xref: /aosp_15_r20/external/libhevc/encoder/ihevce_enc_subpel_gen.c (revision c83a76b084498d55f252f48b2e3786804cdf24b7)
1*c83a76b0SSuyog Pawar /******************************************************************************
2*c83a76b0SSuyog Pawar  *
3*c83a76b0SSuyog Pawar  * Copyright (C) 2018 The Android Open Source Project
4*c83a76b0SSuyog Pawar  *
5*c83a76b0SSuyog Pawar  * Licensed under the Apache License, Version 2.0 (the "License");
6*c83a76b0SSuyog Pawar  * you may not use this file except in compliance with the License.
7*c83a76b0SSuyog Pawar  * You may obtain a copy of the License at:
8*c83a76b0SSuyog Pawar  *
9*c83a76b0SSuyog Pawar  * http://www.apache.org/licenses/LICENSE-2.0
10*c83a76b0SSuyog Pawar  *
11*c83a76b0SSuyog Pawar  * Unless required by applicable law or agreed to in writing, software
12*c83a76b0SSuyog Pawar  * distributed under the License is distributed on an "AS IS" BASIS,
13*c83a76b0SSuyog Pawar  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14*c83a76b0SSuyog Pawar  * See the License for the specific language governing permissions and
15*c83a76b0SSuyog Pawar  * limitations under the License.
16*c83a76b0SSuyog Pawar  *
17*c83a76b0SSuyog Pawar  *****************************************************************************
18*c83a76b0SSuyog Pawar  * Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore
19*c83a76b0SSuyog Pawar */
20*c83a76b0SSuyog Pawar 
21*c83a76b0SSuyog Pawar /*!
22*c83a76b0SSuyog Pawar ******************************************************************************
23*c83a76b0SSuyog Pawar * \file ihevce_enc_subpel_gen.c
24*c83a76b0SSuyog Pawar *
25*c83a76b0SSuyog Pawar * \brief
26*c83a76b0SSuyog Pawar *    This file contains Padding and Subpel plane generation functions
27*c83a76b0SSuyog Pawar *    at CTB level
28*c83a76b0SSuyog Pawar *
29*c83a76b0SSuyog Pawar * \date
30*c83a76b0SSuyog Pawar *    29/12/2012
31*c83a76b0SSuyog Pawar *
32*c83a76b0SSuyog Pawar * \author
33*c83a76b0SSuyog Pawar *    Ittiam
34*c83a76b0SSuyog Pawar *
35*c83a76b0SSuyog Pawar *
36*c83a76b0SSuyog Pawar * List of Functions
37*c83a76b0SSuyog Pawar * - ihevce_suppel_padding()
38*c83a76b0SSuyog Pawar * - ihevce_pad_interp_recon_ctb()
39*c83a76b0SSuyog Pawar *
40*c83a76b0SSuyog Pawar *
41*c83a76b0SSuyog Pawar ******************************************************************************
42*c83a76b0SSuyog Pawar */
43*c83a76b0SSuyog Pawar 
44*c83a76b0SSuyog Pawar /*****************************************************************************/
45*c83a76b0SSuyog Pawar /* File Includes                                                             */
46*c83a76b0SSuyog Pawar /*****************************************************************************/
47*c83a76b0SSuyog Pawar /* System include files */
48*c83a76b0SSuyog Pawar #include <stdio.h>
49*c83a76b0SSuyog Pawar #include <string.h>
50*c83a76b0SSuyog Pawar #include <stdlib.h>
51*c83a76b0SSuyog Pawar #include <assert.h>
52*c83a76b0SSuyog Pawar #include <stdarg.h>
53*c83a76b0SSuyog Pawar #include <math.h>
54*c83a76b0SSuyog Pawar 
55*c83a76b0SSuyog Pawar /* User include files */
56*c83a76b0SSuyog Pawar #include "ihevc_typedefs.h"
57*c83a76b0SSuyog Pawar #include "itt_video_api.h"
58*c83a76b0SSuyog Pawar #include "ihevce_api.h"
59*c83a76b0SSuyog Pawar 
60*c83a76b0SSuyog Pawar #include "rc_cntrl_param.h"
61*c83a76b0SSuyog Pawar #include "rc_frame_info_collector.h"
62*c83a76b0SSuyog Pawar #include "rc_look_ahead_params.h"
63*c83a76b0SSuyog Pawar 
64*c83a76b0SSuyog Pawar #include "ihevc_defs.h"
65*c83a76b0SSuyog Pawar #include "ihevc_debug.h"
66*c83a76b0SSuyog Pawar #include "ihevc_macros.h"
67*c83a76b0SSuyog Pawar #include "ihevc_structs.h"
68*c83a76b0SSuyog Pawar #include "ihevc_platform_macros.h"
69*c83a76b0SSuyog Pawar #include "ihevc_deblk.h"
70*c83a76b0SSuyog Pawar #include "ihevc_itrans_recon.h"
71*c83a76b0SSuyog Pawar #include "ihevc_chroma_itrans_recon.h"
72*c83a76b0SSuyog Pawar #include "ihevc_chroma_intra_pred.h"
73*c83a76b0SSuyog Pawar #include "ihevc_intra_pred.h"
74*c83a76b0SSuyog Pawar #include "ihevc_inter_pred.h"
75*c83a76b0SSuyog Pawar #include "ihevc_mem_fns.h"
76*c83a76b0SSuyog Pawar #include "ihevc_padding.h"
77*c83a76b0SSuyog Pawar #include "ihevc_weighted_pred.h"
78*c83a76b0SSuyog Pawar #include "ihevc_sao.h"
79*c83a76b0SSuyog Pawar #include "ihevc_resi_trans.h"
80*c83a76b0SSuyog Pawar #include "ihevc_quant_iquant_ssd.h"
81*c83a76b0SSuyog Pawar #include "ihevc_cabac_tables.h"
82*c83a76b0SSuyog Pawar #include "ihevc_trans_tables.h"
83*c83a76b0SSuyog Pawar #include "ihevc_trans_macros.h"
84*c83a76b0SSuyog Pawar 
85*c83a76b0SSuyog Pawar #include "ihevce_defs.h"
86*c83a76b0SSuyog Pawar #include "ihevce_lap_enc_structs.h"
87*c83a76b0SSuyog Pawar #include "ihevce_multi_thrd_structs.h"
88*c83a76b0SSuyog Pawar #include "ihevce_multi_thrd_funcs.h"
89*c83a76b0SSuyog Pawar #include "ihevce_me_common_defs.h"
90*c83a76b0SSuyog Pawar #include "ihevce_had_satd.h"
91*c83a76b0SSuyog Pawar #include "ihevce_error_codes.h"
92*c83a76b0SSuyog Pawar #include "ihevce_bitstream.h"
93*c83a76b0SSuyog Pawar #include "ihevce_cabac.h"
94*c83a76b0SSuyog Pawar #include "ihevce_rdoq_macros.h"
95*c83a76b0SSuyog Pawar #include "ihevce_function_selector.h"
96*c83a76b0SSuyog Pawar #include "ihevce_enc_structs.h"
97*c83a76b0SSuyog Pawar #include "ihevce_global_tables.h"
98*c83a76b0SSuyog Pawar #include "ihevce_cmn_utils_instr_set_router.h"
99*c83a76b0SSuyog Pawar #include "ihevce_entropy_structs.h"
100*c83a76b0SSuyog Pawar #include "ihevce_enc_loop_structs.h"
101*c83a76b0SSuyog Pawar #include "ihevce_enc_loop_utils.h"
102*c83a76b0SSuyog Pawar #include "ihevce_inter_pred.h"
103*c83a76b0SSuyog Pawar #include "ihevce_common_utils.h"
104*c83a76b0SSuyog Pawar 
105*c83a76b0SSuyog Pawar /*!
