annotate x86/dsputil_mmx.c @ 12456:a5ddb39627fd libavcodec

Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_ symbols, and move them from libavcodec/avcodec.h to libavutil/cpu.h.
author stefano
date Sat, 04 Sep 2010 09:59:08 +0000
parents f4355cd85faa
children 9fef0a8ddd63
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1 /*
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2 * MMX optimized DSP utils
8629
04423b2f6e0b cosmetics: Remove pointless period after copyright statement non-sentences.
diego
parents: 8596
diff changeset
3 * Copyright (c) 2000, 2001 Fabrice Bellard
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
5 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
6 * This file is part of FFmpeg.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
7 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
8 * FFmpeg is free software; you can redistribute it and/or
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
9 * modify it under the terms of the GNU Lesser General Public
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
10 * License as published by the Free Software Foundation; either
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
11 * version 2.1 of the License, or (at your option) any later version.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
12 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
13 * FFmpeg is distributed in the hope that it will be useful,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
16 * Lesser General Public License for more details.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
17 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
18 * You should have received a copy of the GNU Lesser General Public
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
19 * License along with FFmpeg; if not, write to the Free Software
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
21 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
22 * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
23 */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
24
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
25 #include "libavutil/cpu.h"
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
26 #include "libavutil/x86_cpu.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
27 #include "libavcodec/dsputil.h"
11499
2a4dc3c0b012 Move H264 dsputil functions into their own struct
mru
parents: 11485
diff changeset
28 #include "libavcodec/h264dsp.h"
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
29 #include "libavcodec/mpegvideo.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
30 #include "libavcodec/simple_idct.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
31 #include "dsputil_mmx.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
32 #include "idct_xvid.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
33
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
34 //#undef NDEBUG
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
35 //#include <assert.h>
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
36
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
37 /* pixel operations */
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
38 DECLARE_ALIGNED(8, const uint64_t, ff_bone) = 0x0101010101010101ULL;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
39 DECLARE_ALIGNED(8, const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
40
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
41 DECLARE_ALIGNED(16, const uint64_t, ff_pdw_80000000)[2] =
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
42 {0x8000000080000000ULL, 0x8000000080000000ULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
43
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
44 DECLARE_ALIGNED(8, const uint64_t, ff_pw_3 ) = 0x0003000300030003ULL;
12143
fa452b243aa6 Make ff_pw_4 128 bits
conrad
parents: 12086
diff changeset
45 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_4 ) = {0x0004000400040004ULL, 0x0004000400040004ULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
46 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_5 ) = {0x0005000500050005ULL, 0x0005000500050005ULL};
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
47 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_8 ) = {0x0008000800080008ULL, 0x0008000800080008ULL};
12205
d38e8565ba05 VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents: 12168
diff changeset
48 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_9 ) = {0x0009000900090009ULL, 0x0009000900090009ULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
49 DECLARE_ALIGNED(8, const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
50 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_16 ) = {0x0010001000100010ULL, 0x0010001000100010ULL};
12205
d38e8565ba05 VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents: 12168
diff changeset
51 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_18 ) = {0x0012001200120012ULL, 0x0012001200120012ULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
52 DECLARE_ALIGNED(8, const uint64_t, ff_pw_20 ) = 0x0014001400140014ULL;
12205
d38e8565ba05 VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents: 12168
diff changeset
53 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_27 ) = {0x001B001B001B001BULL, 0x001B001B001B001BULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
54 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_28 ) = {0x001C001C001C001CULL, 0x001C001C001C001CULL};
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
55 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL};
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
56 DECLARE_ALIGNED(8, const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL;
12206
fe243bb5ef61 Move ff_pw_* from vc1dsp_mmx.c to dsputil_mmx.c
conrad
parents: 12205
diff changeset
57 DECLARE_ALIGNED(8, const uint64_t, ff_pw_53 ) = 0x0035003500350035ULL;
12205
d38e8565ba05 VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents: 12168
diff changeset
58 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_63 ) = {0x003F003F003F003FULL, 0x003F003F003F003FULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
59 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_64 ) = {0x0040004000400040ULL, 0x0040004000400040ULL};
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
60 DECLARE_ALIGNED(8, const uint64_t, ff_pw_96 ) = 0x0060006000600060ULL;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
61 DECLARE_ALIGNED(8, const uint64_t, ff_pw_128) = 0x0080008000800080ULL;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
62 DECLARE_ALIGNED(8, const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
63
12454
f4355cd85faa Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents: 12450
diff changeset
64 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_0 ) = {0x0000000000000000ULL, 0x0000000000000000ULL};
12168
b246b214c2e9 VP8 H/V inner loopfilter MMX/MMXEXT/SSE2 optimizations.
rbultje
parents: 12143
diff changeset
65 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_1 ) = {0x0101010101010101ULL, 0x0101010101010101ULL};
11951
afee30fe8c26 16x16 and 8x8c x86 SIMD intra pred functions for VP8 and H.264
darkshikari
parents: 11826
diff changeset
66 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_3 ) = {0x0303030303030303ULL, 0x0303030303030303ULL};
12086
d780ae746855 Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents: 11981
diff changeset
67 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_4 ) = {0x0404040404040404ULL, 0x0404040404040404ULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
68 DECLARE_ALIGNED(8, const uint64_t, ff_pb_7 ) = 0x0707070707070707ULL;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
69 DECLARE_ALIGNED(8, const uint64_t, ff_pb_1F ) = 0x1F1F1F1F1F1F1F1FULL;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
70 DECLARE_ALIGNED(8, const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL;
12086
d780ae746855 Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents: 11981
diff changeset
71 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_80 ) = {0x8080808080808080ULL, 0x8080808080808080ULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
72 DECLARE_ALIGNED(8, const uint64_t, ff_pb_81 ) = 0x8181818181818181ULL;
12454
f4355cd85faa Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents: 12450
diff changeset
73 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_A1 ) = {0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL};
12086
d780ae746855 Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents: 11981
diff changeset
74 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_F8 ) = {0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL};
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
75 DECLARE_ALIGNED(8, const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL;
12086
d780ae746855 Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents: 11981
diff changeset
76 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_FE ) = {0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL};
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
77
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
78 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
79 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
80
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
81 #define JUMPALIGN() __asm__ volatile (ASMALIGN(3)::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
82 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%" #regd ", %%" #regd ::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
83
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
84 #define MOVQ_BFE(regd) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
85 __asm__ volatile ( \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
86 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
87 "paddb %%" #regd ", %%" #regd " \n\t" ::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
88
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
89 #ifndef PIC
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
90 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
91 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
92 #else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
93 // for shared library it's better to use this way for accessing constants
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
94 // pcmpeqd -> -1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
95 #define MOVQ_BONE(regd) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
96 __asm__ volatile ( \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
97 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
98 "psrlw $15, %%" #regd " \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
99 "packuswb %%" #regd ", %%" #regd " \n\t" ::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
100
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
101 #define MOVQ_WTWO(regd) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
102 __asm__ volatile ( \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
103 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
104 "psrlw $15, %%" #regd " \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
105 "psllw $1, %%" #regd " \n\t"::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
106
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
107 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
108
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
109 // using regr as temporary and for the output result
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
110 // first argument is unmodifed and second is trashed
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
111 // regfe is supposed to contain 0xfefefefefefefefe
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
112 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
113 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
114 "pand " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
115 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
116 "pand " #regfe "," #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
117 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
118 "paddb " #regb ", " #regr " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
119
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
120 #define PAVGB_MMX(rega, regb, regr, regfe) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
121 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
122 "por " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
123 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
124 "pand " #regfe "," #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
125 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
126 "psubb " #regb ", " #regr " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
127
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
128 // mm6 is supposed to contain 0xfefefefefefefefe
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
129 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
130 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
131 "movq " #regc ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
132 "pand " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
133 "pand " #regd ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
134 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
135 "pxor " #regc ", " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
136 "pand %%mm6, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
137 "pand %%mm6, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
138 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
139 "psrlq $1, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
140 "paddb " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
141 "paddb " #regd ", " #regp " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
142
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
143 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
144 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
145 "movq " #regc ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
146 "por " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
147 "por " #regd ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
148 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
149 "pxor " #regc ", " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
150 "pand %%mm6, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
151 "pand %%mm6, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
152 "psrlq $1, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
153 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
154 "psubb " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
155 "psubb " #regd ", " #regp " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
156
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
157 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
158 /* MMX no rounding */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
159 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
160 #define SET_RND MOVQ_WONE
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
161 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
162 #define PAVGB(a, b, c, e) PAVGB_MMX_NO_RND(a, b, c, e)
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
163 #define OP_AVG(a, b, c, e) PAVGB_MMX(a, b, c, e)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
164
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
165 #include "dsputil_mmx_rnd_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
166
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
167 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
168 #undef SET_RND
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
169 #undef PAVGBP
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
170 #undef PAVGB
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
171 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
172 /* MMX rounding */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
173
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
174 #define DEF(x, y) x ## _ ## y ##_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
175 #define SET_RND MOVQ_WTWO
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
176 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
177 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
178
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
179 #include "dsputil_mmx_rnd_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
180
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
181 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
182 #undef SET_RND
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
183 #undef PAVGBP
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
184 #undef PAVGB
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
185 #undef OP_AVG
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
186
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
187 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
188 /* 3Dnow specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
189
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
190 #define DEF(x) x ## _3dnow
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
191 #define PAVGB "pavgusb"
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
192 #define OP_AVG PAVGB
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
193
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
194 #include "dsputil_mmx_avg_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
195
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
196 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
197 #undef PAVGB
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
198 #undef OP_AVG
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
199
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
200 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
201 /* MMX2 specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
202
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
203 #define DEF(x) x ## _mmx2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
204
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
205 /* Introduced only in MMX2 set */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
206 #define PAVGB "pavgb"
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
207 #define OP_AVG PAVGB
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
208
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
209 #include "dsputil_mmx_avg_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
210
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
211 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
212 #undef PAVGB
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
213 #undef OP_AVG
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
214
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
215 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
216 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
217 #define put_pixels16_mmx2 put_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
218 #define put_pixels8_mmx2 put_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
219 #define put_pixels4_mmx2 put_pixels4_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
220 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
221 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
222 #define put_pixels16_3dnow put_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
223 #define put_pixels8_3dnow put_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
224 #define put_pixels4_3dnow put_pixels4_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
225 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
226 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
227
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
228 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
229 /* standard MMX */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
230
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
231 void ff_put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
232 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
233 const DCTELEM *p;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
234 uint8_t *pix;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
235
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
236 /* read the pixels */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
237 p = block;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
238 pix = pixels;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
239 /* unrolled loop */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
240 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
241 "movq %3, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
242 "movq 8%3, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
243 "movq 16%3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
244 "movq 24%3, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
245 "movq 32%3, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
246 "movq 40%3, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
247 "movq 48%3, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
248 "movq 56%3, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
249 "packuswb %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
250 "packuswb %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
251 "packuswb %%mm5, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
252 "packuswb %%mm7, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
253 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
254 "movq %%mm2, (%0, %1) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
255 "movq %%mm4, (%0, %1, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
256 "movq %%mm6, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
257 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "m"(*p)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
258 :"memory");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
259 pix += line_size*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
260 p += 32;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
261
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
262 // if here would be an exact copy of the code above
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
263 // compiler would generate some very strange code
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
264 // thus using "r"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
265 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
266 "movq (%3), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
267 "movq 8(%3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
268 "movq 16(%3), %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
269 "movq 24(%3), %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
270 "movq 32(%3), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
271 "movq 40(%3), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
272 "movq 48(%3), %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
273 "movq 56(%3), %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
274 "packuswb %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
275 "packuswb %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
276 "packuswb %%mm5, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
277 "packuswb %%mm7, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
278 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
279 "movq %%mm2, (%0, %1) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
280 "movq %%mm4, (%0, %1, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
281 "movq %%mm6, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
282 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "r"(p)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
283 :"memory");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
284 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
285
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
286 DECLARE_ASM_CONST(8, uint8_t, ff_vector128)[8] =
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
287 { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
288
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
289 #define put_signed_pixels_clamped_mmx_half(off) \
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
290 "movq "#off"(%2), %%mm1 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
291 "movq 16+"#off"(%2), %%mm2 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
292 "movq 32+"#off"(%2), %%mm3 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
293 "movq 48+"#off"(%2), %%mm4 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
294 "packsswb 8+"#off"(%2), %%mm1 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
295 "packsswb 24+"#off"(%2), %%mm2 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
296 "packsswb 40+"#off"(%2), %%mm3 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
297 "packsswb 56+"#off"(%2), %%mm4 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
298 "paddb %%mm0, %%mm1 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
299 "paddb %%mm0, %%mm2 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
300 "paddb %%mm0, %%mm3 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
301 "paddb %%mm0, %%mm4 \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
302 "movq %%mm1, (%0) \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
303 "movq %%mm2, (%0, %3) \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
304 "movq %%mm3, (%0, %3, 2) \n\t"\
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
305 "movq %%mm4, (%0, %1) \n\t"
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
306
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
307 void ff_put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
308 {
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
309 x86_reg line_skip = line_size;
9341
06532529c428 Mark line_skip3 asm argument as output-only instead of using av_uninit.
