/src/libhevc/decoder/ihevcd_deblk.c
Line | Count | Source |
1 | | /****************************************************************************** |
2 | | * |
3 | | * Copyright (C) 2012 Ittiam Systems Pvt Ltd, Bangalore |
4 | | * |
5 | | * Licensed under the Apache License, Version 2.0 (the "License"); |
6 | | * you may not use this file except in compliance with the License. |
7 | | * You may obtain a copy of the License at: |
8 | | * |
9 | | * http://www.apache.org/licenses/LICENSE-2.0 |
10 | | * |
11 | | * Unless required by applicable law or agreed to in writing, software |
12 | | * distributed under the License is distributed on an "AS IS" BASIS, |
13 | | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
14 | | * See the License for the specific language governing permissions and |
15 | | * limitations under the License. |
16 | | * |
17 | | ******************************************************************************/ |
18 | | /** |
19 | | ******************************************************************************* |
20 | | * @file |
21 | | * ihevc_deblk.c |
22 | | * |
23 | | * @brief |
24 | | * Contains definition for the ctb level deblk function |
25 | | * |
26 | | * @author |
27 | | * Srinivas T |
28 | | * |
29 | | * @par List of Functions: |
30 | | * - ihevc_deblk() |
31 | | * |
32 | | * @remarks |
33 | | * None |
34 | | * |
35 | | ******************************************************************************* |
36 | | */ |
37 | | |
38 | | #include <stdio.h> |
39 | | #include <stddef.h> |
40 | | #include <stdlib.h> |
41 | | #include <string.h> |
42 | | #include <assert.h> |
43 | | |
44 | | #include "ihevc_typedefs.h" |
45 | | #include "iv.h" |
46 | | #include "ivd.h" |
47 | | #include "ihevcd_cxa.h" |
48 | | #include "ithread.h" |
49 | | |
50 | | #include "ihevc_defs.h" |
51 | | #include "ihevc_debug.h" |
52 | | #include "ihevc_defs.h" |
53 | | #include "ihevc_structs.h" |
54 | | #include "ihevc_macros.h" |
55 | | #include "ihevc_platform_macros.h" |
56 | | #include "ihevc_cabac_tables.h" |
57 | | |
58 | | #include "ihevc_error.h" |
59 | | #include "ihevc_common_tables.h" |
60 | | |
61 | | #include "ihevcd_trace.h" |
62 | | #include "ihevcd_defs.h" |
63 | | #include "ihevcd_function_selector.h" |
64 | | #include "ihevcd_structs.h" |
65 | | #include "ihevcd_error.h" |
66 | | #include "ihevcd_nal.h" |
67 | | #include "ihevcd_bitstream.h" |
68 | | #include "ihevcd_job_queue.h" |
69 | | #include "ihevcd_utils.h" |
70 | | #include "ihevcd_debug.h" |
71 | | |
72 | | #include "ihevc_deblk.h" |
73 | | #include "ihevc_deblk_tables.h" |
74 | | #include "ihevcd_profile.h" |
75 | | /** |
76 | | ******************************************************************************* |
77 | | * |
78 | | * @brief |
79 | | * Deblock CTB level function. |
80 | | * |
81 | | * @par Description: |
82 | | * For a given CTB, deblocking on both vertical and |
83 | | * horizontal edges is done. Both the luma and chroma |
84 | | * blocks are processed |
85 | | * |
86 | | * @param[in] ps_deblk |
87 | | * Pointer to the deblock context |
88 | | * |
89 | | * @returns |
90 | | * |
91 | | * @remarks |
92 | | * None |
93 | | * |
94 | | ******************************************************************************* |
95 | | */ |
96 | | |
97 | | void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk, |
98 | | WORD32 i4_is_last_ctb_x, |
99 | | WORD32 i4_is_last_ctb_y) |
100 | 121k | { |
101 | 121k | WORD32 ctb_size; |
102 | 121k | WORD32 log2_ctb_size; |
103 | 121k | UWORD32 u4_bs; |
104 | 121k | WORD32 bs_tz; /*Leading zeros in boundary strength*/ |
105 | 121k | WORD32 qp_p, qp_q; |
106 | | |
107 | 121k | WORD32 filter_p, filter_q; |
108 | | |
109 | 121k | UWORD8 *pu1_src; |
110 | 121k | WORD32 qp_strd; |
111 | 121k | UWORD32 *pu4_vert_bs, *pu4_horz_bs; |