106*c83a76b0SSuyog Pawar ******************************************************************************
107*c83a76b0SSuyog Pawar * \if Function name : ihevce_suppel_padding \endif
108*c83a76b0SSuyog Pawar *
109*c83a76b0SSuyog Pawar * \brief
110*c83a76b0SSuyog Pawar *    Subpel Plane planes Padding Function
111*c83a76b0SSuyog Pawar *
112*c83a76b0SSuyog Pawar * \param[in] pu1_dst : pointer to subpel plane
113*c83a76b0SSuyog Pawar *            stride  : subpel plane stride same as recon stride
114*c83a76b0SSuyog Pawar *            tot_wd  : width of the block in subpel plane
115*c83a76b0SSuyog Pawar *            tot_ht  : hieght of the block in subpel plane
116*c83a76b0SSuyog Pawar *            ctb_ctr   : ctb horizontal position
117*c83a76b0SSuyog Pawar *            vert_ctr  : ctb vertical position
118*c83a76b0SSuyog Pawar *            ps_frm_ctb_prms : CTB characteristics parameters
119*c83a76b0SSuyog Pawar * \return
120*c83a76b0SSuyog Pawar *    None
121*c83a76b0SSuyog Pawar *
122*c83a76b0SSuyog Pawar *
123*c83a76b0SSuyog Pawar * \author
124*c83a76b0SSuyog Pawar *  Ittiam
125*c83a76b0SSuyog Pawar *
126*c83a76b0SSuyog Pawar *****************************************************************************
127*c83a76b0SSuyog Pawar */
ihevce_subpel_padding(UWORD8 * pu1_dst,WORD32 stride,WORD32 tot_wd,WORD32 tot_ht,WORD32 pad_subpel_x,WORD32 pad_subpel_y,WORD32 ctb_ctr,WORD32 vert_ctr,WORD32 i4_num_ctbs_horz,WORD32 i4_num_ctbs_vert,func_selector_t * ps_func_selector)128*c83a76b0SSuyog Pawar void ihevce_subpel_padding(
129*c83a76b0SSuyog Pawar     UWORD8 *pu1_dst,
130*c83a76b0SSuyog Pawar     WORD32 stride,
131*c83a76b0SSuyog Pawar     WORD32 tot_wd,
132*c83a76b0SSuyog Pawar     WORD32 tot_ht,
133*c83a76b0SSuyog Pawar     WORD32 pad_subpel_x,
134*c83a76b0SSuyog Pawar     WORD32 pad_subpel_y,
135*c83a76b0SSuyog Pawar     WORD32 ctb_ctr,
136*c83a76b0SSuyog Pawar     WORD32 vert_ctr,
137*c83a76b0SSuyog Pawar     WORD32 i4_num_ctbs_horz,
138*c83a76b0SSuyog Pawar     WORD32 i4_num_ctbs_vert,
139*c83a76b0SSuyog Pawar     func_selector_t *ps_func_selector)
140*c83a76b0SSuyog Pawar {
141*c83a76b0SSuyog Pawar     ihevc_pad_top_ft *pf_pad_top = ps_func_selector->ihevc_pad_top_fptr;
142*c83a76b0SSuyog Pawar     ihevc_pad_bottom_ft *pf_pad_bottom = ps_func_selector->ihevc_pad_bottom_fptr;
143*c83a76b0SSuyog Pawar     ihevc_pad_left_luma_ft *pf_pad_left_luma = ps_func_selector->ihevc_pad_left_luma_fptr;
144*c83a76b0SSuyog Pawar     ihevc_pad_right_luma_ft *pf_pad_right_luma = ps_func_selector->ihevc_pad_right_luma_fptr;
145*c83a76b0SSuyog Pawar 
146*c83a76b0SSuyog Pawar     UWORD8 *pu1_dst_tmp = pu1_dst;
147*c83a76b0SSuyog Pawar     WORD32 cpy_ht = tot_ht;
148*c83a76b0SSuyog Pawar 
149*c83a76b0SSuyog Pawar     /* Top padding*/
150*c83a76b0SSuyog Pawar     if(vert_ctr == 0)
151*c83a76b0SSuyog Pawar     {
152*c83a76b0SSuyog Pawar         PAD_BUF_VER(pu1_dst, stride, tot_wd, pad_subpel_x, pad_subpel_y, pf_pad_top);
153*c83a76b0SSuyog Pawar         /*if curr ctb is 1st ctb in ctb row, update dst pointer for Left padding*/
154*c83a76b0SSuyog Pawar         pu1_dst_tmp = pu1_dst - pad_subpel_y * stride;
155*c83a76b0SSuyog Pawar         cpy_ht += pad_subpel_y;
156*c83a76b0SSuyog Pawar     }
157*c83a76b0SSuyog Pawar     /*bottom padding*/
158*c83a76b0SSuyog Pawar     if(vert_ctr == (i4_num_ctbs_vert - 1))
159*c83a76b0SSuyog Pawar     {
160*c83a76b0SSuyog Pawar         PAD_BUF_VER(
161*c83a76b0SSuyog Pawar             (pu1_dst + (tot_ht * stride)),
162*c83a76b0SSuyog Pawar             stride,
163*c83a76b0SSuyog Pawar             tot_wd,
164*c83a76b0SSuyog Pawar             pad_subpel_x,
165*c83a76b0SSuyog Pawar             pad_subpel_y,
166*c83a76b0SSuyog Pawar             pf_pad_bottom);
167*c83a76b0SSuyog Pawar         /*if curr ctb is 1st ctb in ctb row, update dst pointer for right padding*/
168*c83a76b0SSuyog Pawar         cpy_ht += pad_subpel_y;
169*c83a76b0SSuyog Pawar     }
170*c83a76b0SSuyog Pawar 
171*c83a76b0SSuyog Pawar     /*left padding*/
172*c83a76b0SSuyog Pawar     if(ctb_ctr == 0)
173*c83a76b0SSuyog Pawar     {
174*c83a76b0SSuyog Pawar         PAD_BUF_HOR(pu1_dst_tmp, stride, cpy_ht, pad_subpel_x, pad_subpel_y, pf_pad_left_luma);
175*c83a76b0SSuyog Pawar     }
176*c83a76b0SSuyog Pawar 
177*c83a76b0SSuyog Pawar     /*right padding*/
178*c83a76b0SSuyog Pawar     if(ctb_ctr == (i4_num_ctbs_horz - 1))
179*c83a76b0SSuyog Pawar     {
180*c83a76b0SSuyog Pawar         PAD_BUF_HOR(
181*c83a76b0SSuyog Pawar             pu1_dst_tmp + tot_wd, stride, cpy_ht, pad_subpel_x, pad_subpel_y, pf_pad_right_luma);
182*c83a76b0SSuyog Pawar     }
183*c83a76b0SSuyog Pawar }
184*c83a76b0SSuyog Pawar 
185*c83a76b0SSuyog Pawar /*!
