + 2 >= max_rows)
914 mip2 = mip + mode_info_stride * 2;
915 build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
918 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
919 if (mi_32_col_offset + 2 >= max_cols)
922 build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
925 for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
926 const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
927 const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
928 const int mi_16_col_offset = mi_32_col_offset +
930 const int mi_16_row_offset = mi_32_row_offset +
931 ((idx_16 >> 1) << 1);
933 if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
936 switch (mip[0]->mbmi.sb_type) {
938 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
941 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
942 if (mi_16_row_offset + 1 >= max_rows)
944 mip2 = mip + mode_info_stride;
945 build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
948 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
949 if (mi_16_col_offset +1 >= max_cols)
952 build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
955 const int shift_y = shift_32_y[idx_32] +
958 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
960 for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
961 const int shift_y = shift_32_y[idx_32] +
964 const int mi_8_col_offset = mi_16_col_offset +
966 const int mi_8_row_offset = mi_16_row_offset +
969 if (mi_8_col_offset >= max_cols ||
970 mi_8_row_offset >= max_rows)
972 build_y_mask(lfi_n, mip[0], shift_y, lfm);
983 // The largest loopfilter we have is 16x16 so we use the 16x16 mask
984 // for 32x32 transforms also.
985 lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
986 lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
987 lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
988 lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
990 // We do at least 8 tap filter on every 32x32 even if the transform size
991 // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
992 // remove it from the 4x4.
993 lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
994 lfm->left_y[TX_4X4] &= ~left_border;
995 lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
996 lfm->above_y[TX_4X4] &= ~above_border;
997 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
998 lfm->left_uv[TX_4X4] &= ~left_border_uv;
999 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
1000 lfm->above_uv[TX_4X4] &= ~above_border_uv;
1002 // We do some special edge handling.
1003 if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
1004 const uint64_t rows = cm->mi_rows - mi_row;
1006 // Each pixel inside the border gets a 1,
1007 const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1);
1008 const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1);
1010 // Remove values completely outside our border.
1011 for (i = 0; i < TX_32X32; i++) {
1012 lfm->left_y[i] &= mask_y;
1013 lfm->above_y[i] &= mask_y;
1014 lfm->left_uv[i] &= mask_uv;
1015 lfm->above_uv[i] &= mask_uv;
1017 lfm->int_4x4_y &= mask_y;
1018 lfm->int_4x4_uv &= mask_uv;
1020 // We don't apply a wide loop filter on the last uv block row. If set
1021 // apply the shorter one instead.
1023 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
1024 lfm->above_uv[TX_16X16] = 0;
1027 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
1028 lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
1032 if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
1033 const uint64_t columns = cm->mi_cols - mi_col;
1035 // Each pixel inside the border gets a 1, the multiply copies the border
1036 // to where we need it.
1037 const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL;
1038 const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
1040 // Internal edges are not applied on the last column of the image so
1041 // we mask 1 more for the internal edges
1042 const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
1044 // Remove the bits outside the image edge.
1045 for (i = 0; i < TX_32X32; i++) {
1046 lfm->left_y[i] &= mask_y;
1047 lfm->above_y[i] &= mask_y;
1048 lfm->left_uv[i] &= mask_uv;
1049 lfm->above_uv[i] &= mask_uv;
1051 lfm->int_4x4_y &= mask_y;
1052 lfm->int_4x4_uv &= mask_uv_int;
1054 // We don't apply a wide loop filter on the last uv column. If set
1055 // apply the shorter one instead.
1057 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
1058 lfm->left_uv[TX_16X16] = 0;
1061 lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
1062 lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
1065 // We don't apply a loop filter on the first column in the image, mask that
1068 for (i = 0; i < TX_32X32; i++) {
1069 lfm->left_y[i] &= 0xfefefefefefefefeULL;
1070 lfm->left_uv[i] &= 0xeeee;
1074 // Assert if we try to apply 2 different loop filters at the same position.