112 | 121k | UWORD32 *pu4_ctb_vert_bs, *pu4_ctb_horz_bs; |
113 | 121k | WORD32 bs_strd; |
114 | 121k | WORD32 src_strd; |
115 | 121k | UWORD8 *pu1_qp; |
116 | 121k | UWORD16 *pu2_ctb_no_loop_filter_flag; |
117 | 121k | UWORD16 au2_ctb_no_loop_filter_flag[9]; |
118 | | |
119 | 121k | WORD32 col, row; |
120 | | |
121 | | /* Flag to indicate if QP is constant in CTB |
122 | | * 0 - top_left, 1 - top, 2 - left, 3 - current */ |
123 | 121k | UWORD32 u4_qp_const_in_ctb[4] = { 0, 0, 0, 0 }; |
124 | 121k | WORD32 ctb_indx; |
125 | 121k | WORD32 chroma_yuv420sp_vu = ps_deblk->is_chroma_yuv420sp_vu; |
126 | 121k | sps_t *ps_sps; |
127 | 121k | pps_t *ps_pps; |
128 | 121k | codec_t *ps_codec; |
129 | 121k | slice_header_t *ps_slice_hdr; |
130 | | |
131 | 121k | PROFILE_DISABLE_DEBLK(); |
132 | | |
133 | 121k | ps_sps = ps_deblk->ps_sps; |
134 | 121k | ps_pps = ps_deblk->ps_pps; |
135 | 121k | ps_codec = ps_deblk->ps_codec; |
136 | 121k | ps_slice_hdr = ps_deblk->ps_slice_hdr; |
137 | | |
138 | 121k | log2_ctb_size = ps_sps->i1_log2_ctb_size; |
139 | 121k | ctb_size = (1 << ps_sps->i1_log2_ctb_size); |
140 | | |
141 | | /* strides are in units of number of bytes */ |
142 | | /* ctb_size * ctb_size / 8 / 16 is the number of bytes needed per CTB */ |
143 | 121k | bs_strd = (ps_sps->i2_pic_wd_in_ctb + 1) << (2 * log2_ctb_size - 7); |
144 | | |
145 | 121k | pu4_vert_bs = (UWORD32 *)((UWORD8 *)ps_deblk->s_bs_ctxt.pu4_pic_vert_bs + |
146 | 121k | (ps_deblk->i4_ctb_x << (2 * log2_ctb_size - 7)) + |
147 | 121k | ps_deblk->i4_ctb_y * bs_strd); |
148 | 121k | pu4_ctb_vert_bs = pu4_vert_bs; |
149 | | |
150 | 121k | pu4_horz_bs = (UWORD32 *)((UWORD8 *)ps_deblk->s_bs_ctxt.pu4_pic_horz_bs + |
151 | 121k | (ps_deblk->i4_ctb_x << (2 * log2_ctb_size - 7)) + |
152 | 121k | ps_deblk->i4_ctb_y * bs_strd); |
153 | 121k | pu4_ctb_horz_bs = pu4_horz_bs; |
154 | | |
155 | 121k | qp_strd = ps_sps->i2_pic_wd_in_ctb << (log2_ctb_size - 3); |
156 | 121k | pu1_qp = ps_deblk->s_bs_ctxt.pu1_pic_qp + ((ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * qp_strd) << (log2_ctb_size - 3)); |
157 | | |
158 | 121k | pu2_ctb_no_loop_filter_flag = ps_deblk->au2_ctb_no_loop_filter_flag; |
159 | | |
160 | 121k | ctb_indx = ps_deblk->i4_ctb_x + ps_sps->i2_pic_wd_in_ctb * ps_deblk->i4_ctb_y; |
161 | 121k | if(i4_is_last_ctb_y) |
162 | 299 | { |
163 | 299 | pu4_vert_bs = (UWORD32 *)((UWORD8 *)pu4_vert_bs + bs_strd); |
164 | 299 | pu4_ctb_vert_bs = pu4_vert_bs; |
165 | | /* ctb_size/8 is the number of edges per CTB |
166 | | * ctb_size/4 is the number of BS values needed per edge |
167 | | * divided by 8 for the number of bytes |
168 | | * 2 is the number of bits needed for each BS value */ |
169 | 299 | memset(pu4_vert_bs, 0, 1 << (2 * log2_ctb_size - 7)); |
170 | | |
171 | 299 | pu1_qp += (qp_strd << (log2_ctb_size - 3)); |
172 | 299 | pu2_ctb_no_loop_filter_flag += (ctb_size >> 3); |
173 | 299 | ctb_indx += ps_sps->i2_pic_wd_in_ctb; |
174 | 299 | } |
175 | | |
176 | 121k | if(i4_is_last_ctb_x) |
177 | 4.58k | { |
178 | 4.58k | pu4_horz_bs = (UWORD32 *)((UWORD8 *)pu4_horz_bs + (1 << (2 * log2_ctb_size - 7))); |
179 | 4.58k | pu4_ctb_horz_bs = pu4_horz_bs; |
180 | 4.58k | memset(pu4_horz_bs, 0, 1 << (2 * log2_ctb_size - 7)); |
181 | | |
182 | 4.58k | pu1_qp += (ctb_size >> 3); |
183 | | |
184 | 45.5k | for(row = 0; row < (ctb_size >> 3) + 1; row++) |
185 | 41.0k | au2_ctb_no_loop_filter_flag[row] = ps_deblk->au2_ctb_no_loop_filter_flag[row] >> (ctb_size >> 3); |
186 | 4.58k | pu2_ctb_no_loop_filter_flag = au2_ctb_no_loop_filter_flag; |
187 | 4.58k | ctb_indx += 1; |
188 | 4.58k | } |
189 | | |
190 | 121k | u4_qp_const_in_ctb[3] = ps_deblk->s_bs_ctxt.pu1_pic_qp_const_in_ctb[(ctb_indx) >> 3] & (1 << (ctb_indx & 7)); |
191 | | |
192 | 121k | if(ps_deblk->i4_ctb_x || i4_is_last_ctb_x) |
193 | 114k | { |
