wrapper/libhevc-mirror
1
0
Fork 0
mirror of https://github.com/ittiam-systems/libhevc.git synced 2026-07-09 06:06:22 +07:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

libhevcdec: add support for deblock filter for yuv 444 format

Browse source 
Test: ./hevcdec

Change-Id: I013e44e024258f2f84690fdcc01b4cd56f0fbd19
This commit is contained in:
Vivek R Jadhav 2026-05-08 17:14:44 +05:30 • committed by Harish Mahendrakar
parent 15deccc068
commit 70ee4f3d29
10 changed files with 270 additions and 96 deletions
Download patch file Download diff file Expand all files Collapse all files

16
common/arm/ihevc_deblk_chroma_horz.s
View file

@ -41,6 +41,7 @@
.equ tc_offset_div2_offset, 48
.equ filter_p_offset, 52
.equ filter_q_offset, 56
.equ chroma_fmt, 60
.text
.align 4
@ -76,23 +77,38 @@ ihevc_deblk_chroma_horz_a9q:
ldr r8,[sp,#filter_p_offset]
vld1.8 {d16},[r6]
ldr r9,[sp,#filter_q_offset]
ldr r11, [sp, #chroma_fmt]
adds r1,r10,r2,asr #1
vmovl.u8 q1,d2
ldr r7,[sp,#qp_offset_v_offset]
ldr r3,gai4_ihevc_qp_table_addr
ulbl1:
add r3, r3, pc
cmp r11,#1
bgt qp_u_min_51_horz
bmi l1.3312
cmp r1,#0x39
ldrle r1,[r3,r1,lsl #2]
subgt r1,r1,#6
b qp_u_done_horz
qp_u_min_51_horz:
cmp r1,#0x33
movgt r1,#0x33
qp_u_done_horz:
l1.3312:
adds r2,r7,r2,asr #1
vmovl.u8 q2,d4
cmp r11,#1
bgt qp_v_min_51_horz
bmi l1.3332
cmp r2,#0x39
ldrle r2,[r3,r2,lsl #2]
subgt r2,r2,#6
b qp_v_done_horz
qp_v_min_51_horz:
cmp r2,#0x33
movgt r2,#0x33
qp_v_done_horz:
l1.3332:
add r1,r1,r4,lsl #1
vsub.i16 q3,q0,q1

16
common/arm/ihevc_deblk_chroma_vert.s
View file

@ -42,6 +42,7 @@
.equ tc_offset_div2_offset, 48
.equ filter_p_offset, 52
.equ filter_q_offset, 56
.equ chroma_fmt, 60
.text
.align 4
@ -82,18 +83,33 @@ ulbl1:
add r7,r7,pc
ldr r12,[sp,#filter_p_offset]
ldr r6,[sp,#qp_offset_v_offset]
ldr r10,[sp,#chroma_fmt]
cmp r10,#1
bgt qp_u_min_51_vert
bmi l1.2944
cmp r3,#0x39
ldrle r3,[r7,r3,lsl #2]
subgt r3,r3,#6
b qp_u_done_vert
qp_u_min_51_vert:
cmp r3,#0x33
movgt r3,#0x33
qp_u_done_vert:
l1.2944:
vtrn.16 d5,d16
adds r2,r6,r2,asr #1
vtrn.16 d17,d4
cmp r10,#1
bgt qp_v_min_51_vert
bmi l1.2964
cmp r2,#0x39
ldrle r2,[r7,r2,lsl #2]
subgt r2,r2,#6
b qp_v_done_vert
qp_v_min_51_vert:
cmp r2,#0x33
movgt r2,#0x33
qp_v_done_vert:
l1.2964:
vtrn.32 d5,d17
add r3,r3,r5,lsl #1