1075 assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
1076 assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
1077 assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
1078 assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
1079 assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8]));
1080 assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
1081 assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
1082 assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
1083 assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
1084 assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
1085 assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
1086 assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
1087 assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
1088 assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
1089 assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
1090 assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
1093 static void filter_selectively_vert(uint8_t *s, int pitch,
1094 unsigned int mask_16x16,
1095 unsigned int mask_8x8,
1096 unsigned int mask_4x4,
1097 unsigned int mask_4x4_int,
1098 const loop_filter_info_n *lfi_n,
1099 const uint8_t *lfl) {
1102 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
1104 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
1107 if (mask_16x16 & 1) {
1108 vpx_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
1109 } else if (mask_8x8 & 1) {
1110 vpx_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1111 } else if (mask_4x4 & 1) {
1112 vpx_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1115 if (mask_4x4_int & 1)
1116 vpx_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1126 #if CONFIG_VP9_HIGHBITDEPTH
1127 static void highbd_filter_selectively_vert(uint16_t *s, int pitch,
1128 unsigned int mask_16x16,
1129 unsigned int mask_8x8,
1130 unsigned int mask_4x4,
1131 unsigned int mask_4x4_int,
1132 const loop_filter_info_n *lfi_n,
1133 const uint8_t *lfl, int bd) {
1136 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
1138 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
1141 if (mask_16x16 & 1) {
1142 vpx_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim,
1144 } else if (mask_8x8 & 1) {
1145 vpx_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim,
1146 lfi->hev_thr, 1, bd);
1147 } else if (mask_4x4 & 1) {
1148 vpx_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim,
1149 lfi->hev_thr, 1, bd);
1152 if (mask_4x4_int & 1)
1153 vpx_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
1154 lfi->hev_thr, 1, bd);
1163 #endif // CONFIG_VP9_HIGHBITDEPTH
1165 void vp10_filter_block_plane_non420(VP10_COMMON *cm,
1166 struct macroblockd_plane *plane,
1168 int mi_row, int mi_col) {
1169 const int ss_x = plane->subsampling_x;
1170 const int ss_y = plane->subsampling_y;
1171 const int row_step = 1 << ss_y;
1172 const int col_step = 1 << ss_x;
1173 const int row_step_stride = cm->mi_stride * row_step;
1174 struct buf_2d *const dst = &plane->dst;
1175 uint8_t* const dst0 = dst->buf;
1176 unsigned int mask_16x16[MI_BLOCK_SIZE] = {0};
1177 unsigned int mask_8x8[MI_BLOCK_SIZE] = {0};
1178 unsigned int mask_4x4[MI_BLOCK_SIZE] = {0};
1179 unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
1180 uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
1183 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1184 unsigned int mask_16x16_c = 0;
1185 unsigned int mask_8x8_c = 0;
1186 unsigned int mask_4x4_c = 0;
1187 unsigned int border_mask;
1189 // Determine the vertical edges that need filtering
1190 for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
1191 const MODE_INFO *mi = mi_8x8[c];
1192 const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
1193 const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
1194 // left edge of current unit is block/partition edge -> no skip
1195 const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
1196 !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
1197 const int skip_this_c = skip_this && !block_edge_left;
1198 // top edge of current unit is block/partition edge -> no skip
1199 const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
1200 !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
1201 const int skip_this_r = skip_this && !block_edge_above;
1202 const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
1203 ? get_uv_tx_size(&mi[0].mbmi, plane)
1204 : mi[0].mbmi.tx_size;
1205 const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
1206 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1208 // Filter level can vary per MI
1209 if (!(lfl[(r << 3) + (c >> ss_x)] =
1210 get_filter_level(&cm->lf_info, &mi[0].mbmi)))
1213 // Build masks based on the transform size of each block
1214 if (tx_size == TX_32X32) {
1215 if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
1216 if (!skip_border_4x4_c)
1217 mask_16x16_c |= 1 << (c >> ss_x);
1219 mask_8x8_c |= 1 << (c >> ss_x);
1221 if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
1222 if (!skip_border_4x4_r)
1223 mask_16x16[r] |= 1 << (c >> ss_x);
1225 mask_8x8[r] |= 1 << (c >> ss_x);
1227 } else if (tx_size == TX_16X16) {
1228 if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
1229 if (!skip_border_4x4_c)
1230 mask_16x16_c |= 1 << (c >> ss_x);
1232 mask_8x8_c |= 1 << (c >> ss_x);
1234 if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
1235 if (!skip_border_4x4_r)
1236 mask_16x16[r] |= 1 << (c >> ss_x);
1238 mask_8x8[r] |= 1 << (c >> ss_x);
1241 // force 8x8 filtering on 32x32 boundaries
