annotate x86/h264_deblock_sse2.asm @ 9692:9d103a3236e0 libavcodec

ARM: NEON VP3 Loop Filter
author conrad
date Sat, 23 May 2009 18:36:20 +0000
parents 37ac731fe32c
children c08ca946c80a
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
9006
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
1 ;*****************************************************************************
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
2 ;* MMX/SSE2-optimized H.264 deblocking code
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
3 ;*****************************************************************************
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
4 ;* Copyright (C) 2005-2008 x264 project
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
5 ;*
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
6 ;* Authors: Loren Merritt <lorenm@u.washington.edu>
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
7 ;*
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
8 ;* This program is free software; you can redistribute it and/or modify
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
9 ;* it under the terms of the GNU General Public License as published by
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
10 ;* the Free Software Foundation; either version 2 of the License, or
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
11 ;* (at your option) any later version.
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
12 ;*
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
13 ;* This program is distributed in the hope that it will be useful,
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
14 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
15 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
16 ;* GNU General Public License for more details.
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
17 ;*
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
18 ;* You should have received a copy of the GNU General Public License
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
19 ;* along with this program; if not, write to the Free Software
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
20 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
21 ;*****************************************************************************
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
22
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
23 %include "x86inc.asm"
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
24
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
25 SECTION_RODATA
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
26 pb_00: times 16 db 0x00
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
27 pb_01: times 16 db 0x01
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
28 pb_03: times 16 db 0x03
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
29 pb_a1: times 16 db 0xa1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
30
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
31 SECTION .text
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
32
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
33 ; expands to [base],...,[base+7*stride]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
34 %define PASS8ROWS(base, base3, stride, stride3) \
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
35 [base], [base+stride], [base+stride*2], [base3], \
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
36 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
37
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
38 ; in: 8 rows of 4 bytes in %1..%8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
39 ; out: 4 rows of 8 bytes in m0..m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
40 %macro TRANSPOSE4x8_LOAD 8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
41 movd m0, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
42 movd m2, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
43 movd m1, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
44 movd m3, %4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
45 punpcklbw m0, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
46 punpcklbw m1, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
47 movq m2, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
48 punpcklwd m0, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
49 punpckhwd m2, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
50
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
51 movd m4, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
52 movd m6, %6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
53 movd m5, %7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
54 movd m7, %8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
55 punpcklbw m4, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
56 punpcklbw m5, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
57 movq m6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
58 punpcklwd m4, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
59 punpckhwd m6, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
60
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
61 movq m1, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
62 movq m3, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
63 punpckldq m0, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
64 punpckhdq m1, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
65 punpckldq m2, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
66 punpckhdq m3, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
67 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
68
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
69 ; in: 4 rows of 8 bytes in m0..m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
70 ; out: 8 rows of 4 bytes in %1..%8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
71 %macro TRANSPOSE8x4_STORE 8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
72 movq m4, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
73 movq m5, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
74 movq m6, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
75 punpckhdq m4, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
76 punpckhdq m5, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
77 punpckhdq m6, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
78
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
79 punpcklbw m0, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
80 punpcklbw m2, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
81 movq m1, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
82 punpcklwd m0, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
83 punpckhwd m1, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
84 movd %1, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
85 punpckhdq m0, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
86 movd %2, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
87 movd %3, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
88 punpckhdq m1, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
89 movd %4, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
90
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
91 punpckhdq m3, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
92 punpcklbw m4, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
93 punpcklbw m6, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
94 movq m5, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
95 punpcklwd m4, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
96 punpckhwd m5, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
97 movd %5, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
98 punpckhdq m4, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
99 movd %6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
100 movd %7, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
101 punpckhdq m5, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