21
common/arm64/ihevc_deblk_chroma_horz.s
View file

@ -42,7 +42,8 @@
// WORD32 qp_offset_v,
// WORD32 tc_offset_div2,
// WORD32 filter_flag_p,
// WORD32 filter_flag_q)
// WORD32 filter_flag_q,
// WORD8 chroma_fmt)
//
.include "ihevc_neon_macros.s"
@ -62,6 +63,8 @@ ENTRY ihevc_deblk_chroma_horz_av8
sxtw x6,w6
ldr w9, [sp]
sxtw x9,w9
ldr w11, [sp, #8]
sxtw x11,w11
push_v_regs
stp x19, x20,[sp,#-16]!
mov x10, x4
@ -83,6 +86,8 @@ ENTRY ihevc_deblk_chroma_horz_av8
uxtl v2.8h, v2.8b
adrp x3, :got:gai4_ihevc_qp_table
ldr x3, [x3, #:got_lo12:gai4_ihevc_qp_table]
cmp x11,#1
bgt qp_u_min_51_horz
bmi l1.3312
cmp x1,#0x39
bgt lbl78
@ -90,9 +95,17 @@ ENTRY ihevc_deblk_chroma_horz_av8
lbl78:
sub x20,x1,#6
csel x1, x20, x1,gt
b qp_u_done_horz
qp_u_min_51_horz:
mov x20,#0x33
cmp x1,x20
csel x1, x20, x1,gt
qp_u_done_horz:
l1.3312:
adds x2,x7,x2,asr #1
uxtl v4.8h, v4.8b
cmp x11,#1
bgt qp_v_min_51_horz
bmi l1.3332
cmp x2,#0x39
bgt lbl85
@ -100,6 +113,12 @@ l1.3312:
lbl85:
sub x20,x2,#6
csel x2, x20, x2,gt
b qp_v_done_horz
qp_v_min_51_horz:
mov x20,#0x33
cmp x2,x20
csel x2, x20, x2,gt
qp_v_done_horz:
l1.3332:
add x1,x1,x4,lsl #1
sub v6.8h, v0.8h , v2.8h

21
common/arm64/ihevc_deblk_chroma_vert.s
View file

@ -44,7 +44,8 @@
// WORD32 qp_offset_v,
// WORD32 tc_offset_div2,
// WORD32 filter_flag_p,
// WORD32 filter_flag_q)
// WORD32 filter_flag_q,
// WORD8 chroma_fmt)
.include "ihevc_neon_macros.s"
.text
@ -67,6 +68,8 @@ ENTRY ihevc_deblk_chroma_vert_av8
mov x12, x7
mov x7, x4
ldr w4, [sp]
ldr w11, [sp, #8]
sxtw x11,w11
push_v_regs
stp x19, x20,[sp,#-16]!
@ -89,6 +92,8 @@ ENTRY ihevc_deblk_chroma_vert_av8
adrp x7, :got:gai4_ihevc_qp_table
ldr x7, [x7, #:got_lo12:gai4_ihevc_qp_table]
cmp x11,#1
bgt qp_u_min_51_vert
bmi l1.2944
cmp x3,#0x39
@ -98,6 +103,12 @@ ENTRY ihevc_deblk_chroma_vert_av8
lbl78:
sub x20,x3,#6
csel x3, x20, x3,gt
b qp_u_done_vert
qp_u_min_51_vert:
mov x20,#0x33
cmp x3,x20
csel x3, x20, x3,gt
qp_u_done_vert:
l1.2944:
trn1 v29.4h, v5.4h, v16.4h
trn2 v16.4h, v5.4h, v16.4h
@ -106,6 +117,8 @@ l1.2944:
trn1 v29.4h, v17.4h, v4.4h
trn2 v4.4h, v17.4h, v4.4h
mov v17.d[0], v29.d[0]
cmp x11,#1
bgt qp_v_min_51_vert
bmi l1.2964
cmp x2,#0x39
bgt lbl86
@ -114,6 +127,12 @@ l1.2944:
lbl86:
sub x20,x2,#6
csel x2, x20, x2,gt
b qp_v_done_vert
qp_v_min_51_vert:
mov x20,#0x33
cmp x2,x20
csel x2, x20, x2,gt
qp_v_done_vert:
l1.2964:
trn1 v29.2s, v5.2s, v17.2s
trn2 v17.2s, v5.2s, v17.2s