reimar
parents: 9340
diff changeset
310 x86_reg line_skip3;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
311
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
312 __asm__ volatile (
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
313 "movq "MANGLE(ff_vector128)", %%mm0 \n\t"
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
314 "lea (%3, %3, 2), %1 \n\t"
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
315 put_signed_pixels_clamped_mmx_half(0)
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
316 "lea (%0, %3, 4), %0 \n\t"
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
317 put_signed_pixels_clamped_mmx_half(64)
9341
06532529c428 Mark line_skip3 asm argument as output-only instead of using av_uninit.
reimar
parents: 9340
diff changeset
318 :"+&r" (pixels), "=&r" (line_skip3)
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
319 :"r" (block), "r"(line_skip)
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
320 :"memory");
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
321 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
322
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
323 void ff_add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
324 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
325 const DCTELEM *p;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
326 uint8_t *pix;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
327 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
328
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
329 /* read the pixels */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
330 p = block;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
331 pix = pixels;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
332 MOVQ_ZERO(mm7);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
333 i = 4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
334 do {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
335 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
336 "movq (%2), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
337 "movq 8(%2), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
338 "movq 16(%2), %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
339 "movq 24(%2), %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
340 "movq %0, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
341 "movq %1, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
342 "movq %%mm4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
343 "punpcklbw %%mm7, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
344 "punpckhbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
345 "paddsw %%mm4, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
346 "paddsw %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
347 "movq %%mm6, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
348 "punpcklbw %%mm7, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
349 "punpckhbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
350 "paddsw %%mm6, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
351 "paddsw %%mm5, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
352 "packuswb %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
353 "packuswb %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
354 "movq %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
355 "movq %%mm2, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
356 :"+m"(*pix), "+m"(*(pix+line_size))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
357 :"r"(p)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
358 :"memory");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
359 pix += line_size*2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
360 p += 16;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
361 } while (--i);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
362 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
363
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
364 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
365 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
366 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
367 "lea (%3, %3), %%"REG_a" \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
368 ASMALIGN(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
369 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
370 "movd (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
371 "movd (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
372 "movd %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
373 "movd %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
374 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
375 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
376 "movd (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
377 "movd (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
378 "movd %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
379 "movd %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
380 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
381 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
382 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
383 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
384 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
385 : "r"((x86_reg)line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
386 : "%"REG_a, "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
387 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
388 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
389
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
390 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
391 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
392 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
393 "lea (%3, %3), %%"REG_a" \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
394 ASMALIGN(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
395 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
396 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
397 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
398 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
399 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
400 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
401 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
402 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
403 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
404 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
405 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
406 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
407 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
408 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
409 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
410 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
411 : "r"((x86_reg)line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
412 : "%"REG_a, "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
413 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
414 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
415
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
416 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
417 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
418 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
419 "lea (%3, %3), %%"REG_a" \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
420 ASMALIGN(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
421 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
422 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
423 "movq 8(%1), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
424 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
425 "movq 8(%1, %3), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
426 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
427 "movq %%mm4, 8(%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
428 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
429 "movq %%mm5, 8(%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
430 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
431 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
432 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
433 "movq 8(%1), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
434 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
435 "movq 8(%1, %3), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
436 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
437 "movq %%mm4, 8(%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
438 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
439 "movq %%mm5, 8(%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
440 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
441 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
442 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
443 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
444 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
445 : "r"((x86_reg)line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
446 : "%"REG_a, "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
447 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
448 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
449
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
450 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
451 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
452 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
453 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
454 "movdqu (%1), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
455 "movdqu (%1,%3), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
456 "movdqu (%1,%3,2), %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
457 "movdqu (%1,%4), %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
458 "movdqa %%xmm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
459 "movdqa %%xmm1, (%2,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
460 "movdqa %%xmm2, (%2,%3,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
461 "movdqa %%xmm3, (%2,%4) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
462 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
463 "lea (%1,%3,4), %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
464 "lea (%2,%3,4), %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
465 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
466 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
467 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
468 : "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
469 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
470 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
471
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
472 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
473 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
474 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
475 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
476 "movdqu (%1), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
477 "movdqu (%1,%3), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
478 "movdqu (%1,%3,2), %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
479 "movdqu (%1,%4), %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
480 "pavgb (%2), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
481 "pavgb (%2,%3), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
482 "pavgb (%2,%3,2), %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
483 "pavgb (%2,%4), %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
484 "movdqa %%xmm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
485 "movdqa %%xmm1, (%2,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
486 "movdqa %%xmm2, (%2,%3,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
487 "movdqa %%xmm3, (%2,%4) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
488 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
489 "lea (%1,%3,4), %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
490 "lea (%2,%3,4), %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
491 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
492 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
493 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
494 : "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
495 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
496 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
497
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
498 #define CLEAR_BLOCKS(name,n) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
499 static void name(DCTELEM *blocks)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
500 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
501 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
502 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
503 "mov %1, %%"REG_a" \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
504 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
505 "movq %%mm7, (%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
506 "movq %%mm7, 8(%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
507 "movq %%mm7, 16(%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
508 "movq %%mm7, 24(%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
509 "add $32, %%"REG_a" \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
510 " js 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
511 : : "r" (((uint8_t *)blocks)+128*n),\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
512 "i" (-128*n)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
513 : "%"REG_a\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
514 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
515 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
516 CLEAR_BLOCKS(clear_blocks_mmx, 6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
517 CLEAR_BLOCKS(clear_block_mmx, 1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
518
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
519 static void clear_block_sse(DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
520 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
521 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
522 "xorps %%xmm0, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
523 "movaps %%xmm0, (%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
524 "movaps %%xmm0, 16(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
525 "movaps %%xmm0, 32(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
526 "movaps %%xmm0, 48(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
527 "movaps %%xmm0, 64(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
528 "movaps %%xmm0, 80(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
529 "movaps %%xmm0, 96(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
530 "movaps %%xmm0, 112(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
531 :: "r"(block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
532 : "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
533 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
534 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
535
9861
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
536 static void clear_blocks_sse(DCTELEM *blocks)
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
537 {\
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
538 __asm__ volatile(
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
539 "xorps %%xmm0, %%xmm0 \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
540 "mov %1, %%"REG_a" \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
541 "1: \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
542 "movaps %%xmm0, (%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
543 "movaps %%xmm0, 16(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
544 "movaps %%xmm0, 32(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
545 "movaps %%xmm0, 48(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
546 "movaps %%xmm0, 64(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
547 "movaps %%xmm0, 80(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
548 "movaps %%xmm0, 96(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
549 "movaps %%xmm0, 112(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
550 "add $128, %%"REG_a" \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
551 " js 1b \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
552 : : "r" (((uint8_t *)blocks)+128*6),
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
553 "i" (-128*6)
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
554 : "%"REG_a
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
555 );
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
556 }
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
557
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
558 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
559 x86_reg i=0;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
560 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
561 "jmp 2f \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
562 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
563 "movq (%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
564 "movq (%2, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
565 "paddb %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
566 "movq %%mm1, (%2, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
567 "movq 8(%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
568 "movq 8(%2, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
569 "paddb %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
570 "movq %%mm1, 8(%2, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
571 "add $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
572 "2: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
573 "cmp %3, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
574 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
575 : "+r" (i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
576 : "r"(src), "r"(dst), "r"((x86_reg)w-15)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
577 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
578 for(; i<w; i++)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
579 dst[i+0] += src[i+0];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
580 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
581
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
582 static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
583 x86_reg i=0;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
584 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
585 "jmp 2f \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
586 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
587 "movq (%2, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
588 "movq 8(%2, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
589 "paddb (%3, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
590 "paddb 8(%3, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
591 "movq %%mm0, (%1, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
592 "movq %%mm1, 8(%1, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
593 "add $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
594 "2: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
595 "cmp %4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
596 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
597 : "+r" (i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
598 : "r"(dst), "r"(src1), "r"(src2), "r"((x86_reg)w-15)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
599 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
600 for(; i<w; i++)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
601 dst[i] = src1[i] + src2[i];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
602 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
603
8798
a5c8210814d7 Add check whether the compiler/assembler supports 10 or more operands.