102 movd %8, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
103 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
104
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
105 %macro SBUTTERFLY 4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
106 movq %4, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
107 punpckl%1 %2, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
108 punpckh%1 %4, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
109 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
110
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
111 ; in: 8 rows of 8 (only the middle 6 pels are used) in %1..%8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
112 ; out: 6 rows of 8 in [%9+0*16] .. [%9+5*16]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
113 %macro TRANSPOSE6x8_MEM 9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
114 movq m0, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
115 movq m1, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
116 movq m2, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
117 movq m3, %4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
118 movq m4, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
119 movq m5, %6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
120 movq m6, %7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
121 SBUTTERFLY bw, m0, m1, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
122 SBUTTERFLY bw, m2, m3, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
123 SBUTTERFLY bw, m4, m5, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
124 movq [%9+0x10], m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
125 SBUTTERFLY bw, m6, %8, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
126 SBUTTERFLY wd, m0, m2, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
127 SBUTTERFLY wd, m4, m6, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
128 punpckhdq m0, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
129 movq [%9+0x00], m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
130 SBUTTERFLY wd, m7, [%9+0x10], m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
131 SBUTTERFLY wd, m3, m5, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
132 SBUTTERFLY dq, m7, m3, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
133 SBUTTERFLY dq, m1, m2, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
134 punpckldq m6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
135 movq [%9+0x10], m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
136 movq [%9+0x20], m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
137 movq [%9+0x30], m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
138 movq [%9+0x40], m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
139 movq [%9+0x50], m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
140 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
141
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
142 ; in: 8 rows of 8 in %1..%8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
143 ; out: 8 rows of 8 in %9..%16
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
144 %macro TRANSPOSE8x8_MEM 16
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
145 movq m0, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
146 movq m1, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
147 movq m2, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
148 movq m3, %4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
149 movq m4, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
150 movq m5, %6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
151 movq m6, %7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
152 SBUTTERFLY bw, m0, m1, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
153 SBUTTERFLY bw, m2, m3, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
154 SBUTTERFLY bw, m4, m5, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
155 SBUTTERFLY bw, m6, %8, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
156 movq %9, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
157 SBUTTERFLY wd, m0, m2, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
158 SBUTTERFLY wd, m4, m6, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
159 SBUTTERFLY wd, m7, m1, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
160 movq %11, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
161 movq m2, %9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
162 SBUTTERFLY wd, m2, m5, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
163 SBUTTERFLY dq, m0, m4, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
164 SBUTTERFLY dq, m7, m2, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
165 movq %9, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
166 movq %10, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
167 movq %13, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
168 movq %14, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
169 SBUTTERFLY dq, m3, %11, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
170 SBUTTERFLY dq, m6, m1, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
171 movq %11, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
172 movq %12, m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
173 movq %15, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
174 movq %16, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
175 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
176
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
177 ; out: %4 = |%1-%2|>%3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
178 ; clobbers: %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
179 %macro DIFF_GT 5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
180 mova %5, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
181 mova %4, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
182 psubusb %5, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
183 psubusb %4, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
184 por %4, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
185 psubusb %4, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
186 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
187
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
188 ; out: %4 = |%1-%2|>%3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
189 ; clobbers: %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
190 %macro DIFF_GT2 5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
191 mova %5, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
192 mova %4, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
193 psubusb %5, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
194 psubusb %4, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
195 psubusb %5, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
196 psubusb %4, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
197 pcmpeqb %4, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
198 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
199
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
200 %macro SPLATW 1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
201 %ifidn m0, xmm0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
202 pshuflw %1, %1, 0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
203 punpcklqdq %1, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
204 %else
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
205 pshufw %1, %1, 0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
206 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
207 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
208
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
209 ; in: m0=p1 m1=p0 m2=q0 m3=q1 %1=alpha-1 %2=beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
210 ; out: m5=beta-1, m7=mask, %3=alpha-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
211 ; clobbers: m4,m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