12
common/ihevc_deblk.h
View file

@ -69,7 +69,8 @@ typedef void ihevc_deblk_chroma_vert_ft(
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q);
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc);
typedef void ihevc_deblk_chroma_horz_ft(
UWORD8 *pu1_src,
@ -80,7 +81,8 @@ typedef void ihevc_deblk_chroma_horz_ft(
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q);
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc);
typedef void ihevc_hbd_deblk_luma_vert_ft(
UWORD16 *pu2_src,
@ -116,7 +118,8 @@ typedef void ihevc_hbd_deblk_chroma_vert_ft(
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q,
UWORD8 bit_depth);
UWORD8 bit_depth,
WORD8 chroma_fmt_idc);
typedef void ihevc_hbd_deblk_chroma_horz_ft(
UWORD16 *pu2_src,
@ -128,7 +131,8 @@ typedef void ihevc_hbd_deblk_chroma_horz_ft(
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q,
UWORD8 bit_depth);
UWORD8 bit_depth,
WORD8 chroma_fmt_idc);
ihevc_deblk_luma_vert_ft ihevc_deblk_luma_vert;
ihevc_deblk_luma_horz_ft ihevc_deblk_luma_horz;

37
common/ihevc_deblk_edge_filter.c
View file

@ -46,6 +46,7 @@
#include "ihevc_deblk.h"
#include "ihevc_deblk_tables.h"
#include "ihevc_debug.h"
#include "ihevc_defs.h"
/**
@ -598,7 +599,8 @@ void ihevc_deblk_chroma_vert(UWORD8 *pu1_src,
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q)
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -613,10 +615,19 @@ void ihevc_deblk_chroma_vert(UWORD8 *pu1_src,
/* chroma processing is done only if BS is 2 */
/* this function is assumed to be called only if BS is 2 */
qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
/* 8.7.2.5.5 Filtering process for chroma block edges */
if(chroma_fmt_idc == CHROMA_FMT_IDC_YUV444)
{
qp_chroma_u = MIN(qp_indx_u, 51);
qp_chroma_v = MIN(qp_indx_v, 51);
}
else
{
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
}
tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53);
tc_u = gai4_ihevc_tc_table[tc_indx_u];
@ -718,7 +729,8 @@ void ihevc_deblk_chroma_horz(UWORD8 *pu1_src,
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q)
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -734,10 +746,19 @@ void ihevc_deblk_chroma_horz(UWORD8 *pu1_src,
/* chroma processing is done only if BS is 2 */
/* this function is assumed to be called only if BS is 2 */
qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
/* 8.7.2.5.5. Filtering process for chroma block edges */
if(chroma_fmt_idc == CHROMA_FMT_IDC_YUV444)
{
qp_chroma_u = MIN(qp_indx_u, 51);
qp_chroma_v = MIN(qp_indx_v, 51);
}
else
{
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
}
tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53);
tc_u = gai4_ihevc_tc_table[tc_indx_u];