diego
parents: 8760
diff changeset
604 #if HAVE_7REGS && HAVE_TEN_OPERANDS
10431
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10430
diff changeset
605 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top) {
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
606 x86_reg w2 = -w;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
607 x86_reg x;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
608 int l = *left & 0xff;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
609 int tl = *left_top & 0xff;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
610 int t;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
611 __asm__ volatile(
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
612 "mov %7, %3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
613 "1: \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
614 "movzx (%3,%4), %2 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
615 "mov %2, %k3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
616 "sub %b1, %b3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
617 "add %b0, %b3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
618 "mov %2, %1 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
619 "cmp %0, %2 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
620 "cmovg %0, %2 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
621 "cmovg %1, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
622 "cmp %k3, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
623 "cmovg %k3, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
624 "mov %7, %3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
625 "cmp %2, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
626 "cmovl %2, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
627 "add (%6,%4), %b0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
628 "mov %b0, (%5,%4) \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
629 "inc %4 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
630 "jl 1b \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
631 :"+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
632 :"r"(dst+w), "r"(diff+w), "rm"(top+w)
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
633 );
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
634 *left = l;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
635 *left_top = tl;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
636 }
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
637 #endif
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
638
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
639 #define H263_LOOP_FILTER \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
640 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
641 "movq %0, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
642 "movq %0, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
643 "movq %3, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
644 "movq %3, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
645 "punpcklbw %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
646 "punpckhbw %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
647 "punpcklbw %%mm7, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
648 "punpckhbw %%mm7, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
649 "psubw %%mm2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
650 "psubw %%mm3, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
651 "movq %1, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
652 "movq %1, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
653 "movq %2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
654 "movq %2, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
655 "punpcklbw %%mm7, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
656 "punpckhbw %%mm7, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
657 "punpcklbw %%mm7, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
658 "punpckhbw %%mm7, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
659 "psubw %%mm2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
660 "psubw %%mm3, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
661 "psllw $2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
662 "psllw $2, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
663 "paddw %%mm0, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
664 "paddw %%mm1, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
665 "pxor %%mm6, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
666 "pcmpgtw %%mm4, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
667 "pcmpgtw %%mm5, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
668 "pxor %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
669 "pxor %%mm7, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
670 "psubw %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
671 "psubw %%mm7, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
672 "psrlw $3, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
673 "psrlw $3, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
674 "packuswb %%mm5, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
675 "packsswb %%mm7, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
676 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
677 "movd %4, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
678 "punpcklbw %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
679 "punpcklbw %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
680 "punpcklbw %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
681 "psubusb %%mm4, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
682 "movq %%mm2, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
683 "psubusb %%mm4, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
684 "psubb %%mm3, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
685 "movq %1, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
686 "movq %2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
687 "pxor %%mm6, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
688 "pxor %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
689 "paddusb %%mm2, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
690 "psubusb %%mm2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
691 "pxor %%mm6, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
692 "pxor %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
693 "paddusb %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
694 "packsswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
695 "pcmpgtb %%mm0, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
696 "pxor %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
697 "psubb %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
698 "movq %%mm0, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
699 "psubusb %%mm2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
700 "psubb %%mm0, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
701 "pand %5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
702 "psrlw $2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
703 "pxor %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
704 "psubb %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
705 "movq %0, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
706 "movq %3, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
707 "psubb %%mm1, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
708 "paddb %%mm1, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
709
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
710 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale){
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10645
diff changeset
711 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
712 const int strength= ff_h263_loop_filter_strength[qscale];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
713
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
714 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
715
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
716 H263_LOOP_FILTER
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
717
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
718 "movq %%mm3, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
719 "movq %%mm4, %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
720 "movq %%mm5, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
721 "movq %%mm6, %3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
722 : "+m" (*(uint64_t*)(src - 2*stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
723 "+m" (*(uint64_t*)(src - 1*stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
724 "+m" (*(uint64_t*)(src + 0*stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
725 "+m" (*(uint64_t*)(src + 1*stride))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
726 : "g" (2*strength), "m"(ff_pb_FC)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
727 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
728 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
729 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
730
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
731 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10645
diff changeset
732 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
733 const int strength= ff_h263_loop_filter_strength[qscale];
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
734 DECLARE_ALIGNED(8, uint64_t, temp)[4];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
735 uint8_t *btemp= (uint8_t*)temp;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
736
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
737 src -= 2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
738
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
739 transpose4x4(btemp , src , 8, stride);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
740 transpose4x4(btemp+4, src + 4*stride, 8, stride);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
741 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
742 H263_LOOP_FILTER // 5 3 4 6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
743
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
744 : "+m" (temp[0]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
745 "+m" (temp[1]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
746 "+m" (temp[2]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
747 "+m" (temp[3])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
748 : "g" (2*strength), "m"(ff_pb_FC)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
749 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
750
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
751 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
752 "movq %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
753 "movq %%mm4, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
754 "punpcklbw %%mm3, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
755 "punpcklbw %%mm6, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
756 "punpckhbw %%mm3, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
757 "punpckhbw %%mm6, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
758 "movq %%mm5, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
759 "movq %%mm1, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
760 "punpcklwd %%mm4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
761 "punpcklwd %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
762 "punpckhwd %%mm4, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
763 "punpckhwd %%mm0, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
764 "movd %%mm5, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
765 "punpckhdq %%mm5, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
766 "movd %%mm5, (%0,%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
767 "movd %%mm3, (%0,%2,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
768 "punpckhdq %%mm3, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
769 "movd %%mm3, (%0,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
770 "movd %%mm1, (%1) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
771 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
772 "movd %%mm1, (%1,%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
773 "movd %%mm6, (%1,%2,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
774 "punpckhdq %%mm6, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
775 "movd %%mm6, (%1,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
776 :: "r" (src),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
777 "r" (src + 4*stride),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
778 "r" ((x86_reg) stride ),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
779 "r" ((x86_reg)(3*stride))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
780 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
781 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
782 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
783
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
784 /* draw the edges of width 'w' of an image of size width, height
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
785 this mmx version can only handle w==8 || w==16 */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
786 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
787 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
788 uint8_t *ptr, *last_line;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
789 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
790
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
791 last_line = buf + (height - 1) * wrap;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
792 /* left and right */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
793 ptr = buf;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
794 if(w==8)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
795 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
796 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
797 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
798 "movd (%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
799 "punpcklbw %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
800 "punpcklwd %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
801 "punpckldq %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
802 "movq %%mm0, -8(%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
803 "movq -8(%0, %2), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
804 "punpckhbw %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
805 "punpckhwd %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
806 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
807 "movq %%mm1, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
808 "add %1, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
809 "cmp %3, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
810 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
811 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
812 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
813 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
814 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
815 else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
816 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
817 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
818 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
819 "movd (%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
820 "punpcklbw %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
821 "punpcklwd %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
822 "punpckldq %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
823 "movq %%mm0, -8(%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
824 "movq %%mm0, -16(%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
825 "movq -8(%0, %2), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
826 "punpckhbw %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
827 "punpckhwd %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
828 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
829 "movq %%mm1, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
830 "movq %%mm1, 8(%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
831 "add %1, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
832 "cmp %3, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
833 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
834 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
835 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
836 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
837 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
838
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
839 for(i=0;i<w;i+=4) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
840 /* top and bottom (and hopefully also the corners) */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
841 ptr= buf - (i + 1) * wrap - w;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
842 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
843 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
844 "movq (%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
845 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
846 "movq %%mm0, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
847 "movq %%mm0, (%0, %2, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
848 "movq %%mm0, (%0, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
849 "add $8, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
850 "cmp %4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
851 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
852 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
853 : "r" ((x86_reg)buf - (x86_reg)ptr - w), "r" ((x86_reg)-wrap), "r" ((x86_reg)-wrap*3), "r" (ptr+width+2*w)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
854 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
855 ptr= last_line + (i + 1) * wrap - w;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
856 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
857 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
858 "movq (%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
859 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
860 "movq %%mm0, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
861 "movq %%mm0, (%0, %2, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
862 "movq %%mm0, (%0, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
863 "add $8, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
864 "cmp %4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
865 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
866 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
867 : "r" ((x86_reg)last_line - (x86_reg)ptr - w), "r" ((x86_reg)wrap), "r" ((x86_reg)wrap*3), "r" (ptr+width+2*w)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
868 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
869 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
870 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
871
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
872 #define PAETH(cpu, abs3)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
873 static void add_png_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
874 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
875 x86_reg i = -bpp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
876 x86_reg end = w-3;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
877 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
878 "pxor %%mm7, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
879 "movd (%1,%0), %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
880 "movd (%2,%0), %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
881 "punpcklbw %%mm7, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
882 "punpcklbw %%mm7, %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
883 "add %4, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
884 "1: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
885 "movq %%mm1, %%mm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
886 "movd (%2,%0), %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
887 "movq %%mm2, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
888 "punpcklbw %%mm7, %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
889 "movq %%mm2, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
890 "psubw %%mm1, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
891 "psubw %%mm0, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
892 "movq %%mm3, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
893 "paddw %%mm4, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
894 abs3\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
895 "movq %%mm4, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
896 "pminsw %%mm5, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
897 "pcmpgtw %%mm6, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
898 "pcmpgtw %%mm5, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
899 "movq %%mm4, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
900 "pand %%mm3, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
901 "pandn %%mm3, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
902 "pandn %%mm0, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
903 "movd (%3,%0), %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
904 "pand %%mm1, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
905 "pand %%mm4, %%mm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
906 "punpcklbw %%mm7, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
907 "movq %6, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
908 "paddw %%mm6, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
909 "paddw %%mm2, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
910 "paddw %%mm3, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
911 "pand %%mm5, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
912 "movq %%mm0, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
913 "packuswb %%mm3, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
914 "movd %%mm3, (%1,%0) \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
915 "add %4, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
916 "cmp %5, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
917 "jle 1b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
918 :"+r"(i)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
919 :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
920 "m"(ff_pw_255)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
921 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
922 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
923 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
924
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
925 #define ABS3_MMX2\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
926 "psubw %%mm5, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
927 "pmaxsw %%mm7, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
928 "pxor %%mm6, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
929 "pxor %%mm7, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
930 "psubw %%mm3, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
931 "psubw %%mm4, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
932 "pmaxsw %%mm6, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
933 "pmaxsw %%mm7, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
934 "pxor %%mm7, %%mm7 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
935
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
936 #define ABS3_SSSE3\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
937 "pabsw %%mm3, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
938 "pabsw %%mm4, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
939 "pabsw %%mm5, %%mm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
940
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
941 PAETH(mmx2, ABS3_MMX2)
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
942 #if HAVE_SSSE3
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
943 PAETH(ssse3, ABS3_SSSE3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
944 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
945
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
946 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
947 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
948 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
949 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
950 "movq "#in7", " #m3 " \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
951 "movq "#in0", %%mm5 \n\t" /* D */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
952 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
953 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
954 "movq "#in1", %%mm5 \n\t" /* C */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
955 "movq "#in2", %%mm6 \n\t" /* B */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
956 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
957 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
958 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
959 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
960 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
961 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
962 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
963 "psraw $5, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
964 "packuswb %%mm5, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
965 OP(%%mm5, out, %%mm7, d)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
966
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
967 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
968 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
969 uint64_t temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
970 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
971 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
972 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
973 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
974 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
975 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
976 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
977 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
978 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
979 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
980 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
981 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
982 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
983 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
984 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
985 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
986 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
987 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
988 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
989 "paddw %%mm3, %%mm5 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
990 "paddw %%mm2, %%mm6 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
991 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
992 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
993 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
994 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
995 "paddw %%mm4, %%mm0 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
996 "paddw %%mm1, %%mm5 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
997 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
998 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