212 %macro LOAD_MASK 2-3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
213 movd m4, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
214 movd m5, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
215 SPLATW m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
216 SPLATW m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
217 packuswb m4, m4 ; 16x alpha-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
218 packuswb m5, m5 ; 16x beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
219 %if %0>2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
220 mova %3, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
221 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
222 DIFF_GT m1, m2, m4, m7, m6 ; |p0-q0| > alpha-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
223 DIFF_GT m0, m1, m5, m4, m6 ; |p1-p0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
224 por m7, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
225 DIFF_GT m3, m2, m5, m4, m6 ; |q1-q0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
226 por m7, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
227 pxor m6, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
228 pcmpeqb m7, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
229 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
230
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
231 ; in: m0=p1 m1=p0 m2=q0 m3=q1 m7=(tc&mask)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
232 ; out: m1=p0' m2=q0'
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
233 ; clobbers: m0,3-6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
234 %macro DEBLOCK_P0_Q0 0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
235 mova m5, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
236 pxor m5, m2 ; p0^q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
237 pand m5, [pb_01 GLOBAL] ; (p0^q0)&1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
238 pcmpeqb m4, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
239 pxor m3, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
240 pavgb m3, m0 ; (p1 - q1 + 256)>>1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
241 pavgb m3, [pb_03 GLOBAL] ; (((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
242 pxor m4, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
243 pavgb m4, m2 ; (q0 - p0 + 256)>>1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
244 pavgb m3, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
245 paddusb m3, m4 ; d+128+33
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
246 mova m6, [pb_a1 GLOBAL]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
247 psubusb m6, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
248 psubusb m3, [pb_a1 GLOBAL]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
249 pminub m6, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
250 pminub m3, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
251 psubusb m1, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
252 psubusb m2, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
253 paddusb m1, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
254 paddusb m2, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
255 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
256
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
257 ; in: m1=p0 m2=q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
258 ; %1=p1 %2=q2 %3=[q2] %4=[q1] %5=tc0 %6=tmp
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
259 ; out: [q1] = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
260 ; clobbers: q2, tmp, tc0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
261 %macro LUMA_Q1 6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
262 mova %6, m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
263 pavgb %6, m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
264 pavgb %2, %6 ; avg(p2,avg(p0,q0))
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
265 pxor %6, %3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
266 pand %6, [pb_01 GLOBAL] ; (p2^avg(p0,q0))&1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
267 psubusb %2, %6 ; (p2+((p0+q0+1)>>1))>>1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
268 mova %6, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
269 psubusb %6, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
270 paddusb %5, %1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
271 pmaxub %2, %6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
272 pminub %2, %5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
273 mova %4, %2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
274 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
275
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
276 %ifdef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
277 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
278 ; void x264_deblock_v_luma_sse2( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
279 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
280 INIT_XMM
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
281 cglobal x264_deblock_v_luma_sse2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
282 movd m8, [r4] ; tc0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
283 lea r4, [r1*3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
284 dec r2d ; alpha-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
285 neg r4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
286 dec r3d ; beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
287 add r4, r0 ; pix-3*stride
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
288
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
289 mova m0, [r4+r1] ; p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
290 mova m1, [r4+2*r1] ; p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
291 mova m2, [r0] ; q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
292 mova m3, [r0+r1] ; q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
293 LOAD_MASK r2d, r3d
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
294
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
295 punpcklbw m8, m8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
296 punpcklbw m8, m8 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
297 pcmpeqb m9, m9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
298 pcmpeqb m9, m8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
299 pandn m9, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
300 pand m8, m9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
301
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
302 movdqa m3, [r4] ; p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
303 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
304 pand m6, m9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
305 mova m7, m8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
306 psubb m7, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
307 pand m6, m8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
308 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
309
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
310 movdqa m4, [r0+2*r1] ; q2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
311 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
312 pand m6, m9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
313 pand m8, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
314 psubb m7, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
315 mova m3, [r0+r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
316 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m8, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
317
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
318 DEBLOCK_P0_Q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
319 mova [r4+2*r1], m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
320 mova [r0], m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