194 | 114k | u4_qp_const_in_ctb[2] = ps_deblk->s_bs_ctxt.pu1_pic_qp_const_in_ctb[(ctb_indx - 1) >> 3] & (1 << ((ctb_indx - 1) & 7)); |
195 | 114k | } |
196 | | |
197 | 121k | if((ps_deblk->i4_ctb_x || i4_is_last_ctb_x) && (ps_deblk->i4_ctb_y || i4_is_last_ctb_y)) |
198 | 103k | { |
199 | 103k | u4_qp_const_in_ctb[0] = |
200 | 103k | ps_deblk->s_bs_ctxt.pu1_pic_qp_const_in_ctb[(ctb_indx - ps_sps->i2_pic_wd_in_ctb - 1) >> 3] & |
201 | 103k | (1 << ((ctb_indx - ps_sps->i2_pic_wd_in_ctb - 1) & 7)); |
202 | 103k | } |
203 | | |
204 | | |
205 | | |
206 | 121k | if(ps_deblk->i4_ctb_y || i4_is_last_ctb_y) |
207 | 109k | { |
208 | 109k | u4_qp_const_in_ctb[1] = |
209 | 109k | ps_deblk->s_bs_ctxt.pu1_pic_qp_const_in_ctb[(ctb_indx - ps_sps->i2_pic_wd_in_ctb) >> 3] & |
210 | 109k | (1 << ((ctb_indx - ps_sps->i2_pic_wd_in_ctb) & 7)); |
211 | 109k | } |
212 | | |
213 | 121k | src_strd = ps_codec->i4_strd; |
214 | | |
215 | | /* Luma Vertical Edge */ |
216 | | |
217 | 121k | if(0 == i4_is_last_ctb_x) |
218 | 116k | { |
219 | | /* Top CTB's slice header */ |
220 | 116k | slice_header_t *ps_slice_hdr_top; |
221 | 116k | { |
222 | 116k | WORD32 cur_ctb_indx = ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb; |
223 | 116k | if(i4_is_last_ctb_y) |
224 | 299 | cur_ctb_indx += ps_sps->i2_pic_wd_in_ctb; |
225 | 116k | ps_slice_hdr_top = ps_codec->ps_slice_hdr_base + ps_deblk->pu1_slice_idx[cur_ctb_indx - ps_sps->i2_pic_wd_in_ctb]; |
226 | 116k | } |
227 | | |
228 | 116k | pu1_src = ps_deblk->pu1_cur_pic_luma + ((ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_deblk->ps_codec->i4_strd) << (log2_ctb_size)); |
229 | 116k | pu1_src += i4_is_last_ctb_y ? ps_deblk->ps_codec->i4_strd << log2_ctb_size : 0; |
230 | | |
231 | | /** Deblocking is done on a shifted CTB - |
232 | | * Vertical edge processing is done by shifting the CTB up by four pixels */ |
233 | 116k | pu1_src -= 4 * src_strd; |
234 | | |
235 | 1.01M | for(col = 0; col < ctb_size / 8; col++) |
236 | 902k | { |
237 | 902k | WORD32 shift = 0; |
238 | | |
239 | | /* downshift vert_bs by ctb_size/2 for each column |
240 | | * shift = (col & ((MAX_CTB_SIZE >> log2_ctb_size) - 1)) << (log2_ctb_size - 1); |
241 | | * which will reduce to the following assuming ctb size is one of 16, 32 and 64 |
242 | | * and deblocking is done on 8x8 grid |
243 | | */ |
244 | 902k | if(6 != log2_ctb_size) |
245 | 29.0k | shift = (col & 1) << (log2_ctb_size - 1); |
246 | | |
247 | | /* BS for the column - Last row is excluded and the top row is included*/ |
248 | 902k | u4_bs = (pu4_vert_bs[0] >> shift) << 2; |
249 | | |
250 | 902k | if(ps_deblk->i4_ctb_y || i4_is_last_ctb_y) |
251 | 816k | { |
252 | | /* Picking the last BS of the previous CTB corresponding to the same column */ |
253 | 816k | UWORD32 *pu4_vert_bs_top = (UWORD32 *)((UWORD8 *)pu4_vert_bs - bs_strd); |
254 | 816k | UWORD32 u4_top_bs = (*pu4_vert_bs_top) >> (shift + (1 << (log2_ctb_size - 1)) - 2); |
255 | 816k | u4_bs |= u4_top_bs & 3; |
256 | 816k | } |
257 | | |
258 | 4.22M | for(row = 0; row < ctb_size / 4;) |
259 | 3.31M | { |
260 | 3.31M | WORD8 i1_beta_offset_div2 = ps_slice_hdr->i1_beta_offset_div2; |
261 | 3.31M | WORD8 i1_tc_offset_div2 = ps_slice_hdr->i1_tc_offset_div2; |
262 | | |
263 | | /* Trailing zeros are computed and the corresponding rows are not processed */ |
264 | 3.31M | bs_tz = CTZ(u4_bs) >> 1; |
265 | 3.31M | if(0 != bs_tz) |
266 | 1.62M | { |
267 | 1.62M | u4_bs = u4_bs >> (bs_tz << 1); |
268 | 1.62M | if((row + bs_tz) >= (ctb_size / 4)) |
269 | 801k | pu1_src += 4 * (ctb_size / 4 - row) * src_strd; |
270 | 828k | else |
271 | 828k | pu1_src += 4 * bs_tz * src_strd; |
272 | | |
273 | 1.62M | row += bs_tz; |
274 | 1.62M | continue; |
275 | 1.62M | } |
276 | | |
277 | 1.68M | if(0 == row) |
278 | 101k | { |
279 | 101k | i1_beta_offset_div2 = ps_slice_hdr_top->i1_beta_offset_div2; |