186*c83a76b0SSuyog Pawar ******************************************************************************
187*c83a76b0SSuyog Pawar * \if Function name : ihevce_pad_interp_recon_ctb \endif
188*c83a76b0SSuyog Pawar *
189*c83a76b0SSuyog Pawar * \brief
190*c83a76b0SSuyog Pawar *    Ctb level Subpel Plane generation and padding function
191*c83a76b0SSuyog Pawar *
192*c83a76b0SSuyog Pawar * \param[in]
193*c83a76b0SSuyog Pawar * s_cu_prms : coding unit params structures (recon buffers)
194*c83a76b0SSuyog Pawar *            ctb_ctr   : ctb horizontal position
195*c83a76b0SSuyog Pawar *            vert_ctr  : ctb vertical position
196*c83a76b0SSuyog Pawar *            ps_frm_ctb_prms : CTB characteristics parameters
197*c83a76b0SSuyog Pawar * i4_dist_nbr_mask : nbr-mask for distributed mode. Should be 0 for standalone
198*c83a76b0SSuyog Pawar *                    or distributed-single-client mode
199*c83a76b0SSuyog Pawar *
200*c83a76b0SSuyog Pawar * \return
201*c83a76b0SSuyog Pawar *    None
202*c83a76b0SSuyog Pawar *
203*c83a76b0SSuyog Pawar * \author
204*c83a76b0SSuyog Pawar *  Ittiam
205*c83a76b0SSuyog Pawar *
206*c83a76b0SSuyog Pawar *****************************************************************************
207*c83a76b0SSuyog Pawar */
ihevce_pad_interp_recon_ctb(pad_interp_recon_frm_t * ps_pad_interp_recon,WORD32 ctb_ctr,WORD32 vert_ctr,WORD32 quality_preset,frm_ctb_ctxt_t * ps_frm_ctb_prms,WORD16 * pi2_hxhy_interm,WORD32 i4_bitrate_instance_id,func_selector_t * ps_func_selector)208*c83a76b0SSuyog Pawar void ihevce_pad_interp_recon_ctb(
209*c83a76b0SSuyog Pawar     pad_interp_recon_frm_t *ps_pad_interp_recon,
210*c83a76b0SSuyog Pawar     WORD32 ctb_ctr,
211*c83a76b0SSuyog Pawar     WORD32 vert_ctr,
212*c83a76b0SSuyog Pawar     WORD32 quality_preset,
213*c83a76b0SSuyog Pawar     frm_ctb_ctxt_t *ps_frm_ctb_prms,
214*c83a76b0SSuyog Pawar     WORD16 *pi2_hxhy_interm,
215*c83a76b0SSuyog Pawar     WORD32 i4_bitrate_instance_id,
216*c83a76b0SSuyog Pawar     func_selector_t *ps_func_selector)
217*c83a76b0SSuyog Pawar {
218*c83a76b0SSuyog Pawar     UWORD8 *pu1_src, *pu1_src_uv;
219*c83a76b0SSuyog Pawar     WORD32 stride, stride_uv, wd, ht, wd_uv, ht_uv, pad_x, pad_y, pad_subpel_x, pad_subpel_y;
220*c83a76b0SSuyog Pawar     WORD32 tot_wd, tot_ht, offset, cpy_ht_y, cpy_ht_uv;
221*c83a76b0SSuyog Pawar     WORD32 i4_chroma_vert_pad_default;
222*c83a76b0SSuyog Pawar 
223*c83a76b0SSuyog Pawar     WORD32 ctb_size = ps_frm_ctb_prms->i4_ctb_size;
224*c83a76b0SSuyog Pawar     UWORD8 *pu1_dst_hxfy = ps_pad_interp_recon->pu1_sbpel_hxfy +
225*c83a76b0SSuyog Pawar                            (vert_ctr * ctb_size * ps_pad_interp_recon->i4_luma_recon_stride) +
226*c83a76b0SSuyog Pawar                            (ctb_ctr * ctb_size);
227*c83a76b0SSuyog Pawar     UWORD8 *pu1_dst_fxhy = ps_pad_interp_recon->pu1_sbpel_fxhy +
228*c83a76b0SSuyog Pawar                            (vert_ctr * ctb_size * ps_pad_interp_recon->i4_luma_recon_stride) +
229*c83a76b0SSuyog Pawar                            (ctb_ctr * ctb_size);
230*c83a76b0SSuyog Pawar     UWORD8 *pu1_dst_hxhy = ps_pad_interp_recon->pu1_sbpel_hxhy +
231*c83a76b0SSuyog Pawar                            (vert_ctr * ctb_size * ps_pad_interp_recon->i4_luma_recon_stride) +
232*c83a76b0SSuyog Pawar                            (ctb_ctr * ctb_size);
233*c83a76b0SSuyog Pawar     UWORD8 u1_is_422 = (ps_pad_interp_recon->u1_chroma_array_type == 2);
234*c83a76b0SSuyog Pawar 
235*c83a76b0SSuyog Pawar     ihevc_inter_pred_ft *pf_inter_pred_luma_horz =
236*c83a76b0SSuyog Pawar         ps_func_selector->ihevc_inter_pred_luma_horz_fptr;
237*c83a76b0SSuyog Pawar     ihevc_inter_pred_ft *pf_inter_pred_luma_vert =
238*c83a76b0SSuyog Pawar         ps_func_selector->ihevc_inter_pred_luma_vert_fptr;
239*c83a76b0SSuyog Pawar     ihevc_inter_pred_w16out_ft *pf_inter_pred_luma_horz_w16out =
240*c83a76b0SSuyog Pawar         ps_func_selector->ihevc_inter_pred_luma_horz_w16out_fptr;
241*c83a76b0SSuyog Pawar     ihevc_inter_pred_w16inp_ft *pf_inter_pred_luma_vert_w16inp =
242*c83a76b0SSuyog Pawar         ps_func_selector->ihevc_inter_pred_luma_vert_w16inp_fptr;
243*c83a76b0SSuyog Pawar     stride = ps_pad_interp_recon->i4_luma_recon_stride;
244*c83a76b0SSuyog Pawar     wd = ps_pad_interp_recon->i4_ctb_size;
245*c83a76b0SSuyog Pawar     ht = ps_pad_interp_recon->i4_ctb_size;
246*c83a76b0SSuyog Pawar 
247*c83a76b0SSuyog Pawar     pu1_src = (UWORD8 *)ps_pad_interp_recon->pu1_luma_recon + (vert_ctr * ctb_size * stride) +
248*c83a76b0SSuyog Pawar               (ctb_ctr * ctb_size);
249*c83a76b0SSuyog Pawar 
250*c83a76b0SSuyog Pawar     stride_uv = ps_pad_interp_recon->i4_chrm_recon_stride;
251*c83a76b0SSuyog Pawar     wd_uv = ps_pad_interp_recon->i4_ctb_size;
252*c83a76b0SSuyog Pawar     ht_uv = ps_pad_interp_recon->i4_ctb_size >> (0 == u1_is_422);
253*c83a76b0SSuyog Pawar 
254*c83a76b0SSuyog Pawar     pu1_src_uv = (UWORD8 *)ps_pad_interp_recon->pu1_chrm_recon +
255*c83a76b0SSuyog Pawar                  (vert_ctr * (ctb_size >> (0 == u1_is_422)) * stride_uv) + (ctb_ctr * ctb_size);
256*c83a76b0SSuyog Pawar 
257*c83a76b0SSuyog Pawar     pad_x = ALIGN8(NTAPS_LUMA);
258*c83a76b0SSuyog Pawar     pad_y = ALIGN8(NTAPS_LUMA);
259*c83a76b0SSuyog Pawar     pad_subpel_x = PAD_HORZ - pad_x;
260*c83a76b0SSuyog Pawar     pad_subpel_y = PAD_VERT - pad_y;
261*c83a76b0SSuyog Pawar 
262*c83a76b0SSuyog Pawar     offset = pad_x + (pad_y * stride);
263*c83a76b0SSuyog Pawar 
264*c83a76b0SSuyog Pawar     tot_wd = wd + (pad_x << 1);
265*c83a76b0SSuyog Pawar     tot_ht = ht + (pad_y << 1);
266*c83a76b0SSuyog Pawar 
267*c83a76b0SSuyog Pawar     i4_chroma_vert_pad_default = PAD_VERT >> (0 == u1_is_422);
268*c83a76b0SSuyog Pawar 
269*c83a76b0SSuyog Pawar     if(ctb_ctr == (ps_frm_ctb_prms->i4_num_ctbs_horz - 1))
270*c83a76b0SSuyog Pawar     {
271*c83a76b0SSuyog Pawar         WORD32 last_ctb_x =
272*c83a76b0SSuyog Pawar             ps_frm_ctb_prms->i4_cu_aligned_pic_wd -
273*c83a76b0SSuyog Pawar             ((ps_frm_ctb_prms->i4_num_ctbs_horz - 1) * ps_pad_interp_recon->i4_ctb_size);