321 ret
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
322
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
323 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
324 ; void x264_deblock_h_luma_sse2( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
325 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
326 INIT_MMX
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
327 cglobal x264_deblock_h_luma_sse2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
328 movsxd r10, esi
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
329 lea r11, [r10+r10*2]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
330 lea rax, [r0-4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
331 lea r9, [r0-4+r11]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
332 sub rsp, 0x68
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
333 %define pix_tmp rsp
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
334
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
335 ; transpose 6x16 -> tmp space
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
336 TRANSPOSE6x8_MEM PASS8ROWS(rax, r9, r10, r11), pix_tmp
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
337 lea rax, [rax+r10*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
338 lea r9, [r9 +r10*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
339 TRANSPOSE6x8_MEM PASS8ROWS(rax, r9, r10, r11), pix_tmp+8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
340
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
341 ; vertical filter
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
342 ; alpha, beta, tc0 are still in r2d, r3d, r4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
343 ; don't backup rax, r9, r10, r11 because x264_deblock_v_luma_sse2 doesn't use them
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
344 lea r0, [pix_tmp+0x30]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
345 mov esi, 0x10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
346 call x264_deblock_v_luma_sse2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
347
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
348 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
349 add rax, 2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
350 add r9, 2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
351 movq m0, [pix_tmp+0x18]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
352 movq m1, [pix_tmp+0x28]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
353 movq m2, [pix_tmp+0x38]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
354 movq m3, [pix_tmp+0x48]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
355 TRANSPOSE8x4_STORE PASS8ROWS(rax, r9, r10, r11)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
356
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
357 shl r10, 3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
358 sub rax, r10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
359 sub r9, r10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
360 shr r10, 3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
361 movq m0, [pix_tmp+0x10]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
362 movq m1, [pix_tmp+0x20]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
363 movq m2, [pix_tmp+0x30]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
364 movq m3, [pix_tmp+0x40]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
365 TRANSPOSE8x4_STORE PASS8ROWS(rax, r9, r10, r11)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
366
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
367 add rsp, 0x68
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
368 ret
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
369
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
370 %else
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
371
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
372 %macro DEBLOCK_LUMA 3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
373 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
374 ; void x264_deblock_v8_luma_mmxext( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
375 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
376 cglobal x264_deblock_%2_luma_%1, 5,5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
377 lea r4, [r1*3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
378 dec r2 ; alpha-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
379 neg r4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
380 dec r3 ; beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
381 add r4, r0 ; pix-3*stride
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
382 %assign pad 2*%3+12-(stack_offset&15)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
383 SUB esp, pad
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
384
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
385 mova m0, [r4+r1] ; p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
386 mova m1, [r4+2*r1] ; p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
387 mova m2, [r0] ; q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
388 mova m3, [r0+r1] ; q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
389 LOAD_MASK r2, r3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
390
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
391 mov r3, r4m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
392 movd m4, [r3] ; tc0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
393 punpcklbw m4, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
394 punpcklbw m4, m4 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
395 mova [esp+%3], m4 ; tc
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
396 pcmpeqb m3, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
397 pcmpgtb m4, m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
398 pand m4, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
399 mova [esp], m4 ; mask
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
400
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
401 mova m3, [r4] ; p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
402 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
403 pand m6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
404 pand m4, [esp+%3] ; tc
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
405 mova m7, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
406 psubb m7, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
407 pand m6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
408 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
409
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
410 mova m4, [r0+2*r1] ; q2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
411 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
412 mova m5, [esp] ; mask
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
413 pand m6, m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
414 mova m5, [esp+%3] ; tc
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
415 pand m5, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
416 psubb m7, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
417 mova m3, [r0+r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
418 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m5, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
419
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
420 DEBLOCK_P0_Q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
421 mova [r4+2*r1], m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
422 mova [r0], m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
423 ADD esp, pad
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
424 RET
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
425
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
426 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
427 ; void x264_deblock_h_luma_mmxext( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