280 | 101k | i1_tc_offset_div2 = ps_slice_hdr_top->i1_tc_offset_div2; |
281 | | |
282 | 101k | if(0 == col) |
283 | 24.1k | { |
284 | 24.1k | qp_p = u4_qp_const_in_ctb[0] ? |
285 | 24.0k | pu1_qp[-ctb_size / 8 * qp_strd - ctb_size / 8] : |
286 | 24.1k | pu1_qp[-qp_strd - 1]; |
287 | 24.1k | } |
288 | 77.5k | else |
289 | 77.5k | { |
290 | 77.5k | qp_p = u4_qp_const_in_ctb[1] ? |
291 | 77.0k | pu1_qp[-ctb_size / 8 * qp_strd] : |
292 | 77.5k | pu1_qp[col - 1 - qp_strd]; |
293 | 77.5k | } |
294 | | |
295 | 101k | qp_q = u4_qp_const_in_ctb[1] ? |
296 | 101k | pu1_qp[-ctb_size / 8 * qp_strd] : |
297 | 101k | pu1_qp[col - qp_strd]; |
298 | 101k | } |
299 | 1.58M | else |
300 | 1.58M | { |
301 | 1.58M | if(0 == col) |
302 | 379k | { |
303 | 379k | qp_p = u4_qp_const_in_ctb[2] ? |
304 | 377k | pu1_qp[-ctb_size / 8] : |
305 | 379k | pu1_qp[((row - 1) >> 1) * qp_strd - 1]; |
306 | 379k | } |
307 | 1.20M | else |
308 | 1.20M | { |
309 | 1.20M | qp_p = u4_qp_const_in_ctb[3] ? |
310 | 1.20M | pu1_qp[0] : |
311 | 18.4E | pu1_qp[((row - 1) >> 1) * qp_strd + col - 1]; |
312 | 1.20M | } |
313 | | |
314 | 1.58M | qp_q = u4_qp_const_in_ctb[3] ? |
315 | 1.58M | pu1_qp[0] : |
316 | 1.58M | pu1_qp[((row - 1) >> 1) * qp_strd + col]; |
317 | 1.58M | } |
318 | | |
319 | 1.68M | filter_p = (pu2_ctb_no_loop_filter_flag[(row + 1) >> 1] >> col) & 1; |
320 | 1.68M | filter_q = (pu2_ctb_no_loop_filter_flag[(row + 1) >> 1] >> col) & 2; |
321 | | /* filter_p and filter_q are inverted as they are calculated using no_loop_filter_flags */ |
322 | 1.68M | filter_p = !filter_p; |
323 | 1.68M | filter_q = !filter_q; |
324 | | |
325 | 1.68M | if(filter_p || filter_q) |
326 | 1.69M | { |
327 | 1.69M | DUMP_DEBLK_LUMA_VERT(pu1_src, src_strd, |
328 | 1.69M | u4_bs & 3, qp_p, qp_q, |
329 | 1.69M | ps_slice_hdr->i1_beta_offset_div2, |
330 | 1.69M | ps_slice_hdr->i1_tc_offset_div2, |
331 | 1.69M | filter_p, filter_q); |
332 | 1.69M | ps_codec->s_func_selector.ihevc_deblk_luma_vert_fptr(pu1_src, src_strd, |
333 | 1.69M | u4_bs & 3, qp_p, qp_q, |
334 | 1.69M | i1_beta_offset_div2, |
335 | 1.69M | i1_tc_offset_div2, |
336 | 1.69M | filter_p, filter_q); |
337 | 1.69M | } |
338 | | |
339 | 1.68M | pu1_src += 4 * src_strd; |
340 | 1.68M | u4_bs = u4_bs >> 2; |
341 | 1.68M | row++; |
342 | 1.68M | } |
343 | | |
344 | 902k | if((64 == ctb_size) || |
345 | 902k | ((32 == ctb_size) && (col & 1))) |
346 | 888k | { |
347 | 888k | pu4_vert_bs++; |
348 | 888k | } |
349 | 902k | pu1_src -= (src_strd << log2_ctb_size); |
350 | 902k | pu1_src += 8; |
351 | 902k | } |
352 | 116k | pu4_vert_bs = pu4_ctb_vert_bs; |
353 | 116k | } |
354 | | |
355 | | |
356 | | /* Luma Horizontal Edge */ |
357 | | |
358 | 121k | if(0 == i4_is_last_ctb_y) |
359 | 120k | { |
360 | | |
361 | | /* Left CTB's slice header */ |
362 | 120k | slice_header_t *ps_slice_hdr_left; |
363 | 120k | { |
364 | 120k | WORD32 cur_ctb_indx = ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb; |
365 | 120k | if(i4_is_last_ctb_x) |
366 | 4.58k | cur_ctb_indx += 1; |
367 | 120k | ps_slice_hdr_left = ps_codec->ps_slice_hdr_base + ps_deblk->pu1_slice_idx[cur_ctb_indx - 1]; |
368 | 120k | } |
369 | 120k | pu1_src = ps_deblk->pu1_cur_pic_luma + ((ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_deblk->ps_codec->i4_strd) << log2_ctb_size); |
370 | 120k | pu1_src += i4_is_last_ctb_x ? ctb_size : 0; |
371 | | |
372 | | /** Deblocking is done on a shifted CTB - |
373 | | * Horizontal edge processing is done by shifting the CTB left by four pixels */ |
374 | 120k | pu1_src -= 4; |
375 | 1.05M | for(row = 0; row < ctb_size / 8; row++) |
376 | 936k | { |
377 | 936k | WORD32 shift = 0; |
378 | | |
379 | | /* downshift vert_bs by ctb_size/2 for each column |
380 | | * shift = (row & (MAX_CTB_SIZE / ctb_size - 1)) * ctb_size / 2; |
