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 #if CONFIG_MISC_FIXES
723 uint16_t *const int_4x4_uv = &lfm->left_int_4x4_uv;
725 uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
729 // If filter level is 0 we don't loop filter.
733 const int w = num_8x8_blocks_wide_lookup[block_size];
734 const int h = num_8x8_blocks_high_lookup[block_size];
736 for (i = 0; i < h; i++) {
737 memset(&lfm->lfl_y[index], filter_level, w);
742 // These set 1 in the current block size for the block size edges.
743 // For instance if the block size is 32x16, we'll set:
749 // NOTE : In this example the low bit is left most ( 1000 ) is stored as
752 // U and V set things on a 16 bit scale.
754 *above_y |= above_prediction_mask[block_size] << shift_y;
755 *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
756 *left_y |= left_prediction_mask[block_size] << shift_y;
757 *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
759 // If the block has no coefficients and is not intra we skip applying
760 // the loop filter on block edges.
761 #if CONFIG_MISC_FIXES
762 if ((mbmi->skip || mbmi->has_no_coeffs) && is_inter_block(mbmi))
765 if (mbmi->skip && is_inter_block(mbmi))
769 // Here we are adding a mask for the transform size. The transform
770 // size mask is set to be correct for a 64x64 prediction block size. We
771 // mask to match the size of the block we are working on and then shift it
773 *above_y |= (size_mask[block_size] &
774 above_64x64_txform_mask[tx_size_y]) << shift_y;
775 *above_uv |= (size_mask_uv[block_size] &
776 above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
778 *left_y |= (size_mask[block_size] &
779 left_64x64_txform_mask[tx_size_y]) << shift_y;
780 *left_uv |= (size_mask_uv[block_size] &
781 left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
783 // Here we are trying to determine what to do with the internal 4x4 block
784 // boundaries. These differ from the 4x4 boundaries on the outside edge of
785 // an 8x8 in that the internal ones can be skipped and don't depend on
786 // the prediction block size.
787 if (tx_size_y == TX_4X4)
788 *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
790 if (tx_size_uv == TX_4X4)
791 *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
794 // This function does the same thing as the one above with the exception that
795 // it only affects the y masks. It exists because for blocks < 16x16 in size,
796 // we only update u and v masks on the first block.
797 static void build_y_mask(const loop_filter_info_n *const lfi_n,
798 const MODE_INFO *mi, const int shift_y,
799 LOOP_FILTER_MASK *lfm) {
800 const MB_MODE_INFO *mbmi = &mi->mbmi;
801 const BLOCK_SIZE block_size = mbmi->sb_type;
802 const TX_SIZE tx_size_y = mbmi->tx_size;
803 const int filter_level = get_filter_level(lfi_n, mbmi);
804 uint64_t *const left_y = &lfm->left_y[tx_size_y];
805 uint64_t *const above_y = &lfm->above_y[tx_size_y];
806 uint64_t *const int_4x4_y = &lfm->int_4x4_y;
812 const int w = num_8x8_blocks_wide_lookup[block_size];
813 const int h = num_8x8_blocks_high_lookup[block_size];
815 for (i = 0; i < h; i++) {
816 memset(&lfm->lfl_y[index], filter_level, w);
821 *above_y |= above_prediction_mask[block_size] << shift_y;
822 *left_y |= left_prediction_mask[block_size] << shift_y;
824 #if CONFIG_MISC_FIXES
825 if ((mbmi->skip || mbmi->has_no_coeffs) && is_inter_block(mbmi))
828 if (mbmi->skip && is_inter_block(mbmi))
832 *above_y |= (size_mask[block_size] &
833 above_64x64_txform_mask[tx_size_y]) << shift_y;
835 *left_y |= (size_mask[block_size] &
836 left_64x64_txform_mask[tx_size_y]) << shift_y;
838 if (tx_size_y == TX_4X4)
839 *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
842 // This function sets up the bit masks for the entire 64x64 region represented
843 // by mi_row, mi_col.
844 // TODO(JBB): This function only works for yv12.
845 void vp10_setup_mask(VP10_COMMON *const cm, const int mi_row, const int mi_col,
846 MODE_INFO **mi, const int mode_info_stride,
847 LOOP_FILTER_MASK *lfm) {
848 int idx_32, idx_16, idx_8;
849 const loop_filter_info_n *const lfi_n = &cm->lf_info;
850 MODE_INFO **mip = mi;
851 MODE_INFO **mip2 = mi;
853 // These are offsets to the next mi in the 64x64 block. It is what gets
854 // added to the mi ptr as we go through each loop. It helps us to avoid
855 // setting up special row and column counters for each index. The last step
856 // brings us out back to the starting position.