49
common/ihevc_hbd_deblk_edge_filter.c
View file

@ -52,6 +52,7 @@
#include "ihevc_deblk.h"
#include "ihevc_deblk_tables.h"
#include "ihevc_debug.h"
#include "ihevc_defs.h"
/**
@ -601,7 +602,8 @@ void ihevc_hbd_deblk_chroma_vert(UWORD16 *pu2_src,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q,
UWORD8 bit_depth)
UWORD8 bit_depth,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -616,10 +618,19 @@ void ihevc_hbd_deblk_chroma_vert(UWORD16 *pu2_src,
/* chroma processing is done only if BS is 2 */
/* this function is assumed to be called only if BS is 2 */
qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
/* 8.7.2.5.5 Filtering process for chroma block edges */
if(chroma_fmt_idc == CHROMA_FMT_IDC_YUV444)
{
qp_chroma_u = MIN(qp_indx_u, 51);
qp_chroma_v = MIN(qp_indx_v, 51);
}
else
{
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
}
tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53);
tc_u = gai4_ihevc_tc_table[tc_indx_u] * (1 << (bit_depth - 8));
@ -715,7 +726,8 @@ void ihevc_deblk_422chroma_vert(UWORD8 *pu1_src,
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q)
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -833,7 +845,8 @@ void ihevc_hbd_deblk_422chroma_vert(UWORD16 *pu2_src,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q,
UWORD8 bit_depth)
UWORD8 bit_depth,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -950,7 +963,8 @@ void ihevc_deblk_422chroma_horz
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc
)
{
WORD32 qp_indx_u, qp_chroma_u;
@ -1065,7 +1079,8 @@ void ihevc_hbd_deblk_chroma_horz(UWORD16 *pu2_src,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q,
UWORD8 bit_depth)
UWORD8 bit_depth,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -1081,10 +1096,19 @@ void ihevc_hbd_deblk_chroma_horz(UWORD16 *pu2_src,
/* chroma processing is done only if BS is 2 */
/* this function is assumed to be called only if BS is 2 */
qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
/* 8.7.2.5.5 Filtering process for chroma block edges */
if(chroma_fmt_idc == CHROMA_FMT_IDC_YUV444)
{
qp_chroma_u = MIN(qp_indx_u, 51);
qp_chroma_v = MIN(qp_indx_v, 51);
}
else
{
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
}
tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53);
tc_u = gai4_ihevc_tc_table[tc_indx_u] * (1 << (bit_depth - 8));
@ -1177,7 +1201,8 @@ void ihevc_hbd_deblk_422chroma_horz(UWORD16 *pu2_src,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q,
UWORD8 bit_depth)
UWORD8 bit_depth,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;

37
common/x86/ihevc_deblk_ssse3_intr.c
View file

@ -46,6 +46,7 @@
#include "ihevc_deblk.h"
#include "ihevc_deblk_tables.h"
#include "ihevc_debug.h"
#include "ihevc_defs.h"
#include "ihevc_tables_x86_intr.h"
@ -993,7 +994,8 @@ void ihevc_deblk_chroma_vert_ssse3(UWORD8 *pu1_src,
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q)
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -1006,10 +1008,19 @@ void ihevc_deblk_chroma_vert_ssse3(UWORD8 *pu1_src,
/* chroma processing is done only if BS is 2 */
/* this function is assumed to be called only if BS is 2 */
qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
/* 8.7.2.5.5 Filtering process for chroma block edges */
if(chroma_fmt_idc == CHROMA_FMT_IDC_YUV444 || chroma_fmt_idc == CHROMA_FMT_IDC_YUV422)
{
qp_chroma_u = MIN(qp_indx_u, 51);
qp_chroma_v = MIN(qp_indx_v, 51);
}
else
{
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
}
tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53);
tc_u = gai4_ihevc_tc_table[tc_indx_u];
@ -1144,7 +1155,8 @@ void ihevc_deblk_chroma_horz_ssse3(UWORD8 *pu1_src,
WORD32 qp_offset_v,
WORD32 tc_offset_div2,
WORD32 filter_flag_p,
WORD32 filter_flag_q)
WORD32 filter_flag_q,
WORD8 chroma_fmt_idc)
{
WORD32 qp_indx_u, qp_chroma_u;
WORD32 qp_indx_v, qp_chroma_v;
@ -1159,10 +1171,19 @@ void ihevc_deblk_chroma_horz_ssse3(UWORD8 *pu1_src,
/* chroma processing is done only if BS is 2 */
/* this function is assumed to be called only if BS is 2 */
qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
/* 8.7.2.5.5 Filtering process for chroma block edges */
if(chroma_fmt_idc == CHROMA_FMT_IDC_YUV444 || chroma_fmt_idc == CHROMA_FMT_IDC_YUV422)
{
qp_chroma_u = MIN(qp_indx_u, 51);
qp_chroma_v = MIN(qp_indx_v, 51);
}
else
{
qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]);
qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]);
}
tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53);
tc_u = gai4_ihevc_tc_table[tc_indx_u];