274*c83a76b0SSuyog Pawar         wd = last_ctb_x;
275*c83a76b0SSuyog Pawar         wd_uv = last_ctb_x;
276*c83a76b0SSuyog Pawar     }
277*c83a76b0SSuyog Pawar     if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
278*c83a76b0SSuyog Pawar     {
279*c83a76b0SSuyog Pawar         WORD32 last_ctb_y =
280*c83a76b0SSuyog Pawar             ps_frm_ctb_prms->i4_cu_aligned_pic_ht -
281*c83a76b0SSuyog Pawar             ((ps_frm_ctb_prms->i4_num_ctbs_vert - 1) * ps_pad_interp_recon->i4_ctb_size);
282*c83a76b0SSuyog Pawar         ht = last_ctb_y;
283*c83a76b0SSuyog Pawar         ht_uv = last_ctb_y >> (0 == u1_is_422);
284*c83a76b0SSuyog Pawar     }
285*c83a76b0SSuyog Pawar     tot_ht = ht;
286*c83a76b0SSuyog Pawar     tot_wd = wd;
287*c83a76b0SSuyog Pawar 
288*c83a76b0SSuyog Pawar     /*top padding*/
289*c83a76b0SSuyog Pawar     if(vert_ctr == 0)
290*c83a76b0SSuyog Pawar     {
291*c83a76b0SSuyog Pawar         tot_ht = pad_y + ht - 8;
292*c83a76b0SSuyog Pawar     }
293*c83a76b0SSuyog Pawar     /*bottom padding*/
294*c83a76b0SSuyog Pawar     if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
295*c83a76b0SSuyog Pawar     {
296*c83a76b0SSuyog Pawar         tot_ht = pad_y + ht + 8;
297*c83a76b0SSuyog Pawar     }
298*c83a76b0SSuyog Pawar 
299*c83a76b0SSuyog Pawar     /*Left padding*/
300*c83a76b0SSuyog Pawar     if(ctb_ctr == 0)
301*c83a76b0SSuyog Pawar     {
302*c83a76b0SSuyog Pawar         tot_wd = pad_x + wd - 8;
303*c83a76b0SSuyog Pawar     }
304*c83a76b0SSuyog Pawar     /*right padding*/
305*c83a76b0SSuyog Pawar     if(ctb_ctr == (ps_frm_ctb_prms->i4_num_ctbs_horz - 1))
306*c83a76b0SSuyog Pawar     {
307*c83a76b0SSuyog Pawar         tot_wd = pad_x + wd + 8;
308*c83a76b0SSuyog Pawar     }
309*c83a76b0SSuyog Pawar 
310*c83a76b0SSuyog Pawar     pu1_src -= offset;
311*c83a76b0SSuyog Pawar     pu1_dst_hxhy -= offset;
312*c83a76b0SSuyog Pawar     pu1_dst_hxfy -= offset;
313*c83a76b0SSuyog Pawar     pu1_dst_fxhy -= offset;
314*c83a76b0SSuyog Pawar 
315*c83a76b0SSuyog Pawar     {
316*c83a76b0SSuyog Pawar         tot_wd = ALIGN16(tot_wd);
317*c83a76b0SSuyog Pawar         if(0 ==
318*c83a76b0SSuyog Pawar            i4_bitrate_instance_id)  //do the following subpel calculations for reference bit-rate instance only
319*c83a76b0SSuyog Pawar         {
320*c83a76b0SSuyog Pawar             /* HxFY plane */
321*c83a76b0SSuyog Pawar             pf_inter_pred_luma_horz(
322*c83a76b0SSuyog Pawar                 pu1_src,
323*c83a76b0SSuyog Pawar                 pu1_dst_hxfy,
324*c83a76b0SSuyog Pawar                 stride,
325*c83a76b0SSuyog Pawar                 stride,
326*c83a76b0SSuyog Pawar                 (WORD8 *)gai1_hevc_luma_filter_taps[2],
327*c83a76b0SSuyog Pawar                 tot_ht,
328*c83a76b0SSuyog Pawar                 tot_wd);
329*c83a76b0SSuyog Pawar 
330*c83a76b0SSuyog Pawar             pf_inter_pred_luma_vert(
331*c83a76b0SSuyog Pawar                 pu1_src,
332*c83a76b0SSuyog Pawar                 pu1_dst_fxhy,
333*c83a76b0SSuyog Pawar                 stride,
334*c83a76b0SSuyog Pawar                 stride,
335*c83a76b0SSuyog Pawar                 (WORD8 *)gai1_hevc_luma_filter_taps[2],
336*c83a76b0SSuyog Pawar                 tot_ht,
337*c83a76b0SSuyog Pawar                 tot_wd);
338*c83a76b0SSuyog Pawar 
339*c83a76b0SSuyog Pawar             pf_inter_pred_luma_horz_w16out(
340*c83a76b0SSuyog Pawar                 pu1_src - 3 * stride,
341*c83a76b0SSuyog Pawar                 pi2_hxhy_interm,
342*c83a76b0SSuyog Pawar                 stride,
343*c83a76b0SSuyog Pawar                 tot_wd,
344*c83a76b0SSuyog Pawar                 (WORD8 *)gai1_hevc_luma_filter_taps[2],
345*c83a76b0SSuyog Pawar                 (tot_ht + 7),
346*c83a76b0SSuyog Pawar                 tot_wd);
347*c83a76b0SSuyog Pawar 
348*c83a76b0SSuyog Pawar             /* "Stride" of intermediate buffer in pixels,equals tot_wd */
349*c83a76b0SSuyog Pawar             pf_inter_pred_luma_vert_w16inp(
350*c83a76b0SSuyog Pawar                 pi2_hxhy_interm + (3 * tot_wd),
351*c83a76b0SSuyog Pawar                 pu1_dst_hxhy,
352*c83a76b0SSuyog Pawar                 tot_wd,
353*c83a76b0SSuyog Pawar                 stride,
354*c83a76b0SSuyog Pawar                 (WORD8 *)gai1_hevc_luma_filter_taps[2],
355*c83a76b0SSuyog Pawar                 tot_ht,
356*c83a76b0SSuyog Pawar                 tot_wd);
357*c83a76b0SSuyog Pawar 
358*c83a76b0SSuyog Pawar             ihevce_subpel_padding(
359*c83a76b0SSuyog Pawar                 pu1_dst_fxhy,
360*c83a76b0SSuyog Pawar                 stride,
361*c83a76b0SSuyog Pawar                 tot_wd,
362*c83a76b0SSuyog Pawar                 tot_ht,
363*c83a76b0SSuyog Pawar                 pad_subpel_x,
364*c83a76b0SSuyog Pawar                 pad_subpel_y,
365*c83a76b0SSuyog Pawar                 ctb_ctr,
366*c83a76b0SSuyog Pawar                 vert_ctr,
367*c83a76b0SSuyog Pawar                 ps_frm_ctb_prms->i4_num_ctbs_horz,
368*c83a76b0SSuyog Pawar                 ps_frm_ctb_prms->i4_num_ctbs_vert,
369*c83a76b0SSuyog Pawar                 ps_func_selector);
370*c83a76b0SSuyog Pawar 
371*c83a76b0SSuyog Pawar             ihevce_subpel_padding(
372*c83a76b0SSuyog Pawar                 pu1_dst_hxfy,
373*c83a76b0SSuyog Pawar                 stride,
374*c83a76b0SSuyog Pawar                 tot_wd,
375*c83a76b0SSuyog Pawar                 tot_ht,
376*c83a76b0SSuyog Pawar                 pad_subpel_x,
377*c83a76b0SSuyog Pawar                 pad_subpel_y,
378*c83a76b0SSuyog Pawar                 ctb_ctr,
379*c83a76b0SSuyog Pawar                 vert_ctr,
380*c83a76b0SSuyog Pawar                 ps_frm_ctb_prms->i4_num_ctbs_horz,
381*c83a76b0SSuyog Pawar                 ps_frm_ctb_prms->i4_num_ctbs_vert,
382*c83a76b0SSuyog Pawar                 ps_func_selector);
383*c83a76b0SSuyog Pawar 
384*c83a76b0SSuyog Pawar             ihevce_subpel_padding(
385*c83a76b0SSuyog Pawar                 pu1_dst_hxhy,
386*c83a76b0SSuyog Pawar                 stride,
387*c83a76b0SSuyog Pawar                 tot_wd,
388*c83a76b0SSuyog Pawar                 tot_ht,