428 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
429 INIT_MMX
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
430 cglobal x264_deblock_h_luma_%1, 0,5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
431 mov r0, r0m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
432 mov r3, r1m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
433 lea r4, [r3*3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
434 sub r0, 4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
435 lea r1, [r0+r4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
436 %assign pad 0x78-(stack_offset&15)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
437 SUB esp, pad
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
438 %define pix_tmp esp+12
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
439
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
440 ; transpose 6x16 -> tmp space
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
441 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
442 lea r0, [r0+r3*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
443 lea r1, [r1+r3*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
444 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp+8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
445
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
446 ; vertical filter
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
447 lea r0, [pix_tmp+0x30]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
448 PUSH dword r4m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
449 PUSH dword r3m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
450 PUSH dword r2m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
451 PUSH dword 16
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
452 PUSH dword r0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
453 call x264_deblock_%2_luma_%1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
454 %ifidn %2, v8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
455 add dword [esp ], 8 ; pix_tmp+0x38
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
456 add dword [esp+16], 2 ; tc0+2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
457 call x264_deblock_%2_luma_%1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
458 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
459 ADD esp, 20
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
460
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
461 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
462 mov r0, r0m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
463 sub r0, 2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
464 lea r1, [r0+r4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
465
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
466 movq m0, [pix_tmp+0x10]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
467 movq m1, [pix_tmp+0x20]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
468 movq m2, [pix_tmp+0x30]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
469 movq m3, [pix_tmp+0x40]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
470 TRANSPOSE8x4_STORE PASS8ROWS(r0, r1, r3, r4)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
471
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
472 lea r0, [r0+r3*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
473 lea r1, [r1+r3*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
474 movq m0, [pix_tmp+0x18]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
475 movq m1, [pix_tmp+0x28]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
476 movq m2, [pix_tmp+0x38]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
477 movq m3, [pix_tmp+0x48]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
478 TRANSPOSE8x4_STORE PASS8ROWS(r0, r1, r3, r4)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
479
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
480 ADD esp, pad
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
481 RET
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
482 %endmacro ; DEBLOCK_LUMA
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
483
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
484 INIT_XMM
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
485 DEBLOCK_LUMA sse2, v, 16
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
486
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
487 %endif ; ARCH
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
488
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
489
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
490
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
491 %macro LUMA_INTRA_P012 4 ; p0..p3 in memory
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
492 mova t0, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
493 mova t1, p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
494 pavgb t0, p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
495 pavgb t1, q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
496 pavgb t0, t1 ; ((p2+p1+1)/2 + (p0+q0+1)/2 + 1)/2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
497 mova t5, t1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
498 mova t2, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
499 mova t3, p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
500 paddb t2, p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
501 paddb t3, q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
502 paddb t2, t3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
503 mova t3, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
504 mova t4, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
505 psrlw t2, 1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
506 pavgb t2, mpb_00
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
507 pxor t2, t0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
508 pand t2, mpb_01
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
509 psubb t0, t2 ; p1' = (p2+p1+p0+q0+2)/4;
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
510
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
511 mova t1, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
512 mova t2, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
513 pavgb t1, q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
514 psubb t2, q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
515 paddb t3, t3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
516 psubb t3, t2 ; p2+2*p1+2*p0+2*q0+q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
517 pand t2, mpb_01
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
518 psubb t1, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
519 pavgb t1, p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
520 pavgb t1, t5 ; (((p2+q1)/2 + p1+1)/2 + (p0+q0+1)/2 + 1)/2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
521 psrlw t3, 2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
522 pavgb t3, mpb_00
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
523 pxor t3, t1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
524 pand t3, mpb_01
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
525 psubb t1, t3 ; p0'a = (p2+2*p1+2*p0+2*q0+q1+4)/8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
526
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
527 mova t3, p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
528 mova t2, p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
529 pxor t3, q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
530 pavgb t2, q1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
531 pand t3, mpb_01
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
532 psubb t2, t3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
533 pavgb t2, p1 ; p0'b = (2*p1+p0+q0+2)/4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
534
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
535 pxor t1, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