148
decoder/ihevcd_deblk.c
View file

@ -111,7 +111,7 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
UWORD32 *pu4_vert_bs, *pu4_horz_bs;
UWORD32 *pu4_ctb_vert_bs, *pu4_ctb_horz_bs;
WORD32 bs_strd;
WORD32 src_strd;
WORD32 src_strd, chroma_strd;
UWORD8 *pu1_qp;
UWORD16 *pu2_ctb_no_loop_filter_flag;
UWORD16 au2_ctb_no_loop_filter_flag[9];
@ -135,6 +135,11 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
ps_codec = ps_deblk->ps_codec;
ps_slice_hdr = ps_deblk->ps_slice_hdr;
WORD32 is_yuv444 = ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV444 ? 1 : 0;
WORD32 h_samp_factor = (CHROMA_FMT_IDC_YUV444 == ps_sps->i1_chroma_format_idc) ? 1 : 2;
WORD32 v_samp_factor = (CHROMA_FMT_IDC_YUV420 == ps_sps->i1_chroma_format_idc) ? 2 : 1;
WORD32 chroma_pixel_strd = 2;
log2_ctb_size = ps_sps->i1_log2_ctb_size;
ctb_size = (1 << ps_sps->i1_log2_ctb_size);
@ -211,6 +216,7 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
}
src_strd = ps_codec->i4_strd;
chroma_strd = src_strd * chroma_pixel_strd / h_samp_factor;
/* Luma Vertical Edge */
@ -508,14 +514,14 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
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];
}
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);
pu1_src += i4_is_last_ctb_y ? (ps_deblk->ps_codec->i4_strd / 2) << log2_ctb_size : 0;
pu1_src = ps_deblk->pu1_cur_pic_chroma + ((ps_deblk->i4_ctb_x * chroma_pixel_strd / h_samp_factor + ps_deblk->i4_ctb_y * ps_deblk->ps_codec->i4_strd * chroma_pixel_strd / (h_samp_factor * v_samp_factor)) << log2_ctb_size);
pu1_src += i4_is_last_ctb_y ? ((ps_codec->i4_strd * chroma_pixel_strd) / (h_samp_factor * v_samp_factor)) << log2_ctb_size : 0;
/** Deblocking is done on a shifted CTB -
* Vertical edge processing is done by shifting the CTB up by four pixels */
pu1_src -= 4 * src_strd;
pu1_src -= 4 * chroma_strd;
for(col = 0; col < ctb_size / 16; col++)
for(col = 0; col < ctb_size / (8 * h_samp_factor); col++)
{
/* BS for the column - Last row is excluded and the top row is included*/
@ -529,22 +535,24 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
u4_bs |= u4_top_bs & 3;
}
/* Every alternate boundary strength value is used for chroma */
u4_bs &= 0x22222222;
if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) {
/* Every alternate boundary strength value is used for chroma */
u4_bs &= 0x22222222;
}
for(row = 0; row < ctb_size / 8;)
for(row = 0; row < ctb_size / (4 * v_samp_factor);)
{
WORD8 i1_tc_offset_div2 = ps_slice_hdr->i1_tc_offset_div2;
bs_tz = CTZ(u4_bs) >> 2;
bs_tz = CTZ(u4_bs) >> v_samp_factor;
if(0 != bs_tz)
{
if((row + bs_tz) >= (ctb_size / 8))
pu1_src += 4 * (ctb_size / 8 - row) * src_strd;
if((row + bs_tz) >= (ctb_size / (4 * v_samp_factor)))