857 const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
858 -(mode_info_stride << 2) - 4};
859 const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2,
860 -(mode_info_stride << 1) - 2};
861 const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};
863 // Following variables represent shifts to position the current block
864 // mask over the appropriate block. A shift of 36 to the left will move
865 // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
866 // 4 rows to the appropriate spot.
867 const int shift_32_y[] = {0, 4, 32, 36};
868 const int shift_16_y[] = {0, 2, 16, 18};
869 const int shift_8_y[] = {0, 1, 8, 9};
870 const int shift_32_uv[] = {0, 2, 8, 10};
871 const int shift_16_uv[] = {0, 1, 4, 5};
873 const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ?
874 cm->mi_rows - mi_row : MI_BLOCK_SIZE);
875 const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ?
876 cm->mi_cols - mi_col : MI_BLOCK_SIZE);
879 assert(mip[0] != NULL);
881 // TODO(jimbankoski): Try moving most of the following code into decode
882 // loop and storing lfm in the mbmi structure so that we don't have to go
883 // through the recursive loop structure multiple times.
884 switch (mip[0]->mbmi.sb_type) {
886 build_masks(lfi_n, mip[0] , 0, 0, lfm);
889 build_masks(lfi_n, mip[0], 0, 0, lfm);
890 mip2 = mip + mode_info_stride * 4;
893 build_masks(lfi_n, mip2[0], 32, 8, lfm);
896 build_masks(lfi_n, mip[0], 0, 0, lfm);
900 build_masks(lfi_n, mip2[0], 4, 2, lfm);
903 for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
904 const int shift_y = shift_32_y[idx_32];
905 const int shift_uv = shift_32_uv[idx_32];
906 const int mi_32_col_offset = ((idx_32 & 1) << 2);
907 const int mi_32_row_offset = ((idx_32 >> 1) << 2);
908 if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
910 switch (mip[0]->mbmi.sb_type) {
912 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
915 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
916 if (mi_32_row_offset + 2 >= max_rows)
918 mip2 = mip + mode_info_stride * 2;
919 build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
922 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
923 if (mi_32_col_offset + 2 >= max_cols)
926 build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
929 for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
930 const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
931 const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
932 const int mi_16_col_offset = mi_32_col_offset +
934 const int mi_16_row_offset = mi_32_row_offset +
935 ((idx_16 >> 1) << 1);
937 if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
940 switch (mip[0]->mbmi.sb_type) {
942 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
945 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
946 if (mi_16_row_offset + 1 >= max_rows)
948 mip2 = mip + mode_info_stride;
949 build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
952 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
953 if (mi_16_col_offset +1 >= max_cols)
956 build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
959 const int shift_y = shift_32_y[idx_32] +
962 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
964 for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
965 const int shift_y = shift_32_y[idx_32] +
968 const int mi_8_col_offset = mi_16_col_offset +
970 const int mi_8_row_offset = mi_16_row_offset +
973 if (mi_8_col_offset >= max_cols ||
974 mi_8_row_offset >= max_rows)
976 build_y_mask(lfi_n, mip[0], shift_y, lfm);
987 // The largest loopfilter we have is 16x16 so we use the 16x16 mask
988 // for 32x32 transforms also.
989 lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
990 lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
991 lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
992 lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
994 // We do at least 8 tap filter on every 32x32 even if the transform size
995 // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
996 // remove it from the 4x4.
997 lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
998 lfm->left_y[TX_4X4] &= ~left_border;
999 lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
1000 lfm->above_y[TX_4X4] &= ~above_border;
1001 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
1002 lfm->left_uv[TX_4X4] &= ~left_border_uv;
1003 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
1004 lfm->above_uv[TX_4X4] &= ~above_border_uv;
1006 // We do some special edge handling.
1007 if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
1008 const uint64_t rows = cm->mi_rows - mi_row;
1010 // Each pixel inside the border gets a 1,
1011 const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1);
1012 const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1);
1014 // Remove values completely outside our border.