381 | | * which will reduce to the following assuming ctb size is one of 16, 32 and 64 |
382 | | * and deblocking is done on 8x8 grid |
383 | | */ |
384 | 936k | if(6 != log2_ctb_size) |
385 | 28.6k | shift = (row & 1) << (log2_ctb_size - 1); |
386 | | |
387 | | /* BS for the row - Last column is excluded and the left column is included*/ |
388 | 936k | u4_bs = (pu4_horz_bs[0] >> shift) << 2; |
389 | | |
390 | 936k | if(ps_deblk->i4_ctb_x || i4_is_last_ctb_x) |
391 | 884k | { |
392 | | /** Picking the last BS of the previous CTB corresponding to the same row |
393 | | * UWORD32 *pu4_horz_bs_left = (UWORD32 *)((UWORD8 *)pu4_horz_bs - (ctb_size / 8) * (ctb_size / 4) / 8 * 2); |
394 | | */ |
395 | 884k | UWORD32 *pu4_horz_bs_left = (UWORD32 *)((UWORD8 *)pu4_horz_bs - (1 << (2 * log2_ctb_size - 7))); |
396 | 884k | UWORD32 u4_left_bs = (*pu4_horz_bs_left) >> (shift + (1 << (log2_ctb_size - 1)) - 2); |
397 | 884k | u4_bs |= u4_left_bs & 3; |
398 | 884k | } |
399 | | |
400 | 4.35M | for(col = 0; col < ctb_size / 4;) |
401 | 3.41M | { |
402 | 3.41M | WORD8 i1_beta_offset_div2 = ps_slice_hdr->i1_beta_offset_div2; |
403 | 3.41M | WORD8 i1_tc_offset_div2 = ps_slice_hdr->i1_tc_offset_div2; |
404 | | |
405 | 3.41M | bs_tz = CTZ(u4_bs) >> 1; |
406 | 3.41M | if(0 != bs_tz) |
407 | 1.68M | { |
408 | 1.68M | u4_bs = u4_bs >> (bs_tz << 1); |
409 | | |
410 | 1.68M | if((col + bs_tz) >= (ctb_size / 4)) |
411 | 830k | pu1_src += 4 * (ctb_size / 4 - col); |
412 | 856k | else |
413 | 856k | pu1_src += 4 * bs_tz; |
414 | | |
415 | 1.68M | col += bs_tz; |
416 | 1.68M | continue; |
417 | 1.68M | } |
418 | | |
419 | 1.72M | if(0 == col) |
420 | 106k | { |
421 | 106k | i1_beta_offset_div2 = ps_slice_hdr_left->i1_beta_offset_div2; |
422 | 106k | i1_tc_offset_div2 = ps_slice_hdr_left->i1_tc_offset_div2; |
423 | | |
424 | 106k | if(0 == row) |
425 | 28.3k | { |
426 | 28.3k | qp_p = u4_qp_const_in_ctb[0] ? |
427 | 28.1k | pu1_qp[-ctb_size / 8 * qp_strd - ctb_size / 8] : |
428 | 28.3k | pu1_qp[-qp_strd - 1]; |
429 | 28.3k | } |
430 | 78.5k | else |
431 | 78.5k | { |
432 | 78.5k | qp_p = u4_qp_const_in_ctb[2] ? |
433 | 77.9k | pu1_qp[-ctb_size / 8] : |
434 | 78.5k | pu1_qp[(row - 1) * qp_strd - 1]; |
435 | 78.5k | } |
436 | | |
437 | 106k | qp_q = u4_qp_const_in_ctb[2] ? |
438 | 106k | pu1_qp[-ctb_size / 8] : |
439 | 106k | pu1_qp[row * qp_strd - 1]; |
440 | 106k | } |
441 | 1.62M | else |
442 | 1.62M | { |
443 | 1.62M | if(0 == row) |
444 | 430k | { |
445 | 430k | qp_p = u4_qp_const_in_ctb[1] ? |
446 | 427k | pu1_qp[-ctb_size / 8 * qp_strd] : |
447 | 430k | pu1_qp[((col - 1) >> 1) - qp_strd]; |
448 | 430k | } |
449 | 1.19M | else |
450 | 1.19M | { |
451 | 1.19M | qp_p = u4_qp_const_in_ctb[3] ? |
452 | 1.19M | pu1_qp[0] : |
453 | 1.19M | pu1_qp[((col - 1) >> 1) + (row - 1) * qp_strd]; |
454 | 1.19M | } |
455 | | |
456 | 1.62M | qp_q = u4_qp_const_in_ctb[3] ? |
457 | 1.61M | pu1_qp[0] : |
458 | 1.62M | pu1_qp[((col - 1) >> 1) + row * qp_strd]; |
459 | 1.62M | } |
460 | | |
461 | 1.72M | filter_p = (pu2_ctb_no_loop_filter_flag[row] >> ((col + 1) >> 1)) & 1; |
462 | 1.72M | filter_q = (pu2_ctb_no_loop_filter_flag[row + 1] >> ((col + 1) >> 1)) & 1; |
463 | | /* filter_p and filter_q are inverted as they are calculated using no_loop_filter_flags */ |
464 | 1.72M | filter_p = !filter_p; |
465 | 1.72M | filter_q = !filter_q; |
466 | | |
467 | 1.72M | if(filter_p || filter_q) |
468 | 1.73M | { |
469 | 1.73M | DUMP_DEBLK_LUMA_HORZ(pu1_src, src_strd, |
470 | 1.73M | u4_bs & 3, qp_p, qp_q, |
471 | 1.73M | ps_slice_hdr->i1_beta_offset_div2, |
472 | 1.73M | ps_slice_hdr->i1_tc_offset_div2, |
473 | 1.73M | filter_p, filter_q); |
474 | 1.73M | ps_codec->s_func_selector.ihevc_deblk_luma_horz_fptr(pu1_src, src_strd, |
475 | 1.73M | u4_bs & 3, qp_p, qp_q, |
476 | 1.73M | i1_beta_offset_div2, |