1243 if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
1244 mask_8x8_c |= 1 << (c >> ss_x);
1246 mask_4x4_c |= 1 << (c >> ss_x);
1250 if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
1251 mask_8x8[r] |= 1 << (c >> ss_x);
1253 mask_4x4[r] |= 1 << (c >> ss_x);
1256 if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
1257 mask_4x4_int[r] |= 1 << (c >> ss_x);
1261 // Disable filtering on the leftmost column
1262 border_mask = ~(mi_col == 0);
1263 #if CONFIG_VP9_HIGHBITDEPTH
1264 if (cm->use_highbitdepth) {
1265 highbd_filter_selectively_vert(CONVERT_TO_SHORTPTR(dst->buf),
1267 mask_16x16_c & border_mask,
1268 mask_8x8_c & border_mask,
1269 mask_4x4_c & border_mask,
1271 &cm->lf_info, &lfl[r << 3],
1272 (int)cm->bit_depth);
1274 filter_selectively_vert(dst->buf, dst->stride,
1275 mask_16x16_c & border_mask,
1276 mask_8x8_c & border_mask,
1277 mask_4x4_c & border_mask,
1279 &cm->lf_info, &lfl[r << 3]);
1282 filter_selectively_vert(dst->buf, dst->stride,
1283 mask_16x16_c & border_mask,
1284 mask_8x8_c & border_mask,
1285 mask_4x4_c & border_mask,
1287 &cm->lf_info, &lfl[r << 3]);
1288 #endif // CONFIG_VP9_HIGHBITDEPTH
1289 dst->buf += 8 * dst->stride;
1290 mi_8x8 += row_step_stride;
1293 // Now do horizontal pass
1295 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1296 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1297 const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
1299 unsigned int mask_16x16_r;
1300 unsigned int mask_8x8_r;
1301 unsigned int mask_4x4_r;
1303 if (mi_row + r == 0) {
1308 mask_16x16_r = mask_16x16[r];
1309 mask_8x8_r = mask_8x8[r];
1310 mask_4x4_r = mask_4x4[r];
1312 #if CONFIG_VP9_HIGHBITDEPTH
1313 if (cm->use_highbitdepth) {
1314 highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
1320 &cm->lf_info, &lfl[r << 3],
1321 (int)cm->bit_depth);
1323 filter_selectively_horiz(dst->buf, dst->stride,
1328 &cm->lf_info, &lfl[r << 3]);
1331 filter_selectively_horiz(dst->buf, dst->stride,
1336 &cm->lf_info, &lfl[r << 3]);
1337 #endif // CONFIG_VP9_HIGHBITDEPTH
1338 dst->buf += 8 * dst->stride;
1342 void vp10_filter_block_plane_ss00(VP10_COMMON *const cm,
1343 struct macroblockd_plane *const plane,
1345 LOOP_FILTER_MASK *lfm) {
1346 struct buf_2d *const dst = &plane->dst;
1347 uint8_t *const dst0 = dst->buf;
1349 uint64_t mask_16x16 = lfm->left_y[TX_16X16];
1350 uint64_t mask_8x8 = lfm->left_y[TX_8X8];
1351 uint64_t mask_4x4 = lfm->left_y[TX_4X4];
1352 uint64_t mask_4x4_int = lfm->int_4x4_y;
1354 assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
1356 // Vertical pass: do 2 rows at one time
1357 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
1358 unsigned int mask_16x16_l = mask_16x16 & 0xffff;
1359 unsigned int mask_8x8_l = mask_8x8 & 0xffff;
1360 unsigned int mask_4x4_l = mask_4x4 & 0xffff;
1361 unsigned int mask_4x4_int_l = mask_4x4_int & 0xffff;
1363 // Disable filtering on the leftmost column.
1364 #if CONFIG_VP9_HIGHBITDEPTH
1365 if (cm->use_highbitdepth) {
1366 highbd_filter_selectively_vert_row2(
1367 plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
1368 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1369 &lfm->lfl_y[r << 3], (int)cm->bit_depth);
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]);
1376 filter_selectively_vert_row2(
1377 plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
1378 mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
1379 #endif // CONFIG_VP9_HIGHBITDEPTH
1380 dst->buf += 16 * dst->stride;
1384 mask_4x4_int >>= 16;
1389 mask_16x16 = lfm->above_y[TX_16X16];
1390 mask_8x8 = lfm->above_y[TX_8X8];
1391 mask_4x4 = lfm->above_y[TX_4X4];
1392 mask_4x4_int = lfm->int_4x4_y;
1394 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
1395 unsigned int mask_16x16_r;
1396 unsigned int mask_8x8_r;
1397 unsigned int mask_4x4_r;
1399 if (mi_row + r == 0) {
1404 mask_16x16_r = mask_16x16 & 0xff;
1405 mask_8x8_r = mask_8x8 & 0xff;
1406 mask_4x4_r = mask_4x4 & 0xff;
1409 #if CONFIG_VP9_HIGHBITDEPTH
1410 if (cm->use_highbitdepth) {
1411 highbd_filter_selectively_horiz(
1412 CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
1413 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info, &lfm->lfl_y[r << 3],
1414 (int)cm->bit_depth);
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]);
1421 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1422 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
1423 &lfm->lfl_y[r << 3]);
1424 #endif // CONFIG_VP9_HIGHBITDEPTH
1426 dst->buf += 8 * dst->stride;
1434 void vp10_filter_block_plane_ss11(VP10_COMMON *const cm,
1435 struct macroblockd_plane *const plane,
1437 LOOP_FILTER_MASK *lfm) {
1438 struct buf_2d *const dst = &plane->dst;
1439 uint8_t *const dst0 = dst->buf;
1442 uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
1443 uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
1444 uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
1445 uint16_t mask_4x4_int = lfm->int_4x4_uv;
1447 assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
1449 // Vertical pass: do 2 rows at one time
1450 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
1451 if (plane->plane_type == 1) {
1452 for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
1453 lfm->lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
1454 lfm->lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)];
1459 unsigned int mask_16x16_l = mask_16x16 & 0xff;
1460 unsigned int mask_8x8_l = mask_8x8 & 0xff;
1461 unsigned int mask_4x4_l = mask_4x4 & 0xff;
1462 unsigned int mask_4x4_int_l = mask_4x4_int & 0xff;
1464 // Disable filtering on the leftmost column.