1015 for (i = 0; i < TX_32X32; i++) {
1016 lfm->left_y[i] &= mask_y;
1017 lfm->above_y[i] &= mask_y;
1018 lfm->left_uv[i] &= mask_uv;
1019 lfm->above_uv[i] &= mask_uv;
1021 lfm->int_4x4_y &= mask_y;
1022 #if CONFIG_MISC_FIXES
1023 lfm->above_int_4x4_uv = lfm->left_int_4x4_uv & mask_uv;
1025 lfm->int_4x4_uv &= mask_uv;
1028 // We don't apply a wide loop filter on the last uv block row. If set
1029 // apply the shorter one instead.
1031 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
1032 lfm->above_uv[TX_16X16] = 0;
1035 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
1036 lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
1040 if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
1041 const uint64_t columns = cm->mi_cols - mi_col;
1043 // Each pixel inside the border gets a 1, the multiply copies the border
1044 // to where we need it.
1045 const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL;
1046 const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
1048 // Internal edges are not applied on the last column of the image so
1049 // we mask 1 more for the internal edges
1050 const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
1052 // Remove the bits outside the image edge.
1053 for (i = 0; i < TX_32X32; i++) {
1054 lfm->left_y[i] &= mask_y;
1055 lfm->above_y[i] &= mask_y;
1056 lfm->left_uv[i] &= mask_uv;
1057 lfm->above_uv[i] &= mask_uv;
1059 lfm->int_4x4_y &= mask_y;
1060 #if CONFIG_MISC_FIXES
1061 lfm->left_int_4x4_uv &= mask_uv_int;
1063 lfm->int_4x4_uv &= mask_uv_int;
1066 // We don't apply a wide loop filter on the last uv column. If set
1067 // apply the shorter one instead.
1069 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
1070 lfm->left_uv[TX_16X16] = 0;
1073 lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
1074 lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
1077 // We don't apply a loop filter on the first column in the image, mask that
1080 for (i = 0; i < TX_32X32; i++) {
1081 lfm->left_y[i] &= 0xfefefefefefefefeULL;
1082 lfm->left_uv[i] &= 0xeeee;
1086 // Assert if we try to apply 2 different loop filters at the same position.
1087 assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
1088 assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
1089 assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
1090 assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
1091 assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8]));
1092 assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
1093 assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
1094 #if CONFIG_MISC_FIXES
1095 assert(!(lfm->left_int_4x4_uv & lfm->left_uv[TX_16X16]));
1097 assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
1099 assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
1100 assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
1101 assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
1102 assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
1103 assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
1104 assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
1105 assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
1106 #if CONFIG_MISC_FIXES
1107 assert(!(lfm->above_int_4x4_uv & lfm->above_uv[TX_16X16]));
1109 assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
1113 static void filter_selectively_vert(uint8_t *s, int pitch,
1114 unsigned int mask_16x16,
1115 unsigned int mask_8x8,
1116 unsigned int mask_4x4,
1117 unsigned int mask_4x4_int,
1118 const loop_filter_info_n *lfi_n,
1119 const uint8_t *lfl) {
1122 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
1124 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
1127 if (mask_16x16 & 1) {
1128 vpx_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
1129 } else if (mask_8x8 & 1) {
1130 vpx_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1131 } else if (mask_4x4 & 1) {
1132 vpx_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1135 if (mask_4x4_int & 1)
1136 vpx_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
1146 #if CONFIG_VP9_HIGHBITDEPTH
1147 static void highbd_filter_selectively_vert(uint16_t *s, int pitch,
1148 unsigned int mask_16x16,
1149 unsigned int mask_8x8,
1150 unsigned int mask_4x4,
1151 unsigned int mask_4x4_int,
1152 const loop_filter_info_n *lfi_n,
1153 const uint8_t *lfl, int bd) {
1156 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
1158 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
1161 if (mask_16x16 & 1) {
1162 vpx_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim,
1164 } else if (mask_8x8 & 1) {