477 | 1.73M | i1_tc_offset_div2, filter_p, filter_q); |
478 | 1.73M | } |
479 | | |
480 | 1.72M | pu1_src += 4; |
481 | 1.72M | u4_bs = u4_bs >> 2; |
482 | 1.72M | col++; |
483 | 1.72M | } |
484 | | |
485 | 936k | if((64 == ctb_size) || |
486 | 936k | ((32 == ctb_size) && (row & 1))) |
487 | 922k | { |
488 | 922k | pu4_horz_bs++; |
489 | 922k | } |
490 | 936k | pu1_src -= ctb_size; |
491 | 936k | pu1_src += (src_strd << 3); |
492 | 936k | } |
493 | 120k | pu4_horz_bs = pu4_ctb_horz_bs; |
494 | 120k | } |
495 | | |
496 | | |
497 | | /* Chroma Veritcal Edge */ |
498 | | |
499 | 121k | if(0 == i4_is_last_ctb_x) |
500 | 116k | { |
501 | | |
502 | | /* Top CTB's slice header */ |
503 | 116k | slice_header_t *ps_slice_hdr_top; |
504 | 116k | { |
505 | 116k | WORD32 cur_ctb_indx = ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb; |
506 | 116k | if(i4_is_last_ctb_y) |
507 | 299 | cur_ctb_indx += ps_sps->i2_pic_wd_in_ctb; |
508 | 116k | ps_slice_hdr_top = ps_codec->ps_slice_hdr_base + ps_deblk->pu1_slice_idx[cur_ctb_indx - ps_sps->i2_pic_wd_in_ctb]; |
509 | 116k | } |
510 | | |
511 | 116k | pu1_src = ps_deblk->pu1_cur_pic_chroma + ((ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_deblk->ps_codec->i4_strd / 2) << log2_ctb_size); |
512 | 116k | pu1_src += i4_is_last_ctb_y ? (ps_deblk->ps_codec->i4_strd / 2) << log2_ctb_size : 0; |
513 | | |
514 | | /** Deblocking is done on a shifted CTB - |
515 | | * Vertical edge processing is done by shifting the CTB up by four pixels */ |
516 | 116k | pu1_src -= 4 * src_strd; |
517 | | |
518 | 568k | for(col = 0; col < ctb_size / 16; col++) |
519 | 451k | { |
520 | | |
521 | | /* BS for the column - Last row is excluded and the top row is included*/ |
522 | 451k | u4_bs = pu4_vert_bs[0] << 2; |
523 | | |
524 | 451k | if(ps_deblk->i4_ctb_y || i4_is_last_ctb_y) |
525 | 408k | { |
526 | | /* Picking the last BS of the previous CTB corresponding to the same column */ |
527 | 408k | UWORD32 *pu4_vert_bs_top = (UWORD32 *)((UWORD8 *)pu4_vert_bs - bs_strd); |
528 | 408k | UWORD32 u4_top_bs = (*pu4_vert_bs_top) >> ((1 << (log2_ctb_size - 1)) - 2); |
529 | 408k | u4_bs |= u4_top_bs & 3; |
530 | 408k | } |
531 | | |
532 | | /* Every alternate boundary strength value is used for chroma */ |
533 | 451k | u4_bs &= 0x22222222; |
534 | | |
535 | 1.53M | for(row = 0; row < ctb_size / 8;) |
536 | 1.08M | { |
537 | 1.08M | WORD8 i1_tc_offset_div2 = ps_slice_hdr->i1_tc_offset_div2; |
538 | | |
539 | 1.08M | bs_tz = CTZ(u4_bs) >> 2; |
540 | 1.08M | if(0 != bs_tz) |
541 | 833k | { |
542 | 833k | if((row + bs_tz) >= (ctb_size / 8)) |
543 | 420k | pu1_src += 4 * (ctb_size / 8 - row) * src_strd; |
544 | 413k | else |
545 | 413k | pu1_src += 4 * bs_tz * src_strd; |
546 | 833k | row += bs_tz; |
547 | 833k | u4_bs = u4_bs >> (bs_tz << 2); |
548 | 833k | continue; |
549 | 833k | } |
550 | | |
551 | 248k | if(0 == row) |
552 | 31.1k | { |
553 | 31.1k | i1_tc_offset_div2 = ps_slice_hdr_top->i1_tc_offset_div2; |
554 | | |
555 | 31.1k | if(0 == col) |
556 | 9.88k | { |
557 | 9.88k | qp_p = u4_qp_const_in_ctb[0] ? |
558 | 9.81k | pu1_qp[-ctb_size / 8 * qp_strd - ctb_size / 8] : |
559 | 9.88k | pu1_qp[-qp_strd - 1]; |
560 | 9.88k | } |
561 | 21.2k | else |
562 | 21.2k | { |
563 | 21.2k | qp_p = u4_qp_const_in_ctb[1] ? |
564 | 21.0k | pu1_qp[-ctb_size / 8 * qp_strd] : |
565 | 21.2k | pu1_qp[2 * col - 1 - qp_strd]; |
566 | 21.2k | } |
567 | | |
568 | 31.1k | qp_q = u4_qp_const_in_ctb[1] ? |
569 | 30.9k | pu1_qp[-ctb_size / 8 * qp_strd] : |
570 | 31.1k | pu1_qp[2 * col - qp_strd]; |
571 | 31.1k | } |
572 | 216k | else |
573 | 216k | { |
574 | 216k | if(0 == col) |
575 | 70.7k | { |
576 | 70.7k | qp_p = u4_qp_const_in_ctb[2] ? |
577 | 70.3k | pu1_qp[-ctb_size / 8] : |
578 | 70.7k | pu1_qp[(row - 1) * qp_strd - 1]; |
579 | 70.7k | } |
580 | 146k | else |
581 | 146k | { |