1465 #if CONFIG_VP9_HIGHBITDEPTH
1466 if (cm->use_highbitdepth) {
1467 highbd_filter_selectively_vert_row2(
1468 plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
1469 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1470 &lfm->lfl_uv[r << 1], (int)cm->bit_depth);
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 &lfm->lfl_uv[r << 1]);
1478 filter_selectively_vert_row2(
1479 plane->subsampling_x, dst->buf, dst->stride,
1480 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1481 &lfm->lfl_uv[r << 1]);
1482 #endif // CONFIG_VP9_HIGHBITDEPTH
1484 dst->buf += 16 * dst->stride;
1494 mask_16x16 = lfm->above_uv[TX_16X16];
1495 mask_8x8 = lfm->above_uv[TX_8X8];
1496 mask_4x4 = lfm->above_uv[TX_4X4];
1497 mask_4x4_int = lfm->int_4x4_uv;
1499 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
1500 const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
1501 const unsigned int mask_4x4_int_r =
1502 skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
1503 unsigned int mask_16x16_r;
1504 unsigned int mask_8x8_r;
1505 unsigned int mask_4x4_r;
1507 if (mi_row + r == 0) {
1512 mask_16x16_r = mask_16x16 & 0xf;
1513 mask_8x8_r = mask_8x8 & 0xf;
1514 mask_4x4_r = mask_4x4 & 0xf;
1517 #if CONFIG_VP9_HIGHBITDEPTH
1518 if (cm->use_highbitdepth) {
1519 highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
1520 dst->stride, mask_16x16_r, mask_8x8_r,
1521 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1522 &lfm->lfl_uv[r << 1], (int)cm->bit_depth);
1524 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1525 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1526 &lfm->lfl_uv[r << 1]);
1529 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1530 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1531 &lfm->lfl_uv[r << 1]);
1532 #endif // CONFIG_VP9_HIGHBITDEPTH
1534 dst->buf += 8 * dst->stride;
1542 void vp10_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer,
1544 struct macroblockd_plane planes[MAX_MB_PLANE],
1545 int start, int stop, int y_only) {
1546 const int num_planes = y_only ? 1 : MAX_MB_PLANE;
1548 LOOP_FILTER_MASK lfm;
1553 else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
1555 else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
1558 path = LF_PATH_SLOW;
1560 for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
1561 MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
1563 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
1566 vp10_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
1568 // TODO(JBB): Make setup_mask work for non 420.
1569 vp10_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
1572 vp10_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
1573 for (plane = 1; plane < num_planes; ++plane) {
1576 vp10_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
1579 vp10_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
1582 vp10_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
1591 void vp10_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
1592 VP10_COMMON *cm, MACROBLOCKD *xd,
1593 int frame_filter_level,
1594 int y_only, int partial_frame) {
1595 int start_mi_row, end_mi_row, mi_rows_to_filter;
1596 if (!frame_filter_level) return;
1598 mi_rows_to_filter = cm->mi_rows;
1599 if (partial_frame && cm->mi_rows > 8) {
1600 start_mi_row = cm->mi_rows >> 1;
1601 start_mi_row &= 0xfffffff8;
1602 mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
1604 end_mi_row = start_mi_row + mi_rows_to_filter;
1605 vp10_loop_filter_frame_init(cm, frame_filter_level);
1606 vp10_loop_filter_rows(frame, cm, xd->plane,
1607 start_mi_row, end_mi_row,
1611 void vp10_loop_filter_data_reset(
1612 LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
1613 struct VP10Common *cm,
1614 const struct macroblockd_plane planes[MAX_MB_PLANE]) {
1615 lf_data->frame_buffer = frame_buffer;
1619 lf_data->y_only = 0;
1620 memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
1623 int vp10_loop_filter_worker(LFWorkerData *const lf_data, void *unused) {
1625 vp10_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
1626 lf_data->start, lf_data->stop, lf_data->y_only);