1165 vpx_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim,
1166 lfi->hev_thr, 1, bd);
1167 } else if (mask_4x4 & 1) {
1168 vpx_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim,
1169 lfi->hev_thr, 1, bd);
1172 if (mask_4x4_int & 1)
1173 vpx_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
1174 lfi->hev_thr, 1, bd);
1183 #endif // CONFIG_VP9_HIGHBITDEPTH
1185 void vp10_filter_block_plane_non420(VP10_COMMON *cm,
1186 struct macroblockd_plane *plane,
1188 int mi_row, int mi_col) {
1189 const int ss_x = plane->subsampling_x;
1190 const int ss_y = plane->subsampling_y;
1191 const int row_step = 1 << ss_y;
1192 const int col_step = 1 << ss_x;
1193 const int row_step_stride = cm->mi_stride * row_step;
1194 struct buf_2d *const dst = &plane->dst;
1195 uint8_t* const dst0 = dst->buf;
1196 unsigned int mask_16x16[MI_BLOCK_SIZE] = {0};
1197 unsigned int mask_8x8[MI_BLOCK_SIZE] = {0};
1198 unsigned int mask_4x4[MI_BLOCK_SIZE] = {0};
1199 unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
1200 uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
1203 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1204 unsigned int mask_16x16_c = 0;
1205 unsigned int mask_8x8_c = 0;
1206 unsigned int mask_4x4_c = 0;
1207 unsigned int border_mask;
1209 // Determine the vertical edges that need filtering
1210 for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
1211 const MODE_INFO *mi = mi_8x8[c];
1212 const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
1213 const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
1214 // left edge of current unit is block/partition edge -> no skip
1215 const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
1216 !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
1217 const int skip_this_c = skip_this && !block_edge_left;
1218 // top edge of current unit is block/partition edge -> no skip
1219 const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
1220 !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
1221 const int skip_this_r = skip_this && !block_edge_above;
1222 const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
1223 ? get_uv_tx_size(&mi[0].mbmi, plane)
1224 : mi[0].mbmi.tx_size;
1225 const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
1226 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1228 // Filter level can vary per MI
1229 if (!(lfl[(r << 3) + (c >> ss_x)] =
1230 get_filter_level(&cm->lf_info, &mi[0].mbmi)))
1233 // Build masks based on the transform size of each block
1234 if (tx_size == TX_32X32) {
1235 if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
1236 if (!skip_border_4x4_c)
1237 mask_16x16_c |= 1 << (c >> ss_x);
1239 mask_8x8_c |= 1 << (c >> ss_x);
1241 if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
1242 if (!skip_border_4x4_r)
1243 mask_16x16[r] |= 1 << (c >> ss_x);
1245 mask_8x8[r] |= 1 << (c >> ss_x);
1247 } else if (tx_size == TX_16X16) {
1248 if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
1249 if (!skip_border_4x4_c)
1250 mask_16x16_c |= 1 << (c >> ss_x);
1252 mask_8x8_c |= 1 << (c >> ss_x);
1254 if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
1255 if (!skip_border_4x4_r)
1256 mask_16x16[r] |= 1 << (c >> ss_x);
1258 mask_8x8[r] |= 1 << (c >> ss_x);
1261 // force 8x8 filtering on 32x32 boundaries
1263 if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
1264 mask_8x8_c |= 1 << (c >> ss_x);
1266 mask_4x4_c |= 1 << (c >> ss_x);
1270 if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
1271 mask_8x8[r] |= 1 << (c >> ss_x);
1273 mask_4x4[r] |= 1 << (c >> ss_x);
1276 if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
1277 mask_4x4_int[r] |= 1 << (c >> ss_x);
1281 // Disable filtering on the leftmost column
1282 border_mask = ~(mi_col == 0);
1283 #if CONFIG_VP9_HIGHBITDEPTH
1284 if (cm->use_highbitdepth) {
1285 highbd_filter_selectively_vert(CONVERT_TO_SHORTPTR(dst->buf),
1287 mask_16x16_c & border_mask,
1288 mask_8x8_c & border_mask,
1289 mask_4x4_c & border_mask,
1291 &cm->lf_info, &lfl[r << 3],
1292 (int)cm->bit_depth);
1294 filter_selectively_vert(dst->buf, dst->stride,
1295 mask_16x16_c & border_mask,
1296 mask_8x8_c & border_mask,
1297 mask_4x4_c & border_mask,
1299 &cm->lf_info, &lfl[r << 3]);
1302 filter_selectively_vert(dst->buf, dst->stride,
1303 mask_16x16_c & border_mask,
1304 mask_8x8_c & border_mask,
1305 mask_4x4_c & border_mask,
1307 &cm->lf_info, &lfl[r << 3]);
1308 #endif // CONFIG_VP9_HIGHBITDEPTH
1309 dst->buf += 8 * dst->stride;
1310 mi_8x8 += row_step_stride;
1313 // Now do horizontal pass