582 | 146k | qp_p = u4_qp_const_in_ctb[3] ? |
583 | 145k | pu1_qp[0] : |
584 | 146k | pu1_qp[(row - 1) * qp_strd + 2 * col - 1]; |
585 | 146k | } |
586 | | |
587 | 216k | qp_q = u4_qp_const_in_ctb[3] ? |
588 | 215k | pu1_qp[0] : |
589 | 216k | pu1_qp[(row - 1) * qp_strd + 2 * col]; |
590 | 216k | } |
591 | | |
592 | 248k | filter_p = (pu2_ctb_no_loop_filter_flag[row] >> (col << 1)) & 1; |
593 | 248k | filter_q = (pu2_ctb_no_loop_filter_flag[row] >> (col << 1)) & 2; |
594 | | /* filter_p and filter_q are inverted as they are calculated using no_loop_filter_flags */ |
595 | 248k | filter_p = !filter_p; |
596 | 248k | filter_q = !filter_q; |
597 | | |
598 | 248k | if(filter_p || filter_q) |
599 | 248k | { |
600 | 248k | ASSERT(1 == ((u4_bs & 3) >> 1)); |
601 | 248k | DUMP_DEBLK_CHROMA_VERT(pu1_src, src_strd, |
602 | 248k | u4_bs & 3, qp_p, qp_q, |
603 | 248k | ps_pps->i1_pic_cb_qp_offset, |
604 | 248k | ps_pps->i1_pic_cr_qp_offset, |
605 | 248k | ps_slice_hdr->i1_tc_offset_div2, |
606 | 248k | filter_p, filter_q); |
607 | 248k | if(chroma_yuv420sp_vu) |
608 | 132k | { |
609 | 132k | ps_codec->s_func_selector.ihevc_deblk_chroma_vert_fptr(pu1_src, |
610 | 132k | src_strd, |
611 | 132k | qp_q, |
612 | 132k | qp_p, |
613 | 132k | ps_pps->i1_pic_cr_qp_offset, |
614 | 132k | ps_pps->i1_pic_cb_qp_offset, |
615 | 132k | i1_tc_offset_div2, |
616 | 132k | filter_q, |
617 | 132k | filter_p); |
618 | 132k | } |
619 | 115k | else |
620 | 115k | { |
621 | 115k | ps_codec->s_func_selector.ihevc_deblk_chroma_vert_fptr(pu1_src, |
622 | 115k | src_strd, |
623 | 115k | qp_p, |
624 | 115k | qp_q, |
625 | 115k | ps_pps->i1_pic_cb_qp_offset, |
626 | 115k | ps_pps->i1_pic_cr_qp_offset, |
627 | 115k | i1_tc_offset_div2, |
628 | 115k | filter_p, |
629 | 115k | filter_q); |
630 | 115k | } |
631 | 248k | } |
632 | | |
633 | 248k | pu1_src += 4 * src_strd; |
634 | 248k | u4_bs = u4_bs >> 4; |
635 | 248k | row++; |
636 | 248k | } |
637 | | |
638 | 451k | pu4_vert_bs += (64 == ctb_size) ? 2 : 1; |
639 | 451k | pu1_src -= ((src_strd / 2) << log2_ctb_size); |
640 | 451k | pu1_src += 16; |
641 | 451k | } |
642 | 116k | } |
643 | | |
644 | | /* Chroma Horizontal Edge */ |
645 | | |
646 | 121k | if(0 == i4_is_last_ctb_y) |
647 | 120k | { |
648 | | |
649 | | /* Left CTB's slice header */ |
650 | 120k | slice_header_t *ps_slice_hdr_left; |
651 | 120k | { |
652 | 120k | WORD32 cur_ctb_indx = ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb; |
653 | 120k | if(i4_is_last_ctb_x) |
654 | 4.58k | cur_ctb_indx += 1; |
655 | 120k | ps_slice_hdr_left = ps_codec->ps_slice_hdr_base + ps_deblk->pu1_slice_idx[cur_ctb_indx - 1]; |
656 | 120k | } |
657 | | |
658 | 120k | pu1_src = ps_deblk->pu1_cur_pic_chroma + ((ps_deblk->i4_ctb_x + ps_deblk->i4_ctb_y * ps_deblk->ps_codec->i4_strd / 2) << log2_ctb_size); |
659 | 120k | pu1_src += i4_is_last_ctb_x ? ctb_size : 0; |
660 | | |
661 | | /** Deblocking is done on a shifted CTB - |
662 | | * Vertical edge processing is done by shifting the CTB up by four pixels (8 here beacuse UV are interleaved) */ |
663 | 120k | pu1_src -= 8; |
664 | 589k | for(row = 0; row < ctb_size / 16; row++) |
665 | 469k | { |
666 | | /* BS for the row - Last column is excluded and the left column is included*/ |
667 | 469k | u4_bs = pu4_horz_bs[0] << 2; |
668 | | |
669 | 469k | if(ps_deblk->i4_ctb_x || i4_is_last_ctb_x) |
670 | 442k | { |
671 | | /** Picking the last BS of the previous CTB corresponding to the same row |
672 | | * UWORD32 *pu4_horz_bs_left = (UWORD32 *)((UWORD8 *)pu4_horz_bs - (ctb_size / 8) * (ctb_size / 4) / 8 * 2); |
673 | | */ |
674 | 442k | UWORD32 *pu4_horz_bs_left = (UWORD32 *)((UWORD8 *)pu4_horz_bs - (1 << (2 * log2_ctb_size - 7))); |
675 | 442k | UWORD32 u4_left_bs = (*pu4_horz_bs_left) >> ((1 << (log2_ctb_size - 1)) - 2); |
676 | 442k | u4_bs |= u4_left_bs & 3; |