pu1_src += 4 * (ctb_size / (4 * v_samp_factor) - row) * chroma_strd;
else
pu1_src += 4 * bs_tz * src_strd;
pu1_src += 4 * bs_tz * chroma_strd;
row += bs_tz;
u4_bs = u4_bs >> (bs_tz << 2);
u4_bs = u4_bs >> (bs_tz << v_samp_factor);
continue;
}
@ -562,12 +570,12 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
{
qp_p = u4_qp_const_in_ctb[1] ?
pu1_qp[-ctb_size / 8 * qp_strd] :
pu1_qp[2 * col - 1 - qp_strd];
pu1_qp[h_samp_factor * col - 1 - qp_strd];
}
qp_q = u4_qp_const_in_ctb[1] ?
pu1_qp[-ctb_size / 8 * qp_strd] :
pu1_qp[2 * col - qp_strd];
pu1_qp[h_samp_factor * col - qp_strd];
}
else
{
@ -575,30 +583,32 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
{
qp_p = u4_qp_const_in_ctb[2] ?
pu1_qp[-ctb_size / 8] :
pu1_qp[(row - 1) * qp_strd - 1];
pu1_qp[((row - 1) >> (2 - v_samp_factor)) * qp_strd - 1];
}
else
{
qp_p = u4_qp_const_in_ctb[3] ?
pu1_qp[0] :
pu1_qp[(row - 1) * qp_strd + 2 * col - 1];
pu1_qp[((row - 1) >> (2 - v_samp_factor)) * qp_strd + h_samp_factor * col - 1];
}
qp_q = u4_qp_const_in_ctb[3] ?
pu1_qp[0] :
pu1_qp[(row - 1) * qp_strd + 2 * col];
pu1_qp[((row - 1) >> (2 - v_samp_factor)) * qp_strd + h_samp_factor * col];
}
filter_p = (pu2_ctb_no_loop_filter_flag[row] >> (col << 1)) & 1;
filter_q = (pu2_ctb_no_loop_filter_flag[row] >> (col << 1)) & 2;
filter_p = (pu2_ctb_no_loop_filter_flag[(row + (2 - v_samp_factor)) >> (2 - v_samp_factor)] >> (col << (h_samp_factor - 1))) & 1;
filter_q = (pu2_ctb_no_loop_filter_flag[(row + (2 - v_samp_factor)) >> (2 - v_samp_factor)] >> (col << (h_samp_factor - 1))) & 2;
/* filter_p and filter_q are inverted as they are calculated using no_loop_filter_flags */
filter_p = !filter_p;
filter_q = !filter_q;
if(filter_p || filter_q)
{
ASSERT(1 == ((u4_bs & 3) >> 1));
DUMP_DEBLK_CHROMA_VERT(pu1_src, src_strd,
if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) {
ASSERT(1 == ((u4_bs & 3) >> 1));
}
DUMP_DEBLK_CHROMA_VERT(pu1_src, chroma_strd,
u4_bs & 3, qp_p, qp_q,
ps_pps->i1_pic_cb_qp_offset,
ps_pps->i1_pic_cr_qp_offset,
@ -607,36 +617,45 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
if(chroma_yuv420sp_vu)
{
ps_codec->s_func_selector.ihevc_deblk_chroma_vert_fptr(pu1_src,
src_strd,
chroma_strd,
qp_q,
qp_p,
ps_pps->i1_pic_cr_qp_offset,
ps_pps->i1_pic_cb_qp_offset,
i1_tc_offset_div2,
filter_q,
filter_p);
filter_p,
ps_sps->i1_chroma_format_idc);
}
else
{
ps_codec->s_func_selector.ihevc_deblk_chroma_vert_fptr(pu1_src,
src_strd,
chroma_strd,
qp_p,
qp_q,
ps_pps->i1_pic_cb_qp_offset,
ps_pps->i1_pic_cr_qp_offset,
i1_tc_offset_div2,
filter_p,
filter_q);
filter_q,
ps_sps->i1_chroma_format_idc);
}
}
pu1_src += 4 * src_strd;
u4_bs = u4_bs >> 4;
pu1_src += 4 * chroma_strd;
u4_bs = u4_bs >> (2 * v_samp_factor);
row++;
}
pu4_vert_bs += (64 == ctb_size) ? 2 : 1;