999 "paddw %6, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1000 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1001 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1002 "movq %%mm0, %5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1003 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1004 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1005 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1006 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1007 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1008 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1009 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1010 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1011 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1012 "paddw %%mm0, %%mm2 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1013 "paddw %%mm5, %%mm3 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1014 "paddw %%mm2, %%mm2 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1015 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1016 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1017 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1018 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1019 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1020 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1021 "paddw %%mm2, %%mm1 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1022 "paddw %%mm6, %%mm4 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1023 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1024 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1025 "paddw %6, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1026 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1027 "psraw $5, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1028 "movq %5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1029 "packuswb %%mm3, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1030 OP_MMX2(%%mm1, (%1),%%mm4, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1031 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1032 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1033 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1034 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1035 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1036 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1037 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1038 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1039 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1040 "paddw %%mm1, %%mm5 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1041 "paddw %%mm4, %%mm0 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1042 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1043 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1044 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1045 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1046 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1047 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1048 "paddw %%mm3, %%mm2 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1049 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1050 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1051 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1052 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1053 "paddw %%mm2, %%mm6 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1054 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1055 "paddw %6, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1056 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1057 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1058 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1059 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1060 "paddw %%mm5, %%mm3 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1061 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1062 "paddw %%mm4, %%mm6 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1063 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1064 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1065 "paddw %%mm1, %%mm4 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1066 "paddw %%mm2, %%mm5 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1067 "paddw %%mm6, %%mm6 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1068 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1069 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1070 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1071 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1072 "paddw %6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1073 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1074 "psraw $5, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1075 "packuswb %%mm4, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1076 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1077 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1078 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1079 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1080 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1081 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1082 : "+a"(src), "+c"(dst), "+D"(h)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1083 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1084 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1085 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1086 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1087 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1088 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1089 int i;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1090 int16_t temp[16];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1091 /* quick HACK, XXX FIXME MUST be optimized */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1092 for(i=0; i<h; i++)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1093 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1094 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1095 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1096 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1097 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1098 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1099 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1100 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1101 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1102 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1103 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1104 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1105 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1106 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1107 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1108 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1109 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1110 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1111 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1112 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1113 "paddw %2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1114 "paddw %2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1115 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1116 "psraw $5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1117 "packuswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1118 OP_3DNOW(%%mm0, (%1), %%mm1, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1119 "movq 16(%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1120 "movq 24(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1121 "paddw %2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1122 "paddw %2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1123 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1124 "psraw $5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1125 "packuswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1126 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1127 :: "r"(temp), "r"(dst), "m"(ROUNDER)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1128 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1129 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1130 dst+=dstStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1131 src+=srcStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1132 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1133 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1134 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1135 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1136 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1137 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1138 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1139 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1140 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1141 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1142 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1143 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1144 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1145 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1146 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1147 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1148 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1149 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1150 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1151 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1152 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1153 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1154 "paddw %%mm3, %%mm5 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1155 "paddw %%mm2, %%mm6 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1156 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1157 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1158 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1159 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1160 "paddw %%mm4, %%mm0 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1161 "paddw %%mm1, %%mm5 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1162 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1163 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1164 "paddw %5, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1165 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1166 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1167 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1168 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1169 "movd 5(%0), %%mm5 \n\t" /* FGHI */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1170 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1171 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1172 "paddw %%mm5, %%mm1 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1173 "paddw %%mm6, %%mm2 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1174 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1175 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1176 "paddw %%mm6, %%mm3 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1177 "paddw %%mm5, %%mm4 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1178 "paddw %%mm2, %%mm2 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1179 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1180 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1181 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1182 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1183 "paddw %5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1184 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1185 "psraw $5, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1186 "packuswb %%mm3, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1187 OP_MMX2(%%mm0, (%1), %%mm4, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1188 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1189 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1190 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1191 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1192 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1193 : "+a"(src), "+c"(dst), "+d"(h)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1194 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1195 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1196 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1197 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1198 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1199 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1200 int i;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1201 int16_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1202 /* quick HACK, XXX FIXME MUST be optimized */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1203 for(i=0; i<h; i++)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1204 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1205 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1206 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1207 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1208 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1209 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1210 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1211 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1212 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1213 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1214 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1215 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1216 "paddw %2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1217 "paddw %2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1218 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1219 "psraw $5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1220 "packuswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1221 OP_3DNOW(%%mm0, (%1), %%mm1, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1222 :: "r"(temp), "r"(dst), "m"(ROUNDER)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1223 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1224 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1225 dst+=dstStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1226 src+=srcStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1227 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1228 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1229
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1230 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1231 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1232 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1233 uint64_t temp[17*4];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1234 uint64_t *temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1235 int count= 17;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1236 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1237 /*FIXME unroll */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1238 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1239 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1240 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1241 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1242 "movq (%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1243 "movq 8(%0), %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1244 "movq 8(%0), %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1245 "punpcklbw %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1246 "punpckhbw %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1247 "punpcklbw %%mm7, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1248 "punpckhbw %%mm7, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1249 "movq %%mm0, (%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1250 "movq %%mm1, 17*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1251 "movq %%mm2, 2*17*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1252 "movq %%mm3, 3*17*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1253 "add $8, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1254 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1255 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1256 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1257 : "+r" (src), "+r" (temp_ptr), "+r"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1258 : "r" ((x86_reg)srcStride)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1259 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1260 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1261 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1262 temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1263 count=4;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1264 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1265 /*FIXME reorder for speed */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1266 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1267 /*"pxor %%mm7, %%mm7 \n\t"*/\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1268 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1269 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1270 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1271 "movq 16(%0), %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1272 "movq 24(%0), %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1273 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1274 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1275 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1276 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1277 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1278 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1279 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1280 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1281 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1282 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1283 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1284 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1285 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1286 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1287 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1288 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1289 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1290 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1291 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1292 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1293 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1294 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1295 "add %4, %1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1296 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1297 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1298 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1299 "add $136, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1300 "add %6, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1301 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1302 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1303 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1304 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1305 : "r"((x86_reg)dstStride), "r"(2*(x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(x86_reg)dstStride)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1306 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1307 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1308 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1309 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1310 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1311 uint64_t temp[9*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1312 uint64_t *temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1313 int count= 9;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1314 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1315 /*FIXME unroll */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1316 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1317 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1318 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1319 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1320 "movq (%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1321 "punpcklbw %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1322 "punpckhbw %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1323 "movq %%mm0, (%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1324 "movq %%mm1, 9*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1325 "add $8, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1326 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1327 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1328 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1329 : "+r" (src), "+r" (temp_ptr), "+r"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1330 : "r" ((x86_reg)srcStride)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1331 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1332 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1333 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1334 temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1335 count=2;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1336 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1337 /*FIXME reorder for speed */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1338 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1339 /*"pxor %%mm7, %%mm7 \n\t"*/\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1340 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1341 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1342 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1343 "movq 16(%0), %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1344 "movq 24(%0), %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1345 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1346 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1347 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1348 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1349 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1350 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1351 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1352 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1353 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1354 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1355 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1356 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1357 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1358 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1359 "add $72, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1360 "add %6, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1361 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1362 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1363 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1364 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1365 : "r"((x86_reg)dstStride), "r"(2*(x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(x86_reg)dstStride)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1366 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1367 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1368 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1369 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1370 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1371 OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1372 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1373 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1374 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1375 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1376 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1377 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1378 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1379 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1380 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1381 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1382 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1383 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1384 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1385 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1386 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1387 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1388 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1389 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1390 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1391 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1392 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1393 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1394 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1395 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1396 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1397 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1398 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1399 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1400 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1401 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1402 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1403 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1404 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1405 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1406 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1407 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1408 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1409 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1410 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1411 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1412 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1413 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1414 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1415 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1416 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1417 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1418 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1419 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1420 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1421 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1422 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1423 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1424 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1425 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1426 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1427 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1428 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1429 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1430 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1431 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1432 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1433 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1434 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1435 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1436 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1437 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1438 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1439 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1440 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1441 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1442 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1443 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1444 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1445 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1446 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1447 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1448 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1449 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1450 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1451 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1452 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1453 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1454 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1455 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1456 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1457 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1458 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1459 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1460 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1461 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1462 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1463 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1464 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1465 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1466 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1467 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1468 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1469 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1470 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1471 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1472 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1473 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1474 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1475 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1476 uint64_t half[9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1477 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1478 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1479 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1480 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1481 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1482 OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1483 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1484 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1485 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1486 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1487 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1488 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1489 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1490 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1491 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1492 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1493 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1494 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1495 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1496 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1497 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1498 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1499 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1500 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1501 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1502 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1503 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1504 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1505 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1506 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1507 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1508 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1509 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1510 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1511 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1512 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1513 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1514 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1515 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1516 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1517 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1518 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1519 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1520 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1521 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1522 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1523 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1524 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1525 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1526 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1527 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1528 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1529 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1530 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1531 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1532 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1533 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1534 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1535 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1536 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1537 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1538 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1539 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1540 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1541 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1542 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1543 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1544 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1545 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1546 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1547 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1548 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1549 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1550 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1551 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1552 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1553 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1554 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1555 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1556 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1557 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1558 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1559 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1560 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1561 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1562 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1563 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1564 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1565 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1566 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1567 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1568 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1569 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1570 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1571 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1572 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1573 uint64_t half[17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1574 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1575 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1576 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1577 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1578 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1579 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1580 uint64_t half[17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1581 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1582 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1583 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1584 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1585 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1586 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1587 uint64_t half[17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1588 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1589 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1590 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1591 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1592