536 pxor t2, p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
537 pand t1, mask1p
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
538 pand t2, mask0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
539 pxor t1, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
540 pxor t1, p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
541 mova %1, t1 ; store p0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
542
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
543 mova t1, %4 ; p3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
544 mova t2, t1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
545 pavgb t1, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
546 paddb t2, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
547 pavgb t1, t0 ; (p3+p2+1)/2 + (p2+p1+p0+q0+2)/4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
548 paddb t2, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
549 paddb t2, t4 ; 2*p3+3*p2+p1+p0+q0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
550 psrlw t2, 2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
551 pavgb t2, mpb_00
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
552 pxor t2, t1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
553 pand t2, mpb_01
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
554 psubb t1, t2 ; p2' = (2*p3+3*p2+p1+p0+q0+4)/8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
555
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
556 pxor t0, p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
557 pxor t1, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
558 pand t0, mask1p
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
559 pand t1, mask1p
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
560 pxor t0, p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
561 pxor t1, p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
562 mova %2, t0 ; store p1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
563 mova %3, t1 ; store p2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
564 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
565
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
566 %macro LUMA_INTRA_SWAP_PQ 0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
567 %define q1 m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
568 %define q0 m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
569 %define p0 m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
570 %define p1 m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
571 %define p2 q2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
572 %define mask1p mask1q
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
573 %endmacro
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
574
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
575 %macro DEBLOCK_LUMA_INTRA 2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
576 %define p1 m0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
577 %define p0 m1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
578 %define q0 m2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
579 %define q1 m3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
580 %define t0 m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
581 %define t1 m5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
582 %define t2 m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
583 %define t3 m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
584 %ifdef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
585 %define p2 m8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
586 %define q2 m9
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
587 %define t4 m10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
588 %define t5 m11
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
589 %define mask0 m12
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
590 %define mask1p m13
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
591 %define mask1q [rsp-24]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
592 %define mpb_00 m14
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
593 %define mpb_01 m15
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
594 %else
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
595 %define spill(x) [esp+16*x+((stack_offset+4)&15)]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
596 %define p2 [r4+r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
597 %define q2 [r0+2*r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
598 %define t4 spill(0)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
599 %define t5 spill(1)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
600 %define mask0 spill(2)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
601 %define mask1p spill(3)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
602 %define mask1q spill(4)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
603 %define mpb_00 [pb_00 GLOBAL]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
604 %define mpb_01 [pb_01 GLOBAL]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
605 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
606
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
607 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
608 ; void x264_deblock_v_luma_intra_sse2( uint8_t *pix, int stride, int alpha, int beta )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
609 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
610 cglobal x264_deblock_%2_luma_intra_%1, 4,6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
611 %ifndef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
612 sub esp, 0x60
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
613 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
614 lea r4, [r1*4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
615 lea r5, [r1*3] ; 3*stride
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
616 dec r2d ; alpha-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
617 jl .end
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
618 neg r4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
619 dec r3d ; beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
620 jl .end
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
621 add r4, r0 ; pix-4*stride
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
622 mova p1, [r4+2*r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
623 mova p0, [r4+r5]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
624 mova q0, [r0]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
625 mova q1, [r0+r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
626 %ifdef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
627 pxor mpb_00, mpb_00
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
628 mova mpb_01, [pb_01 GLOBAL]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
629 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
630 SWAP 7, 12 ; m12=mask0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
631 pavgb t5, mpb_00
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
632 pavgb t5, mpb_01 ; alpha/4+1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
633 movdqa p2, [r4+r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
634 movdqa q2, [r0+2*r1]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
635 DIFF_GT2 p0, q0, t5, t0, t3 ; t0 = |p0-q0| > alpha/4+1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
636 DIFF_GT2 p0, p2, m5, t2, t5 ; mask1 = |p2-p0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
637 DIFF_GT2 q0, q2, m5, t4, t5 ; t4 = |q2-q0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
638 pand t0, mask0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
639 pand t4, t0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
640 pand t2, t0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
641 mova mask1q, t4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
642 mova mask1p, t2