389*c83a76b0SSuyog Pawar                 pad_subpel_x,
390*c83a76b0SSuyog Pawar                 pad_subpel_y,
391*c83a76b0SSuyog Pawar                 ctb_ctr,
392*c83a76b0SSuyog Pawar                 vert_ctr,
393*c83a76b0SSuyog Pawar                 ps_frm_ctb_prms->i4_num_ctbs_horz,
394*c83a76b0SSuyog Pawar                 ps_frm_ctb_prms->i4_num_ctbs_vert,
395*c83a76b0SSuyog Pawar                 ps_func_selector);
396*c83a76b0SSuyog Pawar         }
397*c83a76b0SSuyog Pawar     }
398*c83a76b0SSuyog Pawar }
399*c83a76b0SSuyog Pawar 
ihevce_recon_padding(pad_interp_recon_frm_t * ps_pad_interp_recon,WORD32 ctb_ctr,WORD32 vert_ctr,frm_ctb_ctxt_t * ps_frm_ctb_prms,func_selector_t * ps_func_selector)400*c83a76b0SSuyog Pawar void ihevce_recon_padding(
401*c83a76b0SSuyog Pawar     pad_interp_recon_frm_t *ps_pad_interp_recon,
402*c83a76b0SSuyog Pawar     WORD32 ctb_ctr,
403*c83a76b0SSuyog Pawar     WORD32 vert_ctr,
404*c83a76b0SSuyog Pawar     frm_ctb_ctxt_t *ps_frm_ctb_prms,
405*c83a76b0SSuyog Pawar     func_selector_t *ps_func_selector)
406*c83a76b0SSuyog Pawar {
407*c83a76b0SSuyog Pawar     UWORD8 *pu1_src, *pu1_src_uv, *pu1_buf_y, *pu1_buf_uv;
408*c83a76b0SSuyog Pawar     WORD32 stride, stride_uv, wd, ht, wd_uv, ht_uv;
409*c83a76b0SSuyog Pawar     WORD32 cpy_ht_y, cpy_ht_uv;
410*c83a76b0SSuyog Pawar     WORD32 i4_chroma_vert_pad_default;
411*c83a76b0SSuyog Pawar 
412*c83a76b0SSuyog Pawar     WORD32 top_extra_pix = 0, left_extra_pix = 0;
413*c83a76b0SSuyog Pawar     WORD32 ctb_size = ps_frm_ctb_prms->i4_ctb_size;
414*c83a76b0SSuyog Pawar     UWORD8 u1_is_422 = (ps_pad_interp_recon->u1_chroma_array_type == 2);
415*c83a76b0SSuyog Pawar 
416*c83a76b0SSuyog Pawar     ihevc_pad_top_ft *pf_pad_top = ps_func_selector->ihevc_pad_top_fptr;
417*c83a76b0SSuyog Pawar     ihevc_pad_bottom_ft *pf_pad_bottom = ps_func_selector->ihevc_pad_bottom_fptr;
418*c83a76b0SSuyog Pawar     ihevc_pad_left_luma_ft *pf_pad_left_luma = ps_func_selector->ihevc_pad_left_luma_fptr;
419*c83a76b0SSuyog Pawar     ihevc_pad_left_chroma_ft *pf_pad_left_chroma = ps_func_selector->ihevc_pad_left_chroma_fptr;
420*c83a76b0SSuyog Pawar     ihevc_pad_right_luma_ft *pf_pad_right_luma = ps_func_selector->ihevc_pad_right_luma_fptr;
421*c83a76b0SSuyog Pawar     ihevc_pad_right_chroma_ft *pf_pad_right_chroma = ps_func_selector->ihevc_pad_right_chroma_fptr;
422*c83a76b0SSuyog Pawar 
423*c83a76b0SSuyog Pawar     stride = ps_pad_interp_recon->i4_luma_recon_stride;
424*c83a76b0SSuyog Pawar     wd = ps_pad_interp_recon->i4_ctb_size;
425*c83a76b0SSuyog Pawar     ht = ps_pad_interp_recon->i4_ctb_size;
426*c83a76b0SSuyog Pawar 
427*c83a76b0SSuyog Pawar     pu1_src = (UWORD8 *)ps_pad_interp_recon->pu1_luma_recon + (vert_ctr * ctb_size * stride) +
428*c83a76b0SSuyog Pawar               (ctb_ctr * ctb_size);
429*c83a76b0SSuyog Pawar 
430*c83a76b0SSuyog Pawar     stride_uv = ps_pad_interp_recon->i4_chrm_recon_stride;
431*c83a76b0SSuyog Pawar     wd_uv = ps_pad_interp_recon->i4_ctb_size;
432*c83a76b0SSuyog Pawar     ht_uv = ps_pad_interp_recon->i4_ctb_size >> (0 == u1_is_422);
433*c83a76b0SSuyog Pawar 
434*c83a76b0SSuyog Pawar     pu1_src_uv = (UWORD8 *)ps_pad_interp_recon->pu1_chrm_recon +
435*c83a76b0SSuyog Pawar                  (vert_ctr * (ctb_size >> (0 == u1_is_422)) * stride_uv) + (ctb_ctr * ctb_size);
436*c83a76b0SSuyog Pawar 
437*c83a76b0SSuyog Pawar     i4_chroma_vert_pad_default = PAD_VERT >> (0 == u1_is_422);
438*c83a76b0SSuyog Pawar 
439*c83a76b0SSuyog Pawar     if(ctb_ctr == (ps_frm_ctb_prms->i4_num_ctbs_horz - 1))
440*c83a76b0SSuyog Pawar     {
441*c83a76b0SSuyog Pawar         WORD32 last_ctb_x =
442*c83a76b0SSuyog Pawar             ps_frm_ctb_prms->i4_cu_aligned_pic_wd -
443*c83a76b0SSuyog Pawar             ((ps_frm_ctb_prms->i4_num_ctbs_horz - 1) * ps_pad_interp_recon->i4_ctb_size);
444*c83a76b0SSuyog Pawar         wd = last_ctb_x;
445*c83a76b0SSuyog Pawar         wd_uv = last_ctb_x;
446*c83a76b0SSuyog Pawar     }
447*c83a76b0SSuyog Pawar     if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
448*c83a76b0SSuyog Pawar     {
449*c83a76b0SSuyog Pawar         WORD32 last_ctb_y =
450*c83a76b0SSuyog Pawar             ps_frm_ctb_prms->i4_cu_aligned_pic_ht -
451*c83a76b0SSuyog Pawar             ((ps_frm_ctb_prms->i4_num_ctbs_vert - 1) * ps_pad_interp_recon->i4_ctb_size);
452*c83a76b0SSuyog Pawar         ht = last_ctb_y;
453*c83a76b0SSuyog Pawar         ht_uv = last_ctb_y >> (0 == u1_is_422);
454*c83a76b0SSuyog Pawar     }
455*c83a76b0SSuyog Pawar 
456*c83a76b0SSuyog Pawar     pu1_buf_y = pu1_src;
457*c83a76b0SSuyog Pawar     pu1_buf_uv = pu1_src_uv;
458*c83a76b0SSuyog Pawar     cpy_ht_y = ht;
459*c83a76b0SSuyog Pawar     cpy_ht_uv = ht_uv;
460*c83a76b0SSuyog Pawar     if(vert_ctr > 0)
461*c83a76b0SSuyog Pawar     {
462*c83a76b0SSuyog Pawar         top_extra_pix = 8;
463*c83a76b0SSuyog Pawar     }
464*c83a76b0SSuyog Pawar     if(ctb_ctr > 0)
465*c83a76b0SSuyog Pawar     {
466*c83a76b0SSuyog Pawar         left_extra_pix = 8;
467*c83a76b0SSuyog Pawar     }
468*c83a76b0SSuyog Pawar 
469*c83a76b0SSuyog Pawar     /*top padding*/
470*c83a76b0SSuyog Pawar     if(vert_ctr == 0)
471*c83a76b0SSuyog Pawar     {
472*c83a76b0SSuyog Pawar         PAD_BUF_VER(
473*c83a76b0SSuyog Pawar             pu1_src - left_extra_pix, stride, wd + left_extra_pix, PAD_HORZ, PAD_VERT, pf_pad_top);
474*c83a76b0SSuyog Pawar         PAD_BUF_VER(
475*c83a76b0SSuyog Pawar             pu1_src_uv - left_extra_pix,
476*c83a76b0SSuyog Pawar             stride_uv,
477*c83a76b0SSuyog Pawar             wd_uv + left_extra_pix,
478*c83a76b0SSuyog Pawar             PAD_HORZ,
479*c83a76b0SSuyog Pawar             i4_chroma_vert_pad_default,
480*c83a76b0SSuyog Pawar             pf_pad_top);
481*c83a76b0SSuyog Pawar         /*if curr ctb is 1st ctb in ctb row, update dst pointer for Left padding*/
482*c83a76b0SSuyog Pawar         pu1_buf_y = pu1_src - PAD_VERT * stride;
483*c83a76b0SSuyog Pawar         pu1_buf_uv = pu1_src_uv - i4_chroma_vert_pad_default * stride_uv;
484*c83a76b0SSuyog Pawar         cpy_ht_y += PAD_VERT;
485*c83a76b0SSuyog Pawar         cpy_ht_uv += i4_chroma_vert_pad_default;