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1593 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1594 #define AVG_3DNOW_OP(a,b,temp, size) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1595 "mov" #size " " #b ", " #temp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1596 "pavgusb " #temp ", " #a " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1597 "mov" #size " " #a ", " #b " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1598 #define AVG_MMX2_OP(a,b,temp, size) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1599 "mov" #size " " #b ", " #temp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1600 "pavgb " #temp ", " #a " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1601 "mov" #size " " #a ", " #b " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1602
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1603 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1604 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1605 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1606 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1607 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1608 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1609 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1610 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1611 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1612
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1613 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1614 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1615
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1616 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1617 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1618 OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1619 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1620 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1621 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1622 OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1623 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1624
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1625 #define QPEL_2TAP(OPNAME, SIZE, MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1626 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1627 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1628 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1629 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1630 OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1631 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1632 OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1633 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1634 OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1635 static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1636 OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1637 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1638 static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1639 OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1640 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1641 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0, 1, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1642 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1, -1, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1643 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0, stride, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1644 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride, -stride, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1645 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0, stride, 1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1646 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1, stride, -1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1647 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride, -stride, 1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1648 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1649
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1650 QPEL_2TAP(put_, 16, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1651 QPEL_2TAP(avg_, 16, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1652 QPEL_2TAP(put_, 8, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1653 QPEL_2TAP(avg_, 8, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1654 QPEL_2TAP(put_, 16, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1655 QPEL_2TAP(avg_, 16, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1656 QPEL_2TAP(put_, 8, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1657 QPEL_2TAP(avg_, 8, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1658
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1659
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1660 #if 0
8527
f8bf438c6000 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 8519
diff changeset
1661 static void just_return(void) { return; }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1662 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1663
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1664 static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1665 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1666 const int w = 8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1667 const int ix = ox>>(16+shift);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1668 const int iy = oy>>(16+shift);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1669 const int oxs = ox>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1670 const int oys = oy>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1671 const int dxxs = dxx>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1672 const int dxys = dxy>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1673 const int dyxs = dyx>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1674 const int dyys = dyy>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1675 const uint16_t r4[4] = {r,r,r,r};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1676 const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1677 const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1678 const uint64_t shift2 = 2*shift;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1679 uint8_t edge_buf[(h+1)*stride];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1680 int x, y;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1681
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1682 const int dxw = (dxx-(1<<(16+shift)))*(w-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1683 const int dyh = (dyy-(1<<(16+shift)))*(h-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1684 const int dxh = dxy*(h-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1685 const int dyw = dyx*(w-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1686 if( // non-constant fullpel offset (3% of blocks)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1687 ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) |
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1688 (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1689 // uses more than 16 bits of subpel mv (only at huge resolution)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1690 || (dxx|dxy|dyx|dyy)&15 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1691 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1692 //FIXME could still use mmx for some of the rows
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1693 ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1694 return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1695 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1696
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1697 src += ix + iy*stride;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1698 if( (unsigned)ix >= width-w ||
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1699 (unsigned)iy >= height-h )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1700 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1701 ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1702 src = edge_buf;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1703 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1704
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1705 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1706 "movd %0, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1707 "pxor %%mm7, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1708 "punpcklwd %%mm6, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1709 "punpcklwd %%mm6, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1710 :: "r"(1<<shift)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1711 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1712
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1713 for(x=0; x<w; x+=4){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1714 uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1715 oxs - dxys + dxxs*(x+1),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1716 oxs - dxys + dxxs*(x+2),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1717 oxs - dxys + dxxs*(x+3) };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1718 uint16_t dy4[4] = { oys - dyys + dyxs*(x+0),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1719 oys - dyys + dyxs*(x+1),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1720 oys - dyys + dyxs*(x+2),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1721 oys - dyys + dyxs*(x+3) };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1722
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1723 for(y=0; y<h; y++){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1724 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1725 "movq %0, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1726 "movq %1, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1727 "paddw %2, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1728 "paddw %3, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1729 "movq %%mm4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1730 "movq %%mm5, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1731 "psrlw $12, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1732 "psrlw $12, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1733 : "+m"(*dx4), "+m"(*dy4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1734 : "m"(*dxy4), "m"(*dyy4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1735 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1736
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1737 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1738 "movq %%mm6, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1739 "movq %%mm6, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1740 "psubw %%mm4, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1741 "psubw %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1742 "movq %%mm2, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1743 "movq %%mm4, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1744 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1745 "pmullw %%mm5, %%mm3 \n\t" // dx*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1746 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1747 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1748
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1749 "movd %4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1750 "movd %3, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1751 "punpcklbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1752 "punpcklbw %%mm7, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1753 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1754 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1755
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1756 "movd %2, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1757 "movd %1, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1758 "punpcklbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1759 "punpcklbw %%mm7, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1760 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1761 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1762 "paddw %5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1763 "paddw %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1764 "paddw %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1765 "paddw %%mm2, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1766
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1767 "psrlw %6, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1768 "packuswb %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1769 "movd %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1770
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1771 : "=m"(dst[x+y*stride])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1772 : "m"(src[0]), "m"(src[1]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1773 "m"(src[stride]), "m"(src[stride+1]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1774 "m"(*r4), "m"(shift2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1775 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1776 src += stride;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1777 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1778 src += 4-h*stride;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1779 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1780 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1781
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1782 #define PREFETCH(name, op) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1783 static void name(void *mem, int stride, int h){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1784 const uint8_t *p= mem;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1785 do{\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1786 __asm__ volatile(#op" %0" :: "m"(*p));\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1787 p+= stride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1788 }while(--h);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1789 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1790 PREFETCH(prefetch_mmx2, prefetcht0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1791 PREFETCH(prefetch_3dnow, prefetch)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1792 #undef PREFETCH
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1793
12450
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents: 12439
diff changeset
1794 #include "h264_qpel_mmx.c"
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1795
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1796 void ff_put_h264_chroma_mc8_mmx_rnd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1797 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1798 void ff_put_vc1_chroma_mc8_mmx_nornd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1799 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1800 void ff_put_rv40_chroma_mc8_mmx (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1801 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1802 void ff_avg_h264_chroma_mc8_mmx2_rnd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1803 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1804 void ff_avg_vc1_chroma_mc8_mmx2_nornd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1805 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1806 void ff_avg_rv40_chroma_mc8_mmx2 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1807 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1808 void ff_avg_h264_chroma_mc8_3dnow_rnd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1809 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1810 void ff_avg_vc1_chroma_mc8_3dnow_nornd(uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1811 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1812 void ff_avg_rv40_chroma_mc8_3dnow (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1813 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1814
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1815 void ff_put_h264_chroma_mc4_mmx (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1816 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1817 void ff_put_rv40_chroma_mc4_mmx (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1818 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1819 void ff_avg_h264_chroma_mc4_mmx2 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1820 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1821 void ff_avg_rv40_chroma_mc4_mmx2 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1822 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1823 void ff_avg_h264_chroma_mc4_3dnow (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1824 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1825 void ff_avg_rv40_chroma_mc4_3dnow (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1826 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1827
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1828 void ff_put_h264_chroma_mc2_mmx2 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1829 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1830 void ff_avg_h264_chroma_mc2_mmx2 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1831 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1832
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1833 void ff_put_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1834 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1835 void ff_put_vc1_chroma_mc8_ssse3_nornd(uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1836 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1837 void ff_put_h264_chroma_mc4_ssse3 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1838 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1839
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1840 void ff_avg_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1841 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1842 void ff_avg_vc1_chroma_mc8_ssse3_nornd(uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1843 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1844 void ff_avg_h264_chroma_mc4_ssse3 (uint8_t *dst, uint8_t *src,
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1845 int stride, int h, int x, int y);
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
1846
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1847
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1848 /* CAVS specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1849 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1850 put_pixels8_mmx(dst, src, stride, 8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1851 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1852 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1853 avg_pixels8_mmx(dst, src, stride, 8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1854 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1855 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1856 put_pixels16_mmx(dst, src, stride, 16);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1857 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1858 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1859 avg_pixels16_mmx(dst, src, stride, 16);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1860 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1861
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1862 /* VC1 specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1863 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1864 put_pixels8_mmx(dst, src, stride, 8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1865 }
9441
e14cd3ac3806 VC1: extend MMX qpel MC to include MMX2 avg qpel
conrad
parents: 9440
diff changeset
1866 void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) {
e14cd3ac3806 VC1: extend MMX qpel MC to include MMX2 avg qpel
conrad
parents: 9440
diff changeset
1867 avg_pixels8_mmx2(dst, src, stride, 8);
e14cd3ac3806 VC1: extend MMX qpel MC to include MMX2 avg qpel
conrad
parents: 9440
diff changeset
1868 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1869
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1870 /* XXX: those functions should be suppressed ASAP when all IDCTs are
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1871 converted */
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
1872 #if CONFIG_GPL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1873 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1874 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1875 ff_mmx_idct (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1876 ff_put_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1877 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1878 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1879 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1880 ff_mmx_idct (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1881 ff_add_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1882 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1883 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1884 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1885 ff_mmxext_idct (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1886 ff_put_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1887 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1888 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1889 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1890 ff_mmxext_idct (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1891 ff_add_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1892 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1893 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1894 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1895 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1896 ff_idct_xvid_mmx (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1897 ff_put_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1898 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1899 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1900 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1901 ff_idct_xvid_mmx (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1902 ff_add_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1903 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1904 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1905 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1906 ff_idct_xvid_mmx2 (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1907 ff_put_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1908 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1909 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1910 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1911 ff_idct_xvid_mmx2 (block);
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
1912 ff_add_pixels_clamped_mmx(block, dest, line_size);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1913 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1914
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1915 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1916 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1917 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1918 __asm__ volatile("pxor %%mm7, %%mm7":);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1919 for(i=0; i<blocksize; i+=2) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1920 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1921 "movq %0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1922 "movq %1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1923 "movq %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1924 "movq %%mm1, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1925 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1926 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1927 "pslld $31, %%mm2 \n\t" // keep only the sign bit
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1928 "pxor %%mm2, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1929 "movq %%mm3, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1930 "pand %%mm1, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1931 "pandn %%mm1, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1932 "pfadd %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1933 "pfsub %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1934 "movq %%mm3, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1935 "movq %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1936 :"+m"(mag[i]), "+m"(ang[i])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1937 ::"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1938 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1939 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1940 __asm__ volatile("femms");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1941 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1942 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1943 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1944 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1945
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1946 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1947 "movaps %0, %%xmm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1948 ::"m"(ff_pdw_80000000[0])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1949 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1950 for(i=0; i<blocksize; i+=4) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1951 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1952 "movaps %0, %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1953 "movaps %1, %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1954 "xorps %%xmm2, %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1955 "xorps %%xmm3, %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1956 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1957 "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1958 "andps %%xmm5, %%xmm2 \n\t" // keep only the sign bit
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1959 "xorps %%xmm2, %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1960 "movaps %%xmm3, %%xmm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1961 "andps %%xmm1, %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1962 "andnps %%xmm1, %%xmm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1963 "addps %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1964 "subps %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1965 "movaps %%xmm3, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1966 "movaps %%xmm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1967 :"+m"(mag[i]), "+m"(ang[i])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1968 ::"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1969 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1970 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1971 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1972
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1973 #define IF1(x) x
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1974 #define IF0(x)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1975
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1976 #define MIX5(mono,stereo)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1977 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1978 "movss 0(%2), %%xmm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1979 "movss 8(%2), %%xmm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1980 "movss 24(%2), %%xmm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1981 "shufps $0, %%xmm5, %%xmm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1982 "shufps $0, %%xmm6, %%xmm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1983 "shufps $0, %%xmm7, %%xmm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1984 "1: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1985 "movaps (%0,%1), %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1986 "movaps 0x400(%0,%1), %%xmm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1987 "movaps 0x800(%0,%1), %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1988 "movaps 0xc00(%0,%1), %%xmm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1989 "movaps 0x1000(%0,%1), %%xmm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1990 "mulps %%xmm5, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1991 "mulps %%xmm6, %%xmm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1992 "mulps %%xmm5, %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1993 "mulps %%xmm7, %%xmm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1994 "mulps %%xmm7, %%xmm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1995 stereo("addps %%xmm1, %%xmm0 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1996 "addps %%xmm1, %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1997 "addps %%xmm3, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1998 "addps %%xmm4, %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1999 mono("addps %%xmm2, %%xmm0 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2000 "movaps %%xmm0, (%0,%1) \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2001 stereo("movaps %%xmm2, 0x400(%0,%1) \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2002 "add $16, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2003 "jl 1b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2004 :"+&r"(i)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2005 :"r"(samples[0]+len), "r"(matrix)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2006 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2007 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2008