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
643 %else
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
644 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
645 mova m4, t5
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
646 mova mask0, m7
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
647 pavgb m4, [pb_00 GLOBAL]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
648 pavgb m4, [pb_01 GLOBAL] ; alpha/4+1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
649 DIFF_GT2 p0, q0, m4, m6, m7 ; m6 = |p0-q0| > alpha/4+1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
650 pand m6, mask0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
651 DIFF_GT2 p0, p2, m5, m4, m7 ; m4 = |p2-p0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
652 pand m4, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
653 mova mask1p, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
654 DIFF_GT2 q0, q2, m5, m4, m7 ; m4 = |q2-q0| > beta-1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
655 pand m4, m6
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
656 mova mask1q, m4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
657 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
658 LUMA_INTRA_P012 [r4+r5], [r4+2*r1], [r4+r1], [r4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
659 LUMA_INTRA_SWAP_PQ
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
660 LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
661 .end:
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
662 %ifndef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
663 add esp, 0x60
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
664 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
665 RET
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
666
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
667 INIT_MMX
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
668 %ifdef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
669 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
670 ; void x264_deblock_h_luma_intra_sse2( uint8_t *pix, int stride, int alpha, int beta )
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
671 ;-----------------------------------------------------------------------------
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
672 cglobal x264_deblock_h_luma_intra_%1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
673 movsxd r10, r1d
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
674 lea r11, [r10*3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
675 lea rax, [r0-4]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
676 lea r9, [r0-4+r11]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
677 sub rsp, 0x88
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
678 %define pix_tmp rsp
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
679
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
680 ; transpose 8x16 -> tmp space
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
681 TRANSPOSE8x8_MEM PASS8ROWS(rax, r9, r10, r11), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
682 lea rax, [rax+r10*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
683 lea r9, [r9+r10*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
684 TRANSPOSE8x8_MEM PASS8ROWS(rax, r9, r10, r11), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
685
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
686 lea r0, [pix_tmp+0x40]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
687 mov r1, 0x10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
688 call x264_deblock_v_luma_intra_%1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
689
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
690 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
691 lea r9, [rax+r11]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
692 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(rax, r9, r10, r11)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
693 shl r10, 3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
694 sub rax, r10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
695 sub r9, r10
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
696 shr r10, 3
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
697 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(rax, r9, r10, r11)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
698 add rsp, 0x88
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
699 ret
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
700 %else
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
701 cglobal x264_deblock_h_luma_intra_%1, 2,4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
702 lea r3, [r1*3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
703 sub r0, 4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
704 lea r2, [r0+r3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
705 %assign pad 0x8c-(stack_offset&15)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
706 SUB rsp, pad
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
707 %define pix_tmp rsp
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
708
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
709 ; transpose 8x16 -> tmp space
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
710 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
711 lea r0, [r0+r1*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
712 lea r2, [r2+r1*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
713 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
714
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
715 lea r0, [pix_tmp+0x40]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
716 PUSH dword r3m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
717 PUSH dword r2m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
718 PUSH dword 16
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
719 PUSH r0
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
720 call x264_deblock_%2_luma_intra_%1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
721 %ifidn %2, v8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
722 add dword [rsp], 8 ; pix_tmp+8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
723 call x264_deblock_%2_luma_intra_%1
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
724 %endif
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
725 ADD esp, 16
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
726
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
727 mov r1, r1m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
728 mov r0, r0m
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
729 lea r3, [r1*3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
730 sub r0, 4
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
731 lea r2, [r0+r3]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
732 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
733 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
734 lea r0, [r0+r1*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
735 lea r2, [r2+r1*8]
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
736 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
737 ADD rsp, pad
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
738 RET
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
739 %endif ; ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
740 %endmacro ; DEBLOCK_LUMA_INTRA
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
741
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
742 INIT_XMM
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
743 DEBLOCK_LUMA_INTRA sse2, v
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
744 %ifndef ARCH_X86_64
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
745 INIT_MMX
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
746 DEBLOCK_LUMA_INTRA mmxext, v8
37ac731fe32c Convert x264 asm files to proper unix line breaks
darkshikari
parents: 9005
diff changeset
747 %endif