486*c83a76b0SSuyog Pawar     }
487*c83a76b0SSuyog Pawar 
488*c83a76b0SSuyog Pawar     /*bottom padding*/
489*c83a76b0SSuyog Pawar     if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
490*c83a76b0SSuyog Pawar     {
491*c83a76b0SSuyog Pawar         PAD_BUF_VER(
492*c83a76b0SSuyog Pawar             ((pu1_src - left_extra_pix) + (ht * stride)),
493*c83a76b0SSuyog Pawar             stride,
494*c83a76b0SSuyog Pawar             wd + left_extra_pix,
495*c83a76b0SSuyog Pawar             PAD_HORZ,
496*c83a76b0SSuyog Pawar             PAD_VERT,
497*c83a76b0SSuyog Pawar             pf_pad_bottom);
498*c83a76b0SSuyog Pawar         PAD_BUF_VER(
499*c83a76b0SSuyog Pawar             ((pu1_src_uv - left_extra_pix) + (ht_uv * stride_uv)),
500*c83a76b0SSuyog Pawar             stride_uv,
501*c83a76b0SSuyog Pawar             wd_uv + left_extra_pix,
502*c83a76b0SSuyog Pawar             PAD_HORZ,
503*c83a76b0SSuyog Pawar             i4_chroma_vert_pad_default,
504*c83a76b0SSuyog Pawar             pf_pad_bottom);
505*c83a76b0SSuyog Pawar         /*if curr ctb is 1st ctb in ctb row, update dst pointer for right padding*/
506*c83a76b0SSuyog Pawar         cpy_ht_y += PAD_VERT;
507*c83a76b0SSuyog Pawar         cpy_ht_uv += i4_chroma_vert_pad_default;
508*c83a76b0SSuyog Pawar     }
509*c83a76b0SSuyog Pawar 
510*c83a76b0SSuyog Pawar     /*Left padding*/
511*c83a76b0SSuyog Pawar     if(ctb_ctr == 0)
512*c83a76b0SSuyog Pawar     {
513*c83a76b0SSuyog Pawar         PAD_BUF_HOR(
514*c83a76b0SSuyog Pawar             (pu1_buf_y - top_extra_pix * stride),
515*c83a76b0SSuyog Pawar             stride,
516*c83a76b0SSuyog Pawar             cpy_ht_y + top_extra_pix,
517*c83a76b0SSuyog Pawar             PAD_HORZ,
518*c83a76b0SSuyog Pawar             PAD_VERT,
519*c83a76b0SSuyog Pawar             pf_pad_left_luma);
520*c83a76b0SSuyog Pawar         PAD_BUF_HOR(
521*c83a76b0SSuyog Pawar             pu1_buf_uv - (top_extra_pix >> 1) * (u1_is_422 + 1) * stride_uv,
522*c83a76b0SSuyog Pawar             stride_uv,
523*c83a76b0SSuyog Pawar             cpy_ht_uv + (top_extra_pix >> 1) * (u1_is_422 + 1),
524*c83a76b0SSuyog Pawar             PAD_HORZ,
525*c83a76b0SSuyog Pawar             i4_chroma_vert_pad_default,
526*c83a76b0SSuyog Pawar             pf_pad_left_chroma);
527*c83a76b0SSuyog Pawar     }
528*c83a76b0SSuyog Pawar 
529*c83a76b0SSuyog Pawar     /*right padding*/
530*c83a76b0SSuyog Pawar     if(ctb_ctr == (ps_frm_ctb_prms->i4_num_ctbs_horz - 1))
531*c83a76b0SSuyog Pawar     {
532*c83a76b0SSuyog Pawar         PAD_BUF_HOR(
533*c83a76b0SSuyog Pawar             ((pu1_buf_y - (top_extra_pix * stride)) + wd),
534*c83a76b0SSuyog Pawar             stride,
535*c83a76b0SSuyog Pawar             cpy_ht_y + top_extra_pix,
536*c83a76b0SSuyog Pawar             PAD_HORZ,
537*c83a76b0SSuyog Pawar             PAD_VERT,
538*c83a76b0SSuyog Pawar             pf_pad_right_luma);
539*c83a76b0SSuyog Pawar         PAD_BUF_HOR(
540*c83a76b0SSuyog Pawar             ((pu1_buf_uv - ((top_extra_pix >> 1) * (u1_is_422 + 1) * stride_uv)) + wd_uv),
541*c83a76b0SSuyog Pawar             stride_uv,
542*c83a76b0SSuyog Pawar             cpy_ht_uv + (top_extra_pix >> 1) * (u1_is_422 + 1),
543*c83a76b0SSuyog Pawar             PAD_HORZ,
544*c83a76b0SSuyog Pawar             i4_chroma_vert_pad_default,
545*c83a76b0SSuyog Pawar             pf_pad_right_chroma);
546*c83a76b0SSuyog Pawar     }
547*c83a76b0SSuyog Pawar }
548*c83a76b0SSuyog Pawar 
ihevce_pad_interp_recon_src_ctb(pad_interp_recon_frm_t * ps_pad_interp_recon,WORD32 ctb_ctr,WORD32 vert_ctr,frm_ctb_ctxt_t * ps_frm_ctb_prms,WORD32 i4_bitrate_instance_id,func_selector_t * ps_func_selector,WORD32 is_chroma_needs_padding)549*c83a76b0SSuyog Pawar void ihevce_pad_interp_recon_src_ctb(
550*c83a76b0SSuyog Pawar     pad_interp_recon_frm_t *ps_pad_interp_recon,
551*c83a76b0SSuyog Pawar     WORD32 ctb_ctr,
552*c83a76b0SSuyog Pawar     WORD32 vert_ctr,
553*c83a76b0SSuyog Pawar     frm_ctb_ctxt_t *ps_frm_ctb_prms,
554*c83a76b0SSuyog Pawar     WORD32 i4_bitrate_instance_id,
555*c83a76b0SSuyog Pawar     func_selector_t *ps_func_selector,
556*c83a76b0SSuyog Pawar     WORD32 is_chroma_needs_padding)
557*c83a76b0SSuyog Pawar {
558*c83a76b0SSuyog Pawar     UWORD8 *pu1_src, *pu1_src_uv;
559*c83a76b0SSuyog Pawar     WORD32 stride, stride_uv, wd, ht, wd_uv, ht_uv, pad_x, pad_y;
560*c83a76b0SSuyog Pawar     WORD32 tot_wd, tot_ht;
561*c83a76b0SSuyog Pawar     WORD32 i4_chroma_vert_pad_default;
562*c83a76b0SSuyog Pawar 
563*c83a76b0SSuyog Pawar     WORD32 ctb_size = ps_frm_ctb_prms->i4_ctb_size;
564*c83a76b0SSuyog Pawar     UWORD8 u1_is_422 = (ps_pad_interp_recon->u1_chroma_array_type == 2);
565*c83a76b0SSuyog Pawar 
566*c83a76b0SSuyog Pawar     ihevc_pad_top_ft *pf_pad_top = ps_func_selector->ihevc_pad_top_fptr;
567*c83a76b0SSuyog Pawar     ihevc_pad_bottom_ft *pf_pad_bottom = ps_func_selector->ihevc_pad_bottom_fptr;
568*c83a76b0SSuyog Pawar     ihevc_pad_left_luma_ft *pf_pad_left_luma = ps_func_selector->ihevc_pad_left_luma_fptr;
569*c83a76b0SSuyog Pawar     ihevc_pad_left_chroma_ft *pf_pad_left_chroma = ps_func_selector->ihevc_pad_left_chroma_fptr;
570*c83a76b0SSuyog Pawar     ihevc_pad_right_luma_ft *pf_pad_right_luma = ps_func_selector->ihevc_pad_right_luma_fptr;
571*c83a76b0SSuyog Pawar     ihevc_pad_right_chroma_ft *pf_pad_right_chroma = ps_func_selector->ihevc_pad_right_chroma_fptr;
572*c83a76b0SSuyog Pawar 
573*c83a76b0SSuyog Pawar     /* Luma padding */
574*c83a76b0SSuyog Pawar     pu1_src = (UWORD8 *)ps_pad_interp_recon->pu1_luma_recon_src +
575*c83a76b0SSuyog Pawar               (vert_ctr * ctb_size * ps_pad_interp_recon->i4_luma_recon_stride) +
576*c83a76b0SSuyog Pawar               (ctb_ctr * ctb_size);
577*c83a76b0SSuyog Pawar 
578*c83a76b0SSuyog Pawar     stride = ps_pad_interp_recon->i4_luma_recon_stride;
579*c83a76b0SSuyog Pawar     wd = ps_pad_interp_recon->i4_ctb_size;
580*c83a76b0SSuyog Pawar     ht = ps_pad_interp_recon->i4_ctb_size;
581*c83a76b0SSuyog Pawar 
582*c83a76b0SSuyog Pawar     pu1_src_uv =