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2009 #define MIX_MISC(stereo)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2010 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2011 "1: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2012 "movaps (%3,%0), %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2013 stereo("movaps %%xmm0, %%xmm1 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2014 "mulps %%xmm6, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2015 stereo("mulps %%xmm7, %%xmm1 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2016 "lea 1024(%3,%0), %1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2017 "mov %5, %2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2018 "2: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2019 "movaps (%1), %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2020 stereo("movaps %%xmm2, %%xmm3 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2021 "mulps (%4,%2), %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2022 stereo("mulps 16(%4,%2), %%xmm3 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2023 "addps %%xmm2, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2024 stereo("addps %%xmm3, %%xmm1 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2025 "add $1024, %1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2026 "add $32, %2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2027 "jl 2b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2028 "movaps %%xmm0, (%3,%0) \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2029 stereo("movaps %%xmm1, 1024(%3,%0) \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2030 "add $16, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2031 "jl 1b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2032 :"+&r"(i), "=&r"(j), "=&r"(k)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2033 :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2034 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2035 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2036
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2037 static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2038 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2039 int (*matrix_cmp)[2] = (int(*)[2])matrix;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2040 intptr_t i,j,k;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2041
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2042 i = -len*sizeof(float);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2043 if(in_ch == 5 && out_ch == 2 && !(matrix_cmp[0][1]|matrix_cmp[2][0]|matrix_cmp[3][1]|matrix_cmp[4][0]|(matrix_cmp[1][0]^matrix_cmp[1][1])|(matrix_cmp[0][0]^matrix_cmp[2][1]))) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2044 MIX5(IF0,IF1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2045 } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2046 MIX5(IF1,IF0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2047 } else {
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
2048 DECLARE_ALIGNED(16, float, matrix_simd)[in_ch][2][4];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2049 j = 2*in_ch*sizeof(float);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2050 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2051 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2052 "sub $8, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2053 "movss (%2,%0), %%xmm6 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2054 "movss 4(%2,%0), %%xmm7 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2055 "shufps $0, %%xmm6, %%xmm6 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2056 "shufps $0, %%xmm7, %%xmm7 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2057 "movaps %%xmm6, (%1,%0,4) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2058 "movaps %%xmm7, 16(%1,%0,4) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2059 "jg 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2060 :"+&r"(j)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2061 :"r"(matrix_simd), "r"(matrix)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2062 :"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2063 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2064 if(out_ch == 2) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2065 MIX_MISC(IF1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2066 } else {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2067 MIX_MISC(IF0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2068 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2069 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2070 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2071
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2072 static void vector_fmul_3dnow(float *dst, const float *src, int len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2073 x86_reg i = (len-4)*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2074 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2075 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2076 "movq (%1,%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2077 "movq 8(%1,%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2078 "pfmul (%2,%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2079 "pfmul 8(%2,%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2080 "movq %%mm0, (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2081 "movq %%mm1, 8(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2082 "sub $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2083 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2084 "femms \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2085 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2086 :"r"(dst), "r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2087 :"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2088 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2089 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2090 static void vector_fmul_sse(float *dst, const float *src, int len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2091 x86_reg i = (len-8)*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2092 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2093 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2094 "movaps (%1,%0), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2095 "movaps 16(%1,%0), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2096 "mulps (%2,%0), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2097 "mulps 16(%2,%0), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2098 "movaps %%xmm0, (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2099 "movaps %%xmm1, 16(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2100 "sub $32, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2101 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2102 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2103 :"r"(dst), "r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2104 :"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2105 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2106 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2107
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2108 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2109 x86_reg i = len*4-16;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2110 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2111 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2112 "pswapd 8(%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2113 "pswapd (%1), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2114 "pfmul (%3,%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2115 "pfmul 8(%3,%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2116 "movq %%mm0, (%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2117 "movq %%mm1, 8(%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2118 "add $16, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2119 "sub $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2120 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2121 :"+r"(i), "+r"(src1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2122 :"r"(dst), "r"(src0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2123 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2124 __asm__ volatile("femms");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2125 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2126 static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2127 x86_reg i = len*4-32;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2128 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2129 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2130 "movaps 16(%1), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2131 "movaps (%1), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2132 "shufps $0x1b, %%xmm0, %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2133 "shufps $0x1b, %%xmm1, %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2134 "mulps (%3,%0), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2135 "mulps 16(%3,%0), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2136 "movaps %%xmm0, (%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2137 "movaps %%xmm1, 16(%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2138 "add $32, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2139 "sub $32, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2140 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2141 :"+r"(i), "+r"(src1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2142 :"r"(dst), "r"(src0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2143 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2144 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2145
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2146 static void vector_fmul_add_3dnow(float *dst, const float *src0, const float *src1,
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2147 const float *src2, int len){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2148 x86_reg i = (len-4)*4;
10301
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2149 __asm__ volatile(
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2150 "1: \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2151 "movq (%2,%0), %%mm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2152 "movq 8(%2,%0), %%mm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2153 "pfmul (%3,%0), %%mm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2154 "pfmul 8(%3,%0), %%mm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2155 "pfadd (%4,%0), %%mm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2156 "pfadd 8(%4,%0), %%mm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2157 "movq %%mm0, (%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2158 "movq %%mm1, 8(%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2159 "sub $16, %0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2160 "jge 1b \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2161 :"+r"(i)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2162 :"r"(dst), "r"(src0), "r"(src1), "r"(src2)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2163 :"memory"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2164 );
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2165 __asm__ volatile("femms");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2166 }
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2167 static void vector_fmul_add_sse(float *dst, const float *src0, const float *src1,
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2168 const float *src2, int len){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2169 x86_reg i = (len-8)*4;
10301
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2170 __asm__ volatile(
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2171 "1: \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2172 "movaps (%2,%0), %%xmm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2173 "movaps 16(%2,%0), %%xmm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2174 "mulps (%3,%0), %%xmm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2175 "mulps 16(%3,%0), %%xmm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2176 "addps (%4,%0), %%xmm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2177 "addps 16(%4,%0), %%xmm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2178 "movaps %%xmm0, (%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2179 "movaps %%xmm1, 16(%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2180 "sub $32, %0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2181 "jge 1b \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2182 :"+r"(i)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2183 :"r"(dst), "r"(src0), "r"(src1), "r"(src2)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2184 :"memory"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2185 );
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2186 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2187
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2188 static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2189 const float *win, float add_bias, int len){
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2190 #if HAVE_6REGS
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2191 if(add_bias == 0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2192 x86_reg i = -len*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2193 x86_reg j = len*4-8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2194 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2195 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2196 "pswapd (%5,%1), %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2197 "movq (%5,%0), %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2198 "pswapd (%4,%1), %%mm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2199 "movq (%3,%0), %%mm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2200 "movq %%mm0, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2201 "movq %%mm1, %%mm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2202 "pfmul %%mm4, %%mm2 \n" // src0[len+i]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2203 "pfmul %%mm5, %%mm3 \n" // src1[ j]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2204 "pfmul %%mm4, %%mm1 \n" // src0[len+i]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2205 "pfmul %%mm5, %%mm0 \n" // src1[ j]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2206 "pfadd %%mm3, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2207 "pfsub %%mm0, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2208 "pswapd %%mm2, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2209 "movq %%mm1, (%2,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2210 "movq %%mm2, (%2,%1) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2211 "sub $8, %1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2212 "add $8, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2213 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2214 "femms \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2215 :"+r"(i), "+r"(j)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2216 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2217 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2218 }else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2219 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2220 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2221 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2222
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2223 static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2224 const float *win, float add_bias, int len){
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2225 #if HAVE_6REGS
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2226 if(add_bias == 0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2227 x86_reg i = -len*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2228 x86_reg j = len*4-16;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2229 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2230 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2231 "movaps (%5,%1), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2232 "movaps (%5,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2233 "movaps (%4,%1), %%xmm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2234 "movaps (%3,%0), %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2235 "shufps $0x1b, %%xmm1, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2236 "shufps $0x1b, %%xmm5, %%xmm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2237 "movaps %%xmm0, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2238 "movaps %%xmm1, %%xmm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2239 "mulps %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2240 "mulps %%xmm5, %%xmm3 \n" // src1[ j]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2241 "mulps %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2242 "mulps %%xmm5, %%xmm0 \n" // src1[ j]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2243 "addps %%xmm3, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2244 "subps %%xmm0, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2245 "shufps $0x1b, %%xmm2, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2246 "movaps %%xmm1, (%2,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2247 "movaps %%xmm2, (%2,%1) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2248 "sub $16, %1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2249 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2250 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2251 :"+r"(i), "+r"(j)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2252 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2253 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2254 }else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2255 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2256 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2257 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2258
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2259 static void int32_to_float_fmul_scalar_sse(float *dst, const int *src, float mul, int len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2260 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2261 x86_reg i = -4*len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2262 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2263 "movss %3, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2264 "shufps $0, %%xmm4, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2265 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2266 "cvtpi2ps (%2,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2267 "cvtpi2ps 8(%2,%0), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2268 "cvtpi2ps 16(%2,%0), %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2269 "cvtpi2ps 24(%2,%0), %%xmm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2270 "movlhps %%xmm1, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2271 "movlhps %%xmm3, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2272 "mulps %%xmm4, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2273 "mulps %%xmm4, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2274 "movaps %%xmm0, (%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2275 "movaps %%xmm2, 16(%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2276 "add $32, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2277 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2278 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2279 :"r"(dst+len), "r"(src+len), "m"(mul)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2280 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2281 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2282
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2283 static void int32_to_float_fmul_scalar_sse2(float *dst, const int *src, float mul, int len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2284 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2285 x86_reg i = -4*len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2286 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2287 "movss %3, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2288 "shufps $0, %%xmm4, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2289 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2290 "cvtdq2ps (%2,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2291 "cvtdq2ps 16(%2,%0), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2292 "mulps %%xmm4, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2293 "mulps %%xmm4, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2294 "movaps %%xmm0, (%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2295 "movaps %%xmm1, 16(%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2296 "add $32, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2297 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2298 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2299 :"r"(dst+len), "r"(src+len), "m"(mul)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2300 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2301 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2302
10105
7775f6627612 Mark parameter src of vector_clipf() as const
vitor
parents: 10104
diff changeset
2303 static void vector_clipf_sse(float *dst, const float *src, float min, float max,
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2304 int len)
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2305 {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2306 x86_reg i = (len-16)*4;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2307 __asm__ volatile(
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2308 "movss %3, %%xmm4 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2309 "movss %4, %%xmm5 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2310 "shufps $0, %%xmm4, %%xmm4 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2311 "shufps $0, %%xmm5, %%xmm5 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2312 "1: \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2313 "movaps (%2,%0), %%xmm0 \n\t" // 3/1 on intel
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2314 "movaps 16(%2,%0), %%xmm1 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2315 "movaps 32(%2,%0), %%xmm2 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2316 "movaps 48(%2,%0), %%xmm3 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2317 "maxps %%xmm4, %%xmm0 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2318 "maxps %%xmm4, %%xmm1 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2319 "maxps %%xmm4, %%xmm2 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2320 "maxps %%xmm4, %%xmm3 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2321 "minps %%xmm5, %%xmm0 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2322 "minps %%xmm5, %%xmm1 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2323 "minps %%xmm5, %%xmm2 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2324 "minps %%xmm5, %%xmm3 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2325 "movaps %%xmm0, (%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2326 "movaps %%xmm1, 16(%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2327 "movaps %%xmm2, 32(%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2328 "movaps %%xmm3, 48(%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2329 "sub $64, %0 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2330 "jge 1b \n\t"
10107
3b61bc6ce377 Mark "i" parameter of vector_clipf_sse() as early-clobber
vitor
parents: 10105
diff changeset
2331 :"+&r"(i)
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2332 :"r"(dst), "r"(src), "m"(min), "m"(max)
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2333 :"memory"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2334 );
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2335 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2336
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2337 static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2338 x86_reg reglen = len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2339 // not bit-exact: pf2id uses different rounding than C and SSE
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2340 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2341 "add %0 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2342 "lea (%2,%0,2) , %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2343 "add %0 , %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2344 "neg %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2345 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2346 "pf2id (%2,%0,2) , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2347 "pf2id 8(%2,%0,2) , %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2348 "pf2id 16(%2,%0,2) , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2349 "pf2id 24(%2,%0,2) , %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2350 "packssdw %%mm1 , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2351 "packssdw %%mm3 , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2352 "movq %%mm0 , (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2353 "movq %%mm2 , 8(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2354 "add $16 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2355 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2356 "femms \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2357 :"+r"(reglen), "+r"(dst), "+r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2358 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2359 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2360 static void float_to_int16_sse(int16_t *dst, const float *src, long len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2361 x86_reg reglen = len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2362 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2363 "add %0 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2364 "lea (%2,%0,2) , %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2365 "add %0 , %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2366 "neg %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2367 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2368 "cvtps2pi (%2,%0,2) , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2369 "cvtps2pi 8(%2,%0,2) , %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2370 "cvtps2pi 16(%2,%0,2) , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2371 "cvtps2pi 24(%2,%0,2) , %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2372 "packssdw %%mm1 , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2373 "packssdw %%mm3 , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2374 "movq %%mm0 , (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2375 "movq %%mm2 , 8(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2376 "add $16 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2377 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2378 "emms \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2379 :"+r"(reglen), "+r"(dst), "+r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2380 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2381 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2382
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2383 static void float_to_int16_sse2(int16_t *dst, const float *src, long len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2384 x86_reg reglen = len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2385 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2386 "add %0 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2387 "lea (%2,%0,2) , %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2388 "add %0 , %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2389 "neg %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2390 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2391 "cvtps2dq (%2,%0,2) , %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2392 "cvtps2dq 16(%2,%0,2) , %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2393 "packssdw %%xmm1 , %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2394 "movdqa %%xmm0 , (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2395 "add $16 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2396 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2397 :"+r"(reglen), "+r"(dst), "+r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2398 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2399 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2400
12436
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2401 void ff_vp3_idct_mmx(int16_t *input_data);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2402 void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2403 void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2404
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2405 void ff_vp3_idct_dc_add_mmx2(uint8_t *dest, int line_size, const DCTELEM *block);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2406
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2407 void ff_vp3_v_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2408 void ff_vp3_h_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2409
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2410 void ff_vp3_idct_sse2(int16_t *input_data);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2411 void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2412 void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block);
d6d0a43848b4 Move VP3 IDCT functions from inline ASM to YASM. This fixes part of the VP3/5/6
rbultje
parents: 12435
diff changeset
2413
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2414 void ff_float_to_int16_interleave6_sse(int16_t *dst, const float **src, int len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2415 void ff_float_to_int16_interleave6_3dnow(int16_t *dst, const float **src, int len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2416 void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len);
11981
d3551fcf4c1c Add const to some pointer parameters.