pu1_src -= ((src_strd / 2) << log2_ctb_size);
pu1_src -= (((src_strd * chroma_pixel_strd) / (h_samp_factor * v_samp_factor)) << log2_ctb_size);
if(is_yuv444)
{
if((64 ==ctb_size) || ((32 == ctb_size) && (col & 1))) pu4_vert_bs++;
}
else
{
pu4_vert_bs += (64 == ctb_size) ? 2 : 1;
}
pu1_src += 16;
}
}
@ -655,13 +674,13 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
ps_slice_hdr_left = ps_codec->ps_slice_hdr_base + ps_deblk->pu1_slice_idx[cur_ctb_indx - 1];
}
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);
pu1_src += i4_is_last_ctb_x ? ctb_size : 0;
pu1_src = ps_deblk->pu1_cur_pic_chroma + ((ps_deblk->i4_ctb_x * chroma_pixel_strd / h_samp_factor + ps_deblk->i4_ctb_y * ps_deblk->ps_codec->i4_strd * chroma_pixel_strd / (v_samp_factor * h_samp_factor)) << log2_ctb_size);
pu1_src += i4_is_last_ctb_x ? ctb_size * chroma_pixel_strd / h_samp_factor : 0;
/** Deblocking is done on a shifted CTB -
* Vertical edge processing is done by shifting the CTB up by four pixels (8 here beacuse UV are interleaved) */
pu1_src -= 8;
for(row = 0; row < ctb_size / 16; row++)
for(row = 0; row < ctb_size / (8 * v_samp_factor); row++)
{
/* BS for the row - Last column is excluded and the left column is included*/
u4_bs = pu4_horz_bs[0] << 2;
@ -676,20 +695,22 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
u4_bs |= u4_left_bs & 3;
}
/* Every alternate boundary strength value is used for chroma */
u4_bs &= 0x22222222;
if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) {
/* Every alternate boundary strength value is used for chroma */
u4_bs &= 0x22222222;
}
for(col = 0; col < ctb_size / 8;)
for(col = 0; col < ctb_size / (4 * h_samp_factor);)
{
WORD8 i1_tc_offset_div2 = ps_slice_hdr->i1_tc_offset_div2;
bs_tz = CTZ(u4_bs) >> 2;
bs_tz = CTZ(u4_bs) >> h_samp_factor;
if(0 != bs_tz)
{
u4_bs = u4_bs >> (bs_tz << 2);
u4_bs = u4_bs >> (bs_tz << h_samp_factor);
if((col + bs_tz) >= (ctb_size / 8))
pu1_src += 8 * (ctb_size / 8 - col);
if((col + bs_tz) >= (ctb_size / (4 * h_samp_factor)))
pu1_src += 8 * (ctb_size / (4 * h_samp_factor) - col);
else
pu1_src += 8 * bs_tz;
@ -711,12 +732,12 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
{
qp_p = u4_qp_const_in_ctb[2] ?
pu1_qp[-ctb_size / 8] :
pu1_qp[(2 * row - 1) * qp_strd - 1];
pu1_qp[(v_samp_factor * row - 1) * qp_strd - 1];
}
qp_q = u4_qp_const_in_ctb[2] ?
pu1_qp[-ctb_size / 8] :
pu1_qp[(2 * row) * qp_strd - 1];
pu1_qp[(v_samp_factor * row) * qp_strd - 1];
}
else
{
@ -724,30 +745,31 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
{
qp_p = u4_qp_const_in_ctb[1] ?
pu1_qp[-ctb_size / 8 * qp_strd] :
pu1_qp[col - 1 - qp_strd];
pu1_qp[((col - 1) >> (2 - h_samp_factor)) - qp_strd];
}
else
{
qp_p = u4_qp_const_in_ctb[3] ?
pu1_qp[0] :
pu1_qp[(col - 1) + (2 * row - 1) * qp_strd];