583*c83a76b0SSuyog Pawar         (UWORD8 *)ps_pad_interp_recon->pu1_chrm_recon_src +
584*c83a76b0SSuyog Pawar         (vert_ctr * (ctb_size >> (0 == u1_is_422)) * ps_pad_interp_recon->i4_chrm_recon_stride) +
585*c83a76b0SSuyog Pawar         (ctb_ctr * ctb_size);
586*c83a76b0SSuyog Pawar 
587*c83a76b0SSuyog Pawar     stride_uv = ps_pad_interp_recon->i4_chrm_recon_stride;
588*c83a76b0SSuyog Pawar     wd_uv = ps_pad_interp_recon->i4_ctb_size;
589*c83a76b0SSuyog Pawar     ht_uv = ps_pad_interp_recon->i4_ctb_size >> (0 == u1_is_422);
590*c83a76b0SSuyog Pawar 
591*c83a76b0SSuyog Pawar     pad_x = ALIGN8(NTAPS_LUMA);
592*c83a76b0SSuyog Pawar     pad_y = ALIGN8(NTAPS_LUMA);
593*c83a76b0SSuyog Pawar 
594*c83a76b0SSuyog Pawar     tot_wd = wd + (pad_x << 1);
595*c83a76b0SSuyog Pawar     tot_ht = ht + (pad_y << 1);
596*c83a76b0SSuyog Pawar 
597*c83a76b0SSuyog Pawar     i4_chroma_vert_pad_default = PAD_VERT >> (0 == u1_is_422);
598*c83a76b0SSuyog Pawar 
599*c83a76b0SSuyog Pawar     if(ctb_ctr == (ps_frm_ctb_prms->i4_num_ctbs_horz - 1))
600*c83a76b0SSuyog Pawar     {
601*c83a76b0SSuyog Pawar         WORD32 last_ctb_x =
602*c83a76b0SSuyog Pawar             ps_frm_ctb_prms->i4_cu_aligned_pic_wd -
603*c83a76b0SSuyog Pawar             ((ps_frm_ctb_prms->i4_num_ctbs_horz - 1) * ps_pad_interp_recon->i4_ctb_size);
604*c83a76b0SSuyog Pawar         wd = last_ctb_x;
605*c83a76b0SSuyog Pawar         wd_uv = last_ctb_x;
606*c83a76b0SSuyog Pawar     }
607*c83a76b0SSuyog Pawar     if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
608*c83a76b0SSuyog Pawar     {
609*c83a76b0SSuyog Pawar         WORD32 last_ctb_y =
610*c83a76b0SSuyog Pawar             ps_frm_ctb_prms->i4_cu_aligned_pic_ht -
611*c83a76b0SSuyog Pawar             ((ps_frm_ctb_prms->i4_num_ctbs_vert - 1) * ps_pad_interp_recon->i4_ctb_size);
612*c83a76b0SSuyog Pawar         ht = last_ctb_y;
613*c83a76b0SSuyog Pawar         ht_uv = last_ctb_y >> (0 == u1_is_422);
614*c83a76b0SSuyog Pawar     }
615*c83a76b0SSuyog Pawar 
616*c83a76b0SSuyog Pawar     if(ctb_ctr == 0)
617*c83a76b0SSuyog Pawar     {
618*c83a76b0SSuyog Pawar         if(vert_ctr == 0)
619*c83a76b0SSuyog Pawar         {
620*c83a76b0SSuyog Pawar             PAD_BUF_HOR(pu1_src, stride, ht, PAD_HORZ, PAD_VERT, pf_pad_left_luma);
621*c83a76b0SSuyog Pawar             PAD_BUF_VER(pu1_src - PAD_HORZ, stride, PAD_HORZ + wd, PAD_HORZ, PAD_VERT, pf_pad_top);
622*c83a76b0SSuyog Pawar             if(is_chroma_needs_padding)
623*c83a76b0SSuyog Pawar             {
624*c83a76b0SSuyog Pawar                 PAD_BUF_HOR(
625*c83a76b0SSuyog Pawar                     pu1_src_uv,
626*c83a76b0SSuyog Pawar                     stride_uv,
627*c83a76b0SSuyog Pawar                     ht_uv,
628*c83a76b0SSuyog Pawar                     PAD_HORZ,
629*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
630*c83a76b0SSuyog Pawar                     pf_pad_left_chroma);
631*c83a76b0SSuyog Pawar                 PAD_BUF_VER(
632*c83a76b0SSuyog Pawar                     pu1_src_uv - PAD_HORZ,
633*c83a76b0SSuyog Pawar                     stride_uv,
634*c83a76b0SSuyog Pawar                     PAD_HORZ + wd_uv,
635*c83a76b0SSuyog Pawar                     PAD_HORZ,
636*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
637*c83a76b0SSuyog Pawar                     pf_pad_top);
638*c83a76b0SSuyog Pawar             }
639*c83a76b0SSuyog Pawar         }
640*c83a76b0SSuyog Pawar         else if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
641*c83a76b0SSuyog Pawar         {
642*c83a76b0SSuyog Pawar             PAD_BUF_HOR(pu1_src - 8 * stride, stride, ht + 8, PAD_HORZ, PAD_VERT, pf_pad_left_luma);
643*c83a76b0SSuyog Pawar             PAD_BUF_VER(
644*c83a76b0SSuyog Pawar                 (pu1_src - PAD_HORZ + (ht * stride)),
645*c83a76b0SSuyog Pawar                 stride,
646*c83a76b0SSuyog Pawar                 PAD_HORZ + wd,
647*c83a76b0SSuyog Pawar                 PAD_HORZ,
648*c83a76b0SSuyog Pawar                 PAD_VERT,
649*c83a76b0SSuyog Pawar                 pf_pad_bottom);
650*c83a76b0SSuyog Pawar             if(is_chroma_needs_padding)
651*c83a76b0SSuyog Pawar             {
652*c83a76b0SSuyog Pawar                 PAD_BUF_HOR(
653*c83a76b0SSuyog Pawar                     pu1_src_uv - 4 * (u1_is_422 + 1) * stride_uv,
654*c83a76b0SSuyog Pawar                     stride_uv,
655*c83a76b0SSuyog Pawar                     ht_uv + 4 * (u1_is_422 + 1),
656*c83a76b0SSuyog Pawar                     PAD_HORZ,
657*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
658*c83a76b0SSuyog Pawar                     pf_pad_left_chroma);
659*c83a76b0SSuyog Pawar                 PAD_BUF_VER(
660*c83a76b0SSuyog Pawar                     (pu1_src_uv - PAD_HORZ + (ht_uv * stride_uv)),
661*c83a76b0SSuyog Pawar                     stride_uv,
662*c83a76b0SSuyog Pawar                     PAD_HORZ + wd_uv,
663*c83a76b0SSuyog Pawar                     PAD_HORZ,
664*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
665*c83a76b0SSuyog Pawar                     pf_pad_bottom);
666*c83a76b0SSuyog Pawar             }
667*c83a76b0SSuyog Pawar         }
668*c83a76b0SSuyog Pawar         else
669*c83a76b0SSuyog Pawar         {
670*c83a76b0SSuyog Pawar             PAD_BUF_HOR(pu1_src - 8 * stride, stride, ht + 8, PAD_HORZ, PAD_VERT, pf_pad_left_luma);
671*c83a76b0SSuyog Pawar             if(is_chroma_needs_padding)
672*c83a76b0SSuyog Pawar             {
673*c83a76b0SSuyog Pawar                 PAD_BUF_HOR(
674*c83a76b0SSuyog Pawar                     pu1_src_uv - 4 * (u1_is_422 + 1) * stride_uv,
675*c83a76b0SSuyog Pawar                     stride_uv,
676*c83a76b0SSuyog Pawar                     ht_uv + 4 * (u1_is_422 + 1),
677*c83a76b0SSuyog Pawar                     PAD_HORZ,
678*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
679*c83a76b0SSuyog Pawar                     pf_pad_left_chroma);
680*c83a76b0SSuyog Pawar             }
681*c83a76b0SSuyog Pawar         }
682*c83a76b0SSuyog Pawar     }
683*c83a76b0SSuyog Pawar     else if(ctb_ctr == (ps_frm_ctb_prms->i4_num_ctbs_horz - 1))
684*c83a76b0SSuyog Pawar     {
685*c83a76b0SSuyog Pawar         if(vert_ctr == 0)