1315 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1316 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1317 const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
1319 unsigned int mask_16x16_r;
1320 unsigned int mask_8x8_r;
1321 unsigned int mask_4x4_r;
1323 if (mi_row + r == 0) {
1328 mask_16x16_r = mask_16x16[r];
1329 mask_8x8_r = mask_8x8[r];
1330 mask_4x4_r = mask_4x4[r];
1332 #if CONFIG_VP9_HIGHBITDEPTH
1333 if (cm->use_highbitdepth) {
1334 highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
1340 &cm->lf_info, &lfl[r << 3],
1341 (int)cm->bit_depth);
1343 filter_selectively_horiz(dst->buf, dst->stride,
1348 &cm->lf_info, &lfl[r << 3]);
1351 filter_selectively_horiz(dst->buf, dst->stride,
1356 &cm->lf_info, &lfl[r << 3]);
1357 #endif // CONFIG_VP9_HIGHBITDEPTH
1358 dst->buf += 8 * dst->stride;
1362 void vp10_filter_block_plane_ss00(VP10_COMMON *const cm,
1363 struct macroblockd_plane *const plane,
1365 LOOP_FILTER_MASK *lfm) {
1366 struct buf_2d *const dst = &plane->dst;
1367 uint8_t *const dst0 = dst->buf;
1369 uint64_t mask_16x16 = lfm->left_y[TX_16X16];
1370 uint64_t mask_8x8 = lfm->left_y[TX_8X8];
1371 uint64_t mask_4x4 = lfm->left_y[TX_4X4];
1372 uint64_t mask_4x4_int = lfm->int_4x4_y;
1374 assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
1376 // Vertical pass: do 2 rows at one time
1377 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
1378 unsigned int mask_16x16_l = mask_16x16 & 0xffff;
1379 unsigned int mask_8x8_l = mask_8x8 & 0xffff;
1380 unsigned int mask_4x4_l = mask_4x4 & 0xffff;
1381 unsigned int mask_4x4_int_l = mask_4x4_int & 0xffff;
1383 // Disable filtering on the leftmost column.
1384 #if CONFIG_VP9_HIGHBITDEPTH
1385 if (cm->use_highbitdepth) {
1386 highbd_filter_selectively_vert_row2(
1387 plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
1388 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1389 &lfm->lfl_y[r << 3], (int)cm->bit_depth);
1391 filter_selectively_vert_row2(
1392 plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
1393 mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
1396 filter_selectively_vert_row2(
1397 plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
1398 mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
1399 #endif // CONFIG_VP9_HIGHBITDEPTH
1400 dst->buf += 16 * dst->stride;
1404 mask_4x4_int >>= 16;
1409 mask_16x16 = lfm->above_y[TX_16X16];
1410 mask_8x8 = lfm->above_y[TX_8X8];
1411 mask_4x4 = lfm->above_y[TX_4X4];
1412 mask_4x4_int = lfm->int_4x4_y;
1414 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
1415 unsigned int mask_16x16_r;
1416 unsigned int mask_8x8_r;
1417 unsigned int mask_4x4_r;
1419 if (mi_row + r == 0) {
1424 mask_16x16_r = mask_16x16 & 0xff;
1425 mask_8x8_r = mask_8x8 & 0xff;
1426 mask_4x4_r = mask_4x4 & 0xff;
1429 #if CONFIG_VP9_HIGHBITDEPTH
1430 if (cm->use_highbitdepth) {
1431 highbd_filter_selectively_horiz(
1432 CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
1433 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info, &lfm->lfl_y[r << 3],
1434 (int)cm->bit_depth);
1436 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1437 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
1438 &lfm->lfl_y[r << 3]);
1441 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1442 mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
1443 &lfm->lfl_y[r << 3]);
1444 #endif // CONFIG_VP9_HIGHBITDEPTH
1446 dst->buf += 8 * dst->stride;
1454 void vp10_filter_block_plane_ss11(VP10_COMMON *const cm,
1455 struct macroblockd_plane *const plane,
1457 LOOP_FILTER_MASK *lfm) {
1458 struct buf_2d *const dst = &plane->dst;
1459 uint8_t *const dst0 = dst->buf;
1462 uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
1463 uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
1464 uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
1465 #if CONFIG_MISC_FIXES
1466 uint16_t mask_4x4_int = lfm->left_int_4x4_uv;
1468 uint16_t mask_4x4_int = lfm->int_4x4_uv;
1471 assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
1473 // Vertical pass: do 2 rows at one time
1474 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
1475 if (plane->plane_type == 1) {
1476 for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
1477 lfm->lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
1478 lfm->lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)];
1483 unsigned int mask_16x16_l = mask_16x16 & 0xff;
1484 unsigned int mask_8x8_l = mask_8x8 & 0xff;
1485 unsigned int mask_4x4_l = mask_4x4 & 0xff;
1486 unsigned int mask_4x4_int_l = mask_4x4_int & 0xff;
1488 // Disable filtering on the leftmost column.