pu1_qp[((col - 1) >> (2 - h_samp_factor)) + (v_samp_factor * row - 1) * qp_strd];
}
qp_q = u4_qp_const_in_ctb[3] ?
pu1_qp[0] :
pu1_qp[(col - 1) + 2 * row * qp_strd];
pu1_qp[((col - 1) >> (2 - h_samp_factor)) + v_samp_factor * row * qp_strd];
}
filter_p = (pu2_ctb_no_loop_filter_flag[row << 1] >> col) & 1;
filter_q = (pu2_ctb_no_loop_filter_flag[(row << 1) + 1] >> col) & 1;
filter_p = (pu2_ctb_no_loop_filter_flag[row * v_samp_factor] >> ((col + 2 - h_samp_factor) >> (2 - h_samp_factor))) & 1;
filter_q = (pu2_ctb_no_loop_filter_flag[(row * v_samp_factor) + 1] >> ((col + 2 - h_samp_factor) >> (2 - h_samp_factor))) & 1;
/* filter_p and filter_q are inverted as they are calculated using no_loop_filter_flags */
filter_p = !filter_p;
filter_q = !filter_q;
if(filter_p || filter_q)
{
ASSERT(1 == ((u4_bs & 3) >> 1));
DUMP_DEBLK_CHROMA_HORZ(pu1_src, src_strd,
if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) {
ASSERT(1 == ((u4_bs & 3) >> 1));
}
DUMP_DEBLK_CHROMA_HORZ(pu1_src, chroma_strd,
u4_bs & 3, qp_p, qp_q,
ps_pps->i1_pic_cb_qp_offset,
ps_pps->i1_pic_cr_qp_offset,
@ -756,37 +778,45 @@ void ihevcd_deblk_ctb(deblk_ctxt_t *ps_deblk,
if(chroma_yuv420sp_vu)
{
ps_codec->s_func_selector.ihevc_deblk_chroma_horz_fptr(pu1_src,
src_strd,
chroma_strd,
qp_q,
qp_p,
ps_pps->i1_pic_cr_qp_offset,
ps_pps->i1_pic_cb_qp_offset,
i1_tc_offset_div2,
filter_q,
filter_p);
filter_p,
ps_sps->i1_chroma_format_idc);
}
else
{
ps_codec->s_func_selector.ihevc_deblk_chroma_horz_fptr(pu1_src,
src_strd,
chroma_strd,
qp_p,
qp_q,
ps_pps->i1_pic_cb_qp_offset,
ps_pps->i1_pic_cr_qp_offset,
i1_tc_offset_div2,
filter_p,
filter_q);
filter_q,
ps_sps->i1_chroma_format_idc);
}
}
pu1_src += 8;
u4_bs = u4_bs >> 4;
u4_bs = u4_bs >> (2 * h_samp_factor);
col++;
}
pu4_horz_bs += (64 == ctb_size) ? 2 : 1;
pu1_src -= ctb_size;
pu1_src += 8 * src_strd;
if(is_yuv444)
{
if((64 == ctb_size) || ((32 == ctb_size) && (row & 1))) pu4_horz_bs++;
}
else
{
pu4_horz_bs += (64 == ctb_size) ? 2 : 1;
}
pu1_src -= ctb_size * (chroma_pixel_strd / h_samp_factor);
pu1_src += 8 * chroma_strd;
}
}

9
encoder/ihevce_deblk.c
View file

@ -455,7 +455,8 @@ void ihevce_deblk_ctb(
ps_deblk->i4_cr_qp_indx_offset,
ps_deblk->i4_tc_offset_div2,
filter_p,
filter_q);
filter_q,
ps_deblk->u1_chroma_array_type);
u4_bs = u4_bs << (1 << i4_log2_num_bits_per_bs);
pu1_curr_src += (ps_deblk->i4_chroma_pic_stride << 2);
@ -627,7 +628,8 @@ void ihevce_deblk_ctb(
ps_deblk->i4_cr_qp_indx_offset,
ps_deblk->i4_tc_offset_div2,
1,
1);
1,
ps_deblk->u1_chroma_array_type);
}
pu1_src_uv += 8;
@ -662,7 +664,8 @@ void ihevce_deblk_ctb(
ps_deblk->i4_cr_qp_indx_offset,
ps_deblk->i4_tc_offset_div2,
filter_p,
filter_q);
filter_q,
ps_deblk->u1_chroma_array_type);
pu1_src_uv += 8;
u4_bs = u4_bs << 4;