686*c83a76b0SSuyog Pawar         {
687*c83a76b0SSuyog Pawar             PAD_BUF_HOR(pu1_src + wd, stride, ht, PAD_HORZ, PAD_VERT, pf_pad_right_luma);
688*c83a76b0SSuyog Pawar             PAD_BUF_VER(pu1_src - 8, stride, PAD_HORZ + (wd + 8), PAD_HORZ, PAD_VERT, pf_pad_top);
689*c83a76b0SSuyog Pawar             if(is_chroma_needs_padding)
690*c83a76b0SSuyog Pawar             {
691*c83a76b0SSuyog Pawar                 PAD_BUF_HOR(
692*c83a76b0SSuyog Pawar                     pu1_src_uv + wd_uv,
693*c83a76b0SSuyog Pawar                     stride_uv,
694*c83a76b0SSuyog Pawar                     ht_uv,
695*c83a76b0SSuyog Pawar                     PAD_HORZ,
696*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
697*c83a76b0SSuyog Pawar                     pf_pad_right_chroma);
698*c83a76b0SSuyog Pawar                 PAD_BUF_VER(
699*c83a76b0SSuyog Pawar                     pu1_src_uv - 8,
700*c83a76b0SSuyog Pawar                     stride_uv,
701*c83a76b0SSuyog Pawar                     PAD_HORZ + (wd_uv + 8),
702*c83a76b0SSuyog Pawar                     PAD_HORZ,
703*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
704*c83a76b0SSuyog Pawar                     pf_pad_top);
705*c83a76b0SSuyog Pawar             }
706*c83a76b0SSuyog Pawar         }
707*c83a76b0SSuyog Pawar         else if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
708*c83a76b0SSuyog Pawar         {
709*c83a76b0SSuyog Pawar             PAD_BUF_HOR(
710*c83a76b0SSuyog Pawar                 (pu1_src - (8 * stride) + wd),
711*c83a76b0SSuyog Pawar                 stride,
712*c83a76b0SSuyog Pawar                 ht + 8,
713*c83a76b0SSuyog Pawar                 PAD_HORZ,
714*c83a76b0SSuyog Pawar                 PAD_VERT,
715*c83a76b0SSuyog Pawar                 pf_pad_right_luma);
716*c83a76b0SSuyog Pawar             PAD_BUF_VER(
717*c83a76b0SSuyog Pawar                 (pu1_src - 8 + (ht * stride)),
718*c83a76b0SSuyog Pawar                 stride,
719*c83a76b0SSuyog Pawar                 PAD_HORZ + (wd + 8),
720*c83a76b0SSuyog Pawar                 PAD_HORZ,
721*c83a76b0SSuyog Pawar                 PAD_VERT,
722*c83a76b0SSuyog Pawar                 pf_pad_bottom);
723*c83a76b0SSuyog Pawar             if(is_chroma_needs_padding)
724*c83a76b0SSuyog Pawar             {
725*c83a76b0SSuyog Pawar                 PAD_BUF_HOR(
726*c83a76b0SSuyog Pawar                     (pu1_src_uv - (4 * (u1_is_422 + 1) * stride_uv) + wd_uv),
727*c83a76b0SSuyog Pawar                     stride_uv,
728*c83a76b0SSuyog Pawar                     ht_uv + 4 * (u1_is_422 + 1),
729*c83a76b0SSuyog Pawar                     PAD_HORZ,
730*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
731*c83a76b0SSuyog Pawar                     pf_pad_right_chroma);
732*c83a76b0SSuyog Pawar                 PAD_BUF_VER(
733*c83a76b0SSuyog Pawar                     (pu1_src_uv - 8 + (ht_uv * stride_uv)),
734*c83a76b0SSuyog Pawar                     stride_uv,
735*c83a76b0SSuyog Pawar                     PAD_HORZ + (wd_uv + 8),
736*c83a76b0SSuyog Pawar                     PAD_HORZ,
737*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
738*c83a76b0SSuyog Pawar                     pf_pad_bottom);
739*c83a76b0SSuyog Pawar             }
740*c83a76b0SSuyog Pawar         }
741*c83a76b0SSuyog Pawar         else
742*c83a76b0SSuyog Pawar         {
743*c83a76b0SSuyog Pawar             PAD_BUF_HOR(
744*c83a76b0SSuyog Pawar                 (pu1_src - (8 * stride) + wd),
745*c83a76b0SSuyog Pawar                 stride,
746*c83a76b0SSuyog Pawar                 ht + 8,
747*c83a76b0SSuyog Pawar                 PAD_HORZ,
748*c83a76b0SSuyog Pawar                 PAD_VERT,
749*c83a76b0SSuyog Pawar                 pf_pad_right_luma);
750*c83a76b0SSuyog Pawar             if(is_chroma_needs_padding)
751*c83a76b0SSuyog Pawar             {
752*c83a76b0SSuyog Pawar                 PAD_BUF_HOR(
753*c83a76b0SSuyog Pawar                     (pu1_src_uv - (4 * (u1_is_422 + 1) * stride_uv) + wd_uv),
754*c83a76b0SSuyog Pawar                     stride_uv,
755*c83a76b0SSuyog Pawar                     ht_uv + 4 * (u1_is_422 + 1),
756*c83a76b0SSuyog Pawar                     PAD_HORZ,
757*c83a76b0SSuyog Pawar                     i4_chroma_vert_pad_default,
758*c83a76b0SSuyog Pawar                     pf_pad_right_chroma);
759*c83a76b0SSuyog Pawar             }
760*c83a76b0SSuyog Pawar         }
761*c83a76b0SSuyog Pawar     }
762*c83a76b0SSuyog Pawar     else if(vert_ctr == 0)
763*c83a76b0SSuyog Pawar     {
764*c83a76b0SSuyog Pawar         PAD_BUF_VER(pu1_src - 8, stride, (wd + 8), PAD_HORZ, PAD_VERT, pf_pad_top);
765*c83a76b0SSuyog Pawar         if(is_chroma_needs_padding)
766*c83a76b0SSuyog Pawar         {
767*c83a76b0SSuyog Pawar             PAD_BUF_VER(
768*c83a76b0SSuyog Pawar                 pu1_src_uv - 8,
769*c83a76b0SSuyog Pawar                 stride_uv,
770*c83a76b0SSuyog Pawar                 (wd_uv + 8),
771*c83a76b0SSuyog Pawar                 PAD_HORZ,
772*c83a76b0SSuyog Pawar                 i4_chroma_vert_pad_default,
773*c83a76b0SSuyog Pawar                 pf_pad_top);
774*c83a76b0SSuyog Pawar         }
775*c83a76b0SSuyog Pawar     }
776*c83a76b0SSuyog Pawar     else if(vert_ctr == (ps_frm_ctb_prms->i4_num_ctbs_vert - 1))
777*c83a76b0SSuyog Pawar     {
778*c83a76b0SSuyog Pawar         PAD_BUF_VER(
779*c83a76b0SSuyog Pawar             (pu1_src - 8 + (ht * stride)), stride, (wd + 8), PAD_HORZ, PAD_VERT, pf_pad_bottom);
780*c83a76b0SSuyog Pawar         if(is_chroma_needs_padding)
781*c83a76b0SSuyog Pawar         {
782*c83a76b0SSuyog Pawar             PAD_BUF_VER(
783*c83a76b0SSuyog Pawar                 (pu1_src_uv - 8 + (ht_uv * stride_uv)),
784*c83a76b0SSuyog Pawar                 stride_uv,
785*c83a76b0SSuyog Pawar                 (wd_uv + 8),
786*c83a76b0SSuyog Pawar                 PAD_HORZ,
787*c83a76b0SSuyog Pawar                 i4_chroma_vert_pad_default,
788*c83a76b0SSuyog Pawar                 pf_pad_bottom);
789*c83a76b0SSuyog Pawar         }
790*c83a76b0SSuyog Pawar     }
791*c83a76b0SSuyog Pawar }
792