cehoyos
parents: 11951
diff changeset
2417 int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2, int order, int shift);
d3551fcf4c1c Add const to some pointer parameters.
cehoyos
parents: 11951
diff changeset
2418 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order, int shift);
d3551fcf4c1c Add const to some pointer parameters.
cehoyos
parents: 11951
diff changeset
2419 int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);
d3551fcf4c1c Add const to some pointer parameters.
cehoyos
parents: 11951
diff changeset
2420 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);
d3551fcf4c1c Add const to some pointer parameters.
cehoyos
parents: 11951
diff changeset
2421 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);
10431
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10430
diff changeset
2422 void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10430
diff changeset
2423 int ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left);
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10430
diff changeset
2424 int ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);
10645
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2425
12450
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents: 12439
diff changeset
2426 #if !HAVE_YASM
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2427 #define ff_float_to_int16_interleave6_sse(a,b,c) float_to_int16_interleave_misc_sse(a,b,c,6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2428 #define ff_float_to_int16_interleave6_3dnow(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2429 #define ff_float_to_int16_interleave6_3dn2(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2430 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2431 #define ff_float_to_int16_interleave6_sse2 ff_float_to_int16_interleave6_sse
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2432
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2433 #define FLOAT_TO_INT16_INTERLEAVE(cpu, body) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2434 /* gcc pessimizes register allocation if this is in the same function as float_to_int16_interleave_sse2*/\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2435 static av_noinline void float_to_int16_interleave_misc_##cpu(int16_t *dst, const float **src, long len, int channels){\
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11118
diff changeset
2436 DECLARE_ALIGNED(16, int16_t, tmp)[len];\
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2437 int i,j,c;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2438 for(c=0; c<channels; c++){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2439 float_to_int16_##cpu(tmp, src[c], len);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2440 for(i=0, j=c; i<len; i++, j+=channels)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2441 dst[j] = tmp[i];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2442 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2443 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2444 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2445 static void float_to_int16_interleave_##cpu(int16_t *dst, const float **src, long len, int channels){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2446 if(channels==1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2447 float_to_int16_##cpu(dst, src[0], len);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2448 else if(channels==2){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2449 x86_reg reglen = len; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2450 const float *src0 = src[0];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2451 const float *src1 = src[1];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2452 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2453 "shl $2, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2454 "add %0, %1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2455 "add %0, %2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2456 "add %0, %3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2457 "neg %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2458 body\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2459 :"+r"(reglen), "+r"(dst), "+r"(src0), "+r"(src1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2460 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2461 }else if(channels==6){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2462 ff_float_to_int16_interleave6_##cpu(dst, src, len);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2463 }else\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2464 float_to_int16_interleave_misc_##cpu(dst, src, len, channels);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2465 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2466
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2467 FLOAT_TO_INT16_INTERLEAVE(3dnow,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2468 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2469 "pf2id (%2,%0), %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2470 "pf2id 8(%2,%0), %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2471 "pf2id (%3,%0), %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2472 "pf2id 8(%3,%0), %%mm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2473 "packssdw %%mm1, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2474 "packssdw %%mm3, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2475 "movq %%mm0, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2476 "punpcklwd %%mm2, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2477 "punpckhwd %%mm2, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2478 "movq %%mm0, (%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2479 "movq %%mm1, 8(%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2480 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2481 "js 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2482 "femms \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2483 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2484
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2485 FLOAT_TO_INT16_INTERLEAVE(sse,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2486 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2487 "cvtps2pi (%2,%0), %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2488 "cvtps2pi 8(%2,%0), %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2489 "cvtps2pi (%3,%0), %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2490 "cvtps2pi 8(%3,%0), %%mm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2491 "packssdw %%mm1, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2492 "packssdw %%mm3, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2493 "movq %%mm0, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2494 "punpcklwd %%mm2, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2495 "punpckhwd %%mm2, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2496 "movq %%mm0, (%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2497 "movq %%mm1, 8(%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2498 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2499 "js 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2500 "emms \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2501 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2502
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2503 FLOAT_TO_INT16_INTERLEAVE(sse2,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2504 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2505 "cvtps2dq (%2,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2506 "cvtps2dq (%3,%0), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2507 "packssdw %%xmm1, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2508 "movhlps %%xmm0, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2509 "punpcklwd %%xmm1, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2510 "movdqa %%xmm0, (%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2511 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2512 "js 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2513 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2514
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2515 static void float_to_int16_interleave_3dn2(int16_t *dst, const float **src, long len, int channels){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2516 if(channels==6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2517 ff_float_to_int16_interleave6_3dn2(dst, src, len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2518 else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2519 float_to_int16_interleave_3dnow(dst, src, len, channels);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2520 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2521
10964
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2522 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2523
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2524 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2525 {
12414
3fc4c625b6f3 Remove global mm_flags variable
mru
parents: 12366
diff changeset
2526 int mm_flags = mm_support();
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2527
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2528 if (avctx->dsp_mask) {
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2529 if (avctx->dsp_mask & AV_CPU_FLAG_FORCE)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2530 mm_flags |= (avctx->dsp_mask & 0xffff);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2531 else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2532 mm_flags &= ~(avctx->dsp_mask & 0xffff);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2533 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2534
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2535 #if 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2536 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2537 if (mm_flags & AV_CPU_FLAG_MMX)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2538 av_log(avctx, AV_LOG_INFO, " mmx");
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2539 if (mm_flags & AV_CPU_FLAG_MMX2)
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2540 av_log(avctx, AV_LOG_INFO, " mmx2");
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2541 if (mm_flags & AV_CPU_FLAG_3DNOW)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2542 av_log(avctx, AV_LOG_INFO, " 3dnow");
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2543 if (mm_flags & AV_CPU_FLAG_SSE)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2544 av_log(avctx, AV_LOG_INFO, " sse");
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2545 if (mm_flags & AV_CPU_FLAG_SSE2)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2546 av_log(avctx, AV_LOG_INFO, " sse2");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2547 av_log(avctx, AV_LOG_INFO, "\n");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2548 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2549
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2550 if (mm_flags & AV_CPU_FLAG_MMX) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2551 const int idct_algo= avctx->idct_algo;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2552
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2553 if(avctx->lowres==0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2554 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2555 c->idct_put= ff_simple_idct_put_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2556 c->idct_add= ff_simple_idct_add_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2557 c->idct = ff_simple_idct_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2558 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2559 #if CONFIG_GPL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2560 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2561 if(mm_flags & AV_CPU_FLAG_MMX2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2562 c->idct_put= ff_libmpeg2mmx2_idct_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2563 c->idct_add= ff_libmpeg2mmx2_idct_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2564 c->idct = ff_mmxext_idct;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2565 }else{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2566 c->idct_put= ff_libmpeg2mmx_idct_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2567 c->idct_add= ff_libmpeg2mmx_idct_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2568 c->idct = ff_mmx_idct;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2569 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2570 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2571 #endif
9975
d6d7e8d4a04d Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents: 9959
diff changeset
2572 }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) &&
12439
51fc247eed32 Fix compilation failure if yasm is disabled (missing vp3 symbols).