1489 #if CONFIG_VP9_HIGHBITDEPTH
1490 if (cm->use_highbitdepth) {
1491 highbd_filter_selectively_vert_row2(
1492 plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
1493 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1494 &lfm->lfl_uv[r << 1], (int)cm->bit_depth);
1496 filter_selectively_vert_row2(
1497 plane->subsampling_x, dst->buf, dst->stride,
1498 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1499 &lfm->lfl_uv[r << 1]);
1502 filter_selectively_vert_row2(
1503 plane->subsampling_x, dst->buf, dst->stride,
1504 mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
1505 &lfm->lfl_uv[r << 1]);
1506 #endif // CONFIG_VP9_HIGHBITDEPTH
1508 dst->buf += 16 * dst->stride;
1518 mask_16x16 = lfm->above_uv[TX_16X16];
1519 mask_8x8 = lfm->above_uv[TX_8X8];
1520 mask_4x4 = lfm->above_uv[TX_4X4];
1521 #if CONFIG_MISC_FIXES
1522 mask_4x4_int = lfm->above_int_4x4_uv;
1524 mask_4x4_int = lfm->int_4x4_uv;
1527 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
1528 const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
1529 const unsigned int mask_4x4_int_r =
1530 skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
1531 unsigned int mask_16x16_r;
1532 unsigned int mask_8x8_r;
1533 unsigned int mask_4x4_r;
1535 if (mi_row + r == 0) {
1540 mask_16x16_r = mask_16x16 & 0xf;
1541 mask_8x8_r = mask_8x8 & 0xf;
1542 mask_4x4_r = mask_4x4 & 0xf;
1545 #if CONFIG_VP9_HIGHBITDEPTH
1546 if (cm->use_highbitdepth) {
1547 highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
1548 dst->stride, mask_16x16_r, mask_8x8_r,
1549 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1550 &lfm->lfl_uv[r << 1], (int)cm->bit_depth);
1552 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1553 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1554 &lfm->lfl_uv[r << 1]);
1557 filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
1558 mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
1559 &lfm->lfl_uv[r << 1]);
1560 #endif // CONFIG_VP9_HIGHBITDEPTH
1562 dst->buf += 8 * dst->stride;
1570 void vp10_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer,
1572 struct macroblockd_plane planes[MAX_MB_PLANE],
1573 int start, int stop, int y_only) {
1574 const int num_planes = y_only ? 1 : MAX_MB_PLANE;
1576 LOOP_FILTER_MASK lfm;
1581 else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
1583 else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
1586 path = LF_PATH_SLOW;
1588 for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
1589 MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
1591 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
1594 vp10_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
1596 // TODO(JBB): Make setup_mask work for non 420.
1597 vp10_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
1600 vp10_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
1601 for (plane = 1; plane < num_planes; ++plane) {
1604 vp10_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
1607 vp10_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
1610 vp10_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
1619 void vp10_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
1620 VP10_COMMON *cm, MACROBLOCKD *xd,
1621 int frame_filter_level,
1622 int y_only, int partial_frame) {
1623 int start_mi_row, end_mi_row, mi_rows_to_filter;
1624 if (!frame_filter_level) return;
1626 mi_rows_to_filter = cm->mi_rows;
1627 if (partial_frame && cm->mi_rows > 8) {
1628 start_mi_row = cm->mi_rows >> 1;
1629 start_mi_row &= 0xfffffff8;
1630 mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
1632 end_mi_row = start_mi_row + mi_rows_to_filter;
1633 vp10_loop_filter_frame_init(cm, frame_filter_level);
1634 vp10_loop_filter_rows(frame, cm, xd->plane,
1635 start_mi_row, end_mi_row,
1639 void vp10_loop_filter_data_reset(
1640 LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
1641 struct VP10Common *cm,
1642 const struct macroblockd_plane planes[MAX_MB_PLANE]) {
1643 lf_data->frame_buffer = frame_buffer;
1647 lf_data->y_only = 0;
1648 memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
1651 int vp10_loop_filter_worker(LFWorkerData *const lf_data, void *unused) {
1653 vp10_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
1654 lf_data->start, lf_data->stop, lf_data->y_only);