677 | 442k | } |
678 | | |
679 | | /* Every alternate boundary strength value is used for chroma */ |
680 | 469k | u4_bs &= 0x22222222; |
681 | | |
682 | 1.57M | for(col = 0; col < ctb_size / 8;) |
683 | 1.11M | { |
684 | 1.11M | WORD8 i1_tc_offset_div2 = ps_slice_hdr->i1_tc_offset_div2; |
685 | | |
686 | 1.11M | bs_tz = CTZ(u4_bs) >> 2; |
687 | 1.11M | if(0 != bs_tz) |
688 | 871k | { |
689 | 871k | u4_bs = u4_bs >> (bs_tz << 2); |
690 | | |
691 | 871k | if((col + bs_tz) >= (ctb_size / 8)) |
692 | 439k | pu1_src += 8 * (ctb_size / 8 - col); |
693 | 431k | else |
694 | 431k | pu1_src += 8 * bs_tz; |
695 | | |
696 | 871k | col += bs_tz; |
697 | 871k | continue; |
698 | 871k | } |
699 | | |
700 | 238k | if(0 == col) |
701 | 29.2k | { |
702 | 29.2k | i1_tc_offset_div2 = ps_slice_hdr_left->i1_tc_offset_div2; |
703 | | |
704 | 29.2k | if(0 == row) |
705 | 8.73k | { |
706 | 8.73k | qp_p = u4_qp_const_in_ctb[0] ? |
707 | 8.68k | pu1_qp[-ctb_size / 8 * qp_strd - ctb_size / 8] : |
708 | 8.73k | pu1_qp[-qp_strd - 1]; |
709 | 8.73k | } |
710 | 20.4k | else |
711 | 20.4k | { |
712 | 20.4k | qp_p = u4_qp_const_in_ctb[2] ? |
713 | 20.3k | pu1_qp[-ctb_size / 8] : |
714 | 20.4k | pu1_qp[(2 * row - 1) * qp_strd - 1]; |
715 | 20.4k | } |
716 | | |
717 | 29.2k | qp_q = u4_qp_const_in_ctb[2] ? |
718 | 29.0k | pu1_qp[-ctb_size / 8] : |
719 | 29.2k | pu1_qp[(2 * row) * qp_strd - 1]; |
720 | 29.2k | } |
721 | 209k | else |
722 | 209k | { |
723 | 209k | if(0 == row) |
724 | 62.8k | { |
725 | 62.8k | qp_p = u4_qp_const_in_ctb[1] ? |
726 | 62.4k | pu1_qp[-ctb_size / 8 * qp_strd] : |
727 | 62.8k | pu1_qp[col - 1 - qp_strd]; |
728 | 62.8k | } |
729 | 146k | else |
730 | 146k | { |
731 | 146k | qp_p = u4_qp_const_in_ctb[3] ? |
732 | 145k | pu1_qp[0] : |
733 | 146k | pu1_qp[(col - 1) + (2 * row - 1) * qp_strd]; |
734 | 146k | } |
735 | | |
736 | 209k | qp_q = u4_qp_const_in_ctb[3] ? |
737 | 208k | pu1_qp[0] : |
738 | 209k | pu1_qp[(col - 1) + 2 * row * qp_strd]; |
739 | 209k | } |
740 | | |
741 | 238k | filter_p = (pu2_ctb_no_loop_filter_flag[row << 1] >> col) & 1; |
742 | 238k | filter_q = (pu2_ctb_no_loop_filter_flag[(row << 1) + 1] >> col) & 1; |
743 | | /* filter_p and filter_q are inverted as they are calculated using no_loop_filter_flags */ |
744 | 238k | filter_p = !filter_p; |
745 | 238k | filter_q = !filter_q; |
746 | | |
747 | 238k | if(filter_p || filter_q) |
748 | 238k | { |
749 | 238k | ASSERT(1 == ((u4_bs & 3) >> 1)); |
750 | 238k | DUMP_DEBLK_CHROMA_HORZ(pu1_src, src_strd, |
751 | 238k | u4_bs & 3, qp_p, qp_q, |
752 | 238k | ps_pps->i1_pic_cb_qp_offset, |
753 | 238k | ps_pps->i1_pic_cr_qp_offset, |
754 | 238k | ps_slice_hdr->i1_tc_offset_div2, |
755 | 238k | filter_p, filter_q); |
756 | 238k | if(chroma_yuv420sp_vu) |
757 | 122k | { |
758 | 122k | ps_codec->s_func_selector.ihevc_deblk_chroma_horz_fptr(pu1_src, |
759 | 122k | src_strd, |
760 | 122k | qp_q, |
761 | 122k | qp_p, |
762 | 122k | ps_pps->i1_pic_cr_qp_offset, |
763 | 122k | ps_pps->i1_pic_cb_qp_offset, |
764 | 122k | i1_tc_offset_div2, |
765 | 122k | filter_q, |
766 | 122k | filter_p); |
767 | 122k | } |
768 | 116k | else |
769 | 116k | { |
770 | 116k | ps_codec->s_func_selector.ihevc_deblk_chroma_horz_fptr(pu1_src, |
771 | 116k | src_strd, |
772 | 116k | qp_p, |
773 | 116k | qp_q, |
774 | 116k | ps_pps->i1_pic_cb_qp_offset, |
775 | 116k | ps_pps->i1_pic_cr_qp_offset, |
776 | 116k | i1_tc_offset_div2, |
777 | 116k | filter_p, |
778 | 116k | filter_q); |
779 | 116k | } |
780 | 238k | } |
781 | | |
782 | 238k | pu1_src += 8; |
783 | 238k | u4_bs = u4_bs >> 4; |
784 | 238k | col++; |
785 | 238k | } |
786 | | |
787 | 469k | pu4_horz_bs += (64 == ctb_size) ? 2 : 1; |
788 | 469k | pu1_src -= ctb_size; |
789 | 469k | pu1_src += 8 * src_strd; |
790 | | |
791 | 469k | } |
792 | 120k | } |
793 | 121k | } |