rbultje
parents: 12437
diff changeset
2573 idct_algo==FF_IDCT_VP3 && HAVE_YASM){
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2574 if(mm_flags & AV_CPU_FLAG_SSE2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2575 c->idct_put= ff_vp3_idct_put_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2576 c->idct_add= ff_vp3_idct_add_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2577 c->idct = ff_vp3_idct_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2578 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2579 }else{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2580 c->idct_put= ff_vp3_idct_put_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2581 c->idct_add= ff_vp3_idct_add_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2582 c->idct = ff_vp3_idct_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2583 c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2584 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2585 }else if(idct_algo==FF_IDCT_CAVS){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2586 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2587 }else if(idct_algo==FF_IDCT_XVIDMMX){
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2588 if(mm_flags & AV_CPU_FLAG_SSE2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2589 c->idct_put= ff_idct_xvid_sse2_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2590 c->idct_add= ff_idct_xvid_sse2_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2591 c->idct = ff_idct_xvid_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2592 c->idct_permutation_type= FF_SSE2_IDCT_PERM;
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2593 }else if(mm_flags & AV_CPU_FLAG_MMX2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2594 c->idct_put= ff_idct_xvid_mmx2_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2595 c->idct_add= ff_idct_xvid_mmx2_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2596 c->idct = ff_idct_xvid_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2597 }else{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2598 c->idct_put= ff_idct_xvid_mmx_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2599 c->idct_add= ff_idct_xvid_mmx_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2600 c->idct = ff_idct_xvid_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2601 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2602 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2603 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2604
12435
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
2605 c->put_pixels_clamped = ff_put_pixels_clamped_mmx;
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
2606 c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
fe78a4548d12 Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents: 12417
diff changeset
2607 c->add_pixels_clamped = ff_add_pixels_clamped_mmx;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2608 c->clear_block = clear_block_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2609 c->clear_blocks = clear_blocks_mmx;
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2610 if ((mm_flags & AV_CPU_FLAG_SSE) &&
10766
78c2be62260a Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents: 10749
diff changeset
2611 !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){
78c2be62260a Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents: 10749
diff changeset
2612 /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
9861
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
2613 c->clear_block = clear_block_sse;
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
2614 c->clear_blocks = clear_blocks_sse;
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
2615 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2616
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2617 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2618 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2619 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2620 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2621 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2622
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2623 SET_HPEL_FUNCS(put, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2624 SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2625 SET_HPEL_FUNCS(avg, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2626 SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2627 SET_HPEL_FUNCS(put, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2628 SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2629 SET_HPEL_FUNCS(avg, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2630 SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2631
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2632 c->gmc= gmc_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2633
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2634 c->add_bytes= add_bytes_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2635 c->add_bytes_l2= add_bytes_l2_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2636
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2637 c->draw_edges = draw_edges_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2638
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10645
diff changeset
2639 if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2640 c->h263_v_loop_filter= h263_v_loop_filter_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2641 c->h263_h_loop_filter= h263_h_loop_filter_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2642 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2643
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2644 #if HAVE_YASM
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2645 c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_mmx_rnd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2646 c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_mmx;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2647 c->put_no_rnd_vc1_chroma_pixels_tab[0]= ff_put_vc1_chroma_mc8_mmx_nornd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2648
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2649 c->put_rv40_chroma_pixels_tab[0]= ff_put_rv40_chroma_mc8_mmx;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2650 c->put_rv40_chroma_pixels_tab[1]= ff_put_rv40_chroma_mc4_mmx;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2651 #endif
8519
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2652
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2653 if (mm_flags & AV_CPU_FLAG_MMX2) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2654 c->prefetch = prefetch_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2655
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2656 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2657 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2658
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2659 c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2660 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2661 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2662
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2663 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2664 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2665
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2666 c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2667 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2668 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2669
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2670 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2671 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2672 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2673 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2674 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2675 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2676 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2677
12439
51fc247eed32 Fix compilation failure if yasm is disabled (missing vp3 symbols).
rbultje
parents: 12437
diff changeset
2678 if (CONFIG_VP3_DECODER && HAVE_YASM) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2679 c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2680 c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2681 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2682 }
12439
51fc247eed32 Fix compilation failure if yasm is disabled (missing vp3 symbols).
rbultje
parents: 12437
diff changeset
2683 if (CONFIG_VP3_DECODER && HAVE_YASM) {
11637
f7281af560fe vp3: DC-only IDCT
conrad
parents: 11499
diff changeset
2684 c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2;
f7281af560fe vp3: DC-only IDCT
conrad
parents: 11499
diff changeset
2685 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2686
11826
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2687 if (CONFIG_VP3_DECODER
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2688 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2689 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmx2;
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2690 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmx2;
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2691 }
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2692
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2693 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2694 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2695 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2696 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2697 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2698 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2699 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2700 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2701 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2702 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2703 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2704 c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2705 c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2706 c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2707 c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2708 c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2709 c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2710
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2711 SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2712 SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2713 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2714 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2715 SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2716 SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2717
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2718 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2719 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2720 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2721 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2722 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2723 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2724
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2725 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2726 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2727 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2728 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2729
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2730 #if HAVE_YASM
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2731 c->avg_rv40_chroma_pixels_tab[0]= ff_avg_rv40_chroma_mc8_mmx2;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2732 c->avg_rv40_chroma_pixels_tab[1]= ff_avg_rv40_chroma_mc4_mmx2;
8519
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2733
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2734 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= ff_avg_vc1_chroma_mc8_mmx2_nornd;
9440
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
2735
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2736 c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_mmx2_rnd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2737 c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_mmx2;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2738 c->avg_h264_chroma_pixels_tab[2]= ff_avg_h264_chroma_mc2_mmx2;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2739 c->put_h264_chroma_pixels_tab[2]= ff_put_h264_chroma_mc2_mmx2;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2740
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2741 c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2742 #endif
8798
a5c8210814d7 Add check whether the compiler/assembler supports 10 or more operands.
diego
parents: 8760
diff changeset
2743 #if HAVE_7REGS && HAVE_TEN_OPERANDS
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2744 if( mm_flags&AV_CPU_FLAG_3DNOW )
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2745 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2746 #endif
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2747
9995
3141f69e3905 Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents: 9975
diff changeset
2748 if (CONFIG_VC1_DECODER)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2749 ff_vc1dsp_init_mmx(c, avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2750
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2751 c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2;
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2752 } else if (mm_flags & AV_CPU_FLAG_3DNOW) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2753 c->prefetch = prefetch_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2754
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2755 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2756 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2757
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2758 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2759 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2760 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2761
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2762 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2763 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2764
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2765 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2766 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2767 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2768
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2769 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2770 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2771 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2772 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2773 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2774 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2775 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2776 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2777
11826
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2778 if (CONFIG_VP3_DECODER
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2779 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2780 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2781 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2782 }
11c5a87497d3 Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents: 11637
diff changeset
2783
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2784 SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2785 SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2786 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2787 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2788 SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2789 SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2790
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2791 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2792 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2793 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2794 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2795 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2796 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2797
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2798 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2799 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2800 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2801 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2802
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2803 #if HAVE_YASM
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2804 c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_3dnow_rnd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2805 c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_3dnow;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2806
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2807 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= ff_avg_vc1_chroma_mc8_3dnow_nornd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2808
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2809 c->avg_rv40_chroma_pixels_tab[0]= ff_avg_rv40_chroma_mc8_3dnow;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2810 c->avg_rv40_chroma_pixels_tab[1]= ff_avg_rv40_chroma_mc4_3dnow;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2811 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2812 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2813
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2814
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2815 #define H264_QPEL_FUNCS(x, y, CPU)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2816 c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2817 c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2818 c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2819 c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU;
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2820 if((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW)){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2821 // these functions are slower than mmx on AMD, but faster on Intel
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2822 c->put_pixels_tab[0][0] = put_pixels16_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2823 c->avg_pixels_tab[0][0] = avg_pixels16_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2824 H264_QPEL_FUNCS(0, 0, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2825 }
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2826 if(mm_flags & AV_CPU_FLAG_SSE2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2827 H264_QPEL_FUNCS(0, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2828 H264_QPEL_FUNCS(0, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2829 H264_QPEL_FUNCS(0, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2830 H264_QPEL_FUNCS(1, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2831 H264_QPEL_FUNCS(1, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2832 H264_QPEL_FUNCS(1, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2833 H264_QPEL_FUNCS(2, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2834 H264_QPEL_FUNCS(2, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2835 H264_QPEL_FUNCS(2, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2836 H264_QPEL_FUNCS(3, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2837 H264_QPEL_FUNCS(3, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2838 H264_QPEL_FUNCS(3, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2839 }
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2840 #if HAVE_SSSE3
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2841 if(mm_flags & AV_CPU_FLAG_SSSE3){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2842 H264_QPEL_FUNCS(1, 0, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2843 H264_QPEL_FUNCS(1, 1, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2844 H264_QPEL_FUNCS(1, 2, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2845 H264_QPEL_FUNCS(1, 3, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2846 H264_QPEL_FUNCS(2, 0, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2847 H264_QPEL_FUNCS(2, 1, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2848 H264_QPEL_FUNCS(2, 2, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2849 H264_QPEL_FUNCS(2, 3, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2850 H264_QPEL_FUNCS(3, 0, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2851 H264_QPEL_FUNCS(3, 1, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2852 H264_QPEL_FUNCS(3, 2, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2853 H264_QPEL_FUNCS(3, 3, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2854 c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3;
10430
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2855 #if HAVE_YASM
12437
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2856 c->put_no_rnd_vc1_chroma_pixels_tab[0]= ff_put_vc1_chroma_mc8_ssse3_nornd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2857 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= ff_avg_vc1_chroma_mc8_ssse3_nornd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2858 c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_ssse3_rnd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2859 c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_ssse3_rnd;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2860 c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_ssse3;
b242eb86ea9a Move H264 chroma MC from inline asm to yasm. This fixes VP3/5/6 and VC-1
rbultje
parents: 12436
diff changeset
2861 c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_ssse3;
10430
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2862 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2863 if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
10430
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2864 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2865 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2866 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2867 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2868
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2869 if(mm_flags & AV_CPU_FLAG_3DNOW){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2870 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2871 c->vector_fmul = vector_fmul_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2872 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2873 c->float_to_int16 = float_to_int16_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2874 c->float_to_int16_interleave = float_to_int16_interleave_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2875 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2876 }
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2877 if(mm_flags & AV_CPU_FLAG_3DNOWEXT){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2878 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2879 c->vector_fmul_window = vector_fmul_window_3dnow2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2880 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2881 c->float_to_int16_interleave = float_to_int16_interleave_3dn2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2882 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2883 }
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2884 if(mm_flags & AV_CPU_FLAG_MMX2){
10633
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2885 #if HAVE_YASM
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2886 c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2;
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2887 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2;
10633
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2888 #endif
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2889 }
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2890 if(mm_flags & AV_CPU_FLAG_SSE){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2891 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2892 c->ac3_downmix = ac3_downmix_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2893 c->vector_fmul = vector_fmul_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2894 c->vector_fmul_reverse = vector_fmul_reverse_sse;
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2895 c->vector_fmul_add = vector_fmul_add_sse;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2896 c->vector_fmul_window = vector_fmul_window_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2897 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse;
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2898 c->vector_clipf = vector_clipf_sse;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2899 c->float_to_int16 = float_to_int16_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2900 c->float_to_int16_interleave = float_to_int16_interleave_sse;
10964
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2901 #if HAVE_YASM
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2902 c->scalarproduct_float = ff_scalarproduct_float_sse;
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2903 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2904 }
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2905 if(mm_flags & AV_CPU_FLAG_3DNOW)
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2906 c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2907 if(mm_flags & AV_CPU_FLAG_SSE2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2908 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2909 c->float_to_int16 = float_to_int16_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2910 c->float_to_int16_interleave = float_to_int16_interleave_sse2;
10633
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2911 #if HAVE_YASM
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2912 c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2913 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
10633
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2914 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2915 }
12456
a5ddb39627fd Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents: 12454
diff changeset
2916 if((mm_flags & AV_CPU_FLAG_SSSE3) && !(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW)) && HAVE_YASM) // cachesplit
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2917 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2918 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2919
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
2920 if (CONFIG_ENCODERS)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2921 dsputilenc_init_mmx(c, avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2922
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2923 #if 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2924 // for speed testing
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2925 get_pixels = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2926 put_pixels_clamped = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2927 add_pixels_clamped = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2928
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2929 pix_abs16x16 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2930 pix_abs16x16_x2 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2931 pix_abs16x16_y2 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2932 pix_abs16x16_xy2 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2933
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2934 put_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2935 put_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2936 put_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2937 put_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2938
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2939 put_no_rnd_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2940 put_no_rnd_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2941 put_no_rnd_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2942 put_no_rnd_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2943
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2944 avg_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2945 avg_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2946 avg_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2947 avg_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2948
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2949 avg_no_rnd_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2950 avg_no_rnd_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2951 avg_no_rnd_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2952 avg_no_rnd_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2953
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2954 //av_fdct = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2955 //ff_idct = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2956